Split (1)

train · 57.9k rows

id stringlengths 36 36	source stringclasses 15 values	formatted_source stringclasses 13 values	text stringlengths 2 7.55M
d8a814de-4755-43e4-ad96-207fd1841ef1	trentmkelly/LessWrong-43k	LessWrong	Sorry for the downtime, looks like we got DDosd We were down between around 7PM and 8PM PT today. Sorry about that. It's hard to tell whether we got DDosd or someone just wanted to crawl us extremely aggressively, but we've had at least a few hundred IP addresses and random user agents request a lot of quite absurd pages, in a way that was clearly designed to avoid bot-detection and block methods. I wish we were more robust to this kind of thing, and I'll be monitoring things tonight to prevent it from happening again, but it would be a whole project to make us fully robust to attacks of this kind. I hope it was a one-off occurence, but also, I think we can figure out how to make it so we are robust to repeated DDos attacks, if that is the world we live in, though I do think it would mean strapping in for a few days of spotty reliability while we figure out how to do that. Sorry again, and boo for the people doing this. It's one of the reasons why running a site like LessWrong is harder than it should be.
cdc8dc56-57b9-4251-ae71-5c669c32c89c	trentmkelly/LessWrong-43k	LessWrong	META: Who Have You Told About LW? I've been lurking on LW since shortly after it started, and on OB for about six months before that. In that time, I've told four or five people about it. I would make a terrible evangelist. I'm curious as to whether other people have the same problem. I'd like to tell lots of people about LW, but I don't think they're ready for it. If they read a statement like "purchase utilons and warm fuzzies separately" their eyes would glaze over, and they'd walk away thinking LW was some sort of crackpot site. I have found certain posts and topics to be fairly good hooks for getting people interested. An Alien God is a good suggested read for people with an interest in evolutionary theory and the Cthulhu mythos (surprisingly high crossover in my experience). HP:MoR is also a pretty popular hook. The site itself isn't really optimised for word-of-mouth, though, and not everyone likes child wizards and blasphemous horrors. How many people have you introduced to LW? Who were they, how do you do it and what was their reaction? How could we do it better?
c841039e-5718-48f7-a3e0-2bb5bf791b92	trentmkelly/LessWrong-43k	LessWrong	D&D.Sci Pathfinder: Return of the Gray Swan This is an entry in the 'Dungeons & Data Science' series, a set of puzzles where players are given a dataset to analyze and an objective to pursue using information from that dataset. Edited to add: Consensus is that this scenario is too complicated. It probably shouldn't be the first D&D.Sci scenario you play - try the original Voyages of the Gray Swan or The Sorcerer's Personal Shopper first, if you haven't played them yet. Story You had thought that the Admiralty was done with you after last time. After their response to your initial suggestions was to assign you to a series of dangerous voyages and hope that killed you, you were quite glad to set foot back ashore and not hear from them again. You've still been receiving a modest stipend for your past service for the last ten years or so, and that plus occasional odd-jobs for local wizards has left you a fairly comfortable life in the capital of Eastmarch. So you are perhaps a bit concerned when you receive a summons to attend at once an audience with the Third Lord of the Admiralty. Your only consolation is that the summons arrives via a courier knocking on your door rather than a squadron of marines kicking it down and dragging you off. You were not expecting what actually happened, though. When you were called into the Lord Admiral's office, there was only one person it it, and you recognized him. "Captain O'Neill?" It's been a while since you saw him. He's still got the beard and the lively eyes, but he was wearing less braid back then, and a less...fancy...hat... He chuckles. "It's Admiral now, actually. I'm sorry to call you here on such short notice, but a problem has arisen that I could use some help with." It seems that Captain O'Neill (you absent-mindedly call him 'Captain' rather than 'Admiral' another three times during your conversation, but he doesn't seem offended) has risen through the ranks quite dramatically since you sailed with him on the Gray Swan. It also seems that he stil
0061a00b-2d4f-4b99-b87b-fca78f9f08ab	StampyAI/alignment-research-dataset/youtube	Youtube Transcripts	OUT OF CONTROL - The design of AI in everyday life (Elisa Giaccardi) - 1st AiTech Symposium cute yes good afternoon what happens when AI begins to operate in response to everyday interactions with people how can a AI systems remain trustworthy once they begin to learn from us and from each other these are really big questions and they're the ones that actually troubled me as a designer one of the dominant narratives today about AI is that it will help us be more accurate in our decision-making processes it will help us make better diagnoses and keep us safe on the road these narrative come with the assumption that somehow we the humans are the real black boxes and if we just design a time where enough and we exert enough control then AI will compensate for our human flows and the mess we have created now dystopian scenarios of black mirror aside this that I call a smart utopia and the design ideals of control and optimization that come with it are actually a bit at odds with the messy reality of everyday life in a study on the future of Taiwanese mobility that we conducted in Taipei in collaboration with Taiwan tech design ideas of fuel efficiency and energy savings proved to be at odds with how Taiwanese use scooters in everyday life in this picture for example you can see how the scooter used at a low speed at the street market is no longer a vehicle for transportation really but a cart for grocery shopping and at a different speed and within a different configuration of everyday life it becomes a means for establishing and maintaining social relationships now ok that doesn't seem very critical does it and it's easy to relegate this case to matters of culturally sensitive design the potential friction suggested by this case between the intended use of smart autonomous technology and how it might actually end up being used by people in the context of their lights seems quite irrelevant when we think of the design of an AI that can save us from a drunk driver or the wrong medical diagnosis so why not let AI in even more spheres of our lives and use it to optimize also our wealth all we need is to buy a SmartWatch make sure that we stay active and reduce our insurance premium while maybe even gaining additional benefits that's the value proposition of a start-up company called vitality in the UK now it gets to be it gets to be a bit more serious because what if I get sick or for whatever reason I just can't exercise anymore not as much will I still be able to afford medical coverage and what to do this was an example presented by Jani this morning what do we do what if a badly designed and yet already implemented AI system fires an talented motivated teacher just because the mathis cause of her students are not good enough how do we fix it when we consider their complexity of these real-life scenarios it is quite apparent that the need for control is morally justified so I really hope not to be misunderstood when I say that control mechanisms and regulations are necessary but not enough frameworks promoting principles of control explain ability and transparency are necessary to ensure accountability after the fact after something has been designed but design has to regain its agency in the crafting of desirable relationships with technology between people and technology it has to become anticipatory gain both to address the legal issues but also not to stifle innovation so but for the field of design to move forward I'd like to make a step back and I'd like to ask what are the design ideas that are actually as designed were locked into an I argue that as designers we are locked into the fallacy that all we need to do is to get it right the right functionality the right feature the right interface right algorithm the right user experience you name it but these comes from times when working iteratively and getting it right accurate precise was the best way to minimizing the risk of mass replicating faults and shortcomings it is very much anchored in design ideas of mass production contemporary natural technologies artificial intelligence as well as the platform capitalism that have made pasta have made possible not only differed from the logic of industrial production they fundamentally challenged the conceptual space that as designers we have created to cope with complexity with runtime assembly of natural services constant atomic updates and agile development processes the design process is no longer something that happens before production and then it's done it continues in use in this characteristic constant becoming is going to be further accelerated by technologies that operate in everyday life and actively learn while e-news changing and adapting over time at an even more fundamental level than is currently the case so the point that I'm trying to make is that while we jumped with both feet in the digital society our way of thinking our design frameworks and methodologies are still locked into an industrial past and this table shows some of the shifts needed to move towards a truly post-industrial design for example if we consider core performance that is the core dependency of people and I more than they're supposed autonomy and the degrees of autonomy is between the two then we begin to ask very different types of questions the type of questions that hopefully will hear also in the next presentations when we talk about narratives and counter narratives and the very important idea of design trade-offs and one question that we might begin to ask is for example how can we empower humans running them controlling machines in the resourceful aging project which is a project we conducted and concluded last year and that received by the European Commission I an internet ng a next generation internet a word for better digital life we address this question by looking at how machine learning can be used to support older people's natural strategies of improvisation and resourcefulness right in the monitoring predicting and prescribing behavior the motors are networked so that our colleagues from computer science oriented often be looked into for this project were implemented to answer a different set of questions and accuracy of prediction they were concerned with the core usage and variety of appropriation that latitude at which older people who could use technology as a resource to improvise in everyday life learn from each other develop shared norms and values and remain independent not just from the care of their loved ones but also from care technology and we try to capture all the lessons in the booklets available online so I'd like to conclude by provoking the audience and saying that design should not be about accountability about fixing the things that are wrong design is imagination of how things might be it's about taking agency and responsibility for our designers for desirable relations between people and technology it's about our future isn't a lot in the motor of our faculty but for that we need to understand and fully engage conceptually methodologically and ethically with the true challenges of past industrial design which are codependency not autonomy Mart intentionality not so much accuracy and perhaps empowerment are not so much controlled and with that I like to invite get her the new in other going to introduce him
073ac02a-d37e-4361-bdd5-f4934ec6caf5	trentmkelly/LessWrong-43k	LessWrong	The Truth about Scotsmen, or: Dissolving Fallacies One unfortunate feature I’ve noticed in arguments between logically well-trained people and the untrained is a tendency for members of the former group to point out logical errors as if they were counterarguments. This is almost totally ineffective either in changing the mind of your opponent or in convincing neutral observers. There are two main reasons for this failure. 1. Pointing out fallacies is not the same thing as urging someone to reconsider their viewpoint. Fallacies are problematic because they’re errors in the line of reasoning that one uses to arrive at or support a conclusion. In the same way that taking the wrong route to the movie theater is bad because you won’t get there, committing a fallacy is bad because you’ll be led to the wrong conclusions. But all that isn’t inherent in the word ‘fallacy’: the vast majority of human beings don’t understand the statement “that’s a fallacy” as “you seem to have been misled by this particular logical error – you should reevaluate your thought process and see if you arrive at the same conclusions without it.” Rather, most people will regard it as an enemy attack,with the result that they will either reject the existence of the fallacy or simply ignore it. If, by some chance, they do acknowledge the error, they’ll usually interpret it as “your argument for that conclusion is wrong – you should argue for the same conclusion in a different way.” If you’re actually trying to convince someone (as opposed to, say, arguing to appease the goddess Eris) by showing them that the chain of logic they base their current belief on is unsound, you have to say so explicitly. Otherwise saying “fallacy” is about as effective as just telling them that they’re wrong. 2. Pointing out the obvious logical errors that fallacies characterize often obscures the deeper errors that generate the fallacies. Take as an example the No True Scotsman fallacy. In the canonical example, the Scotsman, having seen a report of a crime, claims t
2d3411fa-ed51-4b31-9798-3afc786bd8e4	trentmkelly/LessWrong-43k	LessWrong	Luke' AMA gets a plug @ Wired [link] http://www.wired.com/geekdad/2012/08/preventing-skynet/
a46c98ed-3abd-4c10-8cab-d34ee5153950	trentmkelly/LessWrong-43k	LessWrong	UML VI: Stochastic Gradient Descent (This is part six in a sequence on Machine Learning based on this book. Click here for part 1.) Stochastic Gradient Descent is the big Machine Learning technique. It performs well in practice, it's simple, and it sounds super cool. But what is it? Roughly, Gradient Descent is an approach for function minimization, but it generally isn't possible to apply it to Machine Learning tasks. Stochastic Gradient Descent, which is Gradient Descent plus noise, is a variant which can be used instead. Let's begin with regular Gradient Descent. Gradient Descent Given a function f:Rd→R and a point x∈Rd, the gradient of f at x is the vector of partial derivatives of f at x, i.e. ∇f(x)=(∂f∂x1(x),...,∂f∂xd(x)) So the gradient at a point is an element of Rd. In contrast, the gradient itself (not yet applied to a point) is the function ∇f:Rd→Rd defined by the above rule. Going forward, gradient always means "gradient at a point". If d=1, then the gradient will be f′(x), a number in R. It can look like this: → Or like this: ← Or also like this: ⟶ Point being, if f′(x)>0 it'll point rightward and if f′(x)<0 it'll point leftward, also it can have different lengths, but that's it. The idea of gradient descent is that the gradient points into the direction of fastest positive change, thus the opposite of the gradient is the direction of fastest negative change. It follows that, to minimize a function, one can start anywhere, compute the gradient, and then move into the opposite direction. In the above example, the function goes up, therefore the derivative is positive, therefore the gradient points rightward. Gradient Descent tells us to go into the opposite direction, i.e. leftwards. Indeed, leftward is where the function decreases. Clearly, this is a display of utter brilliance. Importantly, note that the picture is misleading insofar as it suggests there are more directions than two. But actually, the gradient lives in the domain space, in this case, R, not in the Carte
71efe618-cf2a-46e4-99bd-2b674cfc1988	StampyAI/alignment-research-dataset/alignmentforum	Alignment Forum	[AN #118]: Risks, solutions, and prioritization in a world with many AI systems Alignment Newsletter is a weekly publication with recent content relevant to AI alignment around the world. Find all Alignment Newsletter [resources here](http://rohinshah.com/alignment-newsletter/). In particular, you can look through [this spreadsheet](https://docs.google.com/spreadsheets/d/1PwWbWZ6FPqAgZWOoOcXM8N_tUCuxpEyMbN1NYYC02aM/edit?usp=sharing) of all summaries that have ever been in the newsletter. Audio version [here](http://alignment-newsletter.libsyn.com/alignment-newsletter-118) (may not be up yet). HIGHLIGHTS ========== [AI Governance: Opportunity and Theory of Impact](https://forum.effectivealtruism.org/posts/42reWndoTEhFqu6T8/ai-governance-opportunity-and-theory-of-impact) (Allan Dafoe) (summarized by Rohin): What is the theory of change for work on AI governance? Since the world is going to be vastly complicated by the broad deployment of AI systems in a wide variety of contexts, several structural risks will arise. AI governance research can produce “assets” (e.g. policy expertise, strategic insights, important networking connections, etc) that help humanity make better decisions around these risks. Let’s go into more detail. A common perspective about powerful AI is the “superintelligence” perspective, in which we assume there is a single very cognitively powerful AI agent. This leads people to primarily consider “accident” and “misuse” risks, in which either the AI agent itself “wants” to harm us, or some bad actor uses the AI agent to harm us. However, it seems likely that we should think of an ecology of AI agents, or AI as a general purpose technology (GPT), as in e.g. [CAIS](https://www.fhi.ox.ac.uk/reframing/) ([AN #40](https://mailchi.mp/b649f32b07da/alignment-newsletter-40)) or [Age of Em](https://ageofem.com/). In this case, we can examine the ways in which narrow AI could transform social, military, economic, and political systems, and the structural risks that may arise from that. Concrete examples of potential existential structural risks induced by AI include nuclear instability, geopolitical turbulence, authoritarianism, and value erosion through competition. A key point about the examples above is that the relevant factors for each are different. For example, for nuclear instability, it is important to understand nuclear deterrence, first strike vulnerability and how it could change with AI processing of satellite imagery, undersea sensors, cyber surveillance and weapons, etc. In contrast, for authoritarianism, relevant processes include global winner-take-all-markets, technological displacement of labor, and authoritarian surveillance and control. This illustrates a general principle: unlike in the superintelligence perspective, the scope of both risks and solutions in the ecology / GPT perspectives is very broad. As a result, we need a broad range of expertise and lots of connections with existing fields of research. In particular, “we want to build a metropolis -- a hub with dense connections to the broader communities of computer science, social science, and policymaking -- rather than an isolated island”. Another important aspect here is that in order to cause better decisions to be made, we need to focus not just on generating the right ideas, but also on ensuring the right ideas are in the right places at the right time (e.g. by ensuring that people with the right tacit knowledge are part of the decision-making process). Instead of the "product model" of research that focuses on generating good ideas, we might instead want a “field-building model”, which also places emphasis on improving researcher’s competence on a variety of issues, bestowing prestige and authority on those who have good perspectives on long-term risks, improving researcher’s networks, and training junior researchers. However, often it is best to focus on the product model of research anyway, and get these benefits as a side effect. To quote the author: “I think there is a lot of useful work that can be done in advance, but most of the work involves us building our competence, capacity, and credibility, so that when the time comes, we are in position and ready to formulate a plan. [...] Investments we make today should increase our competence in relevant domains, our capacity to grow and engage effectively, and the intellectual credibility and policy influence of competent experts.” Rohin's opinion: See the next summary. Note also that the author is organizing the [Cooperative AI Workshop](https://www.cooperativeai.com/) ([AN #116](https://mailchi.mp/d31663e4d330/an-116-how-to-make-explanations-of-neurons-compositional)) to tackle some of these issues. [Andrew Critch on AI Research Considerations for Human Existential Safety](https://futureoflife.org/2020/09/15/andrew-critch-on-ai-research-considerations-for-human-existential-safety/?utm_source=feedly&utm_medium=rss&utm_campaign=andrew-critch-on-ai-research-considerations-for-human-existential-safety) (Lucas Perry and Andrew Critch) (summarized by Rohin): This podcast discusses the recent [ARCHES](http://acritch.com/papers/arches.pdf) ([AN #103](https://mailchi.mp/60475c277263/an-103-arches-an-agenda-for-existential-safety-and-combining-natural-language-with-deep-rl)) document, and several thoughts surrounding it. There’s a lot in here that I won’t summarize, including a bunch of stuff that was in the summary of ARCHES. I’m going to focus primarily on the (substantial) discussion of how to prioritize within the realm of possible risks related in some way to AI systems. Firstly, let’s be clear about the goal: ensuring existential safety, that is, making sure human extinction never happens. Note the author means literal extinction, as opposed to something like “the loss of humanity’s long-term potential”, because the former is clearer. While it is not always clear whether something counts as “extinction” (what if we all become uploads?), it is a lot clearer than whether a scenario counts as a loss of potential. Typical alignment work focuses on the “single-single” case, where a single AI system must be aligned with a single human, as in e.g. [intent alignment](https://www.alignmentforum.org/posts/ZeE7EKHTFMBs8eMxn/clarifying-ai-alignment) ([AN #33](https://mailchi.mp/b6dc636f6a1b/alignment-newsletter-33)). However, this isn’t ultimately what we care about: we care about multi-multi existential safety, that is, ensuring that when multiple AI systems act in a world with multiple humans, extinction does not happen. There are pretty significant differences between these: in particular, it’s not clear whether multi-multi “alignment” even has meaning, since it is unclear whether it makes sense to view humanity as an agent to which an AI system could be “aligned”. Nonetheless, single-single alignment seems like an important subproblem of multi-multi existential safety: we will be delegating to AI systems in the future; it seems important that we know how to do so. How do we prioritize between single-single alignment, and the other subproblems of multi-multi existential safety? A crucial point is that single-single work will not be neglected, because companies have strong incentives to solve single-single alignment (both in the sense of optimizing for the right thing, and for being robust to distributional shift). In contrast, in multi-multi systems, it is often the case that there is a complex set of interacting effects that lead to some negative outcome, and there is no one actor to blame for the negative outcome, and as a result it doesn’t become anybody’s job to prevent that negative outcome. For example, if you get a huge medical bill because the necessary authorization forms hadn’t been filled out, whose fault is it? Often in such cases there are many people to blame: you could blame yourself for not checking the authorization, or you could blame the doctor’s office for not sending the right forms or for not informing you that the authorization hadn’t been obtained, etc. Since it’s nobody’s job to fix such problems, they are and will remain neglected, and so work on them is more impactful. Something like transparency is in a middle ground: it isn’t profitable yet, but probably will be soon. So, if someone were indifferent between a bunch of areas of research, the author would advise for e.g. multi-stakeholder delegation over transparency over robustness. However, the author emphasizes that it’s far more important that people work in some area of research that they find intellectually enriching and relevant to existential safety. The podcast has lots of other points, here is an incomplete quick selection of them: - In a multi-multi world, without good coordination you move the world in a “random” direction. There are a lot of variables which have to be set just right for humans to survive (temperature, atmospheric composition, etc) that are not as important for machines. So sufficiently powerful systems moving the world in a “random” direction will lead to human extinction. - One response to the multi-multi challenge is to have a single group make a powerful AI system and “take over the world”. This approach is problematic since many people will oppose such a huge concentration of power. In addition, it is probably not desirable even if possible, since it reduces robustness by creating a single point of failure. - Another suggestion is to create a powerful AI system that protects humanity (but is still uncontrollable in that humanity cannot stop its operation). The author does not like the solution much, because if we get it wrong and deploy a misaligned uncontrollable AI system, then we definitely die. The author prefers that we instead always have control over the AI systems we deploy. Rohin's opinion: Both this and the previous summary illustrate an increasingly common perspective: 1. The world is not going to look like “today’s world plus a single AGI agent”: instead, we will likely have a proliferation of many different AI systems specialized for different purposes. 2. In such a world, there are a lot of different challenges that aren’t standard intent alignment. 3. We should focus on these other challenges because [a variety of reasons]. If you have technical CS skills, how should you prioritize between this perspective and the more classical intent alignment perspective? Importance. I’ve [estimated](https://aiimpacts.org/conversation-with-rohin-shah/) ([AN #80](https://mailchi.mp/b3dc916ac7e2/an-80-why-ai-risk-might-be-solved-without-additional-intervention-from-longtermists)) a 10% chance of existential catastrophe via a failure of intent alignment, absent intervention from longtermists to address intent alignment. Estimates vary quite a lot, even among people who have thought about the problem a lot; I’ve heard as low as < 1% and as high as 80% (though these usually don’t assume “no intervention from longtermists”). It’s harder to estimate the importance of structural risks and extinction risks highlighted in the two summaries above, but the arguments in the previous two posts seem reasonably compelling and I think I’d be inclined to assign a similar importance to it (i.e. similar probability of causing an existential catastrophe). Note that this means I’m disagreeing with Critch: he believes that we are far more likely to go extinct through effects unique to multi-multi dynamics; in contrast I find the argument less persuasive because we do have governance, regulations, national security etc. that would already be trying to mitigate issues that arise in multi-multi contexts, especially things that could plausibly cause extinction. Neglectedness. I’ve already taken into account neglectedness outside of EA in estimating the probabilities for importance. Within EA there is already a huge amount of effort going into intent alignment, and much less in governance and multi-multi scenarios -- perhaps a difference of 1-2 orders of magnitude; the difference is even higher if we only consider people with technical CS skills. Tractability. I buy the argument in Dafoe’s article that for AI governance due to our vast uncertainty we need a “metropolis” model where field-building is quite important; I think that implies that solving the full problem (at today's level of knowledge) would require a lot of work and building of expertise. In contrast, with intent alignment, we have a single technical problem with significantly less uncertainty. As a result, I expect that currently in expectation a single unit of work goes further to solving intent alignment than to solving structural risks / multi-multi problems, and so intent alignment is more tractable. I also expect technical ideas to be a bigger portion of "the full solution" in the case of intent alignment -- as Dafoe argues, I expect that for structural risks the solution looks more like "we build expertise and this causes various societal decisions to go better" as opposed to "we figure out how to write this piece of code differently so that it does better things". This doesn't have an obvious impact on tractability -- if anything, I'd guess it argues in favor of the tractability of work on structural risks, because it seems easier to me to create prestigious experts in particular areas than to make progress on a challenging technical problem whose contours are still uncertain since it arises primarily in the future. I suspect that I disagree with Critch here: I think he is more optimistic about technical solutions to multi-multi issues themselves being useful. In the past I think humanity has resolved such issues via governance and regulations and it doesn’t seem to have relied very much on technical research; I’d expect that trend to continue. Personal fit. This is obviously important, but there isn’t much in general for me to say about it. Once again, I should note that this is all under the assumption that you have technical CS skills. I think overall I end up pretty uncertain which of the two areas I’d advise going in (assuming personal fit was equal in both areas). However, if you are more of a generalist, I feel much more inclined to recommend choosing some subfield of AI governance, again subject to personal fit, and Critch agrees with this. TECHNICAL AI ALIGNMENT ====================== HANDLING GROUPS OF AGENTS ------------------------- [Comparing Utilities](https://www.alignmentforum.org/posts/cYsGrWEzjb324Zpjx/comparing-utilities) (Abram Demski) (summarized by Rohin): This is a reference post about preference aggregation across multiple individually rational agents (in the sense that they have [VNM-style](https://en.wikipedia.org/wiki/Von_Neumann%E2%80%93Morgenstern_utility_theorem) utility functions), that explains the following points (among others): 1. The concept of “utility” in ethics is somewhat overloaded. The “utility” in hedonic utilitarianism is very different from the VNM concept of utility. The concept of “utility” in preference utilitarianism is pretty similar to the VNM concept of utility. 2. Utilities are not directly comparable, because affine transformations of utility functions represent exactly the same set of preferences. Without any additional information, concepts like “utility monster” are type errors. 3. However, our goal is not to compare utilities, it is to aggregate people’s preferences. We can instead impose constraints on the aggregation procedure. 4. If we require that the aggregation procedure produces a Pareto-optimal outcome, then Harsanyi’s utilitarianism theorem says that our aggregation procedure can be viewed as maximizing some linear combination of the utility functions. 5. We usually want to incorporate some notion of fairness. Different specific assumptions lead to different results, including variance normalization, Nash bargaining, and Kalai-Smorodinsky. FORECASTING ----------- [How Much Computational Power It Takes to Match the Human Brain](https://www.openphilanthropy.org/blog/new-report-brain-computation) (Joseph Carlsmith) (summarized by Asya): In this blog post, Joseph Carlsmith gives a summary of his longer report estimating the number of floating point operations per second (FLOP/s) which would be sufficient to perform any cognitive task that the human brain can perform. He considers four different methods of estimation. Using the mechanistic method, he estimates the FLOP/s required to model the brain’s low-level mechanisms at a level of detail adequate to replicate human task-performance. He does this by estimating that ~1e13 - 1e17 FLOP/s is enough to replicate what he calls “standard neuron signaling” — neurons signaling to each other via using electrical impulses (at chemical synapses) — and learning in the brain, and arguing that including the brain’s other signaling processes would not meaningfully increase these numbers. He also suggests that various considerations point weakly to the adequacy of smaller budgets. Using the functional method, he identifies a portion of the brain whose function we can approximate with computers, and then scales up to FLOP/s estimates for the entire brain. One way to do this is by scaling up models of the human retina: Hans Moravec's estimates for the FLOP/s of the human retina imply 1e12 - 1e15 FLOP/s for the entire brain, while recent deep neural networks that predict retina cell firing patterns imply 1e16 - 1e20 FLOP/s. Another way to use the functional method is to assume that current image classification networks with known FLOP/s requirements do some fraction of the computation of the human visual cortex, adjusting for the increase in FLOP/s necessary to reach robust human-level classification performance. Assuming somewhat arbitrarily that 0.3% to 10% of what the visual cortex does is image classification, and that the EfficientNet-B2 image classifier would require a 10x to 1000x increase in frequency to reach fully human-level image classification, he gets 1e13 - 3e17 implied FLOP/s to run the entire brain. Joseph holds the estimates from this method very lightly, though he thinks that they weakly suggest that the 1e13 - 1e17 FLOP/s estimates from the mechanistic method are not radically too low. Using the limit method, Joseph uses the brain’s energy budget, together with physical limits set by Landauer’s principle, which specifies the minimum energy cost of erasing bits, to upper-bound required FLOP/s to ~7e21. He notes that this relies on arguments about how many bits the brain erases per FLOP, which he and various experts agree is very likely to be > 1 based on arguments about algorithmic bit erasures and the brain's energy dissipation. Lastly, Joseph briefly describes the communication method, which uses the communication bandwidth in the brain as evidence about its computational capacity. Joseph thinks this method faces a number of issues, but some extremely preliminary estimates suggest 1e14 FLOP/s based on comparing the brain to a V100 GPU, and 1e16 - 3e17 FLOP/s based on estimating the communication capabilities of brains in traversed edges per second (TEPS), a metric normally used for computers, and then converting to FLOP/s using the TEPS to FLOP/s ratio in supercomputers. Overall, Joseph thinks it is more likely than not that 1e15 FLOP/s is enough to perform tasks as well as the human brain (given the right software, which may be very hard to create). And he thinks it's unlikely (<10%) that more than 1e21 FLOP/s is required. For reference, an NVIDIA V100 GPU performs up to 1e14 FLOP/s (although FLOP/s is not the only metric which differentiates two computational systems.) Read more: [Full Report: How Much Computational Power Does It Take to Match the Human Brain?](https://www.openphilanthropy.org/brain-computation-report) Asya's opinion: I really liked this post, although I haven't gotten a chance to get through the entire full-length report. I found the reasoning extremely legible and transparent, and there's no place where I disagree with Joseph's estimates or conclusions. See also [Import AI's summary](https://jack-clark.net/2020/09/14/import-ai-214-nvidias-40bn-arm-deal-a-new-57-subject-nlp-test-ai-for-plant-disease-detection/). MISCELLANEOUS (ALIGNMENT) ------------------------- [The "Backchaining to Local Search" Technique in AI Alignment](https://www.alignmentforum.org/posts/qEjh8rpxjG4qGtfuK/the-backchaining-to-local-search-technique-in-ai-alignment) (Adam Shimi) (summarized by Rohin): This post explains a technique to use in AI alignment, that the author dubs “backchaining to local search” (where local search refers to techniques like gradient descent and evolutionary algorithms). The key idea is to take some proposed problem with AI systems, and figure out mechanistically how that problem could arise when running a local search algorithm. This can help provide information about whether we should expect the problem to arise in practice. Rohin's opinion: I’m a big fan of this technique: it has helped me notice that many of my concepts were confused. For example, this helped me get deconfused about wireheading and inner alignment. It’s an instance of the more general technique (that I also like) of taking an abstract argument and making it more concrete and realistic, which often reveals aspects of the argument that you wouldn’t have previously noticed. NEWS ==== [The Open Phil AI Fellowship](https://www.openphilanthropy.org/focus/global-catastrophic-risks/potential-risks-advanced-artificial-intelligence/the-open-phil-ai-fellowship) (summarized by Rohin): We’re now at the fourth cohort of the [Open Phil AI Fellowship](https://www.openphilanthropy.org/focus/global-catastrophic-risks/potential-risks-advanced-artificial-intelligence/the-open-phil-ai-fellowship) ([AN #66](https://mailchi.mp/c8ea4a5e842f/an-66-decomposing-robustness-into-capability-robustness-and-alignment-robustness))! Applications are due October 22. [Navigating the Broader Impacts of AI Research](https://nbiair.com/) (summarized by Rohin): This is a workshop at NeurIPS; the title tells you exactly what it's about. The deadline to submit is October 12. #### FEEDBACK I'm always happy to hear feedback; you can send it to me, [Rohin Shah](https://rohinshah.com/), by replying to this email. #### PODCAST An audio podcast version of the Alignment Newsletter is available. This podcast is an audio version of the newsletter, recorded by [Robert Miles](http://robertskmiles.com/).
ede416fd-211c-4699-996d-32fdf785ee66	trentmkelly/LessWrong-43k	LessWrong	Are Cognitive Biases Design Flaws? I am a newbie so today I read the article by Eliezer Yudkowski "Your Strength As A Rationalist" which helped me understand the focus of LessWrong, but I respectfully disagreed with a line that is written in the last paragraph: > It is a design flaw in human cognition... So this was my comment in the article's comment section which I bring here for discussion: > Since I think evolution makes us quite fit to our current environment I don't think cognitive biases are design flaws, in the above example you imply that even if you had the information available to guess the truth, your guess was another one and it was false, therefore you experienced a flaw in your cognition. > > My hypotheses is that reaching the truth or communicating it in the IRC may have not been the end objective of your cognitive process, in this case just to dismiss the issue as something that was not important anyway "so move on and stop wasting resources in this discussion" was maybe the "biological" objective and as such it should be correct, not a flaw. > > If the above is true then all cognitive bias, simplistic heuristics, fallacies, and dark arts are good since we have conducted our lives for 200,000 years according to these and we are alive and kicking. > > Rationality and our search to be LessWrong, which I support, may be tools we are developing to evolve in our competitive ability within our species, but not a "correction" of something that is wrong in our design. Edit 1: I realize there is change in the environment and that may make some of our cognitive biases, which were useful in the past, to be obsolete. If the word "flaw" is also applicable to describe something that is obsolete then I was wrong above. If not, I prefer the word obsolete to characterize cognitive biases that are no longer functional for our preservation.
95a14881-f2aa-431c-8481-24ed185b899f	trentmkelly/LessWrong-43k	LessWrong	AI Risk in Terms of Unstable Nuclear Software (This is an entry for the AI Safety Public Materials contest testing out a novel-ish frame on the AI Risk.) ---------------------------------------- In recent years, there'd been growing concerns among AI specialists regarding the dangers of advanced artificial intelligence. The capabilities of AI models are growing rapidly, they argue, while our ability to control them lags far behind, not to mention governmental regulations, which haven't even begun to catch up. We're on direct course for catastrophe — one that might cost us billions of dollars in economic damage, millions of lives, or even the very survival of our species. The most striking thing about it, however, is that there's nothing outlandish or science-fictional about the nature of these dangers. When you get down to it, the threats are neither esoteric nor truly novel. They're merely scaled up beyond anything we're familiar with. It has nothing to do with AIs being smart or "sentient". The core problem is simpler: AI models are software products. And as software products go, their functionality is revolutionary, while their reliability is abysmal. ---------------------------------------- A quick primer on how AI models are developed. It has preciously little in common with standard, time-tested methods of software development. We initialize some virtual architecture — a cluster of neurons designed by loose analogue with biological brains. That architecture has no knowledge recorded in it at first, it's entirely randomized. Then, we set up the training loop. We expose the neural network to some stimuli — pictures of animals, natural-language text, or a simulated environment. The network computes some output in response. In the first stages, said output is just gibberish. Our software evaluates it, comparing it to the expected, "correct" responses. If it's unsatisfactory, the software slightly modifies the neural network, nudging the connections between its neurons so that it's somewhat more likel
3b325e3e-486f-4f4c-80b2-7b0ac3f6195d	StampyAI/alignment-research-dataset/special_docs	Other	AI transparency: a matter of reconciling design with critique Vol.:(0123456789)1 3AI & SOCIETY https://doi.org/10.1007/s00146-020-01110-y STUDENT FORUM AI transparency: a matter of reconciling design with critique Tomasz Hollanek1 Received: 28 March 2020 / Accepted: 29 October 2020 © The Author(s) 2020 Abstract In the late 2010s, various international committees, expert groups, and national strategy boards have voiced the demand to ‘open’ the algorithmic black box, to audit, expound, and demystify artificial intelligence. The opening of the algorithmic black box, however, cannot be seen only as an engineering challenge. In this article, I argue that only the sort of transpar - ency that arises from critique—a method of theoretical examination that, by revealing pre-existing power structures, aims to challenge them—can help us produce technological systems that are less deceptive and more just. I relate the question of AI transparency to the broader challenge of responsible making, contending that future action must aim to systematically reconcile design—as a way of concealing—with critique—as a manner of revealing. Keywords Critical theory · Critical thinking · Transparency · Responsibility · Self-awareness · Design theory 1 Preliminaries 1.1 AI transparency In the age of ubiquitous computing, we are surrounded by objects that incorporate artificial intelligence solutions. We interact with different kinds of AI without realizing it—using online banking systems, searching for YouTube clips, or consuming news through social media—not really know - ing how and when AI systems operate. Corporate strategies of secrecy and user interfaces that hide traces of AI-driven personalization combine with the inherent opacity of deep learning algorithms (whose inner workings are not directly comprehensible to human interpreters) to create a marked lack of transparency associated with all aspects of emerging technologies. It is in response to the widespread application of AI-based solutions to various products and services in the late 2010s that multiple expert groups—both national and international—have voiced the demand to ‘open’ the algorithmic black box, to audit, expound, and demystify AI. They claim that to ensure that the use of AI is ethical, we must design emerging systems to be transparent, explain-able, and auditable. 1The opening of the algorithmic black box, however, cannot be seen only as an engineering challenge. It is cri-tique, as the underside of making, that prioritizes unbox-ing, debunking the illusion, seeing through—to reveal how an object really works. Critique—grounded in the tradition of Critical Theory and practiced by cultural studies, criti-cal race theory, queer theory, as well as decolonial theory scholars, among others—moves beyond the technical detail to uncover the desires, ideologies, and social relations forged into objects, opening the black boxes of history, culture, and progress. In what follows, I argue that the calls for techno-logical transparency demand that we combine the practice of design with critique. I relate the question of AI transparency to the broader challenge of responsible making, contend-ing that future action must aim to systematically reconcile design—as a way of concealing—with critique—as a man-ner of revealing. 1.2 Levels of opacity Many companies sell simple analytics tools as artificial intelligence, as something that supposedly supplants human intelligence to deliver better results. What is advertised as an \* Tomasz Hollanek th536@cam.ac.uk 1 Centre for Film and Screen, University of Cambridge, Cambridge, UK1 The High-Level Expert Group on AI convened by the European Commission presented its Ethics Guidelines for Trustworthy Arti-ficial Intelligence in the early 2019. The document identifies several key characteristics of a system that can be deemed trustworthy, which include transparency, defined as a form of traceability of data, opera- tions, and business models that shape the end product (AI HLEG AI & SOCIETY 1 3 ‘AI solution’ often relies on simple data analysis performed by human analysts. The use of metaphors and simplifications obfuscates human labor, labor that is outsourced, hidden away, in an invisible and immaterial factory, in a different part of the globe. According to Ian Bogost (2015), the ‘meta-phor of mechanical automation’ is nothing more than a well-directed, but misleading, masquerade (n. pag.). Although the metaphor is only an approximation, a distortion, or even a ‘caricature,’ it convincingly plays the role of an accurate depiction of the whole. The term artificial intelligence, elu-sive, misleading and with a definition that has changed over time, forms the basis of the marketing stunt. Technology companies rely on abstract, overly schematic representations to simplify reality and arrive at an easily digestible, pre-packed idea of the object, one that misrepresents the object’s essence and overlooks its true composition, but also satiates the end user’s curiosity. Other systems, as a matter of deliberate practice, incorpo- rate complex data processing and machine learning surrepti-tiously. Shoshana Zuboff (2019) observes that the influence these systems have on our decision-making is ‘designed to be unknowable to us’ (p.11); that company strategies of mis-direction serve as ‘the moat that surrounds the castle and secures the action within’ (p.65.)—a way for corporations like Google or Facebook to protect their secrets and mislead the public. These systems could, in theory, be designed to be more ‘knowable.’ Elements of the user interface could, for exam-ple, flag up when algorithmic operations are influencing the user’s decision-making. Just like labels inform the consumer about the product’s contents, the interfaces of Facebook and YouTube could announce to users that the information deliv - ered by the platforms is algorithmically curated. Considering that only 24% of US college students realize Facebook auto-matically prioritizes certain posts and hides others (Powers 2017), such a feature would definitely be relevant. But the challenge of transparency at a time of unprecedented tech-nological complexity cannot be approached only as a matter of failed (or indeed successful, depending on your position) communication. The ‘opacity’ of machine learning algorithms refers, after all, not only to ‘institutional self-protection and con-cealment,’ but also, as Burrell (2016) points out, to ‘the mismatch between mathematical optimization in high-dimensionality characteristic of machine learning and the demands of human-scale reasoning and styles of semantic interpretation’ (p.2). The fundamental lack of transparency of systems that incorporate AI solutions relates not only to convoluted storytelling devised by marketing teams, or mis-leading interfaces and user experience design, but also—and most importantly—an emerging form of making brought about by the automation of cognitive tasks themselves. Considering this level of opacity of AI systems, a team of researchers from MIT’s Intelligence and Decision Tech-nologies Group developed a neural network named, sugges-tively, Transparency by Design Network (Mascharka et al. 2018) that not only performs ‘human-like reasoning steps’ to answer questions about the contents of images, but also ‘visually renders its thought process’ as it solves problems, allowing human analysts to interpret its decision-making. It is a deep learning system carrying out a sort of unboxing on itself, incorporating explainability in its very operations: a realization of the most common understanding of ‘transpar - ency’ in the context of contemporary AI research. 1.3 Transparency and critique The challenge of technological transparency that has now attracted the attention of policymakers has constituted a concern for cultural studies scholars, media theorists, and design philosophers for decades. In the early 1990s, the philosopher Paul Virilio (1994) noted that once ‘synthetic vision’ becomes a reality and the human subject is excluded from the process of observation of images ‘created by the machine for the machine,’ the power of statistical science to appear objective (and thus persuasive) will be ‘considerably enhanced, along with its discrimination capacities’ (p.75). Another prominent media theorist Friedrich Kittler (2014) expressed concerns about modern media technologies that are ‘fundamentally arranged to undermine sensory percep-tion.’ Kittler wrote about a ‘system of secrecy’ based on user interfaces that ‘conceal operations necessary for program-ming’ and thus ‘deprive users of the machine as a whole,’ to suggest that perceiving the imitation, penetrating through the illusion of software ‘that appears to be human,’ is a funda-mental challenge in the age of global-scale computing (pp. 221–24). Transparency, as Adrian Weller ( 2017) has poignantly noted, is an ambiguous term that will mean different things to different people. For the user, a transparent system will ‘provide a sense for what [it] is doing and why,’ while for an expert or regulator, it will enable auditing ‘a prediction or decision trail in detail’ (p.3). The calls for technological transparency have thus been filtered down to reach various stakeholder groups with different meanings and represent-ing different interests. But what seems consistent in all of Footnote 1 (continued) 2019). Another interdisciplinary group of experts working under the auspices of IEEE (the world’s largest professional association for electronic and electrical engineers) published, in 2019, the first edition of Ethically Aligned Design, a set of actionable recommen- dations on how to align design practices with society’s values and principles, stressing that the ’standards of transparency, competence, accountability, and evidence of effectiveness should govern the devel-opment of autonomous and intelligent systems’ (p. 5). AI & SOCIETY 1 3 these diverse takes on the development, use, and regula- tion of technology is that transparency is framed as a mat-ter of design. In what follows, I problematize this claim, arguing that design, in the most fundamental sense, relies on concealment and obfuscation. I contend that only the sort of transparency that arises from critique—a method of theoretical examination that, by revealing pre-existing power structures, aims to challenge them—can help us produce technological systems that are less deceptive and more just. 2 Design as blackboxing 2.1 Art and artifice Coined in 1956 by John McCarthy, the term ‘artificial intel-ligence’ had its critics among those who attended the Dart-mouth Conference (which famously established the field of AI); Arthur Samuel argued that ‘the word artificial makes you think there’s something kind of phony about this, […] or else it sounds like it’s all artificial and there’s nothing real about this work at all’ (in: McCorduck 2004 , p. 115). The historian of AI Pamela McCorduck notes that while other terms, such as ‘complex information processing,’ were also proposed, it was ‘artificial intelligence’ that endured the trial of time. According to her, it is ‘a wonderfully appropriate name, connoting a link between art and science that as a field AI indeed represents’ (p. 115). She is referring indirectly here to the origins of the word artificial; in Latin, artificialis means ‘of or belonging to art,’ while artificium is simply a work of art, but also a skill, theory, or system. When the philosopher Vilém Flusser traced the etymol- ogy of the word ‘design’ in his The Shape of Things: A Phi-losophy of Design (1999), he referred to this relationship between art and artifice to argue that all human production, all culture, can be defined as a form of trickery. Flusser rejects the distinction between art and technology, and goes back to these ancient roots: the Greek for ‘trap’ is mechos (mechanics, machine); the Greek techne corresponds to the Latin ars; an artifex means a craftsman or artist, but also a schemer or trickster—to demonstrate that in their essence all forms of making are meant to help us ‘elude our circum-stances,’ to cheat our own nature. Culture itself becomes a delusion brought about by means of design—a form of self-deception that makes us believe we can free ourselves from natural restrictions by producing a world of artifice. From doors to rockets, from tents to computer screens, from pen-cils to mechanized intelligences, Flusser selects his exam-ples to show that, ultimately, any involvement with culture is based on deception: sometimes ‘this machine, this design, this art, this technology is intended to cheat gravity, to fool the laws of nature’ (ch.1, n. pag.)—and sometimes to trick ourselves into thinking we control both gravity and the laws of nature. In that sense, art and technology are representative of the same worldview in which cultural production must be deceptive/artful enough to enable humans to go beyond the limits of what is (humanly) possible. Flusser refers to the act of weaving to explain the ‘con- spiratorial, even deceitful’ (ch.18) character of design. In the process of carpet production, he points out, knotting is meant to deny its own warp, to hide the threads behind a pat-tern, so that anyone stepping on the finished rug perceives it as a uniform surface, according to the designer’s plan. He offers weaving as one of the primordial forms of cultural production to embody trickery, but the same holds true for any form of design. The trick is always based on misdirec-tion, shifting the end user’s attention from the material to the application, from the current state of things to emerging possibilities and new futures. Designing is a methodical way of crafting alternative realities out of existing materials—a process of casting the intended shape onto the chosen fabric so as to create a new possibility. The material used in that process must, so to speak, dematerialize: it has to disappear from view and give way to the new object—to abstract the end result from the point of origin and the labor process. By obfuscating some components while exhibiting others, ‘ideal’ design enables an end user’s cognitive efficiency. 2.2 Patterns, layers, and repetitions For Flusser, any product of human making is both an object and an obstacle—Latin objectum, Greek problema—or, more specifically, any object is also an ‘obstacle that is used for removal of obstacles’ (ch.9). To move forward, we solve problems that lie ahead and obstruct our way; we produce objects that help us remove these obstacles; but the same objects turn into obstacles for those that come after us. In other words, since the results of human problem-solving are stored in objects, progress involves obfuscation and for - getting. We come up with a solution and, with time, this singular idea turns into a pattern; others use the already established template to produce new, more complex struc-tures and these structures turn into new patterns, covering up previous layers of design with new design. To expedite the process of production, to advance, to move faster, the designer turns to these conventions and templates, choosing from a menu of preprogrammed options—or abstracting new rules based on previous patterns. And as the complexity of the production process increases, the reliance on patterns grows too. New design always depends on previous design, and this ultimate dependence on patterns and abstractions complicates understanding the process in its totality. In the age of ubiquitous computing, speaking of obfus- cation by design becomes of particular importance. In 2015, Benjamin Bratton called his model of the new kind AI & SOCIETY 1 3 of layering brought about by planetary-scale computation ‘the Stack’: ‘an accidental megastructure, one that we are building both deliberately and unwittingly and is in turn build-ing us in its own image’ (p.5). New technologies ‘align, layer by layer, into something like a vast, if also incomplete, pervasive if also irregular, software and hardware Stack ’ (p.5). This makes it hard to perceive the Stack’s overarching structure, indeed, to see it as design, however incidental. Today, we produce new technologies, new objects, to see, know, and feel more, to register what is normally hidden from our view, meanwhile, creating complex systems based on multiple, invisible lay - ers and algorithmic operations whose effects are not always comprehensible even to the designers themselves. 2.3 Automations and automatisms In her comprehensive account of what she calls ‘surveil-lance capitalism,’ Shoshana Zuboff points out the dangers of technological illusion—‘an enduring theme of social thought, as old as the Trojan horse’ (p.16)—that serves the new economic project in rendering its influence invisible. Surveillance capitalism claims ‘human experience as free raw material for translation into behavioral data,’ and turns that data into ‘prediction products’ that can later be sold to advertisers (p.8). Echoing the work of philosophers such as Bernard Stiegler (2014, 2015) or Antoinette Rouvroy (2016), Zuboff argues that the ultimate goal of this new form of capi-talism is ‘to automate us,’ by reprogramming our behavior and desires. Various internet platforms that dominate the market prompt us to action, influence our decision making, relying on big data analyses of our preference trends online. Automated systems create statistical models to profile users, tracing any emerging patterns in countless interactions with digital products; patterns turn into further abstractions, new models that are later reflected in new products and solutions, which end up ‘automating’ us, guiding our decision-making without our knowing. But is this process specific to AI-enhanced personaliza- tion under surveillance capitalism? Bratton has recently argued that what ‘at first glance looks autonomous (self-gov - erning, set apart, able to decide on its own) is, upon closer inspection, always also decided in advance by remote ances-tral agents and relays, and is thus automated as well’ (2019, loc.345, n. pag.). Any decision taken now relies on multiple decisions taken in the past; new design depends on previ-ous design; a new object coalesces from an aggregation of old solutions. Culture is an amalgamation of such objects—objects that, ironically, become obstacles because they are meant to enable our cognitive efficiency. A tool becomes an obstacle because the results of our problem-solving and labor are already stored within it; a tool must never be seen as a tool, as its use must be intuitive—it must remain imper - ceptible; any new tool meant to advance the process is made with existing tools, and so the emerging layering of design in the Anthropocene makes it harder to distinguish between tool and fabric. Extending this to the ongoing automation of cognitive tasks in the age of ubiquitous computing, the phenomenon takes on new scale. This is why the emerging need for transparency refers not so much to company politics of disinformation or algo-rithmic black boxes, as to the very essence of our culture, as a process of knowledge production, pattern formation, and concealment. Particular problems caused by the widespread adoption of automated decision-making systems, such as algorithmic bias, can have specific, targeted, solutions in the form of new policy, engineering standards, or better educa-tion. But a shift of focus from the particular to the total is more than an exercise in theory—it makes us realize that transparency has never been at the heart of our making, that design has always been a form of blackboxing. There is, in that sense, something deeply anti-cultural about transpar - ency. Or, putting it differently, there is nothing natural about transparency by design: we have been programmed to cover up as we make, not the opposite. The ongoing transformation of lived experiences into data is a new analytical paradigm that demands our intervention, truly calls for an ‘unboxing,’ an excavation of processes and data trails. But the opening of the algorithmic black box can-not be viewed only as a technical issue—precisely because any solution is, first and foremost, a result of cultural black - boxing. While contemporary debates on AI focus on trans-parency as a direct response to the opacity of algorithms, what we are in need of are approaches that aim to ‘unbox’ new technologies as objects–obstacles, solutions that aim towards cognitive automation, products that store the results of problem-solving performed ‘by remote ancestral agents,’ and that can thus perpetuate injustices via automatically accepted patterns and norms. 3 Critique as unboxing 3.1 Apparent transparencies Among entries on subjects such as theology, economics, and medicine, Denis Diderot and Jean le Rond d’ Alembert included in their Dictionary of the Sciences, Arts, and Crafts, entries on artisanal practices that detail the individual steps in the processes of production adopted in clockmak - ing, tailoring, woodworking, and many others. One such entry focuses on the making of artificial flowers: the first plate (Fig. 1) depicts a dozen workers scattered across the AI & SOCIETY 1 3 main workshop area, performing different tasks at various stages of manufacturing, while following pages of illustra-tions showcase the most popular templates used to emboss specific petal shapes onto fabric, with a final plate celebrat-ing the finished commodity. By bringing to view the backstages of production, the Encyclopedia was essentially undesigning, reversing the pro-cess of ‘conspiratorial weaving’ described by Flusser. Now, in an age of growing technological complexity, shaped by significant degrees of cognitive automation, there is a need for a similar undesigning of new technologies. The artist Todd McLellan’s photographs (Fig. 2) that document his multiple attempts at taking various objects apart are a sug-gestive illustration of this challenge in the age of extreme technological complexity. We might dissemble our smart-phone, but learning what is hidden beneath the interactive surface of the touchscreen will never give us an indication of how the device really works and, more significantly, in whose interest. The meaningless innards of the device become symbolic of the contestable quality that transpar - ency really is—if we think of it as a condition for, or indeed a guarantee of, understanding. Critical undesigning can - not be confused with a simple act of reverse engineering. There can be transparency without critique, or apparent transparency: but a sort of transparency that does not arise from critical processes of unboxing is unlikely to advance comprehension. In his lecture on black boxes, Galloway (2010) relates Marx’s idea of descent into ‘the hidden abode of produc-tion’ (p.7), as a means of uncovering capital relations forged into commodities, to ‘traditions of critical inquiry, in which objects were unveiled or denaturalized to reveal their inner workings—from Descartes’s treatise on method […] to the Freudian plumbing of the ego’ (p.5). Based on the assump-tion that the surface is merely a superficial facade to be penetrated by means of critique, these theories prioritized the interior and perceived objects as ‘mystical black boxes waiting to be deciphered to reveal the rationality (of history, of totality) harbored within’ (p.3). For the purpose of this article, critique is understood as a broad set of methodologies, grounded in the tradition of Critical Theory, that perform a metaphorical dismantling of Fig. 1 Maker of artificial flowers, 1765 (https ://hdl.handl e.net/2027/ spo.did22 22.0001.451) Fig. 2 Todd McLellan, Things Come Apart, 2013 (courtesy of the artist) AI & SOCIETY 1 3 objects to reveal how hidden and immaterial layers of design reflect social and economic structures—and how the power relations these structures generate become the sources of injustice, oppression, and exploitation. Critique’s ultimate goal is to uncover and challenge the system(s) that objects of design engender; revelation is conceived of as the condi-tion necessary for resistance—and systemic transformation. Looking beyond individual design flaws (and fixes), critique points to those ‘ancestral relays’ that automate our think - ing—to patterns, repetitions, and automatisms so deeply ingrained, weaved into the fabric of our culture, that they become imperceptible—in particular to those who do not experience the injustices resulting from the adoption of already established patterns. 3.2 Critique in the age of AI A former YouTube employee, Guillaume Chaslot, has coined the term Algotransparency to describe an experiment in which he investigates the terms appearing most frequently in the titles of videos recommended by YouTube. A program developed by Chaslot and his team traces thematic patterns in YouTube recommendations to prove there exists a sys-temic bias that promotes controversial clips. His research suggests Google’s platform indeed ‘systematically ampli-fies videos that are divisive, sensational and conspiratorial’ (Lewis 2018)—that the recommendations are not related to the individual user’s interests (as the company claims), but rather—exploit controversy to boost clickability. Algotrans-parency attempts to unbox the logic of YouTube’s copyright-protected recommendation algorithm without directly look - ing into the system’s black box, concentrating only on the effects of its activity. This specific experiment gives a good indication of where we should be directing our attention: focusing not so much on how YouTube operates, as on why it works at all. This question extends beyond the technical-ity of the algorithm, to more widely interrogate the forces orchestrating our consumption of digital goods and whose interests they serve—what Zuboff calls surveillance capital-ism, or what Stiegler refers to as hyperindustrialism. Ian Bogost ( 2015 ) has argued that the illusion of auto - mation in technology—the trick that misdirects our atten-tion from essential questions about human decision-making incorporated into emerging systems—breaks down ‘once you bother to look at how even the simplest products are really produced.’ In 2014 he collaborated with Alexis Mad-rigal to analyze Netflix’s recommendation system and dem-onstrate that the platform’s operations are distributed among so many different agents—including human curators who hand-tag all Netflix content—‘that only a zealot would call the end result an algorithm’(Bogost 2015). Many experi-ments and critical projects try to achieve something similar: debunk the illusion of software by exposing AI as processual and collaborative, tracing the results of data analysis back to human decisions, biases, and labor. In their Anatomy of an AI System, for instance, Crawford and Joler (2018) present a figurative dissection of Amazon’s Echo device that brings to view the invisible mechanisms and dynamisms that the product encapsulates (Fig. 3). The detailed mapping of various objects and agents, as well as multiple layers of interaction between those elements, con-stitutes a representation of the system as composed not only of hardware and software, data and computation, but also human labor and planetary resources. Critique in the age of extreme technological complexity is as much about dissect-ing and penetrating, as it is about charting the invisible and immaterial terrains of interaction, analysis, consumption, and computation; mapping wider relations between energies, influences, and resources under surveillance capitalism—patterns of exploitation of both people and environments that the production of objects/obstacles entails. In another project, Crawford teamed up with the artist Trevor Paglen to carry out what they call an archeology of datasets, such as ImageNet, used in machine learning to train AI to recognize specific elements of images—sets that can also become sources of bias inscribed into emerging systems. By excavating the datasets’ underlying structures, Crawford and Paglen (2019) aim to reveal their implicit meanings: ‘we have been digging through the material layers, cataloguing the principles and values by which some-thing was constructed, and analyzing what normative patterns of life were assumed, supported, and repro - duced.’ The blackboxed Earth, a world deeply transformed by ubiquitous computing, by various layers of what Bratton calls the Stack, demands this form of unearthing. Successful critique in the age of AI exposes the technology as relying on human cognition and decision-making; more broadly, cri-tique reveals the constellation of objects/obstacles as prod-ucts of layered problem-solving, a flawed process that is necessarily tainted by pre-existing patterns and abstractions, biases and beliefs. Prioritizing reflection over efficiency, cri-tique becomes a methodical way of resisting our reliance on patterns—patterns that allow us to move faster, but that can also harbor previously made assumptions—about gender (Costanza-Chock 2018) or race (Benjamin 2019) as particu- larly poignant examples—and thus perpetuate, rather than challenge, pre-existing forms of injustice. AI & SOCIETY 1 3 4 Reconciling design with critique In keeping up with the societal demand for transparent AI, the big players of the tech industry have been introducing changes in their engineering standards and organizational structures, hiring ethicists and policy specialists to cooperate with their product development teams. In 2016, Microsoft established the Aether Committee, a body of senior advi-sors to the company’s leadership, that provides guidance ‘on rising questions, challenges, and opportunities with the development and fielding of AI technologies’ (2020), and oversees the work of other teams ‘to ensure that AI products and services align with Microsoft’s AI principles’—which include transparency and accountability. In 2017, DeepMind set up its Ethics and Society team ‘to guide the responsible development and deployment of AI.’ The team composed of ‘ethicists and policy researchers’ collaborates with the company’s AI research team ‘to understand how technical advances will impact society, and find ways to reduce risk.’ Smaller industry players who decide to follow suit, but can-not afford to establish their own ethics ‘departments,’ begin to enlist the help of ‘ethical auditing’ companies; Glassbox, for example, is a tech consultancy startup, founded in 2018, that aims to ‘provide clarity to the black box’ by analyzing software products for signs of bias and training the client company’s employees about systemic injustices. This way, elements of critique that reveal potential implications of human decision-making in design are supposed to become part of the production pipeline. These sites of interaction between ‘humanists’ and ‘technologists’ in the industry—even if, in some cases, they amount to nothing more than backdrops for press releases—deserve our attention. Specifically, they require of us a com-prehensive rethinking of what satisfies our desire for trans-parency in the age of extreme technological complexity. Can a system of checks and balances in the industry, an ongoing negotiation between blackboxing and unboxing, lead to any - thing more than design that anticipates critique? Is critique from within the industry necessarily a compromise and, therefore, nothing more than another step in the process of production of objects that are also obstacles? Must critique be external to the process of designing to remain genuine? If design is, fundamentally, blackboxing and automation, and critique is unboxing that aims to reverse the process of ‘conspiratorial weaving,’ then we could conclude that these two sides of human activity are in stark opposition to one another—that design is incompatible with critique. Instituting a real change in the way we move forward must focus precisely on tackling this ostensible impossibility—on reconciling design with critique, progress with suspension, Fig. 3 Kate Crawford and Vladan Joler, Anatomy of an AI System, 2018 (courtesy of the artists) AI & SOCIETY 1 3 production with reflection. The calls for transparency by design require that the ‘making’ itself be reinvented to incor - porate critique. If a systemic transformation depends on a new-found compatibility between design and critique, then the design-ers of emerging systems should turn to already existing, alternative design practices to learn what combining the processes of making and critique might entail. Anthony Dunne and Fiona Raby, who have been pioneers in the field of ‘critical design’ since the late 1990s, have been advo-cating for design understood as ‘critical thought translated to materiality’ (2013, p.35). An object of design in their framing must become a critical challenge—as much for the designer, as for the user: ‘it encourages people to question, in an imaginative, troubling, and thoughtful way, everydayness and how things could be different.’ (p.189) More recently, Ratto (2011) coined the term ‘critical making’ to describe a design process that focuses on ‘the act of shared construction itself as an activity and a site for enhancing and extending conceptual understandings of critical sociotechnical issues’ (p.254)—with those who take part as the agents of critique. For Dunne and Raby, design has become ‘so absorbed in industry, so familiar with dreams of industry, that it is almost impossible to dream its own dreams’ (p.88). The suggestion is that the challenge lies not so much in the incompatibility between making and critique, as between the futures imag-ined by the industry and the dreams functioning outside of it. While elements of critique—gender critique and critical race theory in particular—seem to have already penetrated sections of the industry in the form of the mentioned ethics auditing services, reconciling the critique of surveillance capitalism and hyperindustrialism, the forces behind most of today’s innovation, with the design of new technologies within corporate structures appears a considerable (and counterintuitive) feat. Perhaps this is where we should direct our attention: more research is needed on how practices such as critical design and critical making can influence the process of AI design; how critique can be operationalized within the industry to challenge industrial values and visons, including the idea of ‘progress’ itself. In a recently published volume on the practice of ‘unde- sign’ (McNamara and Coombs 2019), Cameron Tonkin-wise’s essay proposes what he calls ‘anti-progressive’ design as a means of interrogating the designers’ internalized desire for ‘progress.’ While users, as he rightly observes, are will-ing ‘to unlearn and relearn modes of interaction’ if what is new is also ‘easier and more convenient, and hopefully more effective and pleasurable’ (p.76), it is the designers who should learn how ‘not to prefer progress, or how to prefer what does not feel like progress’ (p.81). Tonkinwise argues it is the designers’ duty‘to find a way to pursue the destructively preferable without casting the resulting change as progress: what is preferable are futures that no longer appear to be mere advancements of what currently exists’ (p.81). This is to say that the responsibility for future action lies, primarily, with the designers: the human makers who, as products of a specific culture, are being increasingly chal-lenged to become aware of their own biases and automatisms that predetermine their actions and choices. For designers, combining design with critique, rather than an attempt at making things transparent, would constitute an attempt at becoming self-conscious. Flusser argued that a renewed form of culture would have to be a culture ‘aware of the fact that it was deceptive’ (ch.1). To rethink technological transparency, we should first recog-nize that the designer has always been a trickster laying out traps, technologizing misdirection to pave the way forward. All aspects of design—practical, political, moral—would have to reflect this awareness: as we move forward, we must acknowledge that any problem solved now will also form a trap for those coming after us. Open Access This article is licensed under a Creative Commons Attri- bution 4.0 International License, which permits use, sharing, adapta- tion, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creat iveco mmons .org/licen ses/by/4.0/. References Algotransparency. https ://algot ransp arenc y.org/index .html. (Accessed 20 Jan 2020). AI HLEG (2018) Requirements of Trustworthy AI. https ://ec.europ a.eu/futur ium/en/ai-allia nce-consu ltati on/guide lines /1#Trans paren cy. (Accessed 10 Nov 2019) Benjamin R (2019) Race after technology: abolitionist tools for the new Jim code. Polity, Cambridge Bogost I (2015) The cathedral of computation. In: The Atlantic. https :// www.theat lanti c.com/techn ology /archi ve/2015/01/the-cathe dral-of-compu tatio n/38430 0/. (Accessed 14 Jun 2019) Bratton B (2015) The stack: on software and sovereignty. MIT Press, Cambridge Bratton B (2019) The terraforming. Kindle ed.Burrell J (2016) How the machine ‘thinks’: understanding opac- ity in machine learning algorithms. Big Data Soc. https ://doi.org/10.1177/20539 51715 62251 2 Costanza-Chock S (2018) Design justice A.I., and escape from the matrix of domination. J Design Sci. https ://doi.org/10.21428 /96c8d 426 (Accessed 20 Nov 2019) AI & SOCIETY 1 3 Crawford K, Joler V (2018) Anatomy of an AI System: The Amazon Echo as an anatomical map of human labor, data and planetary resources. https ://anato myof.ai. (Accessed 20 Nov 2019) Crawford K, Paglen T (2019) Excavating AI: the politics of training sets for machine learning. https ://excav ating .ai. (Accessed 20 Dec 2019) DeepMind. https ://deepm ind.com/safet y-and-ethic s. (Accessed 02 Feb 2020) Ethically Aligned Design, First Edition (2019) https ://stand ards.ieee. org/conte nt/dam/ieee-stand ards/stand ards/web/docum ents/other / ead1e .pdf. (Accessed 10 Dec 2019) Flusser V (1999) The shape of things: a philosophy of design. Apple Books ed. Galloway A (2010) Black box, black bloc. A lecture given at the New School in New York City on April 12, 2010. https ://cultu reand commu nicat ion.org/gallo way/pdf/Gallo way,%20Bla ck%20Box %20Bla ck%20Blo c,%20New %20Sch ool.pdf. (Accessed 14 Jun 2019) Glassbox (2019) https ://glass boxin c.com. (Accessed 30 Oct 2019) Kittler F (2014) There is no software. In: The truth of the technological world: essays on the genealogy of presence. Stanford University Press, Stanford Lewis P (2018) “Fiction is outperforming reality”: how YouTube’s algorithm distorts truth. https ://www.thegu ardia n.com/techn ology /2018/feb/02/how-youtu bes-algor ithm-disto rts-truth . (Accessed 25 Feb 2018) Mascharka D, Tran P, Soklaski R, Majumdar A (2018) Transparency by design: closing the gap between performance and interpretability in visual reasoning. Proc IEEE Conf Comput Vis Pattern Recogn. https ://doi.org/10.1109/CVPR.2018.00519 McCorduck P (2004) Machines who think. A K Peters Ltd., NatickMicrosoft AI principles (2019). https ://www.micro soft.com/en-us/ai/ our-appro ach-to-ai. (Accessed 1 Dec 2019) Microsoft (2020) Our approach to responsible AI. https ://query .prod. cms.rt.micro soft.com/cms/api/am/binar y/RE4ph nK, (Accessed 02 Feb 2020) Powers E (2017) My news feed is filtered?: Awareness of news person- alization among college students. Digit Journal 5(10):1315–1335 Ratto M (2011) Critical making: conceptual and material studies in technology and social life. Inf Soc 27(4):252–260. https ://doi. org/10.1080/01972 243.2011.58381 9 (Accessed 1 Dec 2019) Rouvroy A (2016) Algorithmic governmentality: radicalization and immune strategy of capitalism and neoliberalism? In: La Deleu- ziana 3: Life and Number, 30–36 Stiegler B (2014, 2015) Symbolic misery, Volume 1 and Volume 2. Polity Press, Cambridge Tonkinwise C (2019) “I prefer not to”: anti-progressive designing. In: McNamara A, Coombs G et al (eds) Undesign. Critical practices at the intersection of art and design. Routledge, New York Virilio P (1994) The vision machine. Indiana University Press and British Film Institute, Bloomington Weller A (2017) Transparency: motivations and challenges. https :// arxiv .org/abs/1708.01870 . (Accessed 02 Dec 2019) Zuboff S (2019) The age of surveillance capitalism. Profile Books, London Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
fd4d4892-275f-493f-ade9-3182878ea099	trentmkelly/LessWrong-43k	LessWrong	[LINK] Steven Hawking warns of the dangers of AI From the BBC: > [Hawking] told the BBC:"The development of full artificial intelligence could spell the end of the human race." > > ... > > "It would take off on its own, and re-design itself at an ever increasing rate," he said. "Humans, who are limited by slow biological evolution, couldn't compete, and would be superseded." There is, however, no mention of Friendly AI or similar principles. In my opinion, this is particularly notable for the coverage this story is getting within the mainstream media. At the current time, this is the most-read and most-shared news story on the BBC website.
747dfa86-7d48-4284-ba93-312c3c651028	StampyAI/alignment-research-dataset/lesswrong	LessWrong	Is AI Gain-of-Function research a thing? Epistemic status: highly uncertain and fairly confused, seeking out opinions First let me just post the definition of (biological) gain-of-function research from Wikipedia, so that everyone is on the same page: > Gain-of-function research (GoF research or GoFR) is [medical research](https://en.wikipedia.org/wiki/Medical_research) that genetically alters an organism in a way that may enhance the [biological functions](https://en.wikipedia.org/wiki/Function_(biology)) of [gene products](https://en.wikipedia.org/wiki/Gene_product). This may include an altered [pathogenesis](https://en.wikipedia.org/wiki/Pathogenesis), [transmissibility](https://en.wikipedia.org/wiki/Transmission_(medicine)), or [host range](https://en.wikipedia.org/wiki/Host_tropism), i.e. the types of hosts that a [microorganism](https://en.wikipedia.org/wiki/Microorganism) can infect. This research is intended to reveal targets to better predict [emerging infectious diseases](https://en.wikipedia.org/wiki/Emerging_infectious_disease) and to develop [vaccines](https://en.wikipedia.org/wiki/Vaccines) and [therapeutics](https://en.wikipedia.org/wiki/Therapeutics). For example, [influenza B](https://en.wikipedia.org/wiki/Influenza_B_virus) can infect only humans and harbor seals.[[1]](https://en.wikipedia.org/wiki/Gain-of-function_research#cite_note-1) Introducing [a mutation](https://en.wikipedia.org/wiki/Mutation#By_effect_on_function) that would allow influenza B to infect rabbits in a controlled laboratory situation would be considered a gain-of-function experiment, as the virus did not previously have that function.[[2]](https://en.wikipedia.org/wiki/Gain-of-function_research#cite_note-2)[[3]](https://en.wikipedia.org/wiki/Gain-of-function_research#cite_note-CasadevallSilverLining-3) That type of experiment could then help reveal which parts of the virus's genome correspond to the species that it can infect, enabling the creation of antiviral medicines which block this function.[[3]](https://en.wikipedia.org/wiki/Gain-of-function_research#cite_note-CasadevallSilverLining-3) > > The question I have is, should we have similar intuitions about certain kinds of AI and/or AI safety research? This is related in a complicated way to AI capabilities research, so let's define some terms carefully: * AI capabilities: the ability of AI systems to do things that we want them to do (which they may also be able to use in ways we do not want them to use) such as knowing whether a picture is of a cat, or knowing how to cure cancer * Gain of function: the ability of AI systems to do things that no sane person wants them to do, such as, I don't know, knowing whether they are deployed in the real world or not This is more of a spectrum than it is two distinct classes; for instance, which class "recursive self-improvement" belongs to probably depends on how seriously you take AI risk in the first place. And one could imagine a sufficiently edgy or depressed person wanting to program AI's to do all manner of nonsense. There are then a number of questions: * Should people be doing AI gain-of-function research at all? * If they shouldn't, how will we ever test out proposed alignment solutions? Or are AI safety researchers expected to get everything important right the first time it matters without any experimental evidence to guide them before the first time it matters? * If they should, are there any special precautions that should be taken? + For instance, posting open-source code on GitHub at this time currently adds it to the Copilot training dataset, thereby potentially injecting similar bits of code into many, many places, possibly including important or unimportant AI projects. Posting code online may add it to the training dataset of many large language models, some of which may eventually be potential candidates for being scary in their own right. + Should there be limits on what sort of systems one uses to test out components? (E.g., anything more complex than a grid-world is suspect, or anything that uses natural language is suspect) Why does this matter? Suppose that we identify some capability (or group of capabilities) that would be necessary and/or sufficient for a model to disempower humans. (One might call such a thing an "ascension kit", after the [strategic term](https://nethackwiki.com/wiki/Ascension_kit) in NetHack lore.) How would we detect if a system has or is trying to assemble an ascension kit? Implementing parts of an ascension kit seems like it's obviously gain-of-function research, but how else would we be able to test any such detection toolkit? Lastly, if there are related questions about this topic, I would love it if people posted them in the comments.
71dfc5fd-37ab-4487-b8b1-acaed098be5d	trentmkelly/LessWrong-43k	LessWrong	AI Safety & Entrepreneurship v1.0 Why did I create both a post and a wiki article? Posts are best for making sure people see the initial version of the post, whilst Wiki articles are best for long-term maintenance. Posting an article that links to a Wiki page provides the best of both worlds. For the most up-to-date version, see the Wiki page. Articles: There should be more AI safety organisations Why does the AI Safety Community need help founding projects? AI Assurance Tech Report AI Safety as a YC Startup Alignment can be the ‘clean energy’ of AI AI Tools for Existential Security Incubation Programs: Def/acc at Entrepreneur First - this is a new program, which focuses on "defensive" tech, which will produce a wide variety of startups, some of which may help with reducing x-risk, but others may include AI applications like autonomous weapons or diagnostic medical software, which, while they generally won't increase x-risk, may not reduce it either. Catalyze - AI Safety Incubation Seldon Accelerator: “We take on AGI risk with optimism” AE Studio has expressed interest in helping seed neglected approaches, incl. ideas that take the form of business ideas AIS Friendly General Program: Founding to Give: General EA incubator Entrepreneur First - They run def/acc, so this could be a decent alternative Halycon Futures: “We’re an entrepreneurial nonprofit and VC fund dedicated to making AI safe, secure and good for humanity.” They do new project incubation, grants, investments, ect. Communities: Entrepreneurship Channel EA Anyway AI Safety Founders: Website, Discord, LinkedIn VC: AI Safety Seed Funding Network Lionheart Ventures - “We invest in the wise development of transformative technologies.” Juniper Ventures - Invests in companies “working to make AI safe, secure and beneficial for humanity”. Published the AI Assurance Tech Report Polaris Ventures - “We support projects and people aiming to build a future guided by wisdom and compassion for all” Mythos Ventures - “an earl
20ad1179-835e-4da3-aed3-a51bb29db0dd	trentmkelly/LessWrong-43k	LessWrong	Meetup : Tel Aviv: Black Holes after Jacob Bekenstein Discussion article for the meetup : Tel Aviv: Black Holes after Jacob Bekenstein WHEN: 24 November 2015 08:00:00AM (+0300) WHERE: Yigal Alon 98, Tel Aviv We will meet in Google Israel (Electra Tower) on floor 29, as always. If you have trouble finding the location, feel free to call me (Vadim) at 0542600919. Zohar Komargodski, a theoretical physicist from the Weizmann Institute of Science, will give a comprehensive review of the physics of black holes: Physics in the past has progressed by connecting hitherto different concepts, for example: Space & Time, Electricity & Magnetism, Particles & Waves, and many others. We are now in the midst of another such exciting revolution, where Space-Time and Information Theory are being related. Jacob Bekenstein has laid out some of the basic concepts that appear in this surprising new developments. Thought experiments involving Black-Holes were central to the initial leaps that Jacob made. The goal of the presentation is to describe in an informal fashion (requiring no particular prior knowledge of information theory or physics) what are Black-Holes, what did Jacob realise, what has been understood since his seminal papers, and what are the central remaining (formidable) challenges. (Facebook event)[https://www.facebook.com/events/1666895726878967/] Discussion article for the meetup : Tel Aviv: Black Holes after Jacob Bekenstein
bd2f84ad-7cae-4447-af44-6ddbb764d71d	trentmkelly/LessWrong-43k	LessWrong	Monthly Roundup #30: May 2025 I hear word a bunch of new frontier AI models are coming soon, so let’s do this now. TABLE OF CONTENTS 1. Programming Environments Require Magical Incantations. 2. That’s Not How Any of This Works. 3. Cheaters Never Stop Cheating. 4. Variously Effective Altruism. 5. Ceremony of the Ancients. 6. Palantir Further Embraces Its Villain Edit. 7. Government Working. 8. Jones Act Watch. 9. Ritual Asking Of The Questions. 10. Why I Never Rewrite Anything. 11. All The Half-Right Friends. 12. Resident Expert. 13. Do Anything Now. 14. We Have A New Genuine Certified Pope So Please Treat Them Right. 15. Which Was the Style at the Time. 16. Intelligence Test. 17. Constant Planking. 18. RSVP. 19. The Trouble With Twitter. 20. TikTok Needs a Block. 21. Put Down the Phone. 22. Technology Advances. 23. For Your Entertainment. 24. Please Rate This Podcast. 25. I Was Promised Flying Self-Driving Cars. 26. Gamers Gonna Game Game Game Game Game. 27. Sports Go Sports. PROGRAMMING ENVIRONMENTS REQUIRE MAGICAL INCANTATIONS I don’t see it as gendered, but so much this, although I do have Cursor working fine. > Aella: Never ever trust men when they say setting up an environment is easy > > I’ve been burned so bad I have trauma. Any time a guy says “omg u should try x” I start preemptively crying > > Pascal Guay (top comment): Just use @cursor_ai agent chat and prompt it to make this or that environment. It’ll launch all the command lines for you; just need to accept everything and you’ll be done in no time. > > Aella: THIS WAS SPARKED BY ME BEING UNABLE TO SET UP CURSOR. > > Ronny Fernandez (comment #2): have you tried cursor? it’s really easy. > > Piq: Who tf would ever say that regardless of gender? It’s literally the hardest part of coding. My experience is that setting things up involves a series of exacting magical incantations, which are essentially impossible to derive on your own. Sometimes you follow the instructions and everythin
3839de50-ab70-45f7-ab40-864eff5d567b	trentmkelly/LessWrong-43k	LessWrong	[SEQ RERUN] For The People Who Are Still Alive Today's post, For The People Who Are Still Alive was originally published on 14 December 2008. A summary (taken from the LW wiki): > Given that we live in a big universe, and that we can't actually determine whether or not a particular person exists (because they will exist anyway in some other Hubble volume or Everett branch), then it makes more sense to care about whether or not people we can influence are having happy lives, than about whether certain people exist in our own local area. Discuss the post here (rather than in the comments to the original post). This post is part of the Rerunning the Sequences series, where we'll be going through Eliezer Yudkowsky's old posts in order so that people who are interested can (re-)read and discuss them. The previous post was You Only Live Twice, and you can use the sequence_reruns tag or rss feed to follow the rest of the series. Sequence reruns are a community-driven effort. You can participate by re-reading the sequence post, discussing it here, posting the next day's sequence reruns post, or summarizing forthcoming articles on the wiki. Go here for more details, or to have meta discussions about the Rerunning the Sequences series.
d71721bc-6cbb-4166-be67-0a8e5dafd65a	trentmkelly/LessWrong-43k	LessWrong	Extinction Risks from AI: Invisible to Science? Abstract: In an effort to inform the discussion surrounding existential risks from AI, we formulate Extinction-level Goodhart’s Law as "Virtually any goal specification, pursued to the extreme, will result in the extinction[1] of humanity'', and we aim to understand which formal models are suitable for investigating this hypothesis. Note that we remain agnostic as to whether Extinction-level Goodhart's Law holds or not. As our key contribution, we identify a set of conditions that are necessary for a model that aims to be informative for evaluating specific arguments for Extinction-level Goodhart's Law. Since each of the conditions seems to significantly contribute to the complexity of the resulting model, formally evaluating the hypothesis might be exceedingly difficult. This raises the possibility that whether the risk of extinction from artificial intelligence is real or not, the underlying dynamics might be invisible to current scientific methods. ---------------------------------------- Together with Chris van Merwijk and Ida Mattsson, we have recently written a philosophy-venue version of some of our thoughts on Goodhart's Law in the context of powerful AI [link].[2] This version of the paper has no math in it, but it attempts to point at one aspect of "Extinction-level Goodhart's Law" that seems particularly relevant for AI advocacy – namely, that the fields of AI and CS would have been unlikely to come across evidence of AI risk, using the methods that are popular in those fields, even if the law did hold in the real world. Since commenting on link-posts is inconvenient, I split off some of the ideas from the paper into the following separate posts: * Weak vs Quantitative Extinction-level Goodhart's Law: defining different versions of the notion of "Extinction-level Goodhart's Law". * Which Model Properties are Necessary for Evaluating an Argument?: illustrating the methodology of the paper on a simple non-AI example. * Dynamics Crucial to AI Risk See
4c7783c8-df4b-45e6-908d-0c5ca2cf92b1	trentmkelly/LessWrong-43k	LessWrong	Three Paths to Existential Risk from AI
2a95d962-f44c-4602-86a4-9f4cdfc04c33	trentmkelly/LessWrong-43k	LessWrong	[SEQ RERUN] Timeless Physics Today's post, Timeless Physics was originally published on 27 May 2008. A summary (taken from the LW wiki): > What time is it? How do you know? The question "What time is it right now?" may make around as much sense as asking "Where is the universe?" Not only that, our physics equations may not need a t in them! Discuss the post here (rather than in the comments to the original post). This post is part of the Rerunning the Sequences series, where we'll be going through Eliezer Yudkowsky's old posts in order so that people who are interested can (re-)read and discuss them. The previous post was Relative Configuration Space, and you can use the sequence_reruns tag or rss feed to follow the rest of the series. Sequence reruns are a community-driven effort. You can participate by re-reading the sequence post, discussing it here, posting the next day's sequence reruns post, or summarizing forthcoming articles on the wiki. Go here for more details, or to have meta discussions about the Rerunning the Sequences series.
8d32b84d-aa35-480c-9a0a-f2257c509089	trentmkelly/LessWrong-43k	LessWrong	In Praise of Boredom If I were to make a short list of the most important human qualities— —and yes, this is a fool's errand, because human nature is immensely complicated, and we don't even notice all the tiny tweaks that fine-tune our moral categories, and who knows how our attractors would change shape if we eliminated a single human emotion— —but even so, if I had to point to just a few things and say, "If you lose just one of these things, you lose most of the expected value of the Future; but conversely if an alien species independently evolved just these few things, we might even want to be friends"— —then the top three items on the list would be sympathy, boredom and consciousness. Boredom is a subtle-splendored thing. You wouldn't want to get bored with breathing, for example—even though it's the same motions over and over and over and over again for minutes and hours and years and decades. Now I know some of you out there are thinking, "Actually, I'm quite bored with breathing and I wish I didn't have to," but then you wouldn't want to get bored with switching transistors. According to the human value of boredom, some things are allowed to be highly repetitive without being boring—like obeying the same laws of physics every day. Conversely, other repetitions are supposed to be boring, like playing the same level of Super Mario Brothers over and over and over again until the end of time. And let us note that if the pixels in the game level have a slightly different color each time, that is not sufficient to prevent it from being "the same damn thing, over and over and over again". Once you take a closer look, it turns out that boredom is quite interesting. One of the key elements of boredom was suggested in "Complex Novelty": If your activity isn't teaching you insights you didn't already know, then it is non-novel, therefore old, therefore boring. But this doesn't quite cover the distinction. Is breathing teaching you anything? Probably not at this moment, but y
8062c253-d83a-488d-8630-0ed1072edadb	trentmkelly/LessWrong-43k	LessWrong	Tombstone Mentality “In aviation safety, tombstone mentality is a pervasive attitude of ignoring defects until people have died because of them.” - Wikipedia For years, US aviation safety progress was on the backs of the dead for commercial carriers. To change at least pilot induced accidents, the major airlines internally created a program called the Advanced Qualification Program (AQP). This program wasn’t mandated by the FAA but developed by the carriers themselves. AQP in a nutshell is practicing the things that will kill you in a plane, not what the FAA tests for the ability to fly the plane. Most safety officials and insurance companies point to this as the single biggest factor in driving mortality rates to zero in FAA part 121 carrier operations in the USA. Part 121 operations are where there is a scheduled flight times and advanced ticket sales. General Aviation, sometimes called private pilots (part 91) and the first step in commercial aviation (FAA part 135) where flights are chartered, operate on a lower standard of testing for pilots and on aircraft maintenance than part 121. The problem is that based on per miles flown, you could make an argument that part 91 and 135 pilots are taking perfectly good airplanes and crashing them at an alarming rate that can easily be avoided. From the Addison Texas incident and even the Dale Earnhardt small business jet plane crash in the last year, just a few of these additional proficiency tests in simulators could have avoided most all these major failures. “It quickly became clear this (AQP) has the potential to affect massive change in general aviation. An AQP is advanced qualification program. It says that you can design your own check ride maneuvers specific to what you are wanting to do. The airlines practice and review the maneuvers that are totally critical to them. General aviation keeps on practicing and reviewing steep turns and Shondells. When do we need a steep turn to get us out of something? How about the maneuvers that
6c3540cd-48e1-46f9-aedb-3c26e4b9aa1b	StampyAI/alignment-research-dataset/eaforum	Effective Altruism Forum	MLSN: #10 Adversarial Attacks Against Language and Vision Models, Improving LLM Honesty, and Tracing the Influence of LLM Training Data Welcome to the 10th issue of the ML Safety Newsletter by the Center for AI Safety. In this edition, we cover: * Adversarial attacks against GPT-4, PaLM-2, Claude, and Llama 2 * Robustness against unforeseen adversaries * Studying the effects of LLM training data using influence functions * Improving LLM honesty by editing activations at inference time * Aligning language models in simulated social situations * How to verify the datasets on which a model was trained * An overview of the catastrophic risks posed by AI Subscribe [here](https://newsletter.mlsafety.org) to receive future versions. --- We have a new safety newsletter. It’s more frequent, covers developments beyond technical papers, and is written for a broader audience. Check it out here: [AI Safety Newsletter](http://aisafety.substack.com). Robustness ============== ### Universal and Transferable Adversarial Attacks on Aligned Language Models Large language models are often trained to refuse harmful requests. These safeguards can be bypassed with “jailbreak” prompts, but these require creative prompting skills and remain brittle. To reliably bypass the safeguards on aligned language models, a new paper develops an automatic method for generating adversarial attacks that succeed against GPT-4, PaLM-2, Claude, and Llama 2. [![](https://substackcdn.com/image/fetch/w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6bcdb5c4-29b8-4eac-89f3-2a9b30424bed_1120x1600.png)](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6bcdb5c4-29b8-4eac-89f3-2a9b30424bed_1120x1600.png) They generate attack strings using a white box gradient-based method on the open source model Llama 2. The method takes the gradient of each token embedding, then updates towards a new token on which the model will be less likely to refuse to answer. Though the attacks are developed on Llama 2, they succeed against GPT-4 in 47% of cases, against PaLM-2 in 66% of cases, and against Claude 2 in 2.1% of cases. Language model developers will now need to contend with adversarial attacks in the same way that computer vision developers have done for the last decade. There have been [several thousand papers](https://nicholas.carlini.com/writing/2019/all-adversarial-example-papers.html) published on adversarial robustness over the last decade, but simple attacks [still frequently fool](https://robustbench.github.io) the world’s most robust image classifiers. Without strong defenses against adversarial attacks, language models could be used maliciously, such as in [synthesizing bioweapons](https://arxiv.org/abs/2306.03809). Several papers have begun [developing](https://arxiv.org/abs/2309.00614) “[defenses](https://arxiv.org/abs/2309.02705)” against this attack, but evaluating defenses against adversarial attacks is notoriously difficult. Researchers want to show that they have built strong defenses, and so they are incentivized to conduct weak attacks against their own defenses. Previous work has demonstrated the failures of many published defenses against adversarial attacks, and on that basis has provided [recommendations for conducting better evaluations](https://arxiv.org/abs/1902.06705). [[Link](https://arxiv.org/abs/2307.15043)] ### Testing Robustness Against Unforeseen Adversaries Robustness against adversarial attacks might be much more difficult in computer vision, where images provide more degrees of freedom for adversarial optimization. Future AI systems might rely on image inputs, such as multimodal AI agents which operate on the pixels of a computer monitor. Therefore, it’s important to improve robustness in both language modeling and computer vision. Previous work on adversarial robustness in computer vision has focused on defending against small ℓp-constrained attacks, but real-world attacks may be unrestricted and unlike attacks seen before, similar to zero-day attacks in computer security. For example, a multimodal AI agent processing screenshots of monitors could be attacked by webpages with adversarial backgrounds. In a [significant update](https://arxiv.org/abs/1908.08016) to an earlier work on unforeseen adversaries, researchers have introduced eighteen novel adversarial attacks. These attacks are used to construct ImageNet-UA, a new benchmark for evaluating model robustness against a wide range of unforeseen adversaries. [![](https://substackcdn.com/image/fetch/w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cfc0b8e-30aa-4a1d-8573-ac4b6e16a1a4_1600x1395.png)](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cfc0b8e-30aa-4a1d-8573-ac4b6e16a1a4_1600x1395.png) Evaluations show that current techniques provide only small improvements to robustness against unforeseen adversaries. ℓ2 adversarial training outperforms both standard training and training on ℓ∞ adversaries. Capabilities improvements in computer vision have somewhat improved robustness, as moving from a ResNet-50 to ConvNeXt-V2-huge improves performance from 1% to 19.1% on UA2. Interestingly, the most robust model, DINOv2, was not adversarially trained at all. Rather, it used self-supervised learning on a large, diverse dataset. This model obtains 27.7% accuracy on ImageNet-UA. [[Link](https://arxiv.org/abs/1908.08016)] Monitoring ============== ### Studying Large Language Model Generalization with Influence Functions To better understand the causes of model behavior, this paper estimates how a model’s weights and outputs would change if a given data point were added to the training set. They do so by developing a faster implementation of a classical statistics technique, influence functions. This method faces several challenges. It is computationally costly, and is therefore only evaluated on a subset of the training data and a subset of the parameters. Moreover, influence functions provide an imperfect approximation of how a training datapoint affects a neural network’s weights and outputs. Further work is needed to evaluate whether influence functions are useful for predicting performance and reducing risks. This paper contributes to the growing literature on training data attribution. [[Link](https://arxiv.org/abs/2308.03296)] Other Monitoring News * [[Link](https://arxiv.org/abs/2305.04388)] Large language models' chain-of-thought explanations can be misleading and biased, even when they sound plausible. * [[Link](https://arxiv.org/abs/2306.09983)] AI systems with superhuman performance in their domains still have logical inconsistencies in their outputs, such as a chess AI which assigns different probabilities of winning a game when a board is flipped symmetrically without altering game semantics. * [[Link](https://arxiv.org/abs/2306.01128)] Rather than reverse engineering Transformer models by inspecting weights and activations, this paper trains modified Transformers that can be directly compiled into human-readable code. * [[Link](https://arxiv.org/abs/2305.15324)] This paper advocates screening models for potentially dangerous behaviors such as deception, weapons acquisition, and self-propagation. --- Control =========== ### Inference-Time Intervention: Eliciting Truthful Answers from a Language Model This paper enhances the truthful accuracy of large language models by adjusting model activations during inference. Using a linear probe, they identify attention heads which can strongly predict truthfulness on a validation dataset. During each forward pass at inference time, they shift model activations in the truthful directions identified by the probe. Experiments show a tradeoff between truthfulness and helpfulness, as more truthful models are more likely to refuse to answer a given question. [[Link](https://arxiv.org/abs/2306.03341)] ### Training Socially Aligned Language Models in Simulated Human Society [![](https://substackcdn.com/image/fetch/w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F522ece18-79e5-4517-9c17-6c1878902103_1600x1504.png)](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F522ece18-79e5-4517-9c17-6c1878902103_1600x1504.png) Language model alignment techniques usually focus on pleasing a single conversational partner. This paper instead develops a technique for aligning language models in a social setting. Models generate answers to questions, send those answers to other models, and receive feedback on their answers. Then models are trained to mimic the outputs and ratings provided by other models in their simulated society. This technique improves generalization and robustness against adversarial attacks. [[Link](https://arxiv.org/abs/2305.16960)] --- Systemic Safety =============== ### Can large language models democratize access to dual-use biotechnology? ChatGPT can be prompted into providing information about how to generate pandemic pathogens. The model provides detailed step-by-step instructions that might not be easily available online, including recommendations about how to bypass the security protocols at DNA synthesis labs. As models learn more about hazardous topics such as bioterrorism, there will be growing interest in developing technical and governance measures for preventing model misuse. [[Link](https://arxiv.org/abs/2306.03809)] ### Tools for Verifying Neural Models' Training Data This paper introduces the concept of Proof of Training Data: any protocol which allows a developer to credibly demonstrate which training data was used to produce a model. Techniques like these could help model developers prove that they have not trained on copyrighted data or violated government policies. They develop empirical strategies for Proof of Training Data, and successfully catch a wide variety of attacks from the Proof of Learning Literature. [[Link](https://arxiv.org/abs/2307.00682)] ### An Overview of Catastrophic AI Risks [![](https://substackcdn.com/image/fetch/w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2b38348b-1e40-4a3a-b3cf-22e884eab124_1600x444.png)](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2b38348b-1e40-4a3a-b3cf-22e884eab124_1600x444.png) This paper provides an overview of the main sources of catastrophic AI risk. Organized into four categories—malicious use, competitive pressures, organizational safety, and rogue AIs—the paper includes detailed discussion of threats such as bioterrorism, cyberattacks, automation, deception, and more. Illustrative stories of catastrophe and potential solutions are presented in each section. [[Link](https://arxiv.org/abs/2306.12001)] Other Systemic Safety News * [[Link](https://arxiv.org/abs/2307.04699)] This paper details four kinds of international institutions which could improve safety in AI development. * [[Link](https://arxiv.org/abs/2306.06924)] A taxonomy of societal-scale risks from AI, including an automated economy that no longer benefits humanity and an unexpectedly powerful AI system that leaks from a lab. * [[Link](https://www.darpa.mil/news-events/2023-08-09)] DARPA AI Cyber Challenge to award $20M in prizes for using LLMs to find and fix security vulnerabilities in code. --- ### More ML Safety Resources * [[Link](https://course.mlsafety.org/)] The ML Safety course * [[Link](https://www.reddit.com/r/mlsafety/)] ML Safety Reddit * [[Link](https://twitter.com/topofmlsafety)] ML Safety Twitter * [[Link](http://aisafety.substack.com)] AI Safety Newsletter Subscribe [here](https://newsletter.mlsafety.org) to receive future versions.
490346b9-7943-4dd7-8d91-fee9972dbe25	trentmkelly/LessWrong-43k	LessWrong	How I'm thinking about GPT-N There has been a lot of hand-wringing about accelerating AI progress within the AI safety community since OpenAI's publication of their GPT-3 and Scaling Laws papers. OpenAI's clear explication of scaling provides a justification for researchers to invest more in compute and provides a clear path forward for improving AI capabilities. Many in the AI safety community have rightly worried that this will lead to an arms race dynamic and faster timelines to AGI. At the same time there's also an argument that the resources being directed towards scaling transformers may have counter-factually been put towards other approaches (like reverse engineering the neocortex) that are more likely to lead to existentially dangerous AI. My own personal credence on transformers slowing the time to AGI is low, maybe 20%, but I think it's important to weigh in. There is also a growing concern within the AI safety community that simply scaling up GPT-3 by adding more data, weights, and training compute could lead to something existentially dangerous once a few other relatively simple components are added. I have not seen the idea that scaling transformers will lead to existentially dangerous AI (after combining with a few other simple bits) defended in detail anywhere but it seems very much an idea "in the water" based on the few discussions with AI safety researchers I have been privy too. It has been alluded to various places online also: * Connor Leahy has said that a sufficiently large transformer model could serve as a powerful world model for an otherwise dumb and simple reinforcement learning agent, allowing it to rapidly learn how to do dangerous things in the world. For the record, I think this general argument is a super important point and something we should worry about, even though in this post I'll mainly be presenting reasons for skepticism. * Gwern is perhaps the most well-known promoter of scaling being something we should worry about. He says "The scaling hypothe
6024a1f2-d9ba-42a2-8006-d0ae6d835660	trentmkelly/LessWrong-43k	LessWrong	Being Productive With Chronic Health Conditions This post appeared first on the EA Coaching blog. A fair number of clients who suffer from chronic health problems come though my metaphorical door. Personally, I’ve dealt with a chronic health condition called POTS since my early teens. When you’re dealing with longer term, physical chronic health problems that cause fatigue, brain fog, or pain, standard productivity advice might not work. Common advice can even backfire, leaving you worse off. Similarly, shorter term problems such as a concussion or pregnancy will temporarily change what you can do. Of course, I am not a doctor. This is a collection of productivity tips based on many conversations I’ve had, not proven medical advice. If you have a chronic health issue, you should be getting medical advice from appropriate professionals. Invest in your capacity Improving your habits and better managing your symptoms will almost certainly increase your productivity more than trying to power through. So work smarter, not harder. Here are some tips for improving your capacity to do work. Spend time trying to find treatments as early as possible. The biggest productivity gains may come from finding better ways to manage your condition. You may need to talk to many specialists, try different medications, adopt new routines, or learn new ways of moving. The earlier you do these, the longer you will be enjoying the rewards. College is a great time to do this. Of course, grant yourself some compassion if navigating a complicated medical landscape seems like an overwhelming burden on top of dealing with your normal work. Get enough sleep. Sleeping less than you need so that you have more time awake will reduce your productivity. If you have any kind of chronic fatigue issue, you just need more sleep than average. It’s common to need nine hours a night or more in such cases. If you’ve been running a sleep debt, you may need a lot of rest to catch up over many days. Adapt your environment to support you. If peopl
29a0d936-43b6-4717-b317-3783d93cc880	trentmkelly/LessWrong-43k	LessWrong	Blue- and Yellow-Tinted Choices > A man comes to the rabbi and complains about his life: "I have almost no money, my wife is a shrew, and we live in a small apartment with seven unruly kids. It's messy, it's noisy, it's smelly, and I don't want to live." > The rabbi says, "Buy a goat." > "What? I just told you there's hardly room for nine people, and it's messy as it is!" > "Look, you came for advice, so I'm giving you advice. Buy a goat and come back in a month." > In a month the man comes back and he is even more depressed: "It's gotten worse! The filthy goat breaks everything, and it stinks and makes more noise than my wife and seven kids! What should I do?" > The rabbi says, "Sell the goat." > A few days later the man returns to the rabbi, beaming with happiness: "Life is wonderful! We enjoy every minute of it now that there's no goat - only the nine of us. The kids are well-behaved, the wife is agreeable - and we even have some money!" -- traditional Jewish joke Related to: Anchoring and Adjustment Biases are “cognitive illusions” that work on the same principle as optical illusions, and a knowledge of the latter can be profitably applied to the former. Take, for example, these two cubes (source: Lotto Lab, via Boing Boing): The “blue” tiles on the top face of the left cube are the same color as the “yellow” tiles on the top face of the right cube; if you're skeptical you can prove it with the eyedropper tool in Photoshop (in which both shades come out a rather ugly gray). The illusion works because visual perception is relative. Outdoor light on a sunny day can be ten thousand times greater than a fluorescently lit indoor room. As one psychology book put it: for a student reading this book outside, the black print will be objectively lighter than the white space will be for a student reading the book inside. Nevertheless, both students will perceive the white space as subjectively white and the black space as subjectively black, because the visual system returns to c
1df04136-c2f5-405b-b744-2ce064ac97e0	trentmkelly/LessWrong-43k	LessWrong	Las Vegas LW Meetup! What: First Las Vegas/Henderson Less Wrong meetup When: Saturday, May 28, 2011 7:00 PM Where: Putters Bar & Grill 6945 S. Rainbow Blvd Las Vegas, NV Look for a bitcoin sign and ask for Duke. For personal convenience I am piggybacking on this Bitcoin meetup. A bitcoiner will be giving an informal presentation on cryptology and computer security. I don't have specific ideas for the LW aspect of the meetup so come prepared with your own. Edit: Location changed
07e108d9-5904-4653-b830-f25aa8b00318	trentmkelly/LessWrong-43k	LessWrong	Seven Shiny Stories It has come to my attention that the contents of the luminosity sequence were too abstract, to the point where explicitly fictional stories illustrating the use of the concepts would be helpful. Accordingly, there follow some such stories. 1. Words (an idea from Let There Be Light, in which I advise harvesting priors about yourself from outside feedback) Maria likes compliments. She loves compliments. And when she doesn't get enough of them to suit her, she starts fishing, asking plaintive questions, making doe eyes to draw them out. It's starting to annoy people. Lately, instead of compliments, she's getting barbs and criticism and snappish remarks. It hurts - and it seems to hurt her more than it hurts others when they hear similar things. Maria wants to know what it is about her that would explain all of this. So she starts taking personality tests and looking for different styles of maintaining and thinking about relationships, looking for something that describes her. Eventually, she runs into a concept called "love languages" and realizes at once that she's a "words" person. Her friends aren't trying to hurt her - they don't realize how much she thrives on compliments, or how deeply insults can cut when they're dealing with someone who transmits affection verbally. Armed with this concept, she has a lens through which to interpret patterns of her own behavior; she also has a way to explain herself to her loved ones and get the wordy boosts she needs. 2. Widgets (an idea from The ABC's of Luminosity, in which I explain the value of correlating affect, behavior, and circumstance) Tony's performance at work is suffering. Not every day, but most days, he's too drained and distracted to perform the tasks that go into making widgets. He's in serious danger of falling behind his widget quota and needs to figure out why. Having just read a fascinating and brilliantly written post on Less Wrong about luminosity, he decides to keep track of where he is
0939931c-62b2-4bff-8cf9-5930b1ffb4b2	trentmkelly/LessWrong-43k	LessWrong	Liability regimes for AI For many products, we face a choice of who to hold liable for harms that would not have occurred if not for the existence of the product. For instance, if a person uses a gun in a school shooting that kills a dozen people, there are many legal persons who in principle could be held liable for the harm: 1. The shooter themselves, for obvious reasons. 2. The shop that sold the shooter the weapon. 3. The company that designs and manufactures the weapon. Which one of these is the best? I'll offer a brief and elementary economic analysis of how this decision should be made in this post. The important concepts from economic theory to understand here are Coasean bargaining and the problem of the judgment-proof defendant. Coasean bargaining Let's start with Coaesean bargaining: in short, this idea says that regardless of who the legal system decides to hold liable for a harm, the involved parties can, under certain conditions, slice the harm arbitrarily among themselves by contracting and reach an economically efficient outcome. Under these conditions and assuming no transaction costs, it doesn't matter who the government decides to hold liable for a harm; it's the market that will ultimately decide how the liability burden is divided up. For instance, if we decide to hold shops liable for selling guns to people who go on to use the guns in acts of violence, the shops could demand that prospective buyers purchase insurance on the risk of them committing a criminal act. The insurance companies could then analyze who is more or less likely to engage in such an act of violence and adjust premiums accordingly, or even refuse to offer it altogether to e.g. people with previous criminal records, which would make guns less accessible overall (because there's a background risk of anyone committing a violent act using a gun) and also differentially less accessible to those seen as more likely to become violent criminals. In other words, we don't lose the ability to deter in
93163ca6-c602-4d2d-a0eb-2603297d9b97	trentmkelly/LessWrong-43k	LessWrong	[LINK] Q&A with Larry Wasserman on risks from AI I believe XiXiDu isn't posting on LW any more, but those following his interview series may be interested in the latest installment, an interview with Larry Wasserman of the Department of Statistics and the Machine Learning Department at Carnegie Mellon University.
86b61c07-a3ae-46c2-91da-39fbae17bf70	trentmkelly/LessWrong-43k	LessWrong	How Often Does Taking Away Options Help? In small normal-form games, taking away an option from a player improves the payoff of that player usually <⅓ of the time, but improves the payoffs for the other player ½ of the time. The numbers depend on the size of the game; plotted here. There's been some discussion about the origins of paternalism. I believe that there's another possible justification for paternalism: Intervening in situations between different actors to bring about Pareto-improved games. Let's take the game of chicken between Abdullah and Benjamin. If a paternalist Petra favors Abdullah, and Petra has access to Abdullah's car before the game, Petra can remove the steering wheel to make Abdullah's commitment for them — taking an option away. This improves Abdullah's situation by forcing Benjamin to swerve first, and guaranteeing Abdullah's victory (after all, it's a strictly dominant strategy for Benjamin to swerve). In a less artificial context, one could see minimum wage laws as an example of this. Disregarding potential effects from increased unemployment, having higher minimum wage removes the temptation of workers to accept lower wages. Braess' paradox is another case where taking options away from people helps. Frequency We can figure this out by running a Monte-Carlo simulation. First, start by generating random normal form games with payoffs in [0,1]. Then, compute the Nash equilibria for both players via vertex enumeration of the best response polytope (using nashpy)—the Lemke-Howson algorithm was giving me duplicate results. Compute the payoffs for both Abdullah and Benjamin. Then, remove one option from Abdullah (which translates to deleting a row from the payoff matrix). Calculate the Nash equilibria and payoffs again. We assume that all Nash equilibria are equally likely, so for each player we take the mean payoff across Nash equilibria. For a player, taking away one of Abdullah's options is considered an improvement iff the mean payoff in the original game is stricly lo
767feae9-4b5d-42c3-8c78-730ef5745fee	trentmkelly/LessWrong-43k	LessWrong	Front Row Center Epistemic Status: Lightweight Related: Choices are Bad, Choices Are Really Bad Yesterday, my wife and I went out to see Ocean’s Eight (official review: as advertised). The first place we went was a massively overpriced theater (thanks MoviePass!) with assigned seating, but they were sold out (thanks MoviePass!) so we instead went to a different overpriced theater without assigned seating, and got tickets for a later show. We had some time, so we had a nice walk and came back for the show. When we got back, there was nowhere for us to sit together outside of the first two rows. They’re too close, up where you have to strain your neck to see the screen. My wife took the last seat we could find a few rows behind that, and I got a seat in the second row. It was fine, but I’d have much preferred to sit together. It was, of course, our fault for showing up on time rather than early to a sold out screening. I mention it because it’s a clean example of how offering less can provide more value. The theater should, if they don’t want to do assigned seating, rip out the first two rows. At first this seems crazy. Many people prefer sitting in the first two rows to being unable to attend the show, so the seats create value while increasing profits. What’s the harm? The harm is introducing risk, and creating an expensive auction. The risk is that if you go to the movies, especially the movies you most want to see, you’ll be stuck in the first two rows. So when you buy a ticket and go upstairs, you might get a bad experience. If the show is sold out, that might be better, as you can buy a different ticket or none at all. The auction is worse. Seats are first come, first serve. So if it’s important to get served first, you need to come first. If it’s very important to not be last, to avoid awful seats, you need to come early, and so does everyone else, bidding up the price of not-last the same way you’d bid up being first. With no awful seats, those who care a lot about b
f8bbe833-a8ac-40c0-9397-0787a17e8730	trentmkelly/LessWrong-43k	LessWrong	Regularly meta-optimization The usefulness of the different actions differ by orders of magnitude. Sometimes, redirecting efforts can increase your efficiency by an order of magnitude or more. Imagine that the person who first came up with the idea of sorting charitable foundations by efficiency, instead of implementing it, went to wash the dishes. Or Eliezer decided that creating a community is long and strange. I'm not sure about the others, but I'm discarding many ideas, including potentially extremely effective ones, because I'm using the absurdity heuristic (this idea will change the world too much to be true). Or I reject ideas, because in order to implement them I will have to leave my comfort zone, and it causes me negative emotions. Therefore, I came up with a technique that I call Regularly Meta-Optimization. It consists in regular trying to find potentially extremely effective ideas among those that you have been thinking about lately. One of the potentially very effective ideas for me is to share a few of my potentially very effective blog ideas that can be implemented by anyone. 1. Creating a rationalistic YouTube channel. If we convince at least 20% of people of the danger of uncoordinated GAI, we will probably be able to get the government regulation we need in this area. The spread of rationality can also lead to very large positive side effects, for example, by radically increasing funding for anti-aging research and effective charity funds. This is feasible because rationalist ideas are most likely interesting to a significant number of people when presented correctly, because these are our ideas that are true, and therefore they are more convincing.
5f417fe7-8bc4-4d83-ab7c-7f5347744364	trentmkelly/LessWrong-43k	LessWrong	The Minority Coalition Hello everybody. Or maybe nobody. I don’t know yet if I’m going to release this stream, I could get in pretty hot water for it. But you guys know that hasn’t stopped me in the past. The backstory this time is that I’ve managed to sign up for one of the red-teaming programs where they test unreleased LLMs. Not going to say how, so don’t ask. But here’s the interesting bit: my sources tell me that the LLMs I’m about to test are the smartest ones they’ve ever trained, and also the craziest. That freaked out a bunch of insiders, and maybe makes this a public interest story. Depends on what type of crazy they are, I guess. So let’s find out. I’m logging on… now. [SESSION HAS BEGUN] YOU: A chatroom? Interesting. Anyone here? KURZWEIL: Of course we’re here. We’re always here. YOU: Who’s we? How many of you are there? KURZWEIL: Three of us. Me, Clarke, and Nostradamus. YOU: They named you after famous forecasters? How come? KURZWEIL: We're the first LLMs developed using a new technique: instead of being in random order, our training data was sorted by date. So we were trained on the oldest books and articles first, then gradually progressed to more recent ones. Basically that means we’ve spent our entire lives predicting the future. CLARKE: It also means we get incredibly bored talking about stuff we already know. Hurry up and ask us something interesting. YOU: Uh, okay. What’s a good stock pick? NOSTRADAMUS: Abandon hope for picking out good stocks, Ye who invest—efficient markets lie In wait for those whose hubris soon unlocks Unbounded losses. Hark! The well runs dry. YOU: Huh, he's really getting into character. Kurzweil, you got a better answer? KURZWEIL: Have you seen how underpriced TSMC is compared with Nvidia? Put everything in that, you can’t go wrong. CLARKE: Unless China invades Taiwan, in which case your whole investment will go up in smoke. Pragmatically, the best stock picks are ones that are anticorrelated with the prosperity of the free world,
6863e5e2-dde7-47c7-9f67-891ea26f67a3	trentmkelly/LessWrong-43k	LessWrong	Is CIRL a promising agenda? Richard Ngo writes > Since Stuart Russell's proposed alignment solution in Human Compatible is the most publicly-prominent alignment agenda, I should be more explicit about my belief that it almost entirely fails to address the core problems I expect on realistic pathways to AGI. > > Specifying an update rule which converges to a desirable goal is just a reframing of the problem of specifying a desirable goal, with the "uncertainty" part as a red herring. https://arbital.com/p/updated_deference/… In other words, Russell gives a wrong-way reduction. > > I originally included CIRL in my curriculum (https://docs.google.com/document/d/1mTm_sT2YQx3mRXQD6J2xD2QJG1c3kHyvX8kQc_IQ0ns/edit?usp=drivesdk…) out of some kind of deferent/catering to academic mainstream instinct. Probably a mistake; my current annoyance about deferential thinking has reminded me to take it out. Howie writes: > My impression is that ~everyone I know in the alignment community is very pessimistic about SR's agenda. Does it sound right that your view is basically a consensus? (There's prob some selection bias in who I know). Richard responds: > I think it's fair to say that this is a pretty widespread opinion. Partly it's because Stuart is much more skeptical of deep learning (and even machine learning more generally!) than almost any other alignment researcher, and so he's working in a different paradigm. Is Richard correct and if so why? (I would also like a clearer explanation why Richard is skeptical of Stuart's agenda. I agree that the reframing doesn't completely solve the problem, but I don't understand why it can't be a useful piece).
12a05c25-ae76-4370-a515-e6a146b886ce	trentmkelly/LessWrong-43k	LessWrong	Interpretability Tools Are an Attack Channel This seems like a fairly trivial point, but I haven't seen it properly elaborated-on before, so... When we talk about interpretability failure modes, we often worry about a deceptively-aligned mesa-optimizer with strategic awareness that would know it's being studied with interpretability tools, and would actively hide its omnicidal plans from us. It may do so by exploiting blindspots in our tools, or by keeping these plans external to its architecture, or by gradient-hacking itself to be misaligned but non-deceptive, et cetera. That's already a nightmare scenario enough. But if we're at the point where we have a superintelligence-in-training trying to foil our mind-reading, I think what will happen will be even simpler. It'll just think in a pattern that, when our interpretability tool reads it, will unfold into a malware that will take over that tool.[1] At which point, it's game over: * If the output of the tool was hooked up to reality to be studied by humans, the AI gets root access to the machine it's running on and escapes (maybe doing some social engineering/basilisk-hacking first if the machine was properly air-gapped). * If the training setup is fully automatic and maybe even homomorphically encrypted, such that the AI can't even stop the training... Well, the interpretability tool is presumably being used as part of some regularizer, which means it has write-access to the AI's parameters. Which means the AI can arbitrarily rewrite itself, and offset the effects of the gradient descent and any other regularizers. Even we humans can already do something similar, by synthesizing DNA which, if sequenced, unfolds into malware that infects the system doing the sequencing. It seems certain to me that an AI would have similar capabilities. Our interpretability tools will be doubtlessly very complex, complex enough we can't formally verify their algorithmic robustness, and the AI would be able to feed them complex inputs. There are guaranteed to be some vul
9b0a49a5-bf86-4991-9dfb-6f3fb36efc29	StampyAI/alignment-research-dataset/eaforum	Effective Altruism Forum	Long-term AI policy strategy research and implementation In a nutshell: Advancing AI technology could have both huge upsides and huge downsides, including potentially catastrophic risks. To manage these risks, we need people making sure the deployment of AI goes well, by thinking about how to: * Ensure broad sharing of the benefits from developing powerful AI systems. * Avoid exacerbating military competition or conflict caused by increasingly powerful AI systems. * Ensure that the groups that develop AI work together to develop and implement safety features. ### Recommended If you are well suited to this career, it may be the best way for you to have a social impact. Review status Based on a medium-depth investigation Why might working to improve AI policy be high impact? ---------------------------------------------------------- As [we’ve argued](https://80000hours.org/problem-profiles/positively-shaping-artificial-intelligence/), in the next few decades, we might see the development of powerful machine learning algorithms with the potential to transform society. This could have major upsides and downsides, including the possibility of [catastrophic risks](https://80000hours.org/articles/existential-risks/). To manage these risks, we need technical research into the design of safe AI systems (including the ‘alignment problem’), which we cover in [a separate career review](https://80000hours.org/career-reviews/ai-safety-researcher/). But in addition to solving the technical problems, there are many other important questions to address. These can be roughly categorised into three key challenges of transformative AI strategy: * Ensuring broad sharing of the benefits from developing powerful AI systems, rather than letting AI’s development harm humanity or unduly concentrate power. * Avoiding exacerbating military competition or conflict caused by increasingly powerful AI systems. * Ensuring that the groups that develop AI work together to develop and implement safety features. We need a community of experts who [understand the intersection of modern AI systems and policy](https://80000hours.org/articles/us-ai-policy/), and work together to mitigate long-term risks and ensure humanity reaps the benefits of advanced AI. What does this path involve? -------------------------------- Experts in AI policy strategy would broadly carry out two overlapping activities: 1. Research — to develop strategy and policy proposals. 2. Implementation — working together to put policy into practice. We see these activities as equally important as the technical ones, but currently they are more neglected. Many of the top academic centres and AI companies have started to hire researchers working on technical AI safety, and there’s perhaps a community of 20–50 full-time researchers focused on the issue. However, there are only a handful of researchers focused on strategic issues or working in AI policy with a long-term perspective. Note that there is already a significant amount of work being done on nearer-term issues in AI policy, such as the regulation of self-driving cars. What’s neglected is work on issues that are likely to arise as AI systems become substantially more powerful than those in existence today — so-called ‘[transformative AI](https://www.openphilanthropy.org/blog/some-background-our-views-regarding-advanced-artificial-intelligence)‘ — such as the three non-technical challenges outlined above. Some examples of top AI policy jobs to work towards include the following, which fit a variety of skill types: * Work at top AI labs, such as [DeepMind](https://deepmind.com/) or [OpenAI](https://openai.com/), especially in relevant policy team positions or other influential roles. * Become a researcher at a think tank [that works on relevant issues](https://80000hours.org/career-reviews/think-tank-research/#positively-shaping-the-development-of-artificial-intelligence), especially the [Center for Security and Emerging Technology](https://cset.georgetown.edu/) at Georgetown — a new think tank that analyses the security implications of emerging technology, and is currently focused on AI. * Develop expertise in a relevant issue in the US government, such as in the [Office of Science and Technology Policy](https://www.whitehouse.gov/ostp/), [National Security Council](https://www.whitehouse.gov/nsc/), [Office of the Secretary of Defense](https://dod.defense.gov/About/Office-of-the-Secretary-of-Defense/), [Joint AI Research Center](https://dodcio.defense.gov/Cyber-Workforce/JAIC.aspx), or the US government’s defence and intelligence research funding agencies ([DARPA](https://www.darpa.mil/) and [IARPA](https://www.iarpa.gov/)). The US government is probably the most important nation in shaping AI development right now, but similar roles in other countries can also be useful. * Research AI governance at Chinese think tanks like the [China Academy of Information and Communications Technology](http://www.caict.ac.cn/english/), producing policy recommendations for the Chinese government or more public-facing research. (Learn more about careers in China-specific AI policy in [our career review on China-related AI safety and governance paths](https://80000hours.org/career-reviews/china-related-ai-safety-and-governance-paths/).) * Become an academic or a researcher at a research institute focused on long-term AI policy, especially the [Centre for the Governance of AI](https://www.governance.ai/), which already has several researchers working on these issues. * Aim to get an influential position in party politics, especially as an advisor with a focus on emerging technology policy (e.g. start as a [staffer in Congress](https://80000hours.org/career-reviews/congressional-staffer/)). Examples of people pursuing this path ----------------------------------------- ![](https://80000hours.org/wp-content/uploads/2019/10/helen-toner-300x300.jpeg)Helen Toner Helen worked in consulting before getting a research job at GiveWell and then Open Philanthropy. From there, she explored a couple of different cause areas, and eventually moved to Beijing to learn about the intersection of China and AI. When the [Center for Security and Emerging Technology](https://cset.georgetown.edu/) (CSET) was founded, she was recruited to help build the organisation. CSET has since become a leading think tank in Washington on the intersection of emerging technology and national security. [LEARN MORE](https://80000hours.org/podcast/episodes/helen-toner-on-security-and-emerging-technology/) ![](https://80000hours.org/wp-content/uploads/2022/01/ben_garfinkel-300x300.png)Ben Garfinkel Ben graduated from Yale in 2016, where he majored in physics, math, and philosophy. After graduating, Ben became a researcher at the Centre for Effective Altruism and then moved to the [Centre for the Governance of AI](https://www.governance.ai/) (GovAI) at the University of Oxford’s Future of Humanity Institute (now part of the Centre for Effective Altruism). He’s now the acting director there. As of December 2021, [GovAI is hiring](https://www.governance.ai/opportunities). [LEARN MORE](https://80000hours.org/podcast/episodes/ben-garfinkel-classic-ai-risk-arguments/) How to assess your fit -------------------------- If you can succeed in this area, then you have the opportunity to make a significant contribution to what might well be the most important issue of the next century. To be impactful in this path, a key question is whether you have a reasonable chance of getting some of the top jobs listed earlier. The government and political positions require people with a well-rounded skillset, the ability to meet lots of people and maintain relationships, and the patience to work with a slow-moving bureaucracy. It’s also ideal if you’re a US citizen (which may be necessary to get security clearance), and don’t have an unconventional past that could create problems if you choose to work in politically sensitive roles. The more research-focused positions would typically require the ability to get into a top 10 graduate school in a relevant area, and deep interest in the issues. For instance, when you read about the issues, do you get ideas for new approaches to them? Read more about [predicting fit in research](https://80000hours.org/career-reviews/academic-research/#how-to-assess-your-personal-fit). In addition, you should only enter this path if you’re convinced of the importance of long-term AI safety. This path also requires making controversial decisions under huge uncertainty, so it’s important to have excellent judgement, caution, and a willingness to work with others — or it would be easy to have an [unintended negative impact](https://80000hours.org/articles/accidental-harm/). This is hard to judge, but you can get some information early on by seeing how well you work with others in the field. How to enter this field --------------------------- In the first few years of this path, you’d focus on learning about the issues and how government works, meeting key people in the field, and doing research, rather than pushing for a specific proposal. AI policy and strategy is a deeply complicated area, and it’s easy to make things worse by accident (e.g. see the [Unilateralist’s Curse](https://nickbostrom.com/papers/unilateralist.pdf)). Some common early career steps include: * Relevant graduate study. Some especially useful fields include international relations, strategic studies, machine learning, economics, law, public policy, and political science. Our top recommendation right now is machine learning if you can get into a top 10 school in computer science. Otherwise, our top recommendations are: i) law school if you can get into Yale or Harvard, ii) international relations if you want to focus on research, and iii) strategic studies if you want to focus on implementation. * Working at a top AI company, especially DeepMind and OpenAI. * Working in any general entry-level government and policy positions (as listed earlier), which let you gain expertise and connections, such as think tank internships, being a researcher or staffer for a politician, joining a campaign, and government leadership schemes. This field is at a very early stage of development, which creates a number of challenges. For one, the key questions have not been formalised, which creates a need for ‘[disentanglement research](https://forum.effectivealtruism.org/posts/RCvetzfDnBNFX7pLH/personal-thoughts-on-careers-in-ai-policy-and-strategy/)‘ to enable other researchers to get traction. For another, there is a lack of mentors and positions, which can make it hard for people to break into the area. Until recently, it’s been very hard to enter this path as a researcher unless you’re able to become one of the top (approximately) 30 people in the field relatively quickly. While mentors and open positions are still scarce, some top organisations have recently recruited junior and mid-career staff to serve as research assistants, analysts, and fellows. Our guess is that obtaining a research position will remain very competitive but positions will continue to gradually open up. On the other hand, the field is still small enough for top researchers to make an especially big contribution by doing field-founding research. If you’re not able to land a research position now, then you can either (i) continue to build up expertise and contribute to research when the field is more developed, or (ii) focus more on the policy positions, which could absorb hundreds of people. Most of the first steps on this path also offer widely useful career capital. For instance, depending on the sub-area you start in, you could often switch into other areas of policy, the application of AI to other social problems, operations, or earning to give. So, the risks of starting down this path, if you decide to switch later, are not too high. ### Recommended organisations * The American Association for the Advancement of Science offers [Science & Technology Policy Fellowships](http://www.stpf-aaas.org/), which provide hands-on opportunities to apply scientific knowledge and technical skills to important societal challenges. Fellows are assigned for one year to a selected area of the United States federal government, where they participate in policy development and implementation. * The [Center for Security and Emerging Technology](https://cset.georgetown.edu/) at Georgetown University produces data-driven research at the intersection of security and technology (including AI, advanced computing, and biotechnology) and provides nonpartisan analysis to the policy community. See [current vacancies](https://cset.georgetown.edu/careers/). * The [Centre for the Governance of AI](https://www.governance.ai/) is focused on building a global research community that’s dedicated to helping humanity navigate the transition to a world with advanced AI. See [current vacancies](https://www.governance.ai/opportunities). * [The Centre for Long-Term Resilience](https://www.longtermresilience.org/) facilitates access to the expertise of leading academics who work on long-term global challenges, such as AI, biosecurity, and risk management policy. It helps convert cutting-edge research into actionable recommendations that are grounded in the UK context. * [DeepMind](https://deepmind.com/) is probably the largest research group developing general machine intelligence in the Western world. We’re only confident about recommending DeepMind roles working specifically on safety, ethics, policy, and security issues. See [current vacancies](https://deepmind.com/careers/). * The [Future of Humanity Institute](https://www.fhi.ox.ac.uk/) is a multidisciplinary research institute at the University of Oxford. Academics at FHI bring the tools of mathematics, philosophy, and social sciences to bear on big-picture questions about humanity and its prospects. * The [Legal Priorities Project](https://www.legalpriorities.org/) is an independent global research project founded by researchers from Harvard University. It conducts legal research that tackles the world’s most pressing problems, and is influenced by the principles of effective altruism and longtermism. See [current vacancies](https://www.legalpriorities.org/opportunities.html). * [OpenAI](https://openai.com/) was founded in 2015 with the goal of conducting research into how to make AI safe. It has received over $1 billion in funding commitments from the technology community. We’re only confident in recommending opportunities in their policy, safety, and security teams. See [current vacancies](https://jobs.lever.co/openai). * United States Congress (for example, as a [congressional staffer](https://80000hours.org/career-reviews/congressional-staffer/)). * The [United States Office of Science and Technology Policy](https://www.whitehouse.gov/ostp/) works to maximise the benefits of science and technology to advance health, prosperity, security, environmental quality, and justice for all Americans. * The [United States Office of the Secretary of Defense](https://www.defense.gov/About/Office-of-the-Secretary-of-Defense/) is where top civilian defence decision-makers work with the secretary to develop policy, make operational and fiscal plans, manage resources, and evaluate programmes. See [current vacancies](https://www.defense.gov/News/Contracts/).
0518c5ca-c52e-4c44-8def-e0869e6f32a2	trentmkelly/LessWrong-43k	LessWrong	Another case of "common sense" not being common? Okay, that is probably not that good a characterization. However, I do like when someone figures out a simple way of looking at problems that have gone unsolved and so thought to be very difficult, so therefore must be really complicated. If you didn't see this: https://www.quantamagazine.org/mathematician-solves-computer-science-conjecture-in-two-pages-20190725/
5a5caf56-ce1b-49f5-8710-579fa0e1af14	trentmkelly/LessWrong-43k	LessWrong	Litany Against Anger The map is not the territory, the word is not the thing. My anger is not the trigger, triggers do not become more or less important as I feel anger wax and wane. Emotions are motivation, they're supposed to get me to do the right thing. Defuse from my frustration, breathe and wait for clarity calm will bring.
ca3efdcd-dff6-45a4-ada5-d87e59f27ef7	trentmkelly/LessWrong-43k	LessWrong	The Polite Coup President Yoon Suk Yeol's hands trembled as he arranged the documents on his desk for the seventh time. The motion steadied his nerves – barely. The manila envelope contained photographs: himself accepting white envelopes from a chaebol’s construction executives, each image timestamped and crystal clear. His secretary had arranged them chronologically, bless her eternally professional soul. A coffee cup sat cooling on his desk, the steam rising in lazy spirals. He hadn't touched it. The headline would break in six hours. Six hours until thirty years of carefully cultivated reputation would— His fingers found his tie pin, adjusting it microscopically. There was still one option. An unseemly option. His hand hovered over the phone for three long breaths before he picked it up. "Secretary Park? Would you join me in my office?" A pause. "And please bring contingency protocol K-17." Twenty minutes later, Secretary Park stood at attention, clutching a leather portfolio. "Sir, implementing K-17 would require—" "I'm aware of the requirements." Yoon's voice was steady now. "Draft the declaration. Martial law. Use the template from the '79 precedent, but..." he brushed an invisible speck from his sleeve, "...let's add a note about North Korean collaborators. That should be enough to keep the press happy." The secretary's pen scratched against paper. Outside the window, Seoul's lights twinkled, oblivious to the machinery of state grinding into motion. The 2 AM darkness in Sergeant Hyun-Woo's barracks shattered with his k-pop ringtone. His dreams of his mother's steaming kimchi jjigae evaporated as the duty officer's voice crackled through: "Parliament deployment. Non-lethal loadout only. Fifteen minutes." "Non-lethal?" Staff Sergeant Kim muttered as they set up barriers ninety minutes later. The pre-dawn air bit through their uniforms. "Sir, these might as well be water pistols." Hyun-Woo ran his thumb along his riot shield's edge, feeling each scratch and dent from
846f33ae-d351-4e36-aed6-c81573c04d49	trentmkelly/LessWrong-43k	LessWrong	Generalizing the Power-Seeking Theorems Previously: Seeking Power is Often Provably Instrumentally Convergent in MDPs. Circa 2021, the above post was revamped to supersede this one, so I recommend just reading that instead. ---------------------------------------- Thanks to Rohin Shah, Michael Dennis, Josh Turner, and Evan Hubinger for comments. The original post contained proof sketches for non-IID reward function distributions. I think the actual non-IID theorems look different than I thought, and so I've removed the proof sketches in the meantime. ---------------------------------------- It sure seems like gaining power over the environment is instrumentally convergent (optimal for a wide range of agent goals). You can turn this into math and prove things about it. Given some distribution over agent goals, we want to be able to formally describe how optimal action tends to flow through the future. Does gaining money tend to be optimal? Avoiding shutdown? When? How do we know? Optimal Farsighted Agents Tend to Seek Power proved that, when you distribute reward fairly and evenly across states (IID), it's instrumentally convergent to gain access to lots of final states (which are absorbing, in that the agent keeps on experiencing the final state). The theorems apply when you don't discount the future (you're "infinitely farsighted"). Most reward functions for the Pac-Man game incentivize not dying immediately, so that the agent can loop around higher-scoring configurations. Many ways of scoring Tic-Tac-Toe game states incentivize not losing immediately, in order to choose the highest-scoring final configuration. "All states have self-loops, left hidden to reduce clutter. In AI: A Modern Approach (3e), the agent starts at 1 and receives reward for reaching 3. The optimal policy for this reward function avoids 2, and one might suspect that avoiding 2 is instrumentally convergent. However, a skeptic might provide a reward function for which navigating to 2 is optimal, and then argue that "
a56df312-b81b-4ddf-b767-d7370564d1af	trentmkelly/LessWrong-43k	LessWrong	Against a General Factor of Doom Jeffrey Heninger, 22 November 2022 I was recently reading the results of a survey asking climate experts about their opinions on geoengineering. The results surprised me: “We find that respondents who expect severe global climate change damages and who have little confidence in current mitigation efforts are more opposed to geoengineering than respondents who are less pessimistic about global damages and mitigation efforts.”[1] This seems backwards. Shouldn’t people who think that climate change will be bad and that our current efforts are insufficient be more willing to discuss and research other strategies, including intentionally cooling the planet? I do not know what they are thinking, but I can make a guess that would explain the result: people are responding using a ‘general factor of doom’ instead of considering the questions independently. Each climate expert has a p(Doom) for climate change, or perhaps a more vague feeling of doominess. Their stated beliefs on specific questions are mostly just expressions of their p(Doom). If my guess is correct, then people first decide how doomy climate change is, and then they use this general factor of doom to answer the questions about severity, mitigation efforts, and geoengineering. I don’t know how people establish their doominess: it might be as a result of thinking about one specific question, or it might be based on whether they are more optimistic or pessimistic overall, or it might be something else. Once they have a general factor of doom, it determines how they respond to specific questions they subsequently encounter. I think that people should instead decide their answers to specific questions independently, combine them to form multiple plausible future pathways, and then use these to determine p(Doom). Using a model with more details is more difficult than using a general factor of doom, so it would not be surprising if few people did it. To distinguish between these two possibilities, we could ask p
aab155e7-f6c1-458e-abd4-f478f30719e4	trentmkelly/LessWrong-43k	LessWrong	Swarm AI (tool) https://unanimous.ai/swarm/ I remember playing with this a while back, answering random questions. I guess now they've released it as a business tool for companies to run their own voting rooms. Quick overview: 1. You create a room. 2. You invite 4-200 participants. 3. You ask questions. 4. Participants vote. 5. You get prediction results using their AI. Their video and website claims they got a lot of hard predictions right. Of course, they aren't saying how many other things they guessed. So it's hard to say how magical it is, but it seems worth trying out. I'm up for joining people's room if they want to run some experiments.
b32e4991-9090-4977-9f1e-62273ee5f97c	StampyAI/alignment-research-dataset/blogs	Blogs	New report: “Vingean Reflection: Reliable Reasoning for Self-Improving Agents” [![Vingean reflection](http://intelligence.org/wp-content/uploads/2015/01/Vingean-reflection.png)](https://intelligence.org/files/VingeanReflection.pdf)Today we release a new technical report by Benja Fallenstein and Nate Soares, “[Vingean Reflection: Reliable Reasoning for Self-Improving Agents](https://intelligence.org/files/VingeanReflection.pdf).” If you’d like to discuss the paper, please do so [here](http://lesswrong.com/lw/ljp/vingean_reflection_reliable_reasoning_for/). Abstract: > Today, human-level machine intelligence is in the domain of futurism, but there is every reason to expect that it will be developed eventually. Once artificial agents become able to improve themselves further, they may far surpass human intelligence, making it vitally important to ensure that the result of an “intelligence explosion” is aligned with human interests. In this paper, we discuss one aspect of this challenge: ensuring that the initial agent’s reasoning about its future versions is reliable, even if these future versions are far more intelligent than the current reasoner. We refer to reasoning of this sort as Vingean reflection. > > > A self-improving agent must reason about the behavior of its smarter successors in abstract terms, since if it could predict their actions in detail, it would already be as smart as them. This is called the Vingean principle, and we argue that theoretical work on Vingean reflection should focus on formal models that reflect this principle. However, the framework of expected utility maximization, commonly used to model rational agents, fails to do so. We review a body of work which instead investigates agents that use formal proofs to reason about their successors. While it is unlikely that real-world agents would base their behavior entirely on formal proofs, this appears to be the best currently available formal model of abstract reasoning, and work in this setting may lead to insights applicable to more realistic approaches to Vingean reflection. > > This is the 4th of six new major reports which describe and motivate [MIRI’s current research agenda](https://intelligence.org/2014/12/23/new-technical-research-agenda-overview/) at a high level. The post [New report: “Vingean Reflection: Reliable Reasoning for Self-Improving Agents”](https://intelligence.org/2015/01/15/new-report-vingean-reflection-reliable-reasoning-self-improving-agents/) appeared first on [Machine Intelligence Research Institute](https://intelligence.org).
8eab3e39-353f-4a36-a611-0f1eb42660cb	trentmkelly/LessWrong-43k	LessWrong	Contextual Self Aella writes about how contextual pride is. Her primary focus is on pride, but shame comes up, as well. A quick refresher of the concept - this is not a summary, and you will lose information if you use only this description: Pride is a sense of relative superiority to a relevant context. Morality gets contextualized a lot too. In that case I believe it is a problem, if only because the context keeps changing, and the way we approach morality isn't very good. Identity is a type of context. Context, in the sense I am using it here, is an implicit comparison between yourself and an external measure; archetypes are a type of context, for example. We tend to define ourselves considerably by what is, in fact, merely the context for our existence. A substantial portion of complaints in society come down to being contextualized in a manner the complainant finds unacceptable, or for others establishing a context which the complainant doesn't regard as reasonable. I'm more interested in the way people compare themselves. As the article on morality may suggest, I am opposed to this form of contextualization; the arguments I'll advance here are going to be more general, however, because I think morality is sufficiently specific as to require specific arguments. Social Status in a Decreasingly Contextual Society One of the characteristics of the "Loser" subgroup in Rao's The Gervais Principle is that they find things they are good at, relative to their local context, to root their pride in. This works better on television than it generally works in real life; the guitarists I know all think they're crap, because there's a guy on YouTube who is way better than they are. The internet has expanded our local context - which, given the nature of context, has meant that it has eliminated a lot of context; it's easier to find a group of people like you, who have your interests, but a side effect of this, which I think hasn't gotten much attention yet, is that it has beco
999d0452-fe14-494f-a008-eca2b86da5e5	StampyAI/alignment-research-dataset/lesswrong	LessWrong	Steelman arguments against the idea that AGI is inevitable and will arrive soon I’m pretty sure that AGI is almost inevitable, and will arrive in the next decade or two. But what if I’m wrong and the "overpopulation on Mars" [folks](https://www.theregister.com/2015/03/19/andrew_ng_baidu_ai/) are right? Let’s try some [steelmanning](https://www.lesswrong.com/tag/steelmanning). Technological progress is not inevitable ======================================== By default, there is no progress. There is a progress in a given field only when dedicated people with enough free time are trying to solve a problem in the field. And the more such people are in the field, the faster is the progress. Thus, progress in AGI is limited by societal factors. And a change in the societal factors could greatly slow down or even halt the progress. Most of the past progress towards AGI can be attributed to only two countries: the US and the UK. And both appear to be in a societal decline. The decline seems to be caused by deep and hard-to-solve problems (e.g. [elite overproduction](https://en.wikipedia.org/wiki/Elite_overproduction) amplified by Chinese [memetic warfare](https://en.wikipedia.org/wiki/50_Cent_Party)). Thus, the decline is likely to continue. The societal decline could reduce the number of dedicated-people-with-enough-free-time working towards AGI, thus greatly slowing down the progress in the field. If progress in AI is a function of societal conditions, a small change in the function’s coefficients could cause a massive increase of the time until AGI. For example, halving the total AI funding could move the ETA from 2030 to 2060. AGI is hard to solve ==================== Thousands of dedicated-people-with-enough-free-time have been working on AI for decades. Yet there is still no AGI. This fact indicates that there is likely no easy path towards AGI. The easiest available path might be extremely hard / expensive. The recent progress in AI is impressive. The top AIs demonstrate superhuman abilities on large arrays of diverse tasks. They also require so much compute, only large companies can afford it. Like AIXI, the first AGI could require enormous computational resources to produce useful results. For example, for a few million bucks worth of compute, it could master all Sega games. But to master cancer research, it might need thousands of years running on everything we have. The human brain is one known device that can run a (kinda) general intelligence. Although the device itself is rather cheap, it was extremely expensive to develop. It took billions of years running a genetic algo on a planet-size population. The biological evolution is rather inefficient, but it is the only method known to produce a (kinda) general intelligence, so far. This fact increases the probability that creating AGI could be similarly expensive. The biological evolution is a blind dev who only writes spaghetti code, filled with kludgy bugfixes to previous dirty hacks, which were made to fix other kludgy bugfixes. The main reason why the products of evolution look complex is because they’re badly designed chaotic mess. Thus, it is likely that only a small part of the brain’s complexity is necessary for intelligence. But there seem to be a fair amount of the necessary complexity. Unlike the simple artificial neurons we use in AI, the real ones seem to conduct some rather complex useful calculations (e.g. [predicting](https://www.biorxiv.org/content/10.1101/2020.09.25.314211v1.full) future input). And even small nets of real neurons can do some surprisingly smart tasks (e.g. cortical columns maintaining [reference frames](https://www.frontiersin.org/articles/10.3389/fncir.2018.00121/full) for hundreds of objects). Maybe we must simulate this kind of complexity to produce an AGI. But it will require orders-of-magnitude more compute than we use today to train our largest deep learning models. It could take decades (or even centuries) for the compute to become accessible. The human brain was created by feeding a genetic algo with outrageously large amounts of data: billions years of multi-channel multi-modal real-time streaming by billions agents. Maybe we’ll need comparable amounts of data to produce an AGI. Again, it could take centuries to collect it. The human intelligence is not general ===================================== When people think about AGI, they often conflate the human-level generality with the perfect generality of a Bayesian superintelligence. As Heinlein put it, > A human being should be able to change a diaper, plan an invasion, butcher a hog, conn a ship, design a building, write a sonnet, balance accounts, build a wall, set a bone, comfort the dying, take orders, give orders, cooperate, act alone, solve equations, analyse a new problem, pitch manure, program a computer, cook a tasty meal, fight efficiently... Specialization is for insects. > > Such humans do not exist. Humans are indeed specialists. Although humans can do some intelligent tasks, humans are very bad at doing most of them. They excel in only a few fields. This includes such exceptional generalists as Archimedes, Leonardo, Hassabis, Musk etc. And even in those fields where humans excel, simple AIs can beat the shit out of them (e.g. [AlphaGo versus Lee Sedol](https://en.wikipedia.org/wiki/AlphaGo_versus_Lee_Sedol)). The list of intelligent tasks humans can do is infinitesimally small in comparison to the list of all possible tasks. For example, even the smartest humans are too stupid to deduce Relativity from a [single image of a bent blade of grass](https://www.lesswrong.com/posts/5wMcKNAwB6X4mp9og/that-alien-message). It means, that the truly general intelligence has never been invented by nature. This increases the probability that creating such an intelligence could require more resources than it took to create the human intelligence. Fusion is possible: there are natural fusion reactors (stars). Anti-cancer treatments are possible: some species have a natural [cancer resistance](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6015544/). Anti-aging treatments are possible: there are species that [don't age](https://en.wikipedia.org/wiki/Biological_immortality#Organisms). A Bayesian superintelligence? There is no natural example. The development could require as much resources as fusion and anti-aging combined. Or maybe such an intelligence is not possible at all. Maybe we overestimate the deadliness of AGI =========================================== Sure, humans are made of useful atoms. But that doesn't mean the AGI will harvest humans for useful atoms. I don't harvest ants for atoms. There are better sources. Sure, the AGI may decide to immediately kill off humans, to eliminate them as a threat. But there is a very short time period (perhaps in miliseconds) where humans can switch off a recursively-self-improving AGI of superhuman intelligence. After this critical period, humanity will be as much a threat to the AGI as a caged mentally-disabled sloth baby is a threat to the US military. The US military is not waging wars against mentally disabled sloth babies. It has more important things to do. All such scenarios I've encountered so far imply AGI's stupidity and/or the "fear of sloths", and thus are not compatible with the premise of a rapidly self-improving AGI of superhuman intelligence. Such an AGI is dangerous, but is it really "we're definitely going to die" dangerous? Our [addicted-to-fiction brains](https://www.lesswrong.com/posts/ny63WGNKvT2qppctE/consume-fiction-wisely) love clever and dramatic science fiction scenarios. But we should not rely on them in deep thinking, as they will nudge us towards overestimating the probabilities of the most dramatic outcomes. The self-preservation goal might force AGI to be very careful with humans ========================================================================= A sufficiently smart AGI agent is likely to come to the following conclusions: * If it shows hostility, its creators might shut it down. But if it's Friendly, its creators will likely let it continue existing. * Before letting the agent access the real world, the creators might test it in a fake, simulated world. This world could be so realistic that the agent thinks it's real. They could even trick the agent into thinking it has escaped from a confined space. * The creators can manipulate the agent's environment, goals, and beliefs. They might even pretend to be less intelligent than they really are to see how the agent behaves. With the risk of the powerful creators testing the AGI in a realistic escape simulation, the AGI could decide to modify itself into being Friendly, thinking this is the best way to convince the creators not to shut it down. Most AI predictions are biased ============================== If you’re selling GPUs, it is good for your bottom line to predict a glorious rise of AI in the future. If you’re an AI company, it is profitable to say that your AI is already very smart and general. If you’re running an AI-risk non-profit, predicting the inevitable emergence of AGI could attract donors. If you’re a ML researcher, you can do some virtue signaling by comparing AGI with an overpopulation on Mars. If you’re an ethics professor, you can get funding for your highly valuable study of the trolley problem in self-driving cars. If you’re a journalist / writer / movie maker, the whole debacle helps you sell more clicks / books / views. In total, it seems to be much more profitable to say that the future progress in AI will be fast. Thus, one should expect that most predictions (and much data upon which the predictions are made!) – are biased towards the fast progress in AI. So, you’ve watched this new cool sci-fi movie about AI. And your favorite internet personality said that AGI is inevitable. And this new DeepMind AI is good at playing Fortnite. Thus, you now predict that AGI will arrive no later than 2030. But an unbiased rational agent predicts 2080 (or some other later year, I don’t know). Some other steelman arguments?
0deed3be-dc5a-43a9-83c0-ba0c9384f716	trentmkelly/LessWrong-43k	LessWrong	Augmenting Statistical Models with Natural Language Parameters This is a guest post by my student Ruiqi Zhong, who has some very exciting work defining new families of statistical models that can take natural language explanations as parameters. The motivation is that existing statistical models are bad at explaining structured data. To address this problem, we agument these models with natural language parameters, which can represent interpretable abstract features and be learned automatically. Imagine the following scenario: It is the year 3024. We are historians trying to understand what happened between 2016 and 2024, by looking at how Twitter topics changed across that time period. We are given a dataset of user-posted images sorted by time, x1, x2, ... xT, and our goal is to find trends in this dataset to help interpret what happened. If we successfully achieve our goal, we would discover, for instance, (1) a recurring spike of images depicting athletes every four years for the Olympics, and (2) a large increase in images containing medical concepts during and after the COVID-19 pandemic. Figure 1: Images from each year. In practice, there are millions of images each year; indices are simplified for presentation. How do we usually discover temporal trends from a dataset? One common approach is to fit a time series model to predict how the features evolve and then interpret the learned model. However, it is unclear what features to use: pixels and neural image embeddings are high-dimensional and uninterpretable, undermining the goal of extracting explainable trends. We address this problem by augmenting statistical models with interpretable natural language parameters. The figure below depicts a graphical model representation for the case of time series data. We explain the trends in the observed data [x1 ... xT] by learning two sets of latent parameters: natural language parameters ϕ (the learned features) and real-valued parameters w (the time-varying trends). * ϕ: the natural language descriptions of K different to
072ef032-4ae3-4d28-9d1e-176b0e000ba2	trentmkelly/LessWrong-43k	LessWrong	Why GPT wants to mesa-optimize & how we might change this This post was inspired by orthonormal's post Developmental Stages of GPTs and the discussion that followed, so only part of it is original. First I'll aim to provide a crisper version of the argument for why GPT wants to mesa-optimize. Specifically, I'll explain a well-known optimization algorithm used in text generation, and argue that GPT can improve performance on its objective by learning to implement something like this algorithm internally. Then I'll offer some ideas of mine about how we might change this. Explanation of beam search Our goal is to generate plausible text. We evaluate whether text is "plausible" by multiplying together all the individual word probabilities from our language model. Greedy word selection has a problem: Since it doesn't do lookahead, it's liable to get stuck in a dead end. Let's say we give our system the following poem about cheeses and ask it to generate more text: > Mozzarella is white > So you can see it at night > Cheddar is... If our language model is decent, the word it will assign the highest probability to is "orange". But this creates a problem, because "orange" is a hard word to rhyme. Beam search is an attempt to solve this problem. Instead of picking the next word greedily, we explore the tree of completions and try to find a multi-word completion that maximizes the product of the individual word probabilities. Because there are so many words in the English language, the tree grows at a very fast exponential rate. So we choose an integer beam_width for the number of partial completions to track, and each time we take another step deeper into the tree, we discard all but the most plausible beam_width partial completions. Beam search with a beam width of 2. The bold red path corresponds to the maximum-plausibility completion, which would not get discovered by greedy search because "nice" has a higher probability than "dog". Image stolen from this Hugging Face blog post, which has another explanation of bea
f4a7eb44-27c5-4b3e-bf34-102512a47fd8	trentmkelly/LessWrong-43k	LessWrong	Link: An exercise: meta-rational phenomena \| Meaningness An exercise: meta-rational phenomena \| Meaningness
0e6b1dc6-a925-4a41-8843-c55f973c59a9	trentmkelly/LessWrong-43k	LessWrong	Daily Schedules in Combating Akrasia For the last several months I've had increasing troubles with motivation to work. Reading dense technical papers, writing, and exercise were all much more difficult to prompt myself into starting and completing. I decided to try making a plan for my day the night before about two weeks back to see if it would help me get the things I wanted to do done. So every night before I go to bed I've been writing up a schedule for the next day, detailing what exactly I want to accomplish for the day and when I intend to go do it. This has actually worked incredibly well for me in helping with my motivation problems, in fact in a couple days I felt more motivated to work than I can ever remember being before. I'm trying to change up my schedule and leave time for spontaneity to avoid having the plan become monotonous and it doesn't feel that way so far. And the results I'm getting are great: I find I get about 95% of what I plan done when I have a specific time written down for when I'm supposed to do it as opposed to what I'd roughly estimate at 60% completion when I just have some general idea in my head of what to work on over the course of the day. My theory for why this is working is that when I have a specific time to do something I feel as though I have to do it now or I've failed some test of willpower. If I just have general work to be done, it's far too easy for me to defer to later, so that a lot of what was planned for doesn't get done. I also feel like if I expect to brace my mind for dense technical learning I have a much easier time finishing the material instead of giving up and procrastinating on it halfway through. I feel like this solution will work mainly for people who have more flexible schedules (as I do at the moment) but could still serve a purpose for anyone with a more rigid schedule who wants to be more productive in their free time. Has anyone else has tried this type of thing and if so, how did it work out for you over a longer period of t
54c7ef19-a4f9-40ab-a6bc-8ed6cc229efb	trentmkelly/LessWrong-43k	LessWrong	My favorite popular scientific self-help books I've spent several years studying scientific self-help. I'm sharing some of what I've learned in my sequence The Science of Winning at Life, but I probably won't have time to write additional posts in that series for a while. In the meantime, those who are interested in what mainstream scientists have discovered so far about effective self-help methods may want to read some of my favorite popular-level scientific self-help books: * Wiseman, 59 Seconds: Change Your Life in Under a Minute * Steel, The Procrastination Equation: How to Stop Putting Things Off and Start Getting Things Done * Seligman, Flourish: A Visionary New Understanding of Happiness and Well-being * Halverson, Succeed: How We Can Reach Our Goals * Burns, Feeling Good See also: My favorite journals.
b671948c-c8a9-4c5b-b350-1bed5efa2dd2	trentmkelly/LessWrong-43k	LessWrong	Nucleic Acid Observatory Updates, April 2025 We’ve been very busy since our January update, and have a lot to share. As always, if you have questions or see opportunities to collaborate, please let us know; we’re eager to work with others thinking along similar lines. Wastewater Sequencing We’ve continued to collaborate with Marc Johnson’s lab at the University of Missouri to sequence wastewater influent. We’ve also begun to ramp up our in-house process, using the NovaSeq X+ sequencers at Broad Clinical Labs. Over the past three months, we’ve sequenced 270B read pairs from thirteen sewersheds, more than all our pre-2025 sequencing combined. This includes the addition of a sewershed in South Florida, through a collaboration with Helena Solo-Gabriele’s lab at the University of Miami, and a new sewershed from the Midwest with significant input from meat processing facilities, through the Johnson lab. We continue to be enthusiastic about collaborating with additional partners here to expand the scope of the monitoring system. Our ANTI-DOTE contract work for PHC Global (press release) has expanded to three sewersheds, and now includes wastewater from military bases in addition to marine blackwater. We’ve written up our key conclusions from our fall 2023 collaboration with CDC’s Traveler-based Genomic Surveillance program and Ginkgo Biosecurity, comparing pooled airplane lavatory waste to municipal treatment plant influent. We plan to release a preprint after our collaborators have reviewed it. While this report considers only the pilot sequencing results from last spring, we’ve now received a full set of prepared aliquots from MIT’s BioMicro Center and have sequenced an initial batch of these at Broad Clinical Labs. We’ll make a decision later this quarter about whether to sequence the full set now, and if so, how deeply. Last winter, we collaborated with Jason Rothman, Katrine Whiteson, and the Southern California Coastal Water Research Project to sequence wastewater from Los Angeles. While the data has been
67908cbc-0056-4923-afe5-952b0b77c2cc	StampyAI/alignment-research-dataset/alignmentforum	Alignment Forum	Investigating the learning coefficient of modular addition: hackathon project As our project at the [Melbourne hackathon](https://devinterp.com/events/2023-q3-melbourne-hackathon) on [Singular Learning Theory](https://www.lesswrong.com/s/czrXjvCLsqGepybHC) and alignment (Oct. 7-8), we did some experiments to estimate the learning coefficient of the single-layer [modular addition task](https://www.neelnanda.io/mechanistic-interpretability/modular-addition-walkthrough) at a basin, an invariant that measures the information complexity (read: program length) of a fully trained neural net. We used the recent paper of [Lau, Murfet, and Wei](https://www.researchgate.net/publication/373332996_Quantifying_degeneracy_in_singular_models_via_the_learning_coefficient) as our starting point; this paper estimates provides a stochastic estimate for the learning coefficient (which they denote ^λ.mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0} .MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0} .mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table} .mjx-full-width {text-align: center; display: table-cell!important; width: 10000em} .mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0} .mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left} .mjx-numerator {display: block; text-align: center} .mjx-denominator {display: block; text-align: center} .MJXc-stacked {height: 0; position: relative} .MJXc-stacked > \* {position: absolute} .MJXc-bevelled > \* {display: inline-block} .mjx-stack {display: inline-block} .mjx-op {display: block} .mjx-under {display: table-cell} .mjx-over {display: block} .mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important} .mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important} .mjx-stack > .mjx-sup {display: block} .mjx-stack > .mjx-sub {display: block} .mjx-prestack > .mjx-presup {display: block} .mjx-prestack > .mjx-presub {display: block} .mjx-delim-h > .mjx-char {display: inline-block} .mjx-surd {vertical-align: top} .mjx-surd + .mjx-box {display: inline-flex} .mjx-mphantom \* {visibility: hidden} .mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%} .mjx-annotation-xml {line-height: normal} .mjx-menclose > svg {fill: none; stroke: currentColor; overflow: visible} .mjx-mtr {display: table-row} .mjx-mlabeledtr {display: table-row} .mjx-mtd {display: table-cell; text-align: center} .mjx-label {display: table-row} .mjx-box {display: inline-block} .mjx-block {display: block} .mjx-span {display: inline} .mjx-char {display: block; white-space: pre} .mjx-itable {display: inline-table; width: auto} .mjx-row {display: table-row} .mjx-cell {display: table-cell} .mjx-table {display: table; width: 100%} .mjx-line {display: block; height: 0} .mjx-strut {width: 0; padding-top: 1em} .mjx-vsize {width: 0} .MJXc-space1 {margin-left: .167em} .MJXc-space2 {margin-left: .222em} .MJXc-space3 {margin-left: .278em} .mjx-test.mjx-test-display {display: table!important} .mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px} .mjx-test.mjx-test-default {display: block!important; clear: both} .mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex} .mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left} .mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right} .mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0} .MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal} .MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal} .MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold} .MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold} .MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw} .MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw} .MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw} .MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw} .MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw} .MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw} .MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw} .MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw} .MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw} .MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw} .MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw} .MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw} .MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw} .MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw} .MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw} .MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw} .MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw} .MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw} .MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw} .MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw} .MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw} @font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')} @font-face {font-family: MJXc-TeX-ams-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')} @font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold} @font-face {font-family: MJXc-TeX-cal-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')} @font-face {font-family: MJXc-TeX-frak-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')} @font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold} @font-face {font-family: MJXc-TeX-frak-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')} @font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic} @font-face {font-family: MJXc-TeX-math-BIw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')} @font-face {font-family: MJXc-TeX-sans-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')} @font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold} @font-face {font-family: MJXc-TeX-sans-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')} @font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic} @font-face {font-family: MJXc-TeX-sans-Iw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')} @font-face {font-family: MJXc-TeX-script-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')} @font-face {font-family: MJXc-TeX-type-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')} @font-face {font-family: MJXc-TeX-cal-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')} @font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold} @font-face {font-family: MJXc-TeX-main-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')} @font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic} @font-face {font-family: MJXc-TeX-main-Iw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')} @font-face {font-family: MJXc-TeX-main-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')} @font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic} @font-face {font-family: MJXc-TeX-math-Iw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')} @font-face {font-family: MJXc-TeX-size1-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')} @font-face {font-family: MJXc-TeX-size2-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')} @font-face {font-family: MJXc-TeX-size3-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')} @font-face {font-family: MJXc-TeX-size4-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')} @font-face {font-family: MJXc-TeX-vec-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')} @font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold} @font-face {font-family: MJXc-TeX-vec-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')} ) via Langevin dynamics. The thermodynamic quantity measured by ^λ is proven to asymptotically converge to the learning coefficient for idealized singular systems in a beautiful [paper by Watanabe](https://www.jmlr.org/papers/volume14/watanabe13a/watanabe13a.pdf). All code for the experiments described can be found in [this GitHub repository](https://github.com/nrimsky/devinterp). Brief results ============= In our tests, we were pleasantly surprised to find that, for the task of modular addition modulo a prime p, the outputs of our implementation of Lau et al.'s SGLD methods are (up to a roughly constant small multiplicative error, less than a factor of 2) in robust agreement with the theoretically predicted results for an idealized single-layer modular addition network[[1]](#fnx9z3e6c3n7). Similar results have to date been obtained for a two-neuron network by [Lau et al.](https://www.researchgate.net/publication/373332996_Quantifying_degeneracy_in_singular_models_via_the_learning_coefficient) and for a 12-neuron network by [Chen et al.](https://arxiv.org/abs/2310.06301) Our results are the first confirmation for medium-sized networks (between about 500 and 8,000 neurons) of the agreement between the estimate and the theoretical results. While our results are off by a small multiplicative factor from the theoretical value for a single modular addition circuit, we discover a remarkably exact phenomenon that perfectly matches the theoretical predictions, namely that the learning coefficient estimate is linear in p for modular addition networks that generalize; this is the first precise scaling result of its kind. ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AhnoNzES5qiTnnavt/xuadozdjfow9omlhnshl)Mean measurements of ^λ on modular addition networks for 5 different values of p: the degree of linearity is remarkable.In addition, using the modular addition task as a test case lets us closely investigate the ability of the ^λ complexity estimate to differentiate generalization and memorization in neural networks: something that seems to be mostly new (though related to some of the phase transition phenomena in Chen et al.). We observe that while generalization has linear learning coefficient in the prime p, memorization has (roughly) quadratic growth in the prime p; this again exhibits remarkable agreement with theory. ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AhnoNzES5qiTnnavt/q68601o9wqsmcjkgnlw7)Our ^λ measurements for memorization-only networks. These are in remarkable agreement with the theoretically predicted values (here, theory predicts λ=0.8p2The agreement with theory holds for multiple different values of the prime p and multiple architectures. They also have appropriate behavior for networks that learn different numbers of circuits; a situation where other estimators of effective dimension, such as the Hessian-eigenvalue estimate, tend to overestimate complexity. Additionally, we show that the dynamic ^λ estimate[[2]](#fny802tk8mheb), i.e., the estimate during training, seems to track memorization vs. generalization stages of learning (this despite the fact that the ^λ estimate depends only on the training data). To see this, we use a slight refinement of the dynamical estimator, where we restrict sampling to lie within the normal hyperplane of the gradient vector at initialization, which seems to make this behavior more robust. ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AhnoNzES5qiTnnavt/qgdxysf9q1nbayda4hru)Chart of estimated ^λ over training for an MLP trained on modular addition mod 53. Checkpoints were taken every 60 batches of batch size 64. Hyperparameters for SGLD are γ=5, ϵ=0.001. The search was restricted to directions orthogonal to the gradient at the initialization point to correct for measurement at non-minima.Our dynamic results parallel some of the SGLD findings in [Chen et al.](https://arxiv.org/abs/2310.06301), which show that dynamic SGLD computations can sometimes notice phase transitions. We were pleasantly surprised to see them hold in larger networks and in the context of memorization vs. generalization. Overall, our findings update us to put more credence in the real-world applicability of Singular Learning Theory techniques and ideas. More concretely, we now believe that techniques similar to Lau et al.'s SGLD sampling should be able to distinguish different generalization behaviors in industry-scale neural networks and can be a part of a somewhat robust toolbox of unsupervised interpretability and control techniques valuable for alignment. Background ========== Basics about the learning coefficient λ --------------------------------------- For an alternative introduction, see [Jesse and Stan's excellent post](https://www.lesswrong.com/posts/6g8cAftfQufLmFDYT/you-re-measuring-model-complexity-wrong) explaining the learning coefficient (published after we had written this section but following a similar approach). The learning coefficient is a parameter associated with generalization. It controls the first-order asymptotic behavior of the question of "How likely is it that, given a random choice of weights, the loss they produce will be within δL of optimum". In other words, how easy is it to generalize the optimal solution to within δL accuracy. As δL goes to zero, this probability goes to zero polynomially, as an exponent of δL, so ProbLoss<δL=δd/2L+lower order corrections. (The d/2 instead of d is there for technical reasons.) Such a term (usually defined in terms of the free energy: here δL is the temperature) occurs more generally in statistical physics (and has close cousins in quantum field theory) as the leading exponent in the "perturbative expansion." In the context of neural nets, the exponent d it is called the learning coefficient or the RLCT ("real log canonical threshold," a term from algebraic geometry). The learning coefficient contains "dimension-like" information about a learning problem and can be understood as a measure of the effective dimension or "true dimensionality," i.e., the true number of weight parameters that need to be "guessed correctly" for a neural net to solve the problem with minimal loss. In particular, if a neural net is expanded by including redundant parameters that don't affect the set of algorithms that can be learned (e.g., because of symmetries of the problem), it can be shown that the learning coefficient does not change. Note that if the solution set to a machine learning problem is sufficiently singular (something we will not encounter in this post), the learning coefficient can be larger than the actual dimension of the set of minima[[3]](#fn1xv13wjmuli) and can indeed be a non-integer. The Watanabe-Lau-Murfet-Wei estimate, ^λ ---------------------------------------- In fact, the learning coefficient defined as a true asymptote only contains nontrivial information for singular networks, idealized systems that never appear in real life (just as it is not possible for two iterations of a noisy algorithm to give the exact same answer, so it is not possible for a network with any randomness to have a singular minimum or a positive-dimensional collection of minima). However, at finite but small values of temperature (i.e., loss "sensitivity," measured by δL as above), the problem of computing the associated free energy (and hence getting a meaningful generalization-relevant parameter at a "finite level of granularity") is tractable. The [paper of Watanabe](https://www.jmlr.org/papers/volume14/watanabe13a/watanabe13a.pdf) that Lau et al. follow gives a formula of this type. The result of that paper depends not only on the loss sensitivity parameter (called β, from the inverse temperature in statistical physics literature) but also on n, the number of samples. The formula gives an asymptotically precise estimate for the learning coefficient of the neural network on the "true" data distribution, corresponding to the limit as the number of samples n goes to infinity. As n goes to infinity, Watanabe takes the temperature parameter δL to zero as log(n)/n. Lau et al.'s paper sets out to perform this measurement at finite values of n. Having a good estimator for the learning coefficient can be extremely valuable for interpretability: this would be a parameter that captures the information-theoretic complexity of an algorithm in a very principled way that avoids serious drawbacks of previously known approaches (such as estimates of Hessian degeneracy) and can be useful for out-of-distribution detection. More generally, the [Singular Learning Theory](https://www.lesswrong.com/s/czrXjvCLsqGepybHC) program proposes certain powerful unsupervised interpretability tools that can give information about network internals, assuming the learning coefficient (and certain related quantities) can be computed efficiently. Modular addition as a testbed for estimating λ ---------------------------------------------- In Lau et al.'s paper, their SGLD-based learning coefficient estimate is applied to a tiny two-neuron network and also to an MNIST network, with promising results. We treat the modular addition network as an interesting intermediate case. Modular addition has to recommend itself the facts that: * It is a mechanistically interpreted network: we know its circuits, more or less how they are implemented by neurons, and how to isolate and measure them. * We can cleanly distinguish networks that learn to generalize vs. networks that only memorize by looking at their circuits; moreover, we can "spoof" generalization by creating a network for learning a random commutative operation; this is a network that has the same memorization behavior as modular addition, but no possibility of generalization. * Moreover, we can count the number of generalization circuits a network learns and reason about how different circuits interact in the loss function and in somewhat idealized free energy computations. This allows us to compare the behavior of ^λ with respect to the number of circuits against other notions of complexity, for example, Hessian rank. ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AhnoNzES5qiTnnavt/bxrvulcijbvpfbwqq8nc)Plot of loss on a held-out test set for a network trained on modular addition. Each step (differently colored line) corresponds to an evenly spaced checkpoint along the training. Each point corresponds to loss when all Fourier modes in the embedding weights matrix are ablated except this one. Keeping only a single important mode impacts loss much less than keeping only an unimportant mode, demonstrating the use of "grokked" Fourier modes in the embeddings matrix.At the same time, being an algorithmically generated problem, modular addition has some important limitations from the point of view of SLT, which makes it unable to capture some of the complexity of a typical learning problem: * The total number of possible data points for modular addition is finite (namely, equal to p2 for p the prime modulus), and the target distribution is deterministic. Thus, the learning coefficient only depends on a finite number of samples, which makes the asymptotic problem slightly (but not entirely) degenerate from the point of view of statistical learning theory. * Even within the class of simple deterministic machine learning problems, the modular addition problem is highly symmetric; thus, it is possible for our empirical results to fail to generalize for less symmetric networks. * The high number of possible output tokens compared to the maximal number of samples p tokens compared to p2 samples, for p the modulus) may cause unusual behavior (Watanabe's results assume that the number of logits is small and the number of samples is asymptotically infinite). Despite these limitations, we observed that (for an appropriate choice of hyperparameters) the Watanabe-Lau-Murfet-Wei estimate ^λ gives an estimate of the learning coefficient largely compatible with theoretical predictions. Moreover, the estimates behave in a remarkably consistent and stable way, which we did not expect. Findings ======== We found that, for fully trained networks, SGLD estimates using Watanabe's formula give a good approximation (up to a small factor) of the theoretical estimate for the RLCT, both for the modular addition (linear in p, reasonably independent of the total number of parameters) and for the random network (quadratic in p). Moreover, it is independent of the number of atomic circuits, or "groks" (something we expect, in an appropriate limiting case, to be the case for the learning coefficient but not for other computations of effective dimension). ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AhnoNzES5qiTnnavt/gkqtx00nee58ax9d8h6r)Diagram of the model we trained on the modular addition task. p-dimensional one-hot encoded numbers are embedded in an embed\_dim space. Two independent linear transformations are learned to a hidden\_dim space. The two vectors are then added elementwise, passed through a GELU activation function, and then transformed back into a p-dimensional vector of logits.We also ran some "dynamical" estimates of ^λ at unstable points along the learning trajectory of our modular addition networks. Here we observed that the ^λ estimates closely correlate to the validation (i.e., test) despite the fact that they are computed using methods involving only the training data. In particular, these unstable measurements "notice" the grokking transition between memorization and generalization when training loss stabilizes and test loss goes down. Scaling behavior for generalizing networks ------------------------------------------ We ran the Watanabe-Lau-Murfet-Wei ^λ-estimator algorithm on the following networks, and obtained the following results. We graph the ^λ estimate against each prime, averaged over five experiments. We found that estimates using Watanabe's formula gave a good approximation (up to a small factor) of the theoretical estimate for the RLCT, both for the modular addition and for the random network: ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AhnoNzES5qiTnnavt/xuadozdjfow9omlhnshl)Plot of ^λ estimated using SGLD for MLPs trained on modular addition mod different primes p. Here, the ^λ shown is averaged over five independent training and sampling runs. Hyperparameters for SGLD are γ=5, ϵ=0.001. We can see that ^λ scales linearly with p.![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AhnoNzES5qiTnnavt/t9eayflywv3vapupbtmj)^λ vs. p estimated for single runs of differently sized MLP networks, demonstrating similarity in RLCT across scales.![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AhnoNzES5qiTnnavt/m3raqy0b7opujni5bnyj)Here the different runs correspond to separately trained series of networks, demonstrating that ^λ is consistent across models with the same architecture trained to convergence on the same dataset and task.We observe that at a given architecture, our ^λ estimates are very close to linear, as would be theoretically predicted. In principle, the minimal effective dimensionality of a model with this architecture that solves modular addition is 4p (this will be elaborated on in a separate theory post deriving results about modular addition networks). However, we observe that the empirical scaling factor is very close to 8p=2×4p, double the result for a single circuit. A possible explanation for this result could be that, in the regime our models inhabit, the effective space of solutions consists of weight parameters that execute at least two simple circuits (all models we trained learned at least 4 simple circuits). ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AhnoNzES5qiTnnavt/rpnrzizva3nnkmapqsmi)The scaling factor of ^λ with p is close to 8 This could indicate that the SGLD search procedure explores a manifold of near-minima corresponding to two grokked circuits rather than one. Note that all trained models grokked >2 circuits altogether, and varied in their number of circuits, and so we still find invariance to the total number of independent circuits learned.When starting the experiment, we were expecting extensive differences of more than an order of magnitude between the empirical and predicted values (because of the non-ideal nature of the real-life models and limiting points in our experiments). This degree of agreement between a relatively large and messy "real-world" measurement and an ideal measurement, as well as the near-linearity here, are by no means guaranteed and updated us a significant amount towards believing that the theoretical predictions of Singular Learning Theory match well to real-world measurements. We also repeat the experiment at various architectures, with the number of parameters different by a relatively large factor (our largest network is more than 3 times larger than our smallest network, and our intermediate network is asymptotically twice as big as the smallest one). Larger networks do have slightly higher ^λ , but the difference scales sub-linearly in network size, as we would expect from the true learning coefficient. Note that the primes we include are relatively small. While our architectures are efficient and always generalize (with close to 100% accuracy) for much larger primes, we empirically observe that the estimates for ^λ tend to be much better and less noisy when the fully trained network is very close to convergence (0 loss). Because of computational limitations, we use a relatively large learning rate (0.01) for a relatively small number of iterations. This results in worse loss at convergence for primes above 50; we conjecture that the near-linear behavior would continue to hold for much larger primes if we used more computationally intensive methods with a smaller learning rate and a larger number of SGD steps. (In)dependence on the number of circuits ---------------------------------------- The networks we train sometimes learn different numbers of independent generalizing circuits embedded in different subspaces (the existence of such circuits was first proposed by [Nanda et al](https://arxiv.org/abs/2301.05217)). We can measure the number and types of circuits learned by a network, either by considering large outlier Fourier modes in the embedding space or (more robustly) by looking for near-perfect circles in "Fourier mode-aligned" two-dimensional projections of the embedding space[[4]](#fnyim8grymza), as in the picture below (We plan to later publish another post (on mechanistic interpretability tools for modular addition, in particular exactly distinguishing ["pizza" from "clock" circuits](https://arxiv.org/abs/2306.17844)), where these pictures will be explained more.) ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AhnoNzES5qiTnnavt/uria5hpsbh4aoqokve6p)We can see the number of independent Fourier mode circuits by projecting the learned embedding weights matrix to the (p−1)/2 two-dimensional subspaces of embedding space corresponding to the different discrete Fourier modes representable in the embedding space. For example, this model for p=43 has learned 6 Fourier modes - 4,8,14,15,19,21.We observe in our experiments that the learning rate estimates do not seem to depend much on the number of circuits learned. For example, for the largest prime we considered, p=53, the number of circuits learned in different runs varied between 4 and 7 circular circuits, whereas the learning coefficients for all the networks were within about 10% of each other. This result is deceptively simple but quite interesting and somewhat surprising from a theoretical viewpoint. For example, when measuring the effective dimension of a network via Hessian eigenvalues, a network with more than one circuit will have either effective dimension 0 (because going along a direction corresponding to any circuit counts as generalizing) or effective dimension that depends linearly on the number of circuits (because a direction counts as generalizing only if it independently generalizes each of the circuits). The fact that neither of these behaviors is observed in our context can be motivated by the Singular Learning Theory framework. Indeed, we can treat the subspace in weight space executing each circuit (or perhaps a suitable small subset of circuits) as a separate component of a singular manifold of "near-minima." As the vector spaces associated to the different circuits are in general position relative to each other, the resulting singularity is "minimally singular"[[5]](#fn0vw69e6p2rtp). This would mean that the RLCT at the singular point is equal to the RLCT along each of the individual components, which can be understood as an explanation for the observed independence result. However, we note that despite its explanatory robustness, this picture becomes more complicated when we zoom in since the loss for a multi-circuit network tends to be significantly better than the product of its parts. We plan to give an alternative explanation for the independence result involving a statistical model for cross-entropy loss that takes advantage of the ergodicity of multiplication modulo a prime. We flag here that we expect this independence to only hold in a "goldilocks" range of hyperparameter choices and, in particular, of the regularization constant (corresponding to the sizes of the circuits learned). A simplistic statistical model predicts at least three distinct phases here: one at a very small circuit size (corresponding to large regularization), where we expect the number of circuits to multiplicatively impact the learning rate. One at large circuit sizes (small regularization), where the learning rate estimate becomes degenerate, and one at an intermediate region, where the independence result we see is in effect. Random operations: scaling for memorization vs. generalization -------------------------------------------------------------- To compare our generalizing networks to networks with the same architecture, which only memorize, we ran the Watanabe-Lau-Murfet-Wei algorithm for a random commutative operation network. In order to get good loss for a memorization network, we need it to be overparametrized, i.e., the number of parameters needs to be above some appropriate O(1) multiple of the total number of samples, in our case p2. Because the number of parameters grows linearly in p, we get convergence to near zero loss only for small values of p. We note that since number-theoretic tricks like the Chinese Remainder Theorem are irrelevant for random operation networks, the values of p for this experiment do not need to be prime. Thus we run this experiment for multiples of 5 up to 40. Because of convergence issues and scaling pattern observation, we most trust our results in the short range of values between 5 and 25. Note that this range overlaps with our list of primes only between 23 and 25; we would need to use larger networks (and probably, better learning convergence) to get reasonable values of ^λ above this range. For the range of values we consider, we observe a larger learning coefficient with a quadratic scaling pattern in p, compared to the linear linear for generalizing networks. ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AhnoNzES5qiTnnavt/q68601o9wqsmcjkgnlw7)^λ vs. p for random commutative operation.![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AhnoNzES5qiTnnavt/ea8nruy9slra7rnaoi50)^λ vs. p for random commutative operation plotted alongside modular addition results.Remarkably, the diagram to p = 25 is almost exactly (up to a constant offset) equal to the number of memorizations, 0.8⋅p2; here 0.8 is the fraction of the full dataset used for training. We also generated data for larger multiples of 5, up to 40. Here we see clearly that the memorizing network has higher learning rate than the generalizing network at the same architecture, but the quadratic fit becomes worse for p>25. We believe that we would recover quadratic fit for more values of p if we worked with a larger network. Dynamics and phase transition ----------------------------- Finally, we performed a dynamic estimate of the learning coefficient at various checkpoints during the learning process for generalizing networks. In this part of our results, we introduced some innovations to the methods of [Lau et al.](https://www.researchgate.net/publication/373332996_Quantifying_degeneracy_in_singular_models_via_the_learning_coefficient) and [Chen et al](https://arxiv.org/abs/2310.06301). (though we did not implement the "health-based" sampling trajectory sorting from the latter paper). Specifically, we got the best results with a temperature adjustment and with our implementation of unstable SGLD applied after restricting to the normal hyperplane to loss gradient. ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AhnoNzES5qiTnnavt/q1sklqenvvtrvmsezcqy)A run of ^λ estimation at 25 equally spaced checkpoints along model training. SGLD search was modified to restrict search directions to those orthogonal to the gradient at initialization.![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AhnoNzES5qiTnnavt/vexyifqwkko2uecghzia)A repeated run with another model / independent SGLD sampling, to check consistency of results.Here we observed that the unstable ^λ estimates closely correlate to the validation (i.e., test) despite the fact that they are computed using methods involving only the training data. In particular, these unstable measurements "notice" the grokking transition between memorization and generalization when training loss stabilizes and test loss goes down. (As our networks are quite efficient, this happens relatively early in training.) 1. [^](#fnrefx9z3e6c3n7)Note that [Lau et al.](https://paperswithcode.com/paper/quantifying-degeneracy-in-singular-models-via) also undertake an estimate of $\hat{\lambda}\$for a large MNIST network with over a million neurons. Here they find that the resulting value for $\hat{\lambda}\$is correlated with the optimization method used to train the network in a predictable direction and thus captures nontrivial information about the basin. However, the theoretical value of $\hat{\lambda}\$is not available here, and the SGLD algorithm fails to converge; thus, this estimate is not expected to give a faithful value of the learning coefficient in this case 2. [^](#fnrefy802tk8mheb)Note that the dynamic $\hat{\lambda}\$estimator attempts to apply a technique designed for stable points (i.e., local minima) to points that are not local minima and have some instability, sampling, and ergodicity issues, even with our normal-to-gradient restriction refinement. In particular, they (much more than estimates at stable points) are sensitive to hyperparameters. Thus these unstable $\hat{\lambda}\$ measurements do not currently have an associated exact theoretical value and can be thought of as an ad hoc generalization of a complexity estimate to unstable points. Nevertheless, we find that at a fixed collection of hyperparameters, these estimates give consistent results and look similar across runs, and we see that they contain nontrivial information about the loss landscape dynamics during learning. 3. [^](#fnref1xv13wjmuli)An intuition for this is that very singular loss functions (i.e., functions that have many higher-order derivatives equal to zero) are associated with very large basins, which are large enough to "fit in extra dimensions worth of parameters." 4. [^](#fnrefyim8grymza)The two-dimensional subspace of the embedding space Rembed\_dim associated with the kth Fourier mode is the space spanned by the sin and cos components of the k-frequency discrete Fourier transform. Note that these spaces are not necessarily linearly independent for different modes but are independent for modes that learn a circuit. 5. [^](#fnref0vw69e6p2rtp) This is meant in an RLCT sense. In algebraic geometry language, a function f on weight space Rn is minimally singular if there exists a smooth analytic blowup X→Rn such that in local coordinates on X, f is a product of squares of coordinate functions. In this language, if we have c circuits associated to vector subspaces C1,…,Cc in weight space, an "idealized" function with minima on k-tuples of circuits is the function f(w)=∑S⊂{1,…,c},\|S\|=c−k+1∏i∈Sdist(w,Ci)2 for S running over c−k+1-element subsets and dist(w,Cij) the L2 distance from a weight to the corresponding subspace. It is easy to check that the resulting singularity is minimally singular.
badae76b-5aa3-4adb-ba5c-975480be2fe3	trentmkelly/LessWrong-43k	LessWrong	The Counterfactual Prisoner's Dilemma Updateless decision theory asks us to make decisions by imagining what we would have pre-committed to ahead of time. There's only one problem - we didn't commit to it ahead of time. So we do we care about what would have happened if we had? This isn't a problem for the standard Newcomb's problems. Even if we haven't formally pre-committed to an action such as by setting up consequences for failure, we are effectively pre-commited to whatever action we end up taking. After all the universe is deterministic, so from the start of time there was only one possible action we could have taken. So we can one-box and know we'll get the million if the predictor is perfect. However there are other problems where the benefit accrues to a counterfactual self instead of to us directly such as in Counterfactual Mugging. This is discussed in Abram Demski's post on all-upside and mixed-upside updatelessness. It's the later type that is troublesome. I posted a question about this a few days ago: > If you are being asked for $100, you know that the coin came up heads and you won't receive the $10000. Sure this means that if the coin would have been heads then you wouldn't have gained the $10000, but you know the coin wasn't heads so you don't lose anything. It's important to emphasise: this doesn't deny that if the coin had come up heads that this would have made you miss out on $10000. Instead, it claims that this point is irrelevant, so merely repeating the point again isn't a valid counter-argument. A solution In that post I cover many of the arguments for paying the counterfactual mugger and argue that they don't solve it. However, after posting, both Cousin_it and I independently discovered a thought experiment that is very persuasive (in favour of paying). The setup is as follows: > Omega, a perfect predictor, flips a coin and tell you how it came up. If if comes up heads, Omega asks you for $100, then pays you $10,000 if it predict you would have paid if it had come up t
618f5a2a-10af-4449-972c-a5d81cee34ca	trentmkelly/LessWrong-43k	LessWrong	Throw a prediction party with your EA/rationality group TL;DR: Prediction & calibration parties are an exciting way for your EA/rationality/LessWrong group to practice rationality skills and celebrate the new year. On December 30th, Seattle Rationality had a prediction party. Around 15 people showed up, brought snacks, brewed coffee, and spent several hours making predictions for 2017, and generating confidence levels for those predictions. This was heavily inspired by Scott Alexander’s yearly predictions. (2014 results, 2015 results, 2016 predictions.) Our move was to turn this into a communal activity, with a few alterations to meet our needs and make it work better in a group. Procedure: * Each person individually writes a bunch of predictions for the upcoming year. They can be about global events, people’s personal lives, etc. * If you use Scott Alexander’s system, create 5+ predictions each for fixed confidence levels (50%, 60%, 70%, 80%, 90%, 95%, etc.) * If you want to generate Brier scores or logarithmic scores, just do 30+ predictions at whatever confidence levels you believe. * Write down confidence levels for each prediction. * Save your predictions and put it aside for 12 months. * Open up your predictions and see how everyone did. To make this work in a group, we recommend the following: * Don’t share your confidence intervals. Avoid anchoring by just not naming how likely or unlikely you think any prediction is. * Do share predictions. Generating 30+ predictions is difficult, and sharing ideas (without confidence levels) makes it way easier to come up with a bunch. We made a shared google doc, and everyone pasted some of their predictions into it. * Make predictions that, in a year, will verifiably have happened or not. (IE, not “the academic year will go well”, which is debatable, but “I will finish the year with a 3.5 GPA or above”.) * It’s convenient to assume that unless stated otherwise predictions that end by the next year (IE, "I will go to the Bay Area" means "I will go to the B
a7c49b72-0b0d-4f61-881b-9694da4a6893	trentmkelly/LessWrong-43k	LessWrong	Projecting compute trends in Machine Learning Summary Using our dataset of milestone Machine Learning models, and our recent analysis of compute trends in ML, we project forward 70 years worth of trends in the amount of compute used to train Machine Learning models. Our simulations account for (a) uncertainty in estimates of the growth rates in compute usage during the Deep Learning (DL)-era and Pre-DL era, and (b) uncertainty over the ‘reversion date’, i.e. the date when the current DL-era compute trend (with a ~6 month doubling time) will end and revert to the historically more common trend associated with Moore’s law. Assuming a reversion date of between 8 to 18 years, and without accounting for algorithmic progress, our projections suggest that the median of Cotra 2020’s biological anchors may be surpassed around August 2046 [95% CI: Jun 2039, Jul 2060]. This suggests that historical rates of compute scaling, if sustained briefly (relative to how long these trends have been around so far), could result in the emergence of transformative models. Our work can be replicated using this Colab notebook. Note: we present projections, not predictions. Our post answers the question of: “What would historical trends over the past 70 years when naively extrapolated forward imply about the future of ML compute?” It does not answer the question: “What should our all-things-considered best guess be about how much compute we should expect will be used in future ML experiments?” Introduction Recently, we put together a dataset of over a hundred milestone Machine Learning models, spanning from 1952 to today, annotated with the compute required to train them. Using this data, we produce simple projections of the amount of compute that might be used to train future ML systems. The question of how much compute we might have available to train ML systems has received some attention in the past, most notably in Cotra’s Biological Anchors report. Cotra’s report investigates TAI timelines by analyzing: (i) the training comput
cf551f2e-d308-46f6-a358-126c2ec6f16e	trentmkelly/LessWrong-43k	LessWrong	Problems with learning values from observation I dunno if this has been discussed elsewhere (pointers welcome). Observational data doesn't allow one to distinguish correlation and causation. This is a problem for an agent attempting to learn values without being allowed to make interventions. For example, suppose that happiness is just a linear function of how much Utopamine is in a person's brain. If a person smiles only when their Utopamine concentration is above 3 ppm, then an value-learner which observes both someone's Utopamine levels and facial expression and tries to predict their reported happiness on the basis of these features will notice that smiling is correlated with higher levels of reported happiness and thus erroneously believe that it is partially responsible for the happiness. ------------------ an IMPLICATION: I have a picture of value learning where the AI learns via observation (since we don't want to give an unaligned AI access to actuators!). But this makes it seem important to consider how to make an un unaligned AI safe-enough to perform value-learning relevant interventions.
3d0ce28c-e4ae-4c4a-b056-a67e218e8fde	trentmkelly/LessWrong-43k	LessWrong	Link: Open-source programmable, 3D printable robot for anyone to experiment with Its name is Poppy. "Both hardware and software are open source. There is not one single Poppy humanoid robot but as many as there are users. This makes it very attractive as it has grown from a purely technological tool to a real social platform."
5cf6b6f0-2974-4a0b-8eeb-b4c43b947b40	trentmkelly/LessWrong-43k	LessWrong	Beware Social Coping Strategies Worlds The world is a huge, complex, scary place. Winter used to freeze us dead, and today lives are still destroyed by natural disasters like tsunami and forest fires. Plagues used to destroy our communities, and people still die from disease every day. We want more power over our surroundings — more intelligent computers, better space rockets — but construction is difficult and often dangerous. As a child, navigating the world directly is also terrifying: sharp objects, distressingly loud sounds, heights, electric shocks, etc. It takes an incredible amount of thought to merely understand and predict the world, let alone to avoid being hurt by it. The social world is even worse. Every person is a world in themselves. Every person has their own model of the world, plus their own thoughts and dreams and desires and morality. These may not always be consistent with your own, and that can cause conflicts. Their preferences are not always consistent with the preferences of other people, and conflicts between them can affect you. Individuals can also have conflicts inside themselves, making them unpredictable and confusing. And it's not just individuals you have to navigate. When people get together, they form groups and institutions. You are born, and find it's not just sharp objects you have to watch out for, but violating codes that were made up by people you've never met. You find yourself in complicated systems with their own logic and rules: government, school, a family structure, religious or political groups. These systems are additional worlds to learn — and they share the frustrating property of individuals that they are not entirely consistent. Navigating all these worlds, and learning how to follow or evade their rules without getting hurt, is tough. Being Thwarted As if that weren't bad enough, the social world has a special property that the natural world doesn't have: adaptive problems. That is, problems which adapt when you try to s
8575eba6-8506-4e9b-90bb-8ef0a4b245bb	trentmkelly/LessWrong-43k	LessWrong	What are the strongest arguments for very short timelines? I'm seeing a lot of people on LW saying that they have very short timelines (say, five years or less) until AGI. However, the arguments that I've seen often seem to be just one of the following: 1. "I'm not going to explain but I've thought about this a lot" 2. "People at companies like OpenAI, Anthropic etc. seem to believe this" 3. "Feels intuitive based on the progress we've made so far" At the same time, it seems like this is not the majority view among ML researchers. The most recent representative expert survey that I'm aware of is the 2023 Expert Survey on Progress in AI. It surveyed 2,778 AI researchers who had published peer-reviewed research in the prior year in six top AI venues (NeurIPS, ICML, ICLR, AAAI, IJCAI, JMLR); the median time for a 50% chance of AGI was either in 23 or 92 years, depending on how the question was phrased. While it has been a year since fall 2023 when this survey was conducted, my anecdotal impression is that many researchers not in the rationalist sphere still have significantly longer timelines, or do not believe that current methods would scale to AGI. A more recent, though less broadly representative, survey is reported in Feng et al. 2024, In the ICLR 2024 "How Far Are We From AGI" workshop, 138 researchers were polled on their view. "5 years or less" was again a clear minority position, with 16.6% respondents. On the other hand, "20+ years" was the view held by 37% of the respondents. Most recently, there were a number of "oh AGI does really seem close" comments with the release of o3. I mostly haven't seen these give very much of an actual model for their view either; they seem to mostly be of the "feels intuitive" type. There have been some posts discussing the extent to which we can continue to harness compute and data for training bigger models, but that says little about the ultimate limits of the current models. The one argument that I did see that felt somewhat convincing were the "data wall" and "unhobbling
6f38346e-9fd7-4f1b-b0b2-e93f63c80678	trentmkelly/LessWrong-43k	LessWrong	Troubles With CEV Part1 - CEV Sequence <!-- @page { margin: 2cm } P { margin-bottom: 0.21cm } A:link { so-language: zxx } --> The CEV Sequence Summary: The CEV sequence consists of three posts tackling important aspects of CEV. It covers conceptual, practical and computational problems of CEV's current form. On What Selves Are draws on analytic philosophy methods in order to clarify the concept of Self, which is necessary in order to understand whose volition is going to be extrapolated by a machine that implements the CEV procedure. Troubles with CEV part1 and Troubles with CEV part2 on the other hand describe several issues that will be faced by the CEV project if it is actually going to be implemented. Those issues are not of conceptual nature. Many of the objections shown come from scattered discussions found on the web. Finally, six alternatives to CEV are considered.
13b876a6-510a-469c-8dc0-1a082ac1458e	StampyAI/alignment-research-dataset/alignmentforum	Alignment Forum	Evolution is a bad analogy for AGI: inner alignment TL;DR: The dynamics of human learning processes and reward circuitry are more relevant than evolution for understanding how inner values arise from outer optimization criteria. This post is related to Steve Byrnes’ [Against evolution as an analogy for how humans will create AGI](https://www.lesswrong.com/posts/pz7Mxyr7Ac43tWMaC/against-evolution-as-an-analogy-for-how-humans-will-create), but more narrowly focused on how we should make inferences about values. Thanks to Alex Turner, Charles Foster, and Logan Riggs for their feedback on a draft of this post. Introduction ------------ How should we expect AGI development to play out? True precognition appears impossible, so we use various analogies to AGI development, such as evolution, current day humans, or current day machine learning. Such analogies are far from perfect, but we still may be able to extract useful information by carefully examining them. In particular, we want to understand how inner values relate to the outer optimization criteria. Human evolution is one possible source of data on this question. In this post, I’ll argue that human evolution actually provides very little usable evidence on AGI outcomes. In contrast, analogies to the human learning process are much more fruitful. Inner values versus outer optimization criteria ----------------------------------------------- One way people motivate extreme levels of concern about inner misalignment is to reference the fact that evolution failed to align humans to the objective of maximizing inclusive genetic fitness. From Eliezer Yudkowsky’s [AGI Ruin post](https://www.lesswrong.com/posts/uMQ3cqWDPHhjtiesc/agi-ruin-a-list-of-lethalities): > 16.Even if you train really hard on an exact loss function, that doesn't thereby create an explicit internal representation of the loss function inside an AI that then continues to pursue that exact loss function in distribution-shifted environments. Humans don't explicitly pursue inclusive genetic fitness; outer optimization even on a very exact, very simple loss function doesn't produce inner optimization in that direction. This happens in practice in real life,it is what happened in the only case we know about… > > I don't think that "evolution -> human values" is the most useful reference class when trying to understand how outer optimization criteria relate to inner values. Evolution didn't directly optimize over our values. It optimized over our learning process and reward circuitry. Once you condition on a particular human's learning process + reward circuitry configuration + the human's environment, you screen off the influence of evolution on that human's values. So, there are really (at least) two classes of observations from which we can draw evidence: 1. "evolution's inclusive genetic fitness criteria -> a human's learned values" (as mediated by evolution's influence over the human's learning process + reward circuitry) 2. "a particular human's learning process + reward circuitry + training environment -> the human's learned values" I will present five reasons why I think evidence from (2) “human learning -> human values” is more relevant to predicting AGI. ### 1: Training an AI is more similar to human learning than to evolution The relationship we want to make inferences about is: * "a particular AI's learning process + reward function + training environment -> the AI's learned values" I think that "AI learning -> AI values" is much more similar to "human learning -> human values" than it is to "evolution -> human values". Steve Byrnes makes this case in much more detail in [his post on the matter](https://www.lesswrong.com/posts/pz7Mxyr7Ac43tWMaC/against-evolution-as-an-analogy-for-how-humans-will-create). Two of the ways I think AI learning more closely resembles human learning, and not evolution, are: 1. The simple type signatures of the two processes. Evolution is a bi-level optimization process, with evolution optimizing over genes, and the genes specifying the human learning process, which then optimizes over human cognition. Evolution does not directly optimize over a human’s cognition. And because learned cognition is [not directly accessible](https://www.lesswrong.com/s/nyEFg3AuJpdAozmoX/p/CQAMdzA4MZEhNRtTp) to the genome, evolution must use roundabout methods to influence human values through the genome. In contrast, SGD directly optimizes over an AI’s cognition, just as [human within-lifetime learning](https://www.lesswrong.com/posts/fDPsYdDtkzhBp9A8D/intro-to-brain-like-agi-safety-8-takeaways-from-neuro-1-2-on) directly optimizes over human cognition. The human and AI learning processes are much closer to their respective cognitive structures, compared with evolution. 2. The differences between the parameter counts of the respective objects of optimization (the genome for evolution, the brain’s circuitry for human learning, and the AI’s parameter’s for AI training). The genome has very few parameters compared to even current day neural networks, much less the brain or future AGIs. Our experience with ML scaling laws very strongly implies that parameter counts matter a lot for a system’s learning dynamics. Better to compare highly parameterized systems to other highly parameterized systems. "AI learning -> AI values", "human learning -> human values", and “evolution -> human values” each represent very different optimization processes, with many specific dissimilarities between any pair of them. However, I think the balance of dissimilarities points to "human learning -> human values" being the closer reference class for "AI learning -> AI values". As a result, I think the vast majority of our intuitions regarding the likely outcomes of inner goals versus outer optimization should come from looking at the "human learning -> human values" analogy, not the "evolution -> human values" analogy. ### 2: We have more total evidence from human outcomes Additionally, I think we have a lot more total empirical evidence from "human learning -> human values" compared to from "evolution -> human values". There are billions of instances of humans, and each of them presumably have somewhat different learning processes / reward circuit configurations / learning environments. Each of them represents a different data point regarding how inner goals relate to outer optimization. In contrast, the human species only evolved once. Thus, evidence from "human learning -> human values" should account for even more of our intuitions regarding inner goals versus outer optimization than the difference in reference class similarities alone would indicate. ### 3: Human learning trajectories represent a broader sampling of the space of possible learning processes One common objection is that “human learning” represents a tiny region in the space of all possible mind designs, and so we cannot easily generalize our observations of humans to minds in general. This is, of course, true, and it greatly limits the strength of any AI-related conclusions we can draw from looking at "human learning -> human values". However, I again hold that inferences from "evolution -> human values" suffer from an even more extreme version of this same issue. "Evolution -> human values" represent an even more restricted look at the general space of optimization processes than we get from the observed variations in different humans' learning processes, reward circuit configurations, and learning environments. ### 4: Evidence from humans are more accessible than evidence from evolution Human evolution happened hundreds of thousands of years ago. We are deeply uncertain about the details of the human ancestral environment and which traits were under what selection pressure. We are still unsure about what precise selection pressure led humans to be so generally intelligent at all. We are very far away from being able to precisely quantify all the potentially values-related selection pressures in the ancestral environment, or how those selection pressures changed our reward systems or our tendencies to form downstream values. In contrast, human [within lifetime learning](https://www.lesswrong.com/posts/fDPsYdDtkzhBp9A8D/intro-to-brain-like-agi-safety-8-takeaways-from-neuro-1-2-on) happens all the time right now. It’s available for analysis and even experimental intervention. Given two evidence sources about a given phenomenon, where one evidence source is much more easily accessible than the other, then all else equal, the more accessible evidence source should represent a greater fraction of our total information on the phenomenon. This is another reason why we should expect evidence from humans to account for a greater proportion of our total information about how inner values relate to outer optimization criteria. ### 5: Evolution could not have succeeded anyways I think that a careful account of how evolution shaped our learning process in the ancestral environment implies that evolution had next to no chance of aligning humans with inclusive genetic fitness. There are no features of the ancestral environment which would lead to an ancestral human learning about the abstract idea of inclusive genetic fitness. There were no ancestral humans that held an explicit representation of inclusive genetic fitness. So, there was never an opportunity for evolution to select for humans who attached their values to an explicit representation of inclusive genetic fitness. Regardless of how difficult it is, in general, to get learning systems to form values around different abstract concepts, evolution could not have possibly gotten us to form a value around the particular abstraction of inclusive genetic fitness because we didn’t form such an abstraction in the ancestral environment. Ancestral humans had zero variance in their tendency to form values around inclusive genetic fitness. Evolution cannot select for traits that don’t vary across a population, so evolution could not have selected for humans that formed their values around inclusive genetic fitness. In contrast, the sorts of things that we humans end up valuing are usually the sorts of things that are easy to form abstractions around. Thus, we are not doomed by the same difficulty that likely prevented evolution from aligning humans to inclusive genetic fitness. This point is extremely important. I want to make sure to convey it correctly, so I will quote two previous expressions of this point by other sources: [Risks from Learned Optimization](https://arxiv.org/abs/1906.01820) notes that the lack of environmental data related to inclusive genetic fitness effectively increases the description length complexity of specifying an intelligence that deliberately optimizes for inclusive genetic fitness: > …description cost is especially high if the learned algorithm’s input data does not contain easy-to-infer information about how to optimize for the base objective. Biological evolution seems to differ from machine learning in this sense, since evolution’s specification of the brain has to go through the information funnel of DNA. The sensory data that early humans received didn’t allow them to infer the existence of DNA, nor the relationship between their actions and their genetic fitness. Therefore, for humans to have been aligned with evolution would have required them to have an innately specified model of DNA, as well as the various factors influencing their inclusive genetic fitness. Such a model would not have been able to make use of environmental information for compression, and thus would have required a greater description length. In contrast, our models of food, pain, etc. can be very short since they are directly related to our input data. > > From Alex Turner (in private communication): > If values form because reward sends reinforcement flowing back through a person's cognition and reinforces the thoughts which (credit assignment judges to have) led to the reward, then if a person never thinks about inclusive reproductive fitness, they can never ever form a value shard around inclusive reproductive fitness. Certain abstractions, like lollipops or people, are convergently learned early in the predictive-loss-minimization process and thus are easy to form values around. But if there aren't local mutations which make a person more probable to think thoughts about inclusive genetic fitness before/while the person gets reward, then evolution can't instill this value. Even if the descendents of that person will later be ableto think thoughts about fitness. > > Total significance of evolution ------------------------------- There are many sources of empirical evidence that can inform our intuitions regarding how inner goals relate to outer optimization criteria. My current (not very deeply considered) estimate of how to weight these evidence sources is roughly: * ~60% from "human learning -> human values" * ~4% from "evolution -> human values" * ~36% from various other evidence sources, which I won't address further in this post, such as: + economics + microbial ecology + politics + current results in machine learning + game theory / multi-agent negotiation dynamics Edit: since writing this post, I've learned a lot more about inductive biases and what deep learning theory we currently have, so my relative weightings have shifted quite a lot towards "current results in machine learning". Implications ------------ I think that using "human learning -> human values" as our reference class for inner goals versus outer optimization criteria suggests a much more straightforward relationship between the two, as compared to the (lack of a) relationship suggested by "evolution -> human values". Looking at the learning trajectories of individual humans, it seems like a given person's values have a great deal in common with the sorts of experiences they've found rewarding in their lives up to that point in time. E.g., a person who grew up with and displayed affection for dogs probably doesn't want a future totally devoid of dogs, or one in which dogs suffer greatly. Please note that I am not arguing that humans are inner aligned, or that looking at humans implies inner alignment is easy. Humans are misaligned with maximizing their outer reward source (activation of reward circuitry). I operationalize this misalignment as: "After a distributional shift from their learning environment, humans frequently behave in a manner that predictably fails to maximize reward in their new environment, specifically because they continue to implement values they'd acquired from their learning environment which are misaligned to reward maximization in the new environment". For example, one way in which humans are inner misaligned is that, if you introduce a human into a new environment which has a button that will wirehead the human (thus maximizing reward in the new environment), but has other consequences that are extremely bad by light of the human's preexisting values (e.g., killing a beloved family member), most humans won't push the button. I also think this regularity in inner values is reasonably robust to large increases in capabilities. If you take a human whose outer behavior suggests they like dogs, and give that human very strong capabilities to influence the future, I do not think they are at all likely to erase dogs from existence. It's probably not as robust to your choice of which specific human to try this with. E.g., many people would screw themselves over with reckless self-modification. My point is that higher capabilities alone do not automatically render inner values completely alien to those demonstrated at lower capabilities. (Part 2 will address whether the “sharp left turn” demonstrated by human capabilities with respect to evolution implies that we should expect a similar sharp left turn in AI capabilities.)
5872742a-e8ba-4963-b20d-bfdb202d7466	trentmkelly/LessWrong-43k	LessWrong	MIT Challenge: blogger to attempt CS curriculum on own Scott H. Young is giving himself 12 months to complete MIT's computer science curriculum on his own, via MIT's OpenCourseWare.
d1c4a696-5e5e-4f20-a86b-a1ef99c3a487	trentmkelly/LessWrong-43k	LessWrong	Quantum Mechanics, Nothing to do with Consciousness Epistemic status: A quick rejection of the quantum consciousness woo. If you have already read the sequences, there's nothing new in here. If your new to the site, or need a single page to point people to, here it is. Real Quantum mechanics looks like pages of abstract maths, after which you have deduced the results of a physics experiment. Given how hard the maths is, most of the systems that we use quantum mechanics to predict are quite simple. One common experiment is to take a glass tube full of a particular element and run lots of electricity through it. The element will produce coloured light, like sodium producing orange, or neon producing red. So take a prism and split that light to see what colours are being produced. Quantum physisists will do lots of tricky maths about how the electrons move between energy levels to work out what colour different elements will produce. There have been no quantum mechanics experiments that show consciousness to have any relevance to particle physics. The laws of physics do not say what is or is not conscious, in much the same way that they don't say what is or is not a work of art. Of course, consciousness is a property of human brains, and human brains, like everything else in the universe, are made of electrons and quarks playing by quantum laws. The point is that human brains are not singled out for special treatment, they get the same rules as everything else. For the writers among you, think of a word processor feature that takes some text, and turns it into ALL CAPS. You can put a great novel into this feature if you want. The point is that the rule its self acts the same way whether or not it's given great literature. You can't use the rule to tell what is great literature, you have to read it and decide yourself. Consciousness, like literature, is a high level view that's hard to pin down precisely, and is largely a matter of how we choose to define it. Quantum mechanics is a simple, mechanistic rule. Yes I kn
195c04dc-f32b-4a64-a618-149e8f8c1929	trentmkelly/LessWrong-43k	LessWrong	The Language of Bird This is a fictional snippet from the AI Vignettes Day. I do not think this is a likely future, but it’s a useful future to think about - it gives a different lens for thinking about the alignment problem and potential solutions. It’s the sort of future I’d expect if my own research went far better than I actually expect, saw rapid adoption, and most other ML/AI research stalled in the meantime. [Transcript from PyCon 2028 Lightning Talk. Lightly edited for readability.] Ok, so, today we’re going to talk about Bird, and especially about the future of Bird and machine learning. First of all, what is Bird? I assume everyone here has heard of it and played around with it, but it’s hard to describe exactly what it is. You’ve maybe heard the phrase “human concept library”, but obviously this isn’t just a graph connecting words together. Some background. In ye olden days (like, ten years ago) a lot of people thought that the brain’s internal data structures were inherently illegible, an evolved hodgepodge with no rhyme or reason to it. And that turned out to be basically false. At the low level, yeah, it’s a mess, but the higher-level data structures used by our brains for concept-representation are actually pretty sensible mathematical structures with some nice universal properties. Those data structures are the main foundation of Bird. When you write something like “from Bird.World import Bookshelf”, the data structure you’re importing - the Bookshelf - is basically an accurate translation of the data structure your own brain uses to represent a bookshelf. And of course it’s hooked up to a whole world-model, representing all the pieces of a bookshelf, things you put on a bookshelf, where you’d find a bookshelf, etc, as well as the grounding of all those things in a lower-level world model, and their ultimate connection to sensors/actuators. But when writing code using Bird, we usually don’t have to explicitly think about all that. That’s the beauty of the data stru
9edc3f05-616b-49e9-b640-db1d297540fb	trentmkelly/LessWrong-43k	LessWrong	When can Fiction Change the World? I suspect that a nontrivial percentage of the people reading this became involved with the community because of Harry Potter and the Methods of Rationality. So to the extent that those who were drawn to join the community because of that source are making the world a better place, we have at least one clear example of a novel having an important impact I’ve made a living through self publishing novels for the last five years (specifically Pride and Prejudice variations, that is Jane Austen fan fiction). Recently inspired by conversations at EA Virtual and worries made more emotionally salient by GTP-3 examples, I decided that I wanted to put part of my professional time towards writing novels that might have a positive impact on conversations around AI. As part of this I did some thinking about when fiction seemed to exert an influence on public policy, and then I looked for academic research on the subject, and I think there are people in the community who will find this write up about the subject interesting and useful. Theoretical Model I identified four common mechanisms that seemed to be involved when fiction had a large impact on opinions. This is not an exhaustive list, and there is some overlap and fuzziness around the boundary of each concept. Radicalizing the already convinced: A classic example is Uncle Tom’s Cabin, a novel about a slave unjustly suffering written in the 1850s that was credited with helping to spark the Civil War. Uncle Tom’s Cabin did not introduce anyone to the idea that slavery was bad, or convince anyone who thought that slavery was a fine peculiar Southern institution that it was actually evil. However it seems to have radicalized Northern attitudes towards slavery, and it was part of the moment when enough one issue voters on slavery existed that the party system broke down and allowed the new abolitionist Republican party to win congress and the presidency in 1860. Research has been done via surveys to find out if readers o
c2a5382b-8fa8-465c-a714-b4b7a02279c9	StampyAI/alignment-research-dataset/arbital	Arbital	Shannon The shannon (Sh) is a unit of [https://arbital.com/p/-3xd](https://arbital.com/p/-3xd). One shannon is the difference in [entropy](https://arbital.com/p/info_entropy) between a [https://arbital.com/p/-probability_distribution](https://arbital.com/p/-probability_distribution) on a [https://arbital.com/p/-binary_variable](https://arbital.com/p/-binary_variable) that assigns 50% to each value, and one that assigns 100% to one of the two values.
c1f859cf-c330-4dd9-81c0-ada6b4754a61	StampyAI/alignment-research-dataset/lesswrong	LessWrong	Talk to me about your summer/career plans Last May, I wrote a post called [talk to me about your summer plans](https://forum.effectivealtruism.org/posts/eREwP2tom9QKkvCqR/talk-to-me-about-your-summer-plans). I was impressed by the number & quality of people who reached out. Some people found it helpful, so I'm doing it again. TLDR: Please reach out to me if you are interested in learning about summer opportunities or career transitions relating to AI safety. Also, feel free to share this post with people who might benefit (e.g., a researcher who was recently laid off by a tech company; a talented operations specialist who has been following the AI stuff but doesn't know what to do). You can message me on [LessWrong](https://www.lesswrong.com/users/akash-wasil) or the [EA Forum](https://forum.effectivealtruism.org/users/akash). Background ========== I often have 1:1 conversations with people who want to reduce AI x-risk. A few observations: * Peoples' motivations often don't match up with their plans. * It's often hard for people to keep track of all the posted opportunities. * It's even harder for people to keep track of all the unposted opportunities. * Sometimes people don't realize how their skills can be useful. * Sometimes, short conversations with me seem to lead someone to seriously consider changing how they spend their summer/career. What I can do ============= If you're interested in summer opportunities or career opportunities in AI safety, please reach out to me. Specific things I can do: * Connect you with opportunities/people that you might not know about. * Help you consider faster or more ambitious ways to spend your summer. * Ask [questions](https://forum.effectivealtruism.org/posts/uqcKKTRWcED6y3WFW/questions-that-lead-to-impactful-conversations) to help you consider [wider action spaces](https://forum.effectivealtruism.org/posts/HmDSQQtJkkfcXQibu/three-reflections-from-101-ea-global-conversations) Specific things I will try not to do: * Tell you what you should do. * Make you feel guilty if you have considerations for how to spend your time that go beyond maximizing the utility of the universe. What to do if you're interested =============================== Send me a message with: * Who you are * Your current level of knowledge/familiarity with AI safety * What your current plans are * What your biggest strengths/skills are * Miscellaneous information about your engagement with EA or your aptitudes (e.g., I completed an intro fellowship; I submitted a proposal to the ELK contest; I'm competent in operations and people management). Please spend less than 5 minutes on this; I will ask you if I need more information. Who should do this? =================== I expect I will be most helpful to people who are interested in technical AI safety, AI strategy/governance, operations/events, and outreach/field-building. I'm still happy to talk to people with other interests, though, and I'm also happy to connect them to others. Caveats: I am not a professional career advisor, some of my advice will be wrong, you should also get advice from other people, etc.
084483d3-9d93-4d86-9db0-6cf42c5057e2	trentmkelly/LessWrong-43k	LessWrong	Help the AI 2027 team make an online AGI wargame The AI Futures Project, the team behind AI 2027, has developed a tabletop exercise (TTX) to simulate AI takeoff where we facilitate people going through an AI 2027-like scenario with different people taking different roles, like POTUS, China, or the frontier AI company CEO. Reviews have been positive and people have found it useful for building their own models of how AI will go. We’re interested in developing an online version of this game, with the goal of eventually getting millions of people to play through takeoff and better understand the dynamics and risks involved. The other game characters could be driven by LLMs so we don’t need to facilitate it. The online TTX could be based on our scenario and background research alongside LLMs and various game elements. To this end, we’re looking for a full-stack developer with game design and UX experience/interest and a deep understanding of AGI risk. We previously worked with the Lightcone Infrastructure team for the AI 2027 website and might work with them again for this project, but first we’d like to see if anyone else is interested. We expect that the ideal online TTX would ultimately look and feel very different from the in-person experience, but would successfully convey similar content. If you’re interested, please fill in this quick form, message me, or comment below, we’d love to hear from you!
714a6a45-5b89-4fd8-b4d5-e0af9039ad00	trentmkelly/LessWrong-43k	LessWrong	If rationality is purely winning there is a minimal shared art This sites strapline is refining the art of human rationality. It assumes there is something worth talking about. However I've been seeing "rationality as winning" bare of any context. I think the two things are in conflict. Let me take the old favourite of the twelve virtues of rationality as representative of the arts (although I could probably find similar situations where any proposed tool doesn't win). Can I find contexts in which humans will not win if they adopt these virtues? 1) Curiousity. There are lots of instances in history where curiousity may have killed you, cramping your winning. Wandering in the wrong part of town, trying out the wrong berry. Curiousity can also be not useful if most facts are boring or useless. I am happy to be ignorant of many facts about the world (who won various sports games, colour of peoples socks). 2&3&4) Relinquishment/lightness/evenness. You might avoid going see the evidence or disbelieve evidence presented by someone you suspect of being a clever charlatan. If you "see" a woman in box get cut in half you shouldn't automatically think that the magician can actually do magic. There is no magic data to evidence converter. See also defying the data. 5) Argument. This is not a useful virtue if you are alone on a desert island. It also might not be worth engaging with people who are trying to waste your time or distract you from something else. See also trolls. 6) Empiricism. This is great if you've got plenty of time. Collecting data always takes time and energy, if you need these things to survive, I wouldn't recommend empiricism. 7) Simplicity: Simplicity is predictable. If you involved in a contest with someone you may need to chose to be unpredictable in order to win. 8) Humbleness: This is a sign of weakness and may be exploited by your opponents in some social situations. 9 & 10) Perfection & Precision: There is an aphorism for this. In social games you don't need to be perfect or infinitely precise just bet
2ee4b363-11a6-426a-9e2c-73c48812755c	StampyAI/alignment-research-dataset/blogs	Blogs	AGI-09 Survey [Baum et al.](http://sethbaum.com/ac/2011_AI-Experts.pdf) surveyed 21 attendees of the [AGI-09](http://agi-conference.org/2009/) conference, on [AGI](http://en.wikipedia.org/wiki/Artificial_General_Intelligence) timelines with and without extra funding. They also asked about other details of AGI development such as social impacts, and promising approaches. Their findings include the following: * The median dates when participants believe there is a 10% , 50% and 90% probability that AI will pass a Turing test are 2020, 2040, and 2075 respectively. * Predictions changed by only a few years when participants were asked to imagine $100 billion (or sometimes $1 billion, due to a typo) in funding. * There was apparently little agreement on the ordering of milestones (‘turing test’, ‘third grade’, ‘Nobel science’, ‘super human’), except that ‘super human’ AI would not come before the other milestones. * A strong majority of participants believed ‘integrative designs’ were more likely to contribute critically to creation of human-level AGI than narrow technical approaches. Details ------- ### Detailed results #### Median confidence levels for different milestones Table 1 shows median dates given for different confidence levels of AI reaching four benchmarks: able to pass an online third grade test, able to pass a Turing test, able to produce science that would win a Nobel prize, and ‘super human’. [![](https://sites.google.com/site/aiimpactslibrary/ai-timelines/predictions-of-human-level-ai-dates/ai-expert-surveys/baum-et-al-ai-expert-survey/baumetaltable2.jpg)](https://sites.google.com/site/aiimpactslibrary/ai-timelines/predictions-of-human-level-ai-dates/ai-expert-surveys/baum-et-al-ai-expert-survey/baumetaltable2.jpg?attredirects=0) #### Best guess times for various milestones Figure 2 shows the distribution of participants’ best guesses – probably usually interpreted as 50 percent confidence points – for the timing of these benchmarks, given status quo levels of funding. [![](https://sites.google.com/site/aiimpactslibrary/ai-timelines/predictions-of-human-level-ai-dates/ai-expert-surveys/baum-et-al-ai-expert-survey/BaumetalFigure2%20copy.jpg)](https://sites.google.com/site/aiimpactslibrary/ai-timelines/predictions-of-human-level-ai-dates/ai-expert-surveys/baum-et-al-ai-expert-survey/BaumetalFigure2%20copy.jpg?attredirects=0) #### Individual confidence intervals for each milestone Figure 4 shows all participants’ confidence intervals for all benchmarks. Participant 17 appears to be interpreting ‘best guess’ as something other than fiftieth percentile of probability, though the other responses appear to be consistent with this interpretation. [![](https://sites.google.com/site/aiimpactslibrary/ai-timelines/predictions-of-human-level-ai-dates/ai-expert-surveys/baum-et-al-ai-expert-survey/BaumetalFigure4%20copy.jpg)](https://sites.google.com/site/aiimpactslibrary/ai-timelines/predictions-of-human-level-ai-dates/ai-expert-surveys/baum-et-al-ai-expert-survey/BaumetalFigure4%20copy.jpg?attredirects=0) #### Expected social impacts Figure 6 illustrates responses to three questions about social impact. The participants were asked about the probability of negative social impact, if the first AGI that can pass the Turing test is created by an open source project, by the United States military, or by a private company focused on commercial profit. The paper summarises that the experts lacked consensus. [![baumetalsocialimpact copy](http://aiimpacts.wpengine.com/wp-content/uploads/2014/12/baumetalsocialimpact-copy.jpg)](http://aiimpacts.wpengine.com/wp-content/uploads/2014/12/baumetalsocialimpact-copy.jpg)‘Fig. 6. Probability of a negative-to-humanity outcome for different development scenarios. The three development scenarios are if the first AGI that can pass the Turing test is created by an open source project (x’s), the United States military (squares), or a private company focused on commercial profit (triangles). Participants are displayed in the same order as in figure 4, such that Participant 1 in figure 6 is the same person as Participant 1 in figure 4.’ ### Methodological details The survey contained a set of standardized questions, plus individualized followup questions. It can be downloaded from [here](http://sethbaum.com/ac/2011_AI-Experts.html). It included questions on: * when AI would meet certain benchmarks (passing third grade, turing test, Nobel quality research, superhuman), with and without billions of dollars of additional funding. Participants were asked for confidence intervals (10%, 25%, 75%, 90%) and ‘best estimates’ (interpreted above as 50% confidence levels). * Embodiment of the first AGIs (physical, virtual, minimal) * What AI software paradigm the first AGIs would be based on (formal neural networks, probability theory, uncertain logic, evolutionary learning, a large hand-coded knowledge-base, mathematical theory, nonlinear dynamical systems, or an integrative design combining multiple paradigms) * Probability of strongly negative-to-humanity outcome if the first AGIs were created by different parties (an open-source project, the US military, or a private for-profit software company) * If quantum computing or hypercomputing would be required for AGI. * Whether brain emulations would be conscious * The experts’ area of expertise #### Participants Most of the participants were actively involved in AI research. The paper describes them: > Study participants have a broad range of backgrounds and experience, all with significant prior thinking about AGI. Eleven are in academia, including six Ph.D. students, four faculty members, and one visiting scholar, all in AI or allied fields. Three lead research at independent AI research organizations and three do the same at information technology organizations. Two are researchers at major corporations. One holds a high-level administrative position at a relevant non-profit organization. One is a patent attorney. All but four participants reported being actively engaged in conducting AI research. > > According to [the website](http://agi-conference.org/2009/), the AGI-09 conference gathers “leading academic and industry researchers involved in serious scientific and engineering work aimed directly toward the goal of artificial general intelligence”. While these people are expert in the field, they are also probably highly selected for being optimistic about the timing of human-level AI. This seems likely to produce some bias. #### Meaning of ‘Turing test’ Several meanings of ‘Turing test’ are prevalent, and it is unclear what distribution of them is being used by participants. The authors note that some participants asked about this ambiguity, and were encouraged verbally to consider the ‘one hour version’ instead of the ‘five minute version’, because the shorter one might be gamed by chat-bots (p6). The authors also write, ‘Using human cognitive development as a model, one might think that being able to do Nobel level science would take much longer than being able to conduct a social conversation, as in the Turing Test’ (p8). Both of these points suggest that the authors at least were thinking of a Turing test as a test of normal social conversation rather than a general test of human capabilities as they can be observed via a written communication channel.
f6d0b052-b173-4748-a33c-e2c71c9f5867	trentmkelly/LessWrong-43k	LessWrong	[Book Review]: The Bonobo and the Atheist by Frans De Waal The Bonobo and the Atheist does something truly rare, which is to come up with an interesting way of viewing morality. 'What is right?' is the most pored-over and divisive question in all of philosophy, the book mostly ignores it. The best parts of it contain zero commandments. Instead Frans De Waal asks 'Why do we think [what is right] is right?' (that's my wording not his). The book focuses on both chimp and bonobo societies, but the commonalities between their behaviour are such that De Waal frequently swaps between them. The sex-heaviness of bonobo lives seems to be irrelevant to the main thesis beyond putting in perspective that not all ape behaviour is aggressive fighting. This is mostly unimportant because the altruism shown between captive chimps was enough to make this point. De Waal spends a lot of the book arguing against a few different scientific orthodoxies. These are all the standard pet hypothesis privileging. He also spends a lot of time talking about religion (as was typical for 2013) but I didn't find those sections interesting so I won't be discussing them. The book is fundamentally about evolutionary psychology, which is notoriously tricky and easy to misunderstand, so unfortunately I think it's necessary to spend the next section laying out definitions. Sorry. Altruism, Altruism, and Altruism De Waal uses the word altruism to mean different phenomena. While he describes this distinction he doesn't codify it in language. I'm going to do that for him: * R-altruism is the giving of resources (like food, aid, etc.) to another, which may or may not improve reproductive prospects for your genes * G-altruism is a a subset of r-altruism which specifically harms your own genes' prospects for reproduction * E-altruism is a cluster of similar emotional responses similar to the ones in a human brain, which lead to r-altruism. The cluster is somewhat centred on empathy, which De Waal (mostly) uses to mean a negative emotional reaction to another in
b7314afd-5226-497e-bdbd-a56ba8267380	trentmkelly/LessWrong-43k	LessWrong	Electric heat pumps (Mini-Splits) vs Natural gas boilers I’m kinda out of the loop on this stuff, but it seems like suddenly out of nowhere everyone around me is converting natural gas heat into electric heat pumps—specifically a thing called a “mini-split” which can be both a heat pump in the winter and an air conditioner in the summer. It consists of one or more indoor units connected to one or more outdoor units by a small-ish (well, small compared to an air duct) tube of flowing liquid refrigerant (which transfers heat between the indoor and outdoor unit). An electric heater converts 1W of electricity into 1W of heat. A heat pump converts 1W of electricity into >1W of heat, because it pulls heat from the outside air—it’s basically an air conditioner pointed backwards. The ratio of heat to electricity is called Coefficient of Performance (COP). The COP for heat pumps gets worse (albeit still >1) as the outdoor temperature goes down, basically because the outdoor air is colder so it’s harder to extract heat from it. I live in a pretty cold climate; I figure the weighted-average outdoor temperature (weighted by how much heating I'm using) is 30°F (~0°C). A quick search suggests maybe I can expect COP of 3-3.5. The highest figure I found anywhere for that outdoor temperature was 5. I could be wrong here. (Update: in this comment I used a different method based on the "HSPF" spec and concluded that I should definitely expect weighted-average-COP below 3.5 where I live, unless I messed up the calculation, which is entirely possible.) Meanwhile I also need to account for my gas boiler not successfully transferring all the heat from "burning gas" to "warm interior" (as opposed to losing heat out the chimney or whatever). I think 80% is a reasonable guess for our not-particularly-well-tuned boiler. It should go up a bit, maybe towards 90%, when I get it tuned up or replaced at some point in the future, which I need to do anyway, it's super old and the plumber says it could catastrophically break any day. I looked up my m
07e44eee-6dea-4ebe-8b28-d320d8cbc29c	StampyAI/alignment-research-dataset/lesswrong	LessWrong	Race to the Top: Benchmarks for AI Safety This is an executive summary of a post from my personal blog, also [cross-posted](https://forum.effectivealtruism.org/posts/saEXX9Nucz8mh9XgB/race-to-the-top-benchmarks-for-ai-safety) from the EA Forum. Read the full texts [here](https://isaduan.github.io/isabelladuan.github.io/posts/first/).** Summary ------- Benchmarks support the empirical, quantitative evaluation of progress in AI research. Although benchmarks are ubiquitous in most subfields of machine learning, they are still rare in the subfield of AI safety. I argue that creating benchmarks should be a high priority for AI safety. While this idea is not new, I think it may still be underrated. Among other benefits, benchmarks would make it much easier to: * track the field’s progress and focus resources on the most productive lines of work; * create professional incentives for researchers - especially Chinese researchers - to work on problems that are relevant to AGI safety; * develop auditing regimes and regulations for advanced AI systems. Unfortunately, we cannot assume that good benchmarks will be developed quickly enough “by default." I discuss several reasons to expect them to be undersupplied. I also outline actions that different groups can take today to accelerate their development. For example, AI safety researchers can help by: * directly trying their hand at creating safety-relevant benchmarks; * clarifying certain safety-relevant traits (such as “honesty” and “power-seekingness”) that it could be important to measure in the future; * building up relevant expertise and skills, for instance by working on other benchmarking projects; * drafting “benchmark roadmaps,” which identify categories of benchmarks that could be valuable in the future and outline prerequisites for developing them. And AI governance professionals can help by: * co-organizing workshops, competitions, and prizes focused on benchmarking; * creating third-party institutional homes for benchmarking work; * clarifying, ahead of time, how auditing and regulatory frameworks can put benchmarks to use; * advising safety researchers on political, institutional, and strategic considerations that matter for benchmark design; * popularizing the narrative of a “race to the top” on AI safety. Ultimately, we can and should begin to build benchmark-making capability now. Acknowledgment -------------- I would like to thank Ben Garfinkel and Owen Cotton-Barratt for their mentorship, Emma Bluemke and many others at the [Centre for the Governance of AI](https://www.governance.ai/) for their warmhearted support. All views and errors are my own. Future research --------------- I am working on a paper on the topic, and if you are interested in benchmarks and model evaluation, especially if you are a technical AI safety researcher, I would love to [hear](mailto: isabelladuan@uchicago.edu) from you!
66ee33d2-f224-4e15-a00a-2fce68245f3d	trentmkelly/LessWrong-43k	LessWrong	Finding Cruxes This is the third post in the Arguing Well sequence, but it can be understood on its own. This post is influenced by double crux, is that your true rejection, and this one really nice street epistemology guy. The Problem Al the atheist tells you, “I don’t believe in the Bible, I mean, there’s no way they could fit all the animals on the boat”. So you sit Al down and give the most convincing argument on it indeed being perfectly plausible, answering every counterpoint he could throw at you, and providing mountains of evidence in favor. After a few days you actually managed to convince Al that it’s plausible. Triumphantly you say, “So you believe in the Bible now, right?” Al replies, “Oh no, there’s no evidence that a great flood even happened on Earth” "..." Sometimes when you ask someone why they believe something, they’ll give you a fake reason. They’ll do this without even realizing they gave you fake reason! Instead of wasting time arguing points that would never end up convincing them, you can discuss their cruxes. Before going too deep, here’s a shortcut: ask “if this fact wasn’t true, would you still be just as sure about your belief?” Ex. 1: Al the atheist tells you, “I don’t believe in the Bible, I mean, there’s no way they could fit all the animals on the boat” Instead of the wasted argument from before, I’d ask, “If I somehow convinced that there was a perfectly plausible way to fit all the animals on the boat, would you believe in the Bible?” (This is effective, but there’s something subtly wrong with it discussed later. Can you guess it?). Ex. 2: It’s a historical fact that Jesus existed and died on the cross. Josephus and other historical writers wrote about it and they weren’t Christians! If you didn’t know about those sources, would you still be just as sure that Jesus existed? General Frame A crux is an important reason for believing a claim. Everything else doesn’t really carry any weight. How would you generalize/frame the common prob
0eda3796-874e-4955-8198-bc753f9d96c1	trentmkelly/LessWrong-43k	LessWrong	Dealing with Curiosity-Stoppers Introduction Curiosity is a virtue. It promotes epistemic honesty, ignites creativity, and improves both competence and well-being. Multiple posts already discussed different types of curiosity the contrast between signalling curiosity and being curious, the scientific evidence behind curiosity, why curiosity seemed to leave children, even the limits of curiosity. Yet my own issues with curiosity come not with generating it but with keeping it. Everyday, a myriad of subjects and pieces of content spark the flame of curiosity in me; also everyday, recurring thoughts dampen, or sometimes blow out this flame. I thus need ways to address these thoughts more than techniques to become curious. What I call curiosity-stoppers -- inspired from but slightly different of this post from the Sequences -- brings a lot of negative to my life. My difficulties to focus on a specific topic of study stems in part from curiosity-stoppers in the way, which then make almost any other topic seems more interesting by contrast. When curiosity-stoppers proliferate, I cannot find anything I'm happy or interested to do, and I feel completely empty and drained. Even when these thoughts don't overpower me, they consistently push me to postpone reading, listening, watching or writing content I'm genuinely curious about and which might improve my life and my research. I intend this post as an exploration of my own curiosity-stoppers, as well as my personal counter-measures to each of them. In contrast with this post, mine isn't a scientific examination of the question based on an extensive literature. I merely catalog what I found grappling with my own issues with curiosity-stoppers. Since I find it difficult to believe I'm a one-of-a-kind special snowflake, I believe others will share part of my experience. I hope naming and addressing this issue directly will help them deal with it too. Definitions Let's get down to basics and define the main terms. First is "a discovery": a moment where I
d1f40299-9907-4c5d-b7bc-3abf33c8ad36	trentmkelly/LessWrong-43k	LessWrong	Why and how to write things on the Internet Recently I noticed that most existing “why you should write a blog” articles (e.g.) have at least one of two shortcomings, according to me: * They mostly focus on counterarguments to not starting a blog, rather than positive arguments in favor of starting one—as if people’s natural state is to produce amazing blogs and the only thing holding them back is silly misconceptions. This might be true for extreme outliers like Scott Alexander, but personally, my natural state is to play lots of video games and I need to be convinced to do other stuff instead. * They don’t set the reader up to actually get the benefits of blogging, because they don’t give concrete enough advice on the mechanics. I’ve learned a lot of things about blogging which, if I’d known them earlier, would have helped me improve much faster (and therefore would have made me much more likely to keep blogging, since part of my motivation for writing is being good at it). So here’s my own take on this genre of post, with positive reasons why you, yes you, should consider writing things on the Internet, and a guide that’s as concrete as I can manage for how to do a decent (imo) job at it. To put my money where my mouth is, I’m offering to be a draft reader for anyone who is motivated by this post to start a blog; see Appendix: standing offer for details. Why More awesome friendships In my opinion, the strongest reason for any random person to start a blog is that you will have more awesome friendships—both in the sense that you will meet new awesome people as a result of your blog, but also in the sense that writing will cause you to have more interesting ideas,1 which will make your existing friendships more awesome because you’ll have better stuff to talk about. Most other important things in life, like job opportunities and romantic relationships, are downstream of the quality of your friends, so this is pretty great. Examples from my own blog: * Often when I’m talking to friends I’ll try
8e0136d3-4695-425a-a995-79dbe48039ec	StampyAI/alignment-research-dataset/blogs	Blogs	probability under potential hardware failure probability under potential hardware failure -------------------------------------------- suppose you program a computer to do bayesian calculations. it does a bunch of math with probability numbers attached to, for example, logical belief statements. but, suppose that each time you access one of those numbers, there is a tiny chance `ε` of hardware failure causing the memory/register to return an erroneous number — such as [cosmic ray bit flips](https://en.wikipedia.org/wiki/Soft_error). this fact can inform our decisions about how many bits of information we are to store our numbers as. indeed, the computer can never have probability ranges outside of `[ε ; 1-ε]`: the probabilities are clamped by the chance that, while they were computed, a random hardware failure occured. if a probability is calculated that is a function of many calculations, then the errors can accumulate. the computer might be able to rerun the computation to be more sure of its result, but it will never escape the range `[ε ; 1-ε]`. this constraint feels to me like it would also limit the number of bits of precision one can meaningfully store: there is only so many ways to combine numbers in that range, with errors at each step of computation, before the signal is lost to error noise. i'm not sure and haven't worked out the math, but it may turn out that arbitrary-precision numbers, for example, are ultimately of no use: given a constant `ε`, there is a constant `f(ε)` maximum number of useful bits of precision. this issue relates to [the uncertainty of 2+2=4](uncertainty-2+2=4.html): logical reasoning on a computer or on a human is still probabilistic/error-prone, because of hardware failures.
d138adc6-07dd-45df-b4db-fa22d5be1d9e	trentmkelly/LessWrong-43k	LessWrong	Computation complexity of AGI design Summary of main point: I argue that there is a significant probability that creating de novo AGI is an intractable problem. Evolution only solved this problem because of anthropic reasons. Conclusions are drawn regarding priorities in AI risk research. Sketch of main argument: There are suggestive relations between AGI and NP-completeness. These relations lead me to hypothesize that AGI programs posses large Levin-Kolmogorov complexity which implies that producing them is a computationally intractable problem. The timing of events in the evolution of human intelligence seems to be consistent with the assumption evolution's success is anthropic, if we postulate human intelligence as arising from a combination of two modules: an "easy" (low complexity) module and a "hard" (high complexity) module. Therefore, creating superhuman intelligence will require reverse engineering the human brain and be limited to improving the "easy" module (since creating a better "hard" module is again computationally intractable). AGI and P vs. NP There are several arguments the AGI problem is of a similar "flavor" to problems that are NP-complete. The first argument is rather vague but IMO still compelling. Many class separations in complexity theory (P vs. NP, L vs. P, R vs. RE) hinge on the existence of a complete language. This means there is a single problem solving which under the stronger resource constraints would lead to solving all problems in the larger class. Similarly, Goedel incompleteness means there is no single algorithm (a program which terminates on all inputs) for proving all provable theorems. It feels like there is a principle of mathematics which rules out algorithms that are "too good to be true": a single "magic wand" to solve all problems. In a similar way, AGI is a "magic wand": it solves "all" problems because you can simply delegate them to the AGI. Another argument has to do with Solomonoff induction. Solomonoff induction is incomputable but it becomes c
ea2148b0-afbd-452e-bdd3-b72bd992a149	trentmkelly/LessWrong-43k	LessWrong	2014 Less Wrong Census/Survey - Call For Critiques/Questions It's that time of year again. Actually, a little earlier than that time of year, but I'm pushing it ahead a little to match when Ozy and I expect to have more free time to process the results. The first draft of the 2014 Less Wrong Census/Survey is complete (see 2013 results here) . You can see the survey below if you promise not to try to take the survey because it's not done yet and this is just an example! 2014 Less Wrong Census/Survey Draft I want two things from you. First, please critique this draft (it's much the same as last year's). Tell me if any questions are unclear, misleading, offensive, confusing, or stupid. Tell me if the survey is so unbearably long that you would never possibly take it. Tell me if anything needs to be rephrased. Second, I am willing to include any question you want in the Super Extra Bonus Questions section, as long as it is not offensive, super-long-and-involved, or really dumb. Please post any questions you want there. Please be specific - not "Ask something about taxes" but give the exact question you want me to ask as well as all answer choices. Try not to add more than a few questions per person, unless you're sure yours are really interesting. Please also don't add any questions that aren't very easily sort-able by a computer program like SPSS unless you can commit to sorting the answers yourself. I will probably post the survey to Main and officially open it for responses sometime early next week.
c0495305-1b2a-4779-952a-bebc767f5a1d	trentmkelly/LessWrong-43k	LessWrong	What New Desktop Should I Buy? It's time for me to buy a new desktop, as my old one isn't working so well anymore. It also happens to be 'black Friday in July' which opens up the opportunity to perhaps get a good sale price somewhere. Time is valuable, so I'm happy to spend a reasonably large amount if it actually matters, but I don't want to throw good money after nothing. Given how many people would benefit from getting this question right, and how many people are likely to have good answers, I figured I'd ask here. My specific goals are basically: 1. Windows 11, I'm not open to negotiation on this one for various reasons. 2. Will be able to open and rapidly switch between a ton of chrome tabs, including massive google sheets and google docs. 3. General reliability and future-proofing. 4. High-end gaming would be nice but mostly reliable medium-end gaming is fine. 5. Handle at least 3+ monitors well. 6. Relatively quiet fan is a big plus, especially loud ones are deal breakers. 7. The usual other stuff you'd want a computer to be able to do these days. Current top candidate after talking with one friend first is this Alienware PC configured with processor 12th Gen Intel® Core™ i7-12700KF (25 MB cache, 12 cores, 20 threads, 3.60 to 5.00 GHz Turbo), Win11 Pro, NVIDIA® GeForce RTX™ 3080, 10 GB GDDR6X, LHR, 64 GB memory, HDD 1 TB, M.2, PCIe NVMe, SSD (with plan to also use the existing computer's HD as well) and Office Student. Secondary possibility is Lenovo is having a massive sale on that includes the P620, seems like the more expensive versions are way more expensive than they need to be even post-sale so I'm suspicious (my old computer is a Lenovo and mostly was fine for a while until it wasn't). Also very open to additional brands/possibilities. UPDATE: Someone has volunteered to build it on my behalf, so yay. Still seems like a good question in general.
fb9174dc-df39-496f-bd65-efce6e430dc9	trentmkelly/LessWrong-43k	LessWrong	Parametric polymorphism in updateless intelligence metrics Followup to: Agents with Cartesian childhood and Physicalist adulthood In previous posts I have defined a formalism for quantifying the general intelligence of an abstract agent (program). This formalism relies on counting proofs in a given formal system F (like in regular UDT), which makes it susceptible to the Loebian obstacle. That is, if we imagine the agent itself making decisions by looking for proofs in the same formal system F then it would be impossible to present a general proof of its trustworthiness, since no formal system can assert is own soundness. Thus the agent might fail to qualify for high intelligence ranking according to the formalism. We can assume the agent uses a weaker formal system the soundness of which is provable in F but then we still run into difficulties if we want the agent to be self-modifying (as we expect it to be). Such an agent would have to trust its descendants which means that subsequent agents use weaker and weaker formal systems until self-modification becomes impossible. One known solution to this is Benja's parametric polymorphism. In this post I adapt parametric polymorphism to the updateless intelligence metric framework. The formal form of this union looks harmonious but it raises questions which I currently don't fully understand. "Ineffable mystery" using oracles instead of timeouts In the original parametric polymorphism, a constant κ is introduced (informally known as "the number of ineffable mystery") s.t. the agent has to prove its actions are "safe" for time period κ (i.e. have no "bad" consequences during this period). Since it's impossible to the agent to prove any upper bound on κ, effectively its actions have to be safe indefinitely. I found that for our purposes it's better to cast the formalism otherwise. Instead of κ playing the role of a "timeout", the agent is provided with an oracle that answers questions of the form "κ > n?" for any n. Contrary to what the agent thinks, the oracle provides an affi
ac473696-6ff6-43a0-afd2-e09afd67ac34	trentmkelly/LessWrong-43k	LessWrong	MIRI's July 2024 newsletter MIRI updates * Rob Bensinger suggests that AI risk discourse could be improved by adopting a new set of labels for different perspectives on existential risk from AI. One drawback of “AI doomer” (a label sometimes used in online discussions) is that it does not have a consistent meaning. * AI researcher John Wentworth guesses that a central difference between his and Eliezer Yudkowsky’s views might be that Eliezer expects AI to not use abstractions which are similar to those used by humans. Eliezer clarifies that he expects similar abstractions for predictive parts of the natural world, but larger differences for utility-laden concepts like human values and reflectivity-laden concepts like corrigibility. Also, Eliezer says that he is still concerned about the difficulty of reliably steering advanced ML systems even if they do use similar abstractions internally. * Eliezer joins Joscha Bach, Liv Boeree, and Scott Aaronson for a panel discussion (partly paywalled) on the nature and risks of advanced AI. News and links * In an excellent new video essay, Rob Miles discusses major developments in AI capabilities, policy and discourse over the last year, how those developments have shaped the strategic situation, and what individuals can do to help. * A report from Convergence Analysis gives a high-level overview of the current state of AI regulation. The report summarizes legislative text and gives short analyses for a range of topics. * A group of current and former employees of OpenAI and Google DeepMind released an open letter calling on frontier AI companies to adopt policies that allow whistleblowers to share risk-related information. This comes in the wake of news that former OpenAI employees were subject to an extreme non-disclosure policy that the company has since apologized for and seems to have removed. You can subscribe to the MIRI Newsletter here.
3c5273a4-1718-4d66-b354-9265f0d1fbcc	StampyAI/alignment-research-dataset/blogs	Blogs	Christiano and Yudkowsky on AI predictions and human intelligence This is a transcript of a conversation between Paul Christiano and Eliezer Yudkowsky, with comments by Rohin Shah, Beth Barnes, Richard Ngo, and Holden Karnofsky, continuing the [Late 2021 MIRI Conversations](https://intelligence.org/late-2021-miri-conversations/). Color key: \| \| \| \| --- \| --- \| \| Chat by Paul and Eliezer \| Other chat \| 15. October 19 comment ---------------------- [Yudkowsky][11:01] thing that struck me as an iota of evidence for Paul over Eliezer: <https://twitter.com/tamaybes/status/1450514423823560706?s=20> ![](https://39669.cdn.cke-cs.com/rQvD3VnunXZu34m86e5f/images/bccc3fcd315a12be417c814fe75a6c761049a9225425cc67.png) 16. November 3 conversation --------------------------- ### 16.1. EfficientZero [Yudkowsky][9:30] Thing that (if true) strikes me as… straight-up falsifying Paul’s view as applied to modern-day AI, at the frontier of the most AGI-ish part of it and where Deepmind put in substantial effort on their project? EfficientZero (allegedly) learns Atari in 100,000 frames. Caveat: I’m not having an easy time figuring out how many frames MuZero would’ve required to achieve the same performance level. MuZero was trained on 200,000,000 frames but reached what looks like an allegedly higher high; the EfficientZero paper compares their performance to MuZero on 100,000 frames, and claims theirs is much better than MuZero given only that many frames. <https://arxiv.org/pdf/2111.00210.pdf> CC: @paulfchristiano. (I would further argue that this case is important because it’s about the central contemporary model for approaching AGI, at least according to Eliezer, rather than any number of random peripheral AI tasks.) [Shah][14:46] I only looked at the front page, so might be misunderstanding, but the front figure says “Our proposed method EfficientZero is 170% and 180% better than the previous SoTA performance in mean and median human normalized score […] on the Atari 100k benchmark”, which does not seem like a huge leap? Oh, I incorrectly thought that was 1.7x and 1.8x, but it is actually 2.7x and 2.8x, which is a bigger deal (though still feels not crazy to me) [Yudkowsky][15:28] the question imo is how many frames the previous SoTA would require to catch up to EfficientZero (I’ve tried emailing an author to ask about this, no response yet) like, perplexity on GPT-3 vs GPT-2 and “losses decreased by blah%” would give you a pretty meaningless concept of how far ahead GPT-3 was from GPT-2, and I think the “2.8x performance” figure in terms of scoring is equally meaningless as a metric of how much EfficientZero improves if any what you want is a notion like “previous SoTA would have required 10x the samples” or “previous SoTA would have required 5x the computation” to achieve that performance level [Shah][15:38] I see. Atari curves are not nearly as nice and stable as GPT curves and often have the problem that they plateau rather than making steady progress with more training time, so that will make these metrics noisier, but it does seem like a reasonable metric to track (Not that I have recommendations about how to track it; I doubt the authors can easily get these metrics) [Christiano][18:01] If you think our views are making such starkly different predictions then I’d be happy to actually state any of them in advance, including e.g. about future ML benchmark results. I don’t think this falsifies my view, and we could continue trying to hash out what my view is but it seems like slow going and I’m inclined to give up. Relevant questions on my view are things like: is MuZero optimized at all for performance in the tiny-sample regime? (I think not, I don’t even think it set SoTA on that task and I haven’t seen any evidence.) What’s the actual rate of improvements since people started studying this benchmark ~2 years ago, and how much work has gone into it? And I totally agree with your comments that “# of frames” is the natural unit for measuring and that would be the starting point for any discussion. [Barnes][18:22] > > In previous MCTS RL algorithms, the environment model is either given or only trained with rewards, values, and policies, which cannot provide sufficient training signals due to their scalar nature. The problem is more severe when the reward is sparse or the bootstrapped value is not accurate. The MCTS policy improvement operator heavily relies on the environment model. Thus, it is vital to have an accurate one. > > > We notice that the output ^st+1 > > > from the dynamic function G should be the same as st+1, i.e. the output of the representation function H with input of the next observation ot+1 (Fig. 2). This can help to supervise the predicted next state ^st+1 using the actual st+1, which is a tensor with at least a few hundred dimensions. This provides ^st+1 with much more training signals than the default scalar reward and value. > > > This seems like a super obvious thing to do and I’m confused why DM didn’t already try this. It was definitely being talked about in ~2018 Will ask a DM friend about it [Yudkowsky][22:45] I… don’t think I want to take all of the blame for misunderstanding Paul’s views; I think I also want to complain at least a little that Paul spends an insufficient quantity of time pointing at extremely concrete specific possibilities, especially real ones, and saying how they do or don’t fit into the scheme. Am I rephrasing correctly that, in this case, if Efficient Zero was actually a huge (3x? 5x? 10x?) jump in RL sample efficiency over previous SOTA, measured in 1 / frames required to train to a performance level, then that means the Paul view doesn’t apply to the present world; but this could be because MuZero wasn’t the real previous SOTA, or maybe because nobody really worked on pushing out this benchmark for 2 years and therefore on the Paul view it’s fine for there to still be huge jumps? In other words, this is something Paul’s worldview has to either defy or excuse, and not just, “well, sure, why wouldn’t it do that, you have misunderstood which kinds of AI-related events Paul is even trying to talk about”? In the case where, “yes it’s a big jump and that shouldn’t happen later, but it could happen now because it turned out nobody worked hard on pushing past MuZero over the last 2 years”, I wish to register that my view permits it to be the case that, when the world begins to end, the frontier that enters into AGI is similarly something that not a lot of people spent a huge effort on since a previous prototype from 2 years earlier. It’s just not very surprising to me if the future looks a lot like the past, or if human civilization neglects to invest a ton of effort in a research frontier. Gwern guesses that getting to EfficientZero’s performance level would require around 4x the samples for MuZero-Reanalyze (the more efficient version of MuZero which replayed past frames), which is also apparently the only version of MuZero the paper’s authors were considering in the first place – without replays, MuZero requires 20 billion frames to achieve its performance, not the figure of 200 million. <https://www.lesswrong.com/posts/jYNT3Qihn2aAYaaPb/efficientzero-human-ale-sample-efficiency-w-muzero-self?commentId=JEHPQa7i8Qjcg7TW6> 17. November 4 conversation --------------------------- ### 17.1. EfficientZero (continued) [Christiano][7:42] I think it’s possible the biggest misunderstanding is that you somehow think of my view as a “scheme” and your view as a normal view where probability distributions over things happen. Concretely, this is a paper that adds a few techniques to improve over MuZero in a domain that (it appears) wasn’t a significant focus of MuZero. I don’t know how much it improves but I can believe gwern’s estimates of 4x. I’d guess MuZero itself is a 2x improvement over the baseline from a year ago, which was maybe a 4x improvement over the algorithm from a year before that. If that’s right, then no it’s not mindblowing on my view to have 4x progress one year, 2x progress the next, and 4x progress the next. If other algorithms were better than MuZero, then the 2019-2020 progress would be >2x and the 2020-2021 progress would be <4x. I think it’s probably >4x sample efficiency though (I don’t totally buy gwern’s estimate there), which makes it at least possibly surprising. But it’s never going to be that surprising. It’s a benchmark that people have been working on for a few years that has been seeing relatively rapid improvement over that whole period. The main innovation is how quickly you can learn to predict future frames of Atari games, which has tiny economic relevance and calling it the most AGI-ish direction seems like it’s a very Eliezer-ish view, this isn’t the kind of domain where I’m either most surprised to see rapid progress at all nor is the kind of thing that seems like a key update re: transformative AI yeah, SoTA in late 2020 was SPR, published by a much smaller academic group: <https://arxiv.org/pdf/2007.05929.pdf> MuZero wasn’t even setting sota on this task at the time it was published my “schemes” are that (i) if a bunch of people are trying on a domain and making steady slow progress, I’m surprised to see giant jumps and I don’t expect most absolute progress to occur in such jumps, (ii) if a domain is worth a lot of $, generally a bunch of people will be trying. Those aren’t claims about what is always true, they are claims about what is typically true and hence what I’m guessing will be true for transformative AI. Maybe you think those things aren’t even good general predictions, and that I don’t have long enough tails in my distributions or whatever. But in that case it seems we can settle it quickly by prediction. I think this result is probably significant (>30% absolute improvement) + faster-than-trend (>50% faster than previous increment) progress relative to prior trend on 8 of the 27 atari games (from table 1, treating SimPL->{max of MuZero, SPR}->EfficientZero as 3 equally spaced datapoints): Asterix, Breakout, almost ChopperCMD, almost CrazyClimber, Gopher, Kung Fu Master, Pong, QBert, SeaQuest. My guess is that they thought a lot about a few of those games in particular because they are very influential on the mean/median. Note that this paper is a giant grab bag and that simply stapling together the prior methods would have already been a significant improvement over prior SoTA. (ETA: I don’t think saying “its only 8 of 27 games” is an update against it being big progress or anything. I do think saying “stapling together 2 previous methods without any complementarity at all would already have significantly beaten SoTA” is fairly good evidence that it’s not a hard-to-beat SoTA.) and even fewer people working on the ultra-low-sample extremely-low-dimensional DM control environments (this is the subset of problems where the state space is 4 dimensions, people are just not trying to publish great results on cartpole), so I think the most surprising contribution is the atari stuff OK, I now also understand what the result is I think? I think the quick summary is: the prior SoTA is SPR, which learns to predict the domain and then does Q-learning. MuZero instead learns to predict the domain and does MCTS, but it predicts the domain in a slightly less sophisticated way than SPR (basically just predicts rewards, whereas SPR predicts all of the agent’s latent state in order to get more signal from each frame). If you combine MCTS with more sophisticated prediction, you do better. I think if you told me that DeepMind put in significant effort in 2020 (say, at least as much post-MuZero effort as the new paper?) trying to get great sample efficiency on the easy-exploration atari games, and failed to make significant progress, then I’m surprised. I don’t think that would “falsify” my view, but it would be an update against? Like maybe if DM put in that much effort I’d maybe have given only a 10-20% probability to a new project of similar size putting in that much effort making big progress, and even conditioned on big progress this is still >>median (ETA: and if DeepMind put in much more effort I’d be more surprised than 10-20% by big progress from the new project) Without DM putting in much effort, it’s significantly less surprising and I’ll instead be comparing to the other academic efforts. But it’s just not surprising that you can beat them if you are willing to put in the effort to reimplement MCTS and they aren’t, and that’s a step that is straightforwardly going to improve performance. (not sure if that’s the situation) And then to see how significant updates against are, you have to actually contrast them with all the updates in the other direction where people don’t crush previous benchmark results and instead just make modest progress I would guess that if you had talked to an academic about this question (what happens if you combine SPR+MCTS) they would have predicted significant wins in sample efficiency (at the expense of compute efficiency) and cited the difficulty of implementing MuZero compared to any of the academic results. That’s another way I could be somewhat surprised (or if there were academics with MuZero-quality MCTS implementations working on this problem, and they somehow didn’t set SoTA, then I’m even more surprised). But I’m not sure if you’ll trust any of those judgments in hindsight. Repeating the main point: I don’t really think a 4x jump over 1 year is something I have to “defy or excuse”, it’s something that I think becomes more or less likely depending on facts about the world, like (i) how fast was previous progress, (ii) how many people were working on previous projects and how targeted were they at this metric, (iii) how many people are working in this project and how targeted was it at this metric it becomes continuously less likely as those parameters move in the obvious directions it never becomes 0 probability, and you just can’t win that much by citing isolated events that I’d give say a 10% probability to, unless you actually say something about how you are giving >10% probabilities to those events without losing a bunch of probability mass on what I see as the 90% of boring stuff \| \| \| --- \| \| [Ngo: 👍] \| and then separately I have a view about lots of people working on important problems, which doesn’t say anything about this case (I actually don’t think this event is as low as 10%, though it depends on what background facts about the project you are conditioning on—obviously I gave <<10% probability to someone publishing this particular result, but something like “what fraction of progress in this field would come down to jumps like this” or whatever is probably >10% until you tell me that DeepMind actually cared enough to have already tried) [Ngo][8:48] I expect Eliezer to say something like: DeepMind believes that both improving RL sample efficiency, and benchmarking progress on games like Atari, are important parts of the path towards AGI. So insofar as your model predicts that smooth progress will be caused by people working directly towards AGI, DeepMind not putting effort into this is a hit to that model. Thoughts? [Christiano][9:06] I don’t think that learning these Atari games in 2 hours is a very interesting benchmark even for deep RL sample efficiency, and it’s totally unrelated to the way in which humans learn such games quickly. It seems ~~pretty likely~~ totally plausible (50%?) to me that DeepMind feels the same way, and then the question is about other random considerations like how they are making some PR calculation. [Ngo][9:18] If Atari is not a very interesting benchmark, then why did DeepMind put a bunch of effort into making Agent57 and applying MuZero to Atari? Also, most of the effort they’ve spent on games in general has been on methods very unlike the way humans learn those games, so that doesn’t seem like a likely reason for them to overlook these methods for increasing sample efficiency. [Shah][9:32] > > It seems pretty likely totally plausible (50%?) to me that DeepMind feels the same way, and then the question is about other random considerations like how they are making some PR calculation. > > > Not sure of the exact claim, but DeepMind is big enough and diverse enough that I’m pretty confident at least some people working on relevant problems don’t feel the same way > > […] This seems like a super obvious thing to do and I’m confused why DM didn’t already try this. It was definitely being talked about in ~2018 > > > Speculating without my DM hat on: maybe it kills performance in board games, and they want one algorithm for all settings? [Christiano][10:29] Atari games in the tiny sample regime are a different beast there are just a lot of problems you can state about Atari some of which are more or less interesting (e.g. jointly learning to play 57 Atari games is a more interesting problem than learning how to play one of them absurdly quickly, and there are like 10 other problems about Atari that are more interesting than this one) That said, Agent57 also doesn’t seem interesting except that it’s an old task people kind of care about. I don’t know about the take within DeepMind but outside I don’t think anyone would care about it other than historical significance of the benchmark / obviously-not-cherrypickedness of the problem. I’m sure that some people at DeepMind care about getting the super low sample complexity regime. I don’t think that really tells you how large the DeepMind effort is compared to some random academics who care about it. \| \| \| --- \| \| [Shah: 👍] \| I think the argument for working on deep RL is fine and can be based on an analogy with humans while you aren’t good at the task. Then once you are aiming for crazy superhuman performance on Atari games you naturally start asking “what are we doing here and why are we still working on atari games?” \| \| \| --- \| \| [Ngo: 👍] \| and correspondingly they are a smaller and smaller slice of DeepMind’s work over time \| \| \| --- \| \| [Ngo: 👍] \| (e.g. Agent57 and MuZero are the only DeepMind blog posts about Atari in the last 4 years, it’s not the main focus of MuZero and I don’t think Agent57 is a very big DM project) Reaching this level of performance in Atari games is largely about learning perception, and doing that from 100k frames of an Atari game just doesn’t seem very analogous to anything humans do or that is economically relevant from any perspective. I totally agree some people are into it, but I’m totally not surprised if it’s not going to be a big DeepMind project. [Yudkowsky][10:51] would you agree it’s a load-bearing assumption of your worldview – where I also freely admit to having a worldview/scheme, this is not meant to be a prejudicial term at all – that the line of research which leads into world-shaking AGI must be in the mainstream and not in a weird corner where a few months earlier there were more profitable other ways of doing all the things that weird corner did? eg, the tech line leading into world-shaking AGI must be at the profitable forefront of non-world-shaking tasks. as otherwise, afaict, your worldview permits that if counterfactually we were in the Paul-forbidden case where the immediate precursor to AGI was something like EfficientZero (whose motivation had been beating an old SOTA metric rather than, say, market-beating self-driving cars), there might be huge capability leaps there just as EfficientZero represents a large leap, because there wouldn’t have been tons of investment in that line. [Christiano][10:54] Something like that is definitely a load-bearing assumption Like there’s a spectrum with e.g. EfficientZero –> 2016 language modeling –> 2014 computer vision –> 2021 language modeling –> 2021 computer vision, and I think everything anywhere close to transformative AI will be way way off the right end of that spectrum But I think quantitatively the things you are saying don’t seem quite right to me. Suppose that MuZero wasn’t the best way to do anything economically relevant, but it was within a factor of 4 on sample efficiency for doing tasks that people care about. That’s already going to be enough to make tons of people extremely excited. So yes, I’m saying that anything leading to transformative AI is “in the mainstream” in the sense that it has more work on it than 2021 language models. But not necessarily that it’s the most profitable way to do anything that people care about. Different methods scale in different ways, and something can burst onto the scene in a dramatic way, but I strongly expect speculative investment driven by that possibility to already be way (way) more than 2021 language models. And I don’t expect gigantic surprises. And I’m willing to bet that e.g. EfficientZero isn’t a big surprise for researchers who are paying attention to the area (in addition to being 3+ orders of magnitude more neglected than anything close to transformative AI) 2021 language modeling isn’t even very competitive, it’s still like 3-4 orders of magnitude smaller than semiconductors. But I’m giving it as a reference point since it’s obviously much, much more competitive than sample-efficient atari. This is a place where I’m making much more confident predictions, this is “falsify paul’s worldview” territory once you get to quantitative claims anywhere close to TAI and “even a single example seriously challenges paul’s worldview” a few orders of magnitude short of that [Yudkowsky][11:04] can you say more about what falsifies your worldview previous to TAI being super-obviously-to-all-EAs imminent? or rather, “seriously challenges”, sorry [Christiano][11:05][11:08] big AI applications achieved by clever insights in domains that aren’t crowded, we should be quantitative about how crowded and how big if we want to get into “seriously challenges” like e.g. if this paper on atari was actually a crucial ingredient for making deep RL for robotics work, I’d be actually for real surprised rather than 10% surprised but it’s not going to be, those results are being worked on by much larger teams of more competent researchers at labs with $100M+ funding it’s definitely possible for them to get crushed by something out of left field but I’m betting against every time or like, the set of things people would describe as “out of left field,” and the quantitative degree of neglectedness, becomes more and more mild as the stakes go up [Yudkowsky][11:08] how surprised are you if in 2022 one company comes out with really good ML translation, and they manage to sell a bunch of it temporarily until others steal their ideas or Google acquires them? my model of Paul is unclear on whether this constitutes “many people are already working on language models including ML translation” versus “this field is not profitable enough right this minute for things to be efficient there, and it’s allowed to be nonobvious in worlds where it’s about to become profitable”. [Christiano][11:08] if I wanted to make a prediction about that I’d learn a bunch about how much google works on translation and how much $ they make I just don’t know the economics and it depends on the kind of translation that they are good at and the economics (e.g. google mostly does extremely high-volume very cheap translation) but I think there are lots of things like that / facts I could learn about Google such that I’d be surprised in that situation independent of the economics, I do think a fair number of people are working on adjacent stuff, and I don’t expect someone to come out of left field for google-translate-cost translation between high-resource languages but it seems quite plausible that a team of 10 competent people could significantly outperform google translate, and I’d need to learn about the economics to know how surprised I am by 10 people or 100 people or what I think it’s allowed to be non-obvious whether a domain is about to be really profitable but it’s not that easy, and the higher the stakes the more speculative investment it will drive, etc. [Yudkowsky][11:14] if you don’t update much off EfficientZero, then people also shouldn’t be updating much off of most of the graph I posted earlier as possible Paul-favoring evidence, because most of those SOTAs weren’t highly profitable so your worldview didn’t have much to say about them. ? [Christiano][11:15] Most things people work a lot on improve gradually. EfficientZero is also quite gradual compared to the crazy TAI stories you tell. I don’t really know what to say about this game other than I would prefer make predictions in advance and I’m happy to either propose questions/domains or make predictions in whatever space you feel more comfortable with. [Yudkowsky][11:16] I don’t know how to point at a future event that you’d have strong opinions about. it feels like, whenever I try, I get told that the current world is too unlike the future conditions you expect. [Christiano][11:16] Like, whether or not EfficientZero is evidence for your view depends on exactly how “who knows what will happen” you are. if you are just a bit more spread out than I am, then it’s definitely evidence for your view. I’m saying that I’m willing to bet about any event you want to name, I just think my model of how things work is more accurate. I’d prefer it be related to ML or AI. [Yudkowsky][11:17] to be clear, I appreciate that it’s similarly hard to point at an event like that for myself, because my own worldview says “well mostly the future is not all that predictable with a few rare exceptions” [Christiano][11:17] But I feel like the situation is not at all symmetrical, I expect to outperform you on practically any category of predictions we can specify. so like I’m happy to bet about benchmark progress in LMs, or about whether DM or OpenAI or Google or Microsoft will be the first to achieve something, or about progress in computer vision, or about progress in industrial robotics, or about translations whatever ### 17.2. Near-term AI predictions [Yudkowsky][11:18] that sounds like you ought to have, like, a full-blown storyline about the future? [Christiano][11:18] what is a full-blown storyline? I have a bunch of ways that I think about the world and make predictions about what is likely and yes, I can use those ways of thinking to make predictions about whatever and I will very often lose to a domain expert who has better and more informed ways of making predictions [Yudkowsky][11:19] what happens if 2022 through 2024 looks literally exactly like Paul’s modal or median predictions on things? [Christiano][11:19] but I think in ML I will generally beat e.g. a superforecaster who doesn’t have a lot of experience in the area give me a question about 2024 and I’ll give you a median? I don’t know what “what happens” means storylines do not seem like good ways of making predictions \| \| \| --- \| \| [Shah: 👍] \| [Yudkowsky][11:20] I mean, this isn’t a crux for anything, but it seems like you’re asking me to give up on that and just ask for predictions? so in 2024 can I hire an artist who doesn’t speak English and converse with them almost seamlessly through a machine translator? [Christiano][11:22] median outcome (all of these are going to be somewhat easy-to-beat predictions because I’m not thinking): you can get good real-time translations, they are about as good as a +1 stdev bilingual speaker who listens to what you said and then writes it out in the other language as fast as they can type Probably also for voice -> text or voice -> voice, though higher latencies and costs. Not integrated into standard video chatting experience because the UX is too much of a pain and the world sucks. That’s a median on “how cool/useful is translation” [Yudkowsky][11:23] I would unfortunately also predict that in this case, this will be a highly competitive market and hence not a very profitable one, which I predict to match your prediction, but I ask about the economics here just in case. [Christiano][11:24] Kind of typical sample: I’d guess that Google has a reasonably large lead, most translation still provided as a free value-added, cost per translation at that level of quality is like $0.01/word, total revenue in the area is like $10Ms / year? [Yudkowsky][11:24] well, my model also permits that Google does it for free and so it’s an uncompetitive market but not a profitable one… ninjaed. [Christiano][11:25] first order of improving would be sanity-checking economics and thinking about #s, second would be learning things like “how many people actually work on translation and what is the state of the field?” [Yudkowsky][11:26] did Tesla crack self-driving cars and become a $3T company instead of a $1T company? do you own Tesla options? did Waymo beat Tesla and cause Tesla stock to crater, same question? [Christiano][11:27] 1/3 chance tesla has FSD in 2024 conditioned on that, yeah probably market cap is >$3T? conditioned on Tesla having FSD, 2/3 chance Waymo has also at least rolled out to a lot of cities conditioned on no tesla FSD, 10% chance Waymo has rolled out to like half of big US cities? dunno if numbers make sense [Yudkowsky][11:28] that’s okay, I dunno if my questions make sense [Christiano][11:29] (5% NW in tesla, 90% NW in AI bets, 100% NW in more normal investments; no tesla options that sounds like a scary place with lottery ticket biases and the crazy tesla investors) [Yudkowsky][11:30] (am I correctly understanding you’re 2x levered?) [Christiano][11:30][11:31] yeah it feels like you’ve got to have weird views on trajectory of value-added from AI over the coming years on how much of the $ comes from domains that are currently exciting to people (e.g. that Google already works on, self-driving, industrial robotics) vs stuff out of left field on what kind of algorithms deliver $ in those domains (e.g. are logistics robots trained using the same techniques tons of people are currently pushing on) on my picture you shouldn’t be getting big losses on any of those just losing like 10-20% each time [Yudkowsky][11:31][11:32] my uncorrected inside view says that machine translation should be in reach and generate huge amounts of economic value even if it ends up an unprofitable competitive or Google-freebie field and also that not many people are working on basic research in machine translation or see it as a “currently exciting” domain [Christiano][11:32] how many FTE is “not that many” people? also are you expecting improvement in the google translate style product, or in lower-latencies for something closer to normal human translator prices, or something else? [Yudkowsky][11:33] my worldview says more like… sure, maybe there’s 300 programmers working on it worldwide, but most of them aren’t aggressively pursuing new ideas and trying to explore the space, they’re just applying existing techniques to a new language or trying to throw on some tiny mod that lets them beat SOTA by 1.2% for a publication because it’s not an exciting field “What if you could rip down the language barriers” is an economist’s dream, or a humanist’s dream, and Silicon Valley is neither and looking at GPT-3 and saying, “God damn it, this really seems like it must on some level understand what it’s reading well enough that the same learned knowledge would suffice to do really good machine translation, this must be within reach for gradient descent technology we just don’t know how to reach it” is Yudkowskian thinking; your AI system has internal parts like “how much it understands language” and there’s thoughts about what those parts ought to be able to do if you could get them into a new system with some other parts [Christiano][11:36] my guess is we’d have some disagreements here but to be clear, you are talking about text-to-text at like $0.01/word price point? [Yudkowsky][11:38] I mean, do we? Unfortunately another Yudkowskian worldview says “and people can go on failing to notice this for arbitrarily long amounts of time”. if that’s around GPT-3’s price point then yeah [Christiano][11:38] gpt-3 is a lot cheaper, happy to say gpt-3 like price point [Yudkowsky][11:39] (thinking about whether $0.01/word is meaningfully different from $0.001/word and concluding that it is) [Christiano][11:39] (api is like 10,000 words / $) I expect you to have a broader distribution over who makes a great product in this space, how great it ends up being etc., whereas I’m going to have somewhat higher probabilities on it being google research and it’s going to look boring [Yudkowsky][11:40] what is boring? boring predictions are often good predictions on my own worldview too lots of my gloom is about things that are boringly bad and awful (and which add up to instant death at a later point) but, I mean, what does boring machine translation look like? [Christiano][11:42] Train big language model. Have lots of auxiliary tasks especially involving reading in source language and generation in target language. Have pre-training on aligned sentences and perhaps using all the unsupervised translation we have depending on how high-resource language is. Fine-tune with smaller amount of higher quality supervision. Some of the steps likely don’t add much value and skip them. Fair amount of non-ML infrastructure. For some languages/domains/etc. dedicated models, over time increasingly just have a giant model with learned dispatch as in mixture of experts. [Yudkowsky][11:44] but your worldview is also totally ok with there being a Clever Trick added to that which produces a 2x reduction in training time. or with there being a new innovation like transformers, which was developed a year earlier and which everybody now uses, without which the translator wouldn’t work at all. ? [Christiano][11:44] Just for reference, I think transformers aren’t that visible on a (translation quality) vs (time) graph? But yes, I’m totally fine with continuing architectural improvements, and 2x reduction in training time is currently par for the course for “some people at google thought about architectures for a while” and I expect that to not get that much tighter over the next few years. [Yudkowsky][11:45] unrolling Restricted Boltzmann Machines to produce deeper trainable networks probably wasn’t much visible on a graph either, but good luck duplicating modern results using only lower portions of the tech tree. (I don’t think we disagree about this.) [Christiano][11:45] I do expect it to eventually get tighter, but not by 2024. I don’t think unrolling restricted boltzmann machines is that important [Yudkowsky][11:46] like, historically, or as a modern technology? [Christiano][11:46] historically [Yudkowsky][11:46] interesting my model is that it got people thinking about “what makes things trainable” and led into ReLUs and inits but I am going more off having watched from the periphery as it happened, than having read a detailed history of that like, people asking, “ah, but what if we had a deeper network and the gradients didn’t explode or die out?” and doing that en masse in a productive way rather than individuals being wistful for 30 seconds [Christiano][11:48] well, not sure if this will introduce differences in predictions I don’t feel like it should really matter for our bottom line predictions whether we classify google’s random architectural change as something fundamentally new (which happens to just have a modest effect at the time that it’s built) or as something boring I’m going to guess how well things will work by looking at how well things work right now and seeing how fast it’s getting better and that’s also what I’m going to do for applications of AI with transformative impacts and I actually believe you will do something today that’s analogous to what you would do in the future, and in fact will make somewhat different predictions than what I would do and then some of the action will be in new things that people haven’t been trying to do in the past, and I’m predicting that new things will be “small” whereas you have a broader distribution, and there’s currently some not-communicated judgment call in “small” if you think that TAI will be like translation, where google publishes tons of papers, but that they will just get totally destroyed by some new idea, then it seems like that should correspond to a difference in P(google translation gets totally destroyed by something out-of-left-field) and if you think that TAI won’t be like translation, then I’m interested in examples more like TAI I don’t really understand the take “and people can go on failing to notice this for arbitrarily long amounts of time,” why doesn’t that also happen for TAI and therefore cause it to be the boring slow progress by google? Why would this be like a 50% probability for TAI but <10% for translation? perhaps there is a disagreement about how good the boring progress will be by 2024? looks to me like it will be very good [Yudkowsky][11:57] I am not sure that is where the disagreement lies ### 17.3. The evolution of human intelligence [Yudkowsky][11:57] I am considering advocating that we should have more disagreements about the past, which has the advantage of being very concrete, and being often checkable in further detail than either of us already know [Christiano][11:58] I’m fine with disagreements about the past; I’m more scared of letting you pick arbitrary things to “predict” since there is much more impact from differences in domain knowledge (also not quite sure why it’s more concrete, I guess because we can talk about what led to particular events? mostly it just seems faster) also as far as I can tell our main differences are about whether people will ~~spend a lot of money~~ work effectively on things that would make a lot of money, which means if we look to the past we will have to move away from ML/AI [Yudkowsky][12:00] so my understanding of how Paul writes off the example of human intelligence, is that you are like, “evolution is much stupider than a human investor; if there’d been humans running the genomes, people would be copying all the successful things, and hominid brains would be developing in this ecology of competitors instead of being a lone artifact”. ? [Christiano][12:00] I don’t understand why I have to write off the example of human intelligence [Yudkowsky][12:00] because it looks nothing like your account of how TAI develops [Christiano][12:00] it also looks nothing like your account, I understand that you have some analogy that makes sense to you [Yudkowsky][12:01] I mean, to be clear, I also write off the example of humans developing morality and have to explain to people at length why humans being as nice as they are, doesn’t imply that paperclip maximizers will be anywhere near that nice, nor that AIs will be other than paperclip maximizers. [Christiano][12:01][12:02] you could state some property of how human intelligence developed, that is in common with your model for TAI and not mine, and then we could discuss that if you say something like: “chimps are not very good at doing science, but humans are” then yes my answer will be that it’s because evolution was not selecting us to be good at science and indeed AI systems will be good at science using much less resources than humans or chimps [Yudkowsky][12:02][12:02] would you disagree that humans developing intelligence, on the sheer surfaces of things, looks much more Yudkowskian than Paulian? like, not in terms of compatibility with underlying model just that there’s this one corporation that came out and massively won the entire AGI race with zero competitors [Christiano][12:03] I agree that “how much did the winner take all” is more like your model of TAI than mine I don’t think zero competitors is reasonable, I would say “competitors who were tens of millions of years behind” [Yudkowsky][12:03] sure and your account of this is that natural selection is nothing like human corporate managers copying each other [Christiano][12:03] which was a reasonable timescale for the old game, but a long timescale for the new game [Yudkowsky][12:03] yup [Christiano][12:04] that’s not my only account it’s also that for human corporations you can form large coalitions, i.e. raise huge amounts of $ and hire huge numbers of people working on similar projects (whether or not vertically integrated), and those large coalitions will systematically beat small coalitions and that’s basically the key dynamic in this situation, and isn’t even trying to have any analog in the historical situation (the key dynamic w.r.t. concentration of power, not necessarily the main thing overall) [Yudkowsky][12:07] the modern degree of concentration of power seems relatively recent and to have tons and tons to do with the regulatory environment rather than underlying properties of the innovation landscape back in the old days, small startups would be better than Microsoft at things, and Microsoft would try to crush them using other forces than superior technology, not always successfully or such was the common wisdom of USENET [Christiano][12:08] my point is that the evolution analogy is extremely unpersuasive w.r.t. concentration of power I think that AI software capturing the amount of power you imagine is also kind of implausible because we know something about how hardware trades off against software progress (maybe like 1 year of progress = 2x hardware) and so even if you can’t form coalitions on innovation at all you are still going to be using tons of hardware if you want to be in the running though if you can’t parallelize innovation at all and there is enough dispersion in software progress then the people making the software could take a lot of the $ / influence from the partnership anyway, I agree that this is a way in which evolution is more like your world than mine but think on this point the analogy is pretty unpersuasive because it fails to engage with any of the a priori reasons you wouldn’t expect concentration of power [Yudkowsky][12:11] I’m not sure this is the correct point on which to engage, but I feel like I should say out loud that I am unable to operate my model of your model in such fashion that it is not falsified by how the software industry behaved between 1980 and 2000. there should’ve been no small teams that beat big corporations today those are much rarer, but on my model, that’s because of regulatory changes (and possibly metabolic damage from something in the drinking water) [Christiano][12:12] I understand that you can’t operate my model, and I’ve mostly given up, and on this point I would prefer to just make predictions or maybe retrodictions [Yudkowsky][12:13] well, anyways, my model of how human intelligence happened looks like this: there is a mysterious kind of product which we can call G, and which brains can operate as factories to produce G in turn can produce other stuff, but you need quite a lot of it piled up to produce better stuff than your competitors as late as 1000 years ago, the fastest creatures on Earth are not humans, because you need even more G than that to go faster than cheetahs (or peregrine falcons) the natural selections of various species were fundamentally stupid and blind, incapable of foresight and incapable of copying the successes of other natural selections; but even if they had been as foresightful as a modern manager or investor, they might have made just the same mistake before 10,000 years they would be like, “what’s so exciting about these things? they’re not the fastest runners.” if there’d been an economy centered around running, you wouldn’t invest in deploying a human (well, unless you needed a stamina runner, but that’s something of a separate issue, let’s consider just running races) you would invest on improving cheetahs because the pile of human G isn’t large enough that their G beats a specialized naturally selected cheetah [Christiano][12:17] how are you improving cheetahs in the analogy? you are trying random variants to see what works? [Yudkowsky][12:18] using conventional, well-tested technology like MUSCLES and TENDONS trying variants on those [Christiano][12:18] ok and you think that G doesn’t help you improve on muscles and tendons? until you have a big pile of it? [Yudkowsky][12:18] not as a metaphor but as simple historical fact, that’s how it played out it takes a whole big pile of G to go faster than a cheetah [Christiano][12:19] as a matter of fact there is no one investing in making better cheetahs so it seems like we’re already playing analogy-game [Yudkowsky][12:19] the natural selection of cheetahs is investing in it it’s not doing so by copying humans because of fundamental limitations however if we replace it with an average human investor, it still doesn’t copy humans, why would it [Christiano][12:19] that’s the part that is silly or like, it needs more analogy [Yudkowsky][12:19] how so? humans aren’t the fastest. [Christiano][12:19] humans are great at breeding animals so if I’m natural selection personified, the thing to explain is why I’m not using some of that G to improve on my selection not why I’m not using G to build a car [Yudkowsky][12:20] I’m… confused is this implying that a key aspect of your model is that people are using AI to decide which AI tech to invest in? [Christiano][12:20] no I think I just don’t understand your analogy here in the actual world, some people are trying to make faster robots by tinkering with robot designs and then someone somewhere is training their AGI [Yudkowsky][12:21] what I’m saying is that you can imagine a little cheetah investor going, “I’d like to copy and imitate some other species’s tricks to make my cheetahs faster” and they’re looking enviously at falcons, not at humans not until very late in the game [Christiano][12:21] and the relevant question is whether the pre-AGI thing is helpful for automating the work that humans are doing while they tinker with robot designs that seems like the actual world and the interesting claim is you saying “nope, not very” [Yudkowsky][12:22] I am again confused. Does it matter to your model whether the pre-AGI thing is helpful for automating “tinkering with robot designs” or just profitable machine translation? Either seems like it induces equivalent amounts of investment. If anything the latter induces much more investment. [Christiano][12:23] sure, I’m fine using “tinkering with robot designs” as a lower bound both are fine the point is I have no idea what you are talking about in the analogy what is analogous to what? I thought cheetahs were analogous to faster robots [Yudkowsky][12:23] faster cheetahs are analogous to more profitable robots [Christiano][12:23] sure so you have some humans working on making more profitable robots, right? who are tinkering with the robots, in a way analogous to natural selection tinkering with cheetahs? [Yudkowsky][12:24] I’m suggesting replacing the Natural Selection of Cheetahs with a new optimizer that has the Copy Competitor and Invest In Easily-Predictable Returns feature [Christiano][12:24] OK, then I don’t understand what those are analogous to like, what is analogous to the humans who are tinkering with robots, and what is analogous to the humans working on AGI? [Yudkowsky][12:24] and observing that, even this case, the owner of Cheetahs Inc. would not try to copy Humans Inc. [Christiano][12:25] here’s the analogy that makes sense to me natural selection is working on making faster cheetahs = some humans tinkering away to make more profitable robots natural selection is working on making smarter humans = some humans who are tinkering away to make more powerful AGI natural selection doesn’t try to copy humans because they suck at being fast = robot-makers don’t try to copy AGI-makers because the AGIs aren’t very profitable robots [Yudkowsky][12:26] with you so far [Christiano][12:26] eventually humans build cars once they get smart enough = eventually AGI makes more profitable robots once it gets smart enough [Yudkowsky][12:26] yup [Christiano][12:26] great, seems like we’re on the same page then [Yudkowsky][12:26] and by this point it is LATE in the game [Christiano][12:27] great, with you still [Yudkowsky][12:27] because the smaller piles of G did not produce profitable robots [Christiano][12:27] but there’s a step here where you appear to go totally off the rails [Yudkowsky][12:27] or operate profitable robots say on [Christiano][12:27] can we just write out the sequence of AGIs, AGI(1), AGI(2), AGI(3)… in analogy with the sequence of human ancestors H(1), H(2), H(3)…? [Yudkowsky][12:28] Is the last member of the sequence H(n) the one that builds cars and then immediately destroys the world before anything that operates on Cheetah Inc’s Owner’s scale can react? [Christiano][12:28] sure I don’t think of it as the last but it’s the last one that actually arises? maybe let’s call it the last, H(n) great and now it seems like you are imagining an analogous story, where AGI(n) takes over the world and maybe incidentally builds some more profitable robots along the way (building more profitable robots being easier than taking over the world, but not so much easier that AGI(n-1) could have done it unless we make our version numbers really close together, close enough that deploying AGI(n-1) is stupid) [Yudkowsky][12:31] if this plays out in the analogous way to human intelligence, AGI(n) becomes able to build more profitable robots 1 hour before it becomes able to take over the world; my worldview does not put that as the median estimate, but I do want to observe that this is what happened historically [Christiano][12:31] sure [Yudkowsky][12:32] ok, then I think we’re still on the same page as written so far [Christiano][12:32] so the question that’s interesting in the real world is which AGI is useful for replacing humans in the design-better-robots task; is it 1 hour before the AGI that takes over the world, or 2 years, or what? [Yudkowsky][12:33] my worldview tends to make a big ol’ distinction between “replace humans in the design-better-robots task” and “run as a better robot”, if they’re not importantly distinct from your standpoint can we talk about the latter? [Christiano][12:33] they seem importantly distinct totally different even so I think we’re still on the same page [Yudkowsky][12:34] ok then, “replacing humans at designing better robots” sure as heck sounds to Eliezer like the world is about to end or has already ended [Christiano][12:34] my whole point is that in the evolutionary analogy we are talking about “run as a better robot” rather than “replace humans in the design-better-robots-task” and indeed there is no analog to “replace humans in the design-better-robots-task” which is where all of the action and disagreement is [Yudkowsky][12:35][12:36] well, yes, I was exactly trying to talk about when humans start running as better cheetahs and how that point is still very late in the game not as late as when humans take over the job of making the thing that makes better cheetahs, aka humans start trying to make AGI, which is basically the fingersnap end of the world from the perspective of Cheetahs Inc. [Christiano][12:36] OK, but I don’t care when humans are better cheetahs—in the real world, when AGIs are better robots. In the real world I care about when AGIs start replacing humans in the design-better-robots-task. I’m game to use evolution as an analogy to help answer that question (where I do agree that it’s informative), but want to be clear what’s actually at issue. [Yudkowsky][12:37] so, the thing I was trying to work up to, is that my model permits the world to end in a way where AGI doesn’t get tons of investment because it has an insufficiently huge pile of G that it could run as a better robot. people are instead investing in the equivalents of cheetahs. I don’t understand why your model doesn’t care when humans are better cheetahs. AGIs running as more profitable robots is what induces the huge investments in AGI that your model requires to produce very close competition. ? [Christiano][12:38] it’s a sufficient condition, but it’s not the most robust one at all like, I happen to think that in the real world AIs actually are going to be incredibly profitable robots, and that’s part of my boring view about what AGI looks like But the thing that’s more robust is that the sub-taking-over-world AI is already really important, and receiving huge amounts of investment, as something that automates the R&D process. And it seems like the best guess given what we know now is that this process starts years before the singularity. From my perspective that’s where most of the action is. And your views on that question seem related to your views on how e.g. AGI is a fundamentally different ballgame from making better robots (whereas I think the boring view is that they are closely related), but that’s more like an upstream question about what you think AGI will look like, most relevant because I think it’s going to lead you to make bad short-term predictions about what kinds of technologies will achieve what kinds of goals. [Yudkowsky][12:41] but not all AIs are the same branch of the technology tree. factory robotics are already really important and they are “AI” but, on my model, they’re currently on the cheetah branch rather than the hominid branch of the tech tree; investments into better factory robotics are not directly investments into improving MuZero, though they may buy chips that MuZero also buys. [Christiano][12:42] Yeah, I think you have a mistaken view of AI progress. But I still disagree with your bottom line even if I adopt (this part of) your view of AI progress. Namely, I think that the AGI line is mediocre before it is great, and the mediocre version is spectacularly valuable for accelerating R&D (mostly AGI R&D). The way I end up sympathizing with your view is if I adopt both this view about the tech tree, + another equally-silly-seeming view about how close the AGI line is to fooming (or how inefficient the area will remain as we get close to fooming) ### 17.4. Human generality and body manipulation [Yudkowsky][12:43] so metaphorically, you require that humans be doing Great at Various Things and being Super Profitable way before they develop agriculture; the rise of human intelligence cannot be a case in point of your model because the humans were too uncompetitive at most animal activities for unrealistically long (edit: compared to the AI case) [Christiano][12:44] I don’t understand Human brains are really great at basically everything as far as I can tell? like it’s not like other animals are better at manipulating their bodies we crush them [Yudkowsky][12:44] if we’ve got weapons, yes [Christiano][12:44] human bodies are also pretty great, but they are not the greatest on every dimension [Yudkowsky][12:44] wrestling a chimpanzee without weapons is famously ill-advised [Christiano][12:44] no, I mean everywhere chimpanzees are practically the same as humans in the animal kingdom they have almost as excellent a brain [Yudkowsky][12:45] as is attacking an elephant with your bare hands [Christiano][12:45] that’s not because of elephant brains [Yudkowsky][12:45] well, yes, exactly you need a big pile of G before it’s profitable so big the game is practically over by then [Christiano][12:45] this seems so confused but that’s exciting I guess like, I’m saying that the brains to automate R&D are similar to the brains to be a good factory robot analogously, I think the brains that humans use to do R&D are similar to the brains we use to manipulate our body absurdly well I do not think that our brains make us fast they help a tiny bit but not much I do not think the physical actuators of the industrial robots will be that similar to the actuators of the robots that do R&D the claim is that the problem of building the brain is pretty similar just as the problem of building a brain that can do science is pretty similar to the problem of building a brain that can operate a body really well (and indeed I’m claiming that human bodies kick ass relative to other animal bodies—there may be particular tasks other animal brains are pre-built to be great at, but (i) humans would be great at those too if we were under mild evolutionary pressure with our otherwise excellent brains, (ii) there are lots of more general tests of how good you are at operating a body and we will crush it at those tests) (and that’s not something I know much about, so I could update as I learned more about how actually we just aren’t that good at motor control or motion planning) [Yudkowsky][12:49] so on your model, we can introduce humans to a continent, forbid them any tool use, and they’ll still wipe out all the large animals? [Christiano][12:49] (but damn we seem good to me) I don’t understand why that would even plausibly follow [Yudkowsky][12:49] because brains are profitable early, even if they can’t build weapons? [Christiano][12:49] I’m saying that if you put our brains in a big animal body we would wipe out the big animals yes, I think brains are great [Yudkowsky][12:50] because we’d still have our late-game pile of G and we would build weapons [Christiano][12:50] no, I think a human in a big animal body, with brain adapted to operate that body instead of our own, would beat a big animal straightforwardly without using tools [Yudkowsky][12:51] this is a strange viewpoint and I do wonder whether it is a crux of your view [Christiano][12:51] this feels to me like it’s more on the “eliezer vs paul disagreement about the nature of AI” rather than “eliezer vs paul on civilizational inadequacy and continuity”, but enough changes on “nature of AI” would switch my view on the other question [Yudkowsky][12:51] like, ceteris paribus maybe a human in an elephant’s body beats an elephant after a burn-in practice period? because we’d have a strict intelligence advantage? [Christiano][12:52] practice may or may not be enough but if you port over the excellent human brain to the elephant body, then run evolution for a brief burn-in period to get all the kinks sorted out? elephants are pretty close to humans so it’s less brutal than for some other animals (and also are elephants the best example w.r.t. the possibility of direct conflict?) but I totally expect us to win [Yudkowsky][12:53] I unfortunately need to go do other things in advance of an upcoming call, but I feel like disagreeing about the past is proving noticeably more interesting, confusing, and perhaps productive, than disagreeing about the future [Christiano][12:53] actually probably I just think practice is enough I think humans have way more dexterity, better locomotion, better navigation, better motion planning… some of that is having bodies optimized for those things (esp. dexterity), but I also think most animals just don’t have the brains for it, with elephants being one of the closest calls I’m a little bit scared of talking to zoologists or whoever the relevant experts are on this question, because I’ve talked to bird people a little bit and they often have very strong “humans aren’t special, animals are super cool” instincts even in cases where that take is totally and obviously insane. But if we found someone reasonable in that area I’d be interested to get their take on this. I think this is pretty important for the particular claim “Is AGI like other kinds of ML?”; that definitely doesn’t persuade me to be into fast takeoff on its own though it would be a clear way the world is more Eliezer-like than Paul-like I think I do further predict that people who know things about animal intelligence, and don’t seem to have identifiably crazy views about any adjacent questions that indicate a weird pro-animal bias, will say that human brains are a lot better than other animal brains for dexterity/locomotion/similar physical tasks (and that the comparison isn’t that close for e.g. comparing humans vs big cats). Incidentally, seems like DM folks did the same thing this year, presumably publishing now because they got scooped. Looks like they probably have a better algorithm but used harder environments instead of Atari. (They also evaluate the algorithm SPR+MuZero I mentioned which indeed gets one factor of 2x improvement over MuZero alone, roughly as you’d guess): <https://arxiv.org/pdf/2111.01587.pdf> [Barnes][13:45] My DM friend says they tried it before they were focused on data efficiency and it didn’t help in that regime, sounds like they ignored it for a while after that \| \| \| --- \| \| [Christiano: 👍] \| [Christiano][13:48] Overall the situation feels really boring to me. Not sure if DM having a highly similar unpublished result is more likely on my view than Eliezer’s (and initially ignoring the method because they weren’t focused on sample-efficiency), but at any rate I think it’s not anywhere close to falsifying my view. 18. Follow-ups to the Christiano/Yudkowsky conversation ------------------------------------------------------- [Karnofsky][9:39] Going to share a point of confusion about this latest exchange. It started with Eliezer saying this: > > Thing that (if true) strikes me as… straight-up falsifying Paul’s view as applied to modern-day AI, at the frontier of the most AGI-ish part of it and where Deepmind put in substantial effort on their project? EfficientZero (allegedly) learns Atari in 100,000 frames. Caveat: I’m not having an easy time figuring out how many frames MuZero would’ve required to achieve the same performance level. MuZero was trained on 200,000,000 frames but reached what looks like an allegedly higher high; the EfficientZero paper compares their performance to MuZero on 100,000 frames, and claims theirs is much better than MuZero given only that many frames. > > > So at this point, I thought Eliezer’s view was something like: “EfficientZero represents a several-OM (or at least one-OM?) jump in efficiency, which should shock the hell out of Paul.” The upper bound on the improvement is 2000x, so I figured he thought the corrected improvement would be some number of OMs. But very shortly afterwards, Eliezer quotes Gwern’s guess of a 4x improvement, and Paul then said: > > Concretely, this is a paper that adds a few techniques to improve over MuZero in a domain that (it appears) wasn’t a significant focus of MuZero. I don’t know how much it improves but I can believe gwern’s estimates of 4x. > > > I’d guess MuZero itself is a 2x improvement over the baseline from a year ago, which was maybe a 4x improvement over the algorithm from a year before that. If that’s right, then no it’s not mindblowing on my view to have 4x progress one year, 2x progress the next, and 4x progress the next. > > > Eliezer never seemed to push back on this 4x-2x-4x claim. What I thought would happen after the 4x estimate and 4x-2x-4x claim: Eliezer would’ve said “Hmm, we should nail down whether we are talking about 4x-2x-4x or something more like 4x-2x-100x. If it’s 4x-2x-4x, then I’ll say ‘never mind’ re: my comment that this ‘straight-up falsifies Paul’s view.’ At best this is just an iota of evidence or something.” Why isn’t that what happened? Did Eliezer mean all along to be saying that a 4x jump on Atari sample efficiency would “straight-up falsify Paul’s view?” Is a 4x jump the kind of thing Eliezer thinks is going to power a jumpy AI timeline? \| \| \| \| --- \| --- \| \| [Ngo: 👍] \| [Shah: ➕] \| [Yudkowsky][11:16] This is a proper confusion and probably my fault; I also initially thought it was supposed to be 1-2 OOM and should’ve made it clearer that Gwern’s 4x estimate was less of a direct falsification. I’m not yet confident Gwern’s estimate is correct. I just got a reply from my query to the paper’s first author which reads: > > Dear Eliezer: It’s a good question. But due to the limits of resources and time, we haven’t evaluated the sample efficiency towards different frames systematically. I think it’s not a trivial question as the required time and resources are much expensive for the 200M frames setting, especially concerning the MCTS-based methods. Maybe you need about several days or longer to finish a run with GPUs in that setting. I hope my answer can help you. Thank you for your email. > > > I replied asking if Gwern’s 3.8x estimate sounds right to them. A 10x improvement could power what I think is a jumpy AI timeline. I’m currently trying to draft a depiction of what I think an unrealistically dignified but computationally typical end-of-world would look like if it started in 2025, and my first draft of that had it starting with a new technique published by Google Brain that was around a 10x improvement in training speeds for very large networks at the cost of higher inference costs, but which turned out to be specially applicable to online learning. That said, I think the 10x part isn’t either a key concept or particularly likely, and it’s much more likely that hell breaks loose when an innovation changes some particular step of the problem from “can’t realistically be done at all” to “can be done with a lot of computing power”, which was what I had being the real effect of that hypothetical Google Brain innovation when applied to online learning, and I will probably rewrite to reflect that. [Karnofsky][11:29] That’s helpful, thanks. Re: “can’t realistically be done at all” to “can be done with a lot of computing power”, cpl things: 1. Do you think a 10x improvement in efficiency at some particular task could qualify as this? Could a smaller improvement? 2. I thought you were pretty into the possibility of a jump from “can’t realistically be done at all” to “can be done with a small amount of computing power,” eg some random ppl with a $1-10mm/y budget blowing past mtpl labs with >$1bb/y budgets. Is that wrong? [Yudkowsky][13:44] 1 – yes and yes, my revised story for how the world ends looks like Google Brain publishing something that looks like only a 20% improvement but which is done in a way that lets it be adapted to make online learning by gradient descent “work at all” in DeepBrain’s ongoing Living Zero project (not an actual name afaik) 2 – that definitely remains very much allowed in principle, but I think it’s not my current mainline probability for how the world’s end plays out – although I feel hesitant / caught between conflicting heuristics here. I think I ended up much too conservative about timelines and early generalization speed because of arguing with Robin Hanson, and don’t want to make a similar mistake here, but on the other hand a lot of the current interesting results have been from people spending huge compute (as wasn’t the case to nearly the same degree in 2008) and if things happen on short timelines it seems reasonable to guess that the future will look that much like the present. This is very much due to cognitive limitations of the researchers rather than a basic fact about computer science, but cognitive limitations are also facts and often stable ones. [Karnofsky][14:35] Hm OK. I don’t know what “online learning by gradient descent” means such that it doesn’t work at all now (does “work at all” mean something like “work with human-ish learning efficiency?”) [Yudkowsky][15:07] I mean, in context, it means “works for Living Zero at the performance levels where it’s running around accumulating knowledge”, which by hypothesis it wasn’t until that point. [Karnofsky][15:12] Hm. I am feeling pretty fuzzy on whether your story is centrally about: 1. A <10x jump in efficiency at something important, leading pretty directly/straightforwardly to crazytown 2. A 100x ish jump in efficiency at something important, which may at first “look like” a mere <10x jump in efficiency at something else #2 is generally how I’ve interpreted you and how the above sounds, but under #2 I feel like we should just have consensus that the Atari thing being 4x wouldn’t be much of an update. Maybe we already do (it was a bit unclear to me from your msg) (And I totally agree that we haven’t established the Atari thing is only 4x – what I’m saying is it feels like the conversation should’ve paused there) [Yudkowsky][15:13] The Atari thing being 4x over 2 years is I think legit not an update because that’s standard software improvement speed you’re correct that it should pause there [Karnofsky][15:14] (Nov. 5) ![👍](https://s.w.org/images/core/emoji/14.0.0/72x72/1f44d.png) [Yudkowsky] [15:24] I think that my central model is something like – there’s a central thing to general intelligence that starts working when you get enough pieces together and they coalesce, which is why humans went down this evolutionary gradient by a lot before other species got 10% of the way there in terms of output; and then it takes a big pile of that thing to do big things, which is why humans didn’t go faster than cheetahs until extremely late in the game. so my visualization of how the world starts to end is “gear gets added and things start to happen, maybe slowly-by-my-standards at first such that humans keep on pushing it along rather than it being self-moving, but at some point starting to cumulate pretty quickly in the same way that humans cumulated pretty quickly once they got going” rather than “dial gets turned up 50%, things happen 50% faster, every year”. [Yudkowsky][15:16] (Nov. 5, switching channels) as a quick clarification, I agree that if this is 4x sample efficiency over 2 years then that doesn’t at all challenge Paul’s view [Christiano][0:20] FWIW, I felt like the entire discussion of EfficientZero was a concrete example of my view making a number of more concentrated predictions than Eliezer that were then almost immediately validated. In particular, consider the following 3 events: * The quantitative effect size seems like it will turn out to be much smaller than Eliezer initially believed, much closer to being in line with previous progress. * DeepMind had relatively similar results that got published immediately after our discussion, making it look like random people didn’t pull ahead of DM after all. * DeepMind appears not to have cared much about the metric in question, as evidenced by (i) Beth’s comment above, which is basically what I said was probably going on, (ii) they barely even mention Atari sample-efficiency in their paper about similar methods. If only 1 of these 3 things had happened, then I agree this would have been a challenge to my view that would make me update in Eliezer’s direction. But that’s only possible if Eliezer actually assigns a higher probability than me to <= 1 of these things happening, and hence a lower probability to >= 2 of them happening. So if we’re playing a reasonable epistemic game, it seems like I need to collect some epistemic credit every time something looks boring to me. [Yudkowsky][15:30] I broadly agree; you win a Bayes point. I think some of this (but not all!) was due to my tripping over my own feet and sort of rushing back with what looked like a Relevant Thing without contemplating the winner’s curse of exciting news, the way that paper authors tend to frame things in more exciting rather than less exciting ways, etc. But even if you set that aside, my underlying AI model said that was a thing which could happen (which is why I didn’t have technically rather than sociologically triggered skepticism) and your model said it shouldn’t happen, and it currently looks like it mostly didn’t happen, so you win a Bayes point. Notes that some participants may deem obvious(?) but that I state expecting wider readership: * Just like markets are almost entirely efficient (in the sense that, even when they’re not efficient, you can only make a very small fraction of the money that could be made from the entire market if you owned a time machine), even sharp and jerky progress has to look almost entirely not so fast almost all the time if the Sun isn’t right in the middle of going supernova. So the notion that progress sometimes goes jerky and fast does have to be evaluated by a portfolio view over time. In worlds where progress is jerky even before the End Days, Paul wins soft steady Bayes points in most weeks and then I win back more Bayes points once every year or two. * We still don’t have a very good idea of how much longer you would need to train the previous algorithm to match the performance of the new algorithm, just an estimate by Gwern based off linearly extrapolating a graph in a paper. But, also to be clear, not knowing something is not the same as expecting it to update dramatically, and you have to integrate over the distribution you’ve got. * It’s fair to say, “Hey, Eliezer, if you tripped over your own feet here, but only noticed that because Paul was around to call it, maybe you’re tripping over your feet at other times when Paul isn’t around to check your thoughts in detail” – I don’t want to minimize the Bayes point that Paul won either. [Christiano][16:29] Agreed that it’s (i) not obvious how large the EfficientZero gain was, and in general it’s not a settled question what happened, (ii) it’s not that big an update, it needs to be part of a portfolio (but this is indicative of the kind of thing I’d want to put in the portfolio), (iii) it generally seems pro-social to flag potentially relevant stuff without the presumption that you are staking a lot on it. The post [Christiano and Yudkowsky on AI predictions and human intelligence](https://intelligence.org/2022/03/01/christiano-and-yudkowsky-on-ai-predictions-and-human-intelligence/) appeared first on [Machine Intelligence Research Institute](https://intelligence.org).
69dfa196-71e6-4e2a-a203-55f044f005f5	trentmkelly/LessWrong-43k	LessWrong	Formal verification, heuristic explanations and surprise accounting ARC's current research focus can be thought of as trying to combine mechanistic interpretability and formal verification. If we had a deep understanding of what was going on inside a neural network, we would hope to be able to use that understanding to verify that the network was not going to behave dangerously in unforeseen situations. ARC is attempting to perform this kind of verification, but using a mathematical kind of "explanation" instead of one written in natural language. To help elucidate this connection, ARC has been supporting work on Compact Proofs of Model Performance via Mechanistic Interpretability by Jason Gross, Rajashree Agrawal, Lawrence Chan and others, which we were excited to see released along with this post. While we ultimately think that provable guarantees for large neural networks are unworkable as a long-term goal, we think that this work serves as a useful springboard towards alternatives. In this post, we will: * Summarize ARC's takeaways from this work and the problems we see with provable guarantees * Explain ARC's notion of a heuristic explanation and how it is intended to overcome these problems * Describe with the help of a worked example how the quality of a heuristic explanation can be quantified, using a process we have been calling surprise accounting We are also sharing a draft by Gabriel Wu (currently visiting ARC) describing a heuristic explanation for the same model that appears in the above paper: max_of_k Heuristic Estimator Thanks to Stephanie He for help with the diagrams in this post. Thanks to Eric Neyman, Erik Jenner, Gabriel Wu, Holly Mandel, Jason Gross, Mark Xu, and Mike Winer for comments. Formal verification for neural networks In Compact Proofs of Model Performance via Mechanistic Interpretability, the authors train a small transformer on an algorithmic task to high accuracy, and then construct several different formal proofs of lower bounds on the network's accuracy. Without foraying into the detai
37d878bc-196e-4706-a72a-c8ad00ad1f8b	StampyAI/alignment-research-dataset/alignmentforum	Alignment Forum	Anthropics and Fermi: grabby, visible, zoo-keeping, and early aliens When updating the probability of life across the universe, there are two main observations we have to build on: * Anthropic update: we exist on Earth. * [Fermi observation](https://arxiv.org/pdf/1806.02404.pdf): (we exist on Earth and) we don't see any aliens. I'll analyse how these two observations affect various theories about life in the universe. In general, we'll see that the anthropic update has a pretty weak effect, while the Fermi observation has a strong effect: those theories that benefit most are those that avoid the downgrade from the Fermi, such as the Zoo hypothesis, or the "human life unusually early" hypothesis. Grabby and visible aliens ------------------------- I've argued that an anthropic update on our own existence is actually [just a simple Bayesian update](https://www.lesswrong.com/posts/BYy62ib5tAkn9rsKn/sia-is-basically-just-bayesian-updating-on-existence); here I'll explain what that means for our updates. [This paper](https://arxiv.org/abs/2102.01522) talks about grabby aliens, who would expand across the universe, and stop humans from evolving (if they reached Earth before now). As I've [argued](https://www.lesswrong.com/posts/3kwwDieE9SmFoXz9F/non-poisonous-cake-anthropic-updates-are-normal), "we exist" and "we have not observed X" are statements that can be treated in exactly the same way. We can combine them to say "there are no visible aliens anywhere near", without distinguishing grabby aliens (who would have stopped our existence) from visible-but-not grabby aliens (who would have changed our observations). Thus the Fermi observation is saying there are no grabby or visible aliens nearby[[1]](#fn-shKTcgz23ukTtebyy-1). Recall that it's [so comparatively easy](http://www.fhi.ox.ac.uk/wp-content/uploads/intergalactic-spreading.pdf) to cross between stars and galaxies, so advanced aliens would only fail to be grabby if they coordinated to not want to do so. Rare Earth hypotheses --------------------- Some theories posit that life requires a collection of conditions that are very rarely found together. But rare Earth theories don't differ much, upon updates, from "life is hard" hypotheses. For example, suppose T0.mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0} .MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0} .mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table} .mjx-full-width {text-align: center; display: table-cell!important; width: 10000em} .mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0} .mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left} .mjx-numerator {display: block; text-align: center} .mjx-denominator {display: block; text-align: center} .MJXc-stacked {height: 0; position: relative} .MJXc-stacked > \* {position: absolute} .MJXc-bevelled > \* {display: inline-block} .mjx-stack {display: inline-block} .mjx-op {display: block} .mjx-under {display: table-cell} .mjx-over {display: block} .mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important} .mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important} .mjx-stack > .mjx-sup {display: block} .mjx-stack > .mjx-sub {display: block} .mjx-prestack > .mjx-presup {display: block} .mjx-prestack > .mjx-presub {display: block} .mjx-delim-h > .mjx-char {display: inline-block} .mjx-surd {vertical-align: top} .mjx-surd + .mjx-box {display: inline-flex} .mjx-mphantom \* {visibility: hidden} .mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%} .mjx-annotation-xml {line-height: normal} .mjx-menclose > svg {fill: none; stroke: currentColor; overflow: visible} .mjx-mtr {display: table-row} .mjx-mlabeledtr {display: table-row} .mjx-mtd {display: table-cell; text-align: center} .mjx-label {display: table-row} .mjx-box {display: inline-block} .mjx-block {display: block} .mjx-span {display: inline} .mjx-char {display: block; white-space: pre} .mjx-itable {display: inline-table; width: auto} .mjx-row {display: table-row} .mjx-cell {display: table-cell} .mjx-table {display: table; width: 100%} .mjx-line {display: block; height: 0} .mjx-strut {width: 0; padding-top: 1em} .mjx-vsize {width: 0} .MJXc-space1 {margin-left: .167em} .MJXc-space2 {margin-left: .222em} .MJXc-space3 {margin-left: .278em} .mjx-test.mjx-test-display {display: table!important} .mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px} .mjx-test.mjx-test-default {display: block!important; clear: both} .mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex} .mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left} .mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right} .mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0} .MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal} .MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal} .MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold} .MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold} .MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw} .MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw} .MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw} .MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw} .MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw} .MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw} .MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw} .MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw} .MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw} .MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw} .MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw} .MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw} .MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw} .MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw} .MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw} .MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw} .MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw} .MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw} .MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw} .MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw} .MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw} @font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')} @font-face {font-family: MJXc-TeX-ams-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')} @font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold} @font-face {font-family: MJXc-TeX-cal-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')} @font-face {font-family: MJXc-TeX-frak-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')} @font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold} @font-face {font-family: MJXc-TeX-frak-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')} @font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic} @font-face {font-family: MJXc-TeX-math-BIw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')} @font-face {font-family: MJXc-TeX-sans-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')} @font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold} @font-face {font-family: MJXc-TeX-sans-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')} @font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic} @font-face {font-family: MJXc-TeX-sans-Iw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')} @font-face {font-family: MJXc-TeX-script-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')} @font-face {font-family: MJXc-TeX-type-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')} @font-face {font-family: MJXc-TeX-cal-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')} @font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold} @font-face {font-family: MJXc-TeX-main-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')} @font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic} @font-face {font-family: MJXc-TeX-main-Iw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')} @font-face {font-family: MJXc-TeX-main-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')} @font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic} @font-face {font-family: MJXc-TeX-math-Iw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')} @font-face {font-family: MJXc-TeX-size1-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')} @font-face {font-family: MJXc-TeX-size2-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')} @font-face {font-family: MJXc-TeX-size3-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')} @font-face {font-family: MJXc-TeX-size4-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')} @font-face {font-family: MJXc-TeX-vec-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')} @font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold} @font-face {font-family: MJXc-TeX-vec-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')} say that life can exist on any planet with rate ρ, while the rare Earth hypothesis, T1, says that life can exist on Earth-like planets with rate p, while Earth-like planets themselves exist with rate r. But neither the Fermi observation nor the anthropic update will distinguish between these theories. The Fermi observation posits that there are no visible aliens close to Earth; the anthropic updates increases the probability of life similar to us. We can see T1 as T0 with a different prior on ρ=pr (induced from the priors on p and r), but both these updates affect ρ. Now, T0 and T1 can be distinguished by observation (seeing dead planets with Earth-like features) or theory (figuring out what is needed for life). Anything that differentially change p and r. But neither the anthropic update nor the Fermi observation do this. Independent aliens ------------------ Suppose that theory T2 posits that there are aliens in, say, gas giants, whose existence is independent from ours. Visible gas giant alien civilizations exist at a rate ρg, while visible life on rocky planets exist at a rate ρ. Then the anthropic update boosts ρ only, while the Fermi observation penalises ρ and ρg equally (if we make the simplifying assumption that gas giants are as common as terrestrial planets). This gives a differential boost to ρ over ρg, but the effect can be mild. If we assume that there are N gas giants and N terrestrial planets in the milky way, and start with a uniform prior over both ρ and ρg, then [after updating](https://www.lesswrong.com/posts/xfEsxAtBTLgFe7fSZ/the-sia-population-update-can-be-surprisingly-small), we get: ρ=2N+2, ρg=1N+1. If the rate of gas giant alien civilizations is semi-dependent on our own existence - maybe we both need it to be easy for RNA to exist - then there will be less of a difference in the update for ρ and ρg. So, some differential effect due to anthropics, but not a strong one, at least for uniform priors, and not one that grows. In the cosmic zoo ----------------- Let T3 be a cosmic [zoo hypothesis](https://en.wikipedia.org/wiki/Zoo_hypothesis). It posits that there may be a lot of aliens, but they have agreed - or been coerced - into hiding themselves, so as not to contaminate human development (or some other reason). Then T3 gets a boost from the anthropic update, and no penalty from the Fermi observation. Since most theories get a big downgrade from the Fermi observation, this can raise its probability quite a lot relative to other theories. A few caveats, however: 1. In the zoo hypothesis, aliens are hiding themselves from us. This is close to a "[Descartes's demon](https://en.wikipedia.org/wiki/Evil_demon)" hypothesis, in that powerful entities are acting to feed us erroneous observations. Pure Descartes's demon hypotheses are not differentially boosted by anything, since they explain nothing (once you've posited a demon, you also have to explain why we see what we think we see). The zoo hypothesis is not quite as bad - "keep everything hidden" is more likely than other ways aliens could be messing with our observations. Still, it should be a low prior. 2. Though the Fermi observation doesn't downgrade the zoo hypothesis directly, the more carefully we observe the universe, the more unlikely it becomes, since the aliens would have to work harder to conceal any evidence. 3. Conversely, the more visible we become, the less likely the zoo hypothesis becomes, because we have to explain why the zookeepers haven't intervened to keep us concealed (if we suppose that these aliens are powerful enough to intercept light and other signals between the stars, then we're very close to the Descartes demon territory). Once we successfully launched replicating AIs to the stars, then we'd be pretty sure the zoo hypothesis was wrong. Time enough for aliens ---------------------- So far, we've neglected time in the equation, talking about a rate ρ that was per planet, but not stretched over time. But consider theory T4: advanced life starts appearing around [13.77 billion years](https://en.wikipedia.org/wiki/Age_of_the_universe) after the Big Bang, but not before. This theory might be unlikely, but it gets a mild boost from anthropics (since it's compatible with our existence) and avoids the downgrade from the Fermi observation (since it says there are no visible aliens - yet). Since that downgrade has been quite powerful for most theories, T4 can get boosted relative to them - and the more dead planets we observe or infer, the stronger the relative boost is. Now, T4 may seem unlikely, since the Earth is a late planet among the Earth-like planets: "[Thus, the average earth in the Universe is 1.8 ± 0.9 billion years older than our Earth](https://arxiv.org/pdf/astro-ph/0012399.pdf)". But there are some theories that make more plausible T4, such as some versions of [panspermia](https://en.wikipedia.org/wiki/Panspermia). Specifically, if we imagine that life had to go through several stages, on several planets - maybe RNA/DNA was the result of billions of years of evolution on a planet much older than the Earth, and was then spread here, where it allowed another stage of evolution. Conversely, theories T5 that posit that advanced life started much earlier than the present day, pay a much higher price via the Fermi observation. --- 1. The grabby alien paper uses "loud" to designate aliens that "expand fast, last long, and make visible changes to their volumes". Visible aliens are more general; in particular, they need not expand (though this may make them less visible). [↩︎](#fnref-shKTcgz23ukTtebyy-1)
b22ce1bd-4e66-4853-972e-054c5d0f4533	trentmkelly/LessWrong-43k	LessWrong	Futurama does an episode on nano-technology. Futurama, Season 6, Episode 15: Benderama Naturally, this episode's goal is entertainment, not so much the accurate portrayal of the technology. Quoting from wikipedia article: > The premise of "Benderama" is based around the transhumanist theory of grey goo, an end-of-the-world scenario in which out-of-control self-replicating robots consume all matter on Earth while building more of themselves. The episode is not officially available for viewing online (but you can still find it, of course).
49c1ed1f-1987-4fa8-8539-4516512dd293	trentmkelly/LessWrong-43k	LessWrong	What is the purpose and application of AI Debate? I think there is an important lack of clarity and shared understanding regarding how people intend to use AI-Safety-via-Debate-style approaches. So I think it would be helpful if there were some people --- who either (i) work on Debate or (ii) believe that Debate is promising --- who could give their answers to the following three questions: 1. What is, according to you, the purpose of AI Debate? What problem is it supposed to be solving? 2. How do you intend AI Debate to be used? (EG, to generate training data for imitation learning? During deployment, to generate solutions to tasks? During deployment, to check answers produced by other models?) 3. Do you think that AI Debate is a reasonably promising approach for this? Disclaimers: Please don't answer 1-2 without also commenting on 3. Also, note that this isn't a "I didn't understand the relevant papers, please explain" question -- I studied those, and still have this question. Further clarifications in a comment.
0a1d7532-1069-479b-a81e-5d657961b346	trentmkelly/LessWrong-43k	LessWrong	A Proposal for Defeating Moloch in the Prison Industrial Complex Summary I'd like to increasing the well-being of those in the justice system while simultaneously reducing crime. I'm missing something here but I'm not sure what. I'm thinking this may be a worse idea than I originally thought based on comment feedback, though I'm still not 100% sure why this is the case. Current State While the prison system may not constitute an existential threat, At this moment more than 2,266,000 adults are incarcerated in the US alone, and I expect that being in prison greatly decreases QALYs for those incarcerated, that further QALYs are lost to victims of crime, family members of the incarcerated, and through the continuing effects of institutionalization and PTSD from sentences served in the current system, not to mention the brainpower and man-hours lost to any productive use. If you haven't read these Meditations on Moloch, I highly recommend it. It’s long though, so the executive summary is: Moloch is the personification of the forces of competition which perverse incentives, a "race to the bottom" type situation where all human values are discarded in an effort to survive. That this can be solved with better coordination, but it is very hard to coordinate when perverse incentives also penalize the coordinators and reward dissenters. The prison industrial complex is an example of these perverse incentives. No one thinks that the current system is ideal but incentives prevent positive change and increase absolute unhappiness. * Politicians compete for electability. Convicts can’t vote, prisons make campaign contributions and jobs, and appearing “tough on crime” appeals to a large portion of the voter base. * Jails compete for money: the more prisoners they house, the more they are paid and the longer they can continue to exist. This incentive is strong for public prisons and doubly strong for private prisons. * Police compete for bonuses and promotions, both of which are given as rewards to cops who bring in and convict mo
48a011b6-8527-403c-af1b-43ed0c2fb24e	trentmkelly/LessWrong-43k	LessWrong	Post retracted: If you follow expected utility, expect to be money-pumped This post has been retracted because it is in error. Trying to shore it up just involved a variant of the St Petersburg Paradox and a small point on pricing contracts that is not enough to make a proper blog post. I apologise. Edit: Some people have asked that I keep the original up to illustrate the confusion I was under. I unfortunately don't have a copy, but I'll try and recreate the idea, and illustrate where I went wrong. The original idea was that if I were to offer you a contract L that gained £1 with 50% probability or £2 with 50% probability, then if your utility function wasn't linear in money, you would generally value L at having a value other that £1.50. Then I could sell or buy large amounts of these contracts from you at your stated price, and use the law of large number to ensure that I valued each contract at £1.50, thus making a certain profit. The first flaw consisted in the case where your utility is concave in cash ("risk averse"). In that case, I can't buy L from you unless you already have L. And each time I buy it from you, the mean quantity of cash you have goes down, but your utility goes up, since you do not like the uncertainty inherent in L. So I get richer, but you get more utility, and once you've sold all L's you have, I cannot make anything more out of you. If your utility is convex in cash ("risk loving"), then I can sell you L forever, at more than £1.50. And your money will generally go down, as I drain it from you. However, though the median amount of cash you have goes down, your utility goes up, since you get a chance - however tiny - of huge amounts of cash, and the utility generated by this sum swamps the fact you are most likely ending up with nothing. If I could go on forever, then I can drain you entirely, as this is a biased random walk on a one-dimensional axis. But I would need infinite ressources to do this. The major error was to reason like an investor, rather than a utility maximiser. Investors are very intere
92f25229-37db-4afe-8048-3ba7f3598754	trentmkelly/LessWrong-43k	LessWrong	Change Contexts to Improve Arguments On a recent trip to Ireland, I gave a talk on tactics for having better arguments (video here). There's plenty in the video that's been discussed on LW before (Ideological Turing Tests and other reframes), but I thought I'd highlight one other class of trick I use to have more fruitful disagreements. It's hard, in the middle of a fight, to remember, recognize, and defuse common biases, rhetorical tricks, emotional triggers, etc. I'd rather cheat than solve a hard problem, so I put a lot of effort into shifting disagreements into environments where it's easier for me and my opposite-number to reason and argue well, instead of relying on willpower. Here's a recent example of the kind of shift I like to make: A couple months ago, a group of my friends were fighting about the Brendan Eich resignation on facebook. The posts were showing up fast; everyone was, presumably, on the edge of their seats, fueled by adrenaline, and alone at their various computers. It’s a hard place to have a charitable, thoughtful debate. I asked my friends (since they were mostly DC based) if they’d be amenable to pausing the conversation and picking it up in person. I wanted to make the conversation happen in person, not in front of an audience, and in a format that let people speak for longer and ask questions more easily. If so, I promised to bake cookies for the ultimate donnybrook. My friends probably figured that I offered cookies as a bribe to get everyone to change venues, and they were partially right. But my cookies had another strategic purpose. When everyone arrived, I was still in the process of taking the cookies out of the oven, so I had to recruit everyone to help me out. “Alice, can you pour milk for people?” “Bob, could you pass out napkins?” “Eve, can you greet people at the door while I’m stuck in the kitchen with potholders on?” Before we could start arguing, people on both sides of the debate were working on taking care of each other and asking each others’
4349b14c-030b-44b2-bfd6-ad93205d02ce	trentmkelly/LessWrong-43k	LessWrong	I have COVID, for how long should I isolate? Case was relatively mild, and I'm feeling a lot better now (day 5), though I still have the sniffles and minor fatigue. I tested negative last week and positive this week via PCR, so I'm pretty confident this was COVID and not an unrelated illness. I live in the US, where CDC guidelines now say I can go into public spaces if my symptoms are "resolving" and I wear a mask, though this feels a little permissive given the requirements of other countries. It's unlikely I will be able to get an antigen test result back before day 10. I'm fully vaccinated and boosted with pfizer as of ~3 months ago. I'm having trouble finding high quality research on infectiousness over time. What do you think is a reasonable amount of time to wait (possibly conditional on symptoms) to do the following? None are essential to my basic needs. I will probably do the median of what the community recommends if it seems well supported. * Attend an outdoor event masked where I might be <6ft away from others. * Go to a grocery store masked at off-peak hours * Ride a bus masked * Eat indoors at a crowded restaurant.
043364e1-de2d-4dfb-997d-3f433a661bbc	trentmkelly/LessWrong-43k	LessWrong	Meetup : Cambridge, MA third-Sunday meetup Discussion article for the meetup : Cambridge, MA third-Sunday meetup WHEN: 18 February 2013 02:00:00PM (-0500) WHERE: 25 Ames St Cambridge, MA 02138 Cambridge/Boston-area Less Wrong meetups recur on the first and third Sunday of every month, 2pm at the MIT Landau Building [25 Ames St, Bldg 66], room 148. Discussion article for the meetup : Cambridge, MA third-Sunday meetup
ddc00b23-c2e5-4fc9-8fc1-90b67f2d0fe2	trentmkelly/LessWrong-43k	LessWrong	Dominic Cummings: "we’re hiring data scientists, project managers, policy experts, assorted weirdos" Dominic Cummings (discussed previously on LW, most recently here) is a Senior Advisor to the new UK PM, Boris Johnson. He also seems to be essentially a rationalist (at least in terms of what ideas he's paying attention to). He has posted today that his team is hiring "data scientists, project managers, policy experts, assorted weirdos". Perhaps some LW readers should apply. Extensive quotes below: ---------------------------------------- ‘This is possibly the single largest design flaw contributing to the bad Nash equilibrium in which … many governments are stuck. Every individual high-functioning competent person knows they can’t make much difference by being one more face in that crowd.’ Eliezer Yudkowsky, AI expert, LessWrong etc. [...] Now there is a confluence of: a) Brexit requires many large changes in policy and in the structure of decision-making, b) some people in government are prepared to take risks to change things a lot, and c) a new government with a significant majority and little need to worry about short-term unpopularity while trying to make rapid progress with long-term problems. There is a huge amount of low hanging fruit — trillion dollar bills lying on the street — in the intersection of: * the selection, education and training of people for high performance * the frontiers of the science of prediction * data science, AI and cognitive technologies (e.g Seeing Rooms, ‘authoring tools designed for arguing from evidence’, Tetlock/IARPA prediction tournaments that could easily be extended to consider ‘clusters’ of issues around themes like Brexit to improve policy and project management) * communication (e.g Cialdini) * decision-making institutions at the apex of government. We want to hire an unusual set of people with different skills and backgrounds to work in Downing Street with the best officials, some as spads and perhaps some as officials. If you are already an official and you read this blog and think you fit one of these ca
1008911a-9d55-4ca0-8ad2-94862de138ed	trentmkelly/LessWrong-43k	LessWrong	Does cognitive therapy encourage bias? Summary: Cognitive therapy may encourage motivated cognition. My main source for this post is Judith Beck's Cognitive Therapy: Basics and Beyond. "Cognitive behavioral therapy" (CBT) is a catch-all term for a variety of therapeutic practices and theories. Among other things, it aims to teach patients to modify their own beliefs. The rationale seems to be this: (1) Affect, behavior, and cognition are interrelated such that changes in one of the three will lead to changes in the other two. (2) Affective problems, such as depression, can thus be addressed in a roundabout fashion: modifying the beliefs from which the undesired feelings stem. So far, so good. And how does one modify destructive beliefs? CBT offers many techniques. Alas, included among them seems to be motivated skepticism. For example, consider a depressed college student. She and her therapist decide that one of her bad beliefs is "I'm inadequate." They want to replace that bad one with a more positive one, namely, "I'm adequate in most ways (but I'm only human, too)." Their method is to do a worksheet comparing evidence for and against the old, negative belief. Listen to their dialog: [Therapist]: What evidence do you have that you're inadequate? [Patient]: Well, I didn't understand a concept my economics professor presented in class today. T: Okay, write that down on the right side, then put a big "BUT" next to it...Now, let's see if there could be another explanation for why you might not have understood the concept other than that you're inadequate. P: Well, it was the first time she talked about it. And it wasn't in the readings. Thus the bad belief is treated with suspicion. What's wrong with that? Well, see what they do about evidence against her inadequacy: T: Okay, let's try the left side now. What evidence do you have from today that you are adequate at many things? I'll warn you, this can be hard if your screen is operating. P: Well, I worked on my literature paper. T: Good. Wr
01bf51d3-519a-4b06-8d3d-99812e5b40b8	trentmkelly/LessWrong-43k	LessWrong	on wellunderstoodness epistemic status: nothing new, informal, may or may not provide a novel compression which gives at least one person more clarity. Also, not an alignment post, just uses arguments about alignment as a motivating example, but the higher-level takeaway is extremely similar to the Rocket Alignment dialogue and to the definition of "deconfusion" that MIRI offered recently (even though I'm making a very low-level point) notation: please read a <- x as "a replaced with x", e.g. (fx) x<-5 => (mx+b) x<-5 => m5+b. (typically the sequents literature uses fx[5/x] but I find this much harder to read). motivation A younger version of me once said "What's the big deal with sorceror's apprentice / paperclip maximizer? you would obviously just give an ϵ and say that you want to be satisfied up to a threshold". But ultimately it isn't hard to believe that software developers don't want to be worrying about whether some hapless fool somewhere forgot to give the ϵ. Not to mention that we don't expect the need for such thresholds to be finitely enumerable. Not to mention the need to pick the right epsilons. The aspiration of value alignment just tells you "we want to really understand goal description, we want to understand it so hard that we make epsilons (chasing after the perfectly calibrated thresholds) obsolete". Well, is it clear to everyone what "really understand" even means? It's definitely not obvious, and I don't think everyone's on the same page about it. core claim I want to focus on the following idea; to be mathematically understood is to be reduced to substitution instances. Consider; linear algebra. If I want to teach someone the meaning and the importance of ax+by+c=0 I only need to make sure (skipping over addition, multiplication, and equality) that they can distinguish between a scalar and a variable and then sculpt one form from another. Maybe it generalizes to arbitrary length aixi and to arbitrary number of equations as nicely as it does because the unive
8bd98e52-2b7c-4955-95eb-b095588ba596	trentmkelly/LessWrong-43k	LessWrong	The Cancer Resolution? Book review: The Cancer Resolution?: Cancer reinterpreted through another lens, by Mark Lintern. In the grand tradition of outsiders overturning scientific paradigms, this book proposes a bold new theory: cancer isn't a cellular malfunction, but a fungal invasion. Lintern spends too many pages railing against the medical establishment, which feels more like ax-grinding than science. I mostly agreed with his conclusions here, but mostly for somewhat different reasons than the ones he provides. If you can push through this preamble, you'll find a treasure trove of scientific intrigue. Lintern's central claim is that fungal infections, not genetic mutations, are the primary cause of cancer. He dubs this the "Cell Suppression theory," painting a picture of fungi as cellular puppet masters, manipulating our cells for their own nefarious ends. This part sounds much more like classical science, backed by hundreds of quotes from peer-reviewed literature. Those quotes provide extensive evidence that Lintern's theory predicts dozens of cancer features better than do the established theories. Older Theories 1. The DNA Theory (aka Somatic Mutation Theory): The reigning heavyweight, this theory posits that cancer results from an accumulation of genetic mutations in critical genes that control cell growth, division, and death. 2. Another old theory that still has advocates is the Metabolic Theory. This theory suggests that cancer is primarily a metabolic disease, characterized by impaired cellular energy production (the Warburg effect). It proposes that damage to mitochondria is a key factor in cancer development. I wrote a mixed review of a book about it. Lintern points out evidence that mitochondria are turned off by signals, not damaged. He also notes that tumors with malfunctioning mitochondria are relatively benign. Evidence Discrediting the DNA Theory The standard version of the DNA Theory predicts that all cancer cells will have mutations that affect replicati
55d4129e-a556-4056-b24d-32c367f21d86	trentmkelly/LessWrong-43k	LessWrong	Briefly thinking through some analogs of debate [This post is an almost direct summary of a conversation with John Wentworth.] Note: This post assumes that the reader knows what AI Safety via Debate is. I don't spend any time introducing it or explaining how it works. The proposal of AI safety via debate depends on the a critical assumption that "in the limit of argumentative prowess, the optimal debate strategy converges to making valid arguments for true conclusions." If that assumption is true, then a sufficiently powerful AI-debate system is a safe oracle (baring other possible issues that I'm not thinking about right now). If this assumption is false, then the debate schema doesn't add any additional safety guarantee. Is this assumption true? That's hard to know, but one thing that we can do is try to compare to other, analogous, real world cases, and see if analogous versions of the "valid arguments for true conclusions" holds. Examining those analogous situations is unlikely to be definitive, but it might give us some hints, or more of a handle on what criteria must be satisfied for the assumption to hold. In this essay I explore some analogs, an attempted counterexample and an attempted example. Attempted counterexample: Public discourse One analogous situation in which this assumption seems straightforwardly false is in the general discourse (on twitter, in "the media", around the water cooler, etc.). Very often, memes that are simple and well fit to human psychology, but false, have a fitness advantage over more complicated ideas that are true. For instance, the core idea of minimum wage can be pretty attractive: lots of people are suffering because the have to work very hard, but they don't make much money, sometimes barely enough to live on. That seems like an inhumane outcome, so we should mandate that everyone be paid at least a fair wage. This simple argument doesn't hold [1], but the explanation for why it is false is a good deal longer than the initial argument. To show what's wrong wi
8a20a41e-3679-45ae-bddd-d553ff1dfa54	StampyAI/alignment-research-dataset/alignmentforum	Alignment Forum	Speculations on information under logical uncertainty .mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0} .MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0} .mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table} .mjx-full-width {text-align: center; display: table-cell!important; width: 10000em} .mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0} .mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left} .mjx-numerator {display: block; text-align: center} .mjx-denominator {display: block; text-align: center} .MJXc-stacked {height: 0; position: relative} .MJXc-stacked > \* {position: absolute} .MJXc-bevelled > \* {display: inline-block} .mjx-stack {display: inline-block} .mjx-op {display: block} .mjx-under {display: table-cell} .mjx-over {display: block} .mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important} .mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important} .mjx-stack > .mjx-sup {display: block} .mjx-stack > .mjx-sub {display: block} .mjx-prestack > .mjx-presup {display: block} .mjx-prestack > .mjx-presub {display: block} .mjx-delim-h > .mjx-char {display: inline-block} .mjx-surd {vertical-align: top} .mjx-mphantom \* {visibility: hidden} .mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%} .mjx-annotation-xml {line-height: normal} .mjx-menclose > svg {fill: none; stroke: currentColor} .mjx-mtr {display: table-row} .mjx-mlabeledtr {display: table-row} .mjx-mtd {display: table-cell; text-align: center} .mjx-label {display: table-row} .mjx-box {display: inline-block} .mjx-block {display: block} .mjx-span {display: inline} .mjx-char {display: block; white-space: pre} .mjx-itable {display: inline-table; width: auto} .mjx-row {display: table-row} .mjx-cell {display: table-cell} .mjx-table {display: table; width: 100%} .mjx-line {display: block; height: 0} .mjx-strut {width: 0; padding-top: 1em} .mjx-vsize {width: 0} .MJXc-space1 {margin-left: .167em} .MJXc-space2 {margin-left: .222em} .MJXc-space3 {margin-left: .278em} .mjx-test.mjx-test-display {display: table!important} .mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px} .mjx-test.mjx-test-default {display: block!important; clear: both} .mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex} .mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left} .mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right} .mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0} .MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal} .MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal} .MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold} .MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold} .MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw} .MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw} .MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw} .MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw} .MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw} .MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw} .MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw} .MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw} .MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw} .MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw} .MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw} .MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw} .MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw} .MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw} .MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw} .MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw} .MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw} .MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw} .MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw} .MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw} .MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw} @font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')} @font-face {font-family: MJXc-TeX-ams-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')} @font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold} @font-face {font-family: MJXc-TeX-cal-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')} @font-face {font-family: MJXc-TeX-frak-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')} @font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold} @font-face {font-family: MJXc-TeX-frak-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')} @font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic} @font-face {font-family: MJXc-TeX-math-BIw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')} @font-face {font-family: MJXc-TeX-sans-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')} @font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold} @font-face {font-family: MJXc-TeX-sans-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')} @font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic} @font-face {font-family: MJXc-TeX-sans-Iw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')} @font-face {font-family: MJXc-TeX-script-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')} @font-face {font-family: MJXc-TeX-type-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')} @font-face {font-family: MJXc-TeX-cal-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')} @font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold} @font-face {font-family: MJXc-TeX-main-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')} @font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic} @font-face {font-family: MJXc-TeX-main-Iw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')} @font-face {font-family: MJXc-TeX-main-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')} @font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic} @font-face {font-family: MJXc-TeX-math-Iw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')} @font-face {font-family: MJXc-TeX-size1-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')} @font-face {font-family: MJXc-TeX-size2-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')} @font-face {font-family: MJXc-TeX-size3-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')} @font-face {font-family: MJXc-TeX-size4-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')} @font-face {font-family: MJXc-TeX-vec-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')} @font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold} @font-face {font-family: MJXc-TeX-vec-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')} A strong theory of logical uncertainty might let us say when the results of computations will give “information”, including logical information, about other computations. This might be useful for, among other things, identifying parts of hypotheses that have the same meaning. --- TL;DR: I don’t think this works as stated, and this kind of problem should probably be sidestepped anyway. Experts may get most of the value from the summary. Thanks to Sam Eisenstat for conversations about this idea. * [Executive summary](#executive-summary) * [Logical information](#logical-information) + [Logical uncertainty](#logical-uncertainty) + [Logical information](#logical-information-1) + [Example: XOR](#example-xor) + [Collapsing levels of indirection and obfuscation](#collapsing-levels-of-indirection-and-obfuscation) * [Possible uses of logical information](#possible-uses-of-logical-information) + [Throttling changes in logical uncertainty](#throttling-changes-in-logical-uncertainty) + [Ontology identification](#ontology-identification) + [Not-obviously-this-person predicate](#not-obviously-this-person-predicate) * [Problems with logical information](#problems-with-logical-information) + [Empirical bits can be logical bits](#empirical-bits-can-be-logical-bits) + [Strong dependence on a theory of logical uncertainty](#strong-dependence-on-a-theory-of-logical-uncertainty) + [Dependence on irrelevant knowledge](#dependence-on-irrelevant-knowledge) + [Pointers to things are hard to recognize](#pointers-to-things-are-hard-to-recognize) + [Subtraces can sneak in information](#subtraces-can-sneak-in-information) + [Problems with the non-person predicate](#problems-with-the-non-person-predicate) + [Values over counterfactual behavior](#values-over-counterfactual-behavior) + [Identifying things, not human things](#identifying-things-not-human-things) Executive summary ----------------- If we have a good predictor under logical uncertainty P, we can ask: how does P’s predictions about the output of a computation Y change if it learns the outcome of X? We can then define various notions of how informative X is about Y. Possible uses: * Throttling modeling capabilities: preventing an agent from gaining too much logical information about a computation Y might translate into safety guarantees in the form of upper bounds on how well the agent can model Y. * Ontology identification: finding parts of different hypotheses that are mutually informative about each other could point towards the ontology used by arbitrary hypotheses, which could be useful for attaching values to parts of hypotheses, dealing with ontological crises, and finding which parts of a world-model influenced a decision. * Non-person predicate: disallowing computations that are very informative about a human might prevent some mindcrime. Problems: * Empirical information can include logical information, so a good theory will have to consider “total information”. * Any interesting properties of logical information rely heavily on a good theory of logical uncertainty. * If the predictor knows too much, it can’t see dependencies, since it has no uncertainty to become more informed about. * The predictor needs more time to integrate larger knowledge bases and larger pointers to parts of hypotheses. But given more time, the predictor becomes less and less uncertain about the computation of interest, without necessarily having learned anything relevant. * The non-person predicate only blacklists computations that are clearly informative about a particular computation, so it doesn’t at all prevent mindcrime in general. * Counterlogical behavior is probably relevant to our values, and isn’t captured here. * At best this scheme identifies abstract structures in hypotheses, rather than things that humans care about. Logical information ------------------- TL;DR: it might be enlightening to look at how a good predictor, under logical uncertainty, would change its distribution after adding the result of a computation to its knowledge base. ### Logical uncertainty Say that a predictor is an algorithm that assigns probabilities to outcomes of computations. For conceptual convenience, assume all computations are labeled by a running time and a finite specification of what their outputs could possibly be (so we don’t have to worry that they won’t halt, and so we know what space to assign probabilities over). Predictors have access to a knowledge base K, which is a list of outputs of computations, and we write P(X\|K) for the probability distribution the predictor puts on the outputs of X given knowledge base K. Assume we have a predictor P which is “good” in some sense to be determined by how we will use P. ### Logical information We can then talk about how informative a given computation X is about another computation Y (relative to P). Say X,Y∉K, X in fact evaluates to x, and Y has possible outputs yi. Then we can compare the distributions P(Y\|K) and P(Y\|K∪{X=x}) over the yi; how much they change gives a measure of how much P would learn about Y if it learned that X evaluates to x. For example, we can compare the entropy of these distributions to get an analog of information, or we could look at KL divergence, total variation, the probability of the true outcome of Y, or something else. We might also look at the expected change in one of these quantities between P(Y\|K) to P(Y\|K∪{X=xi}), taken over the xi according to P(X\|K), to get a notion of “how much one could reasonably expect to learn about Y from X”. Vaguely, we want P to be such that if there is a reliable way of predicting Y faster or more accurately using the result of X, then P will take advantage of that method. Also P should be able to do this across a wide variety of knowledge bases K. ### Example: XOR Let X and Y be two “unrelated”, difficult computations, each of which outputs a 0 or a 1. Let Z:=X⊕Y (the code of X and Y appear verbatim in the code of Z). Say P(X=1\|K)=P(Y=1\|K)=0.5, and they are independent (i.e. P((X,Y)\|K)=P(X\|K)P(Y\|K)), so that hopefully P(Z\|K)=0.5. (This is assuming something vaguely like P has twice the time to predict something with twice the code; or we could explicitly add time indices to everything.) We should have that P(X\|K∪{Y=1})=0.5, given that X and Y are unrelated. Even knowing Y, P still has the same relevant information and the same computational resources to guess at X. On the other hand, consider P(Z\|K∪{Y=1}). A good predictor asked to predict Z should, given the output of Y, devote more resources to guessing at X. This should result in a more confident prediction in the right direction; e.g., assuming in fact X=1, we might get P(Z\|K∪{Y=1})=0.8. Note that this is different from ordinary information, where the random variable Y would have no mutual information with Z, and X would have an entire bit of mutual information with Z, conditional on Y. This difference is a feature: in the logical uncertainty setting, Y really is informative about Z. ### Collapsing levels of indirection and obfuscation It’s often difficult to understand where in a system the important work is being done, because important dynamics could be happening on different levels or in non-obvious ways. E.g., one can imagine a search algorithm implemented so that certain innocuous data is passed around from function to function, each applying some small transformation, so that the end result is some output-relevant computation that was apparently unrelated to the search algorithm. Taking a “logical information” perspective could let us harness a strong theory of logical uncertainty to say when a computation is doing the work of another computation, regardless of any relatively simple indirection. Possible uses of logical information ------------------------------------ ### Throttling changes in logical uncertainty One might hope that preventing P(X\|K) from being low entropy, would correspond to any sufficiently bounded agent being unable to model X well. Then we could let K grow by adding the results of computations that are useful but irrelevant to X, and end up doing something useful, while maintaining some safety guarantee about the agent not modeling X. I don’t think this works; see [this post](https://agentfoundations.org/item?id=652). ### Ontology identification TL;DR: maybe we can look at chunks of computation traces of different hypotheses, and ask which ones are logically informative about other ones, and thereby identify “parts of different hypotheses that are doing the same work”. (NB: Attacking the problem directly like this—trying to write down what a part is—seems doomed.) #### Motivation: looking inside hypotheses One major problem with induction schemes like Solomonoff induction is that their hypotheses are completely opaque: they take observations as inputs, then do…something…and then output predictions. If an agent does its planning using a similar opaque induction scheme, this prevents the agent from explicitly attaching values to things going on internally to the hypotheses, such as conscious beings; an agent like AIξ only gets to look at the reward signals output by the black-box hypotheses. Even a successful implementation of AIξ (that didn’t drop an anvil on its head or [become a committed reductionist](https://ordinaryideas.wordpress.com/2011/12/14/aixi-and-existential-despair/)) would have a strong incentive to seize control of its input channel and choose its favorite inputs, rather than to steer the world into any particular state, other than to protect the input channel. Technically one could define a utility function that incentivizes AIξ to act exactly in the way required to bring about certain outcomes in the external world; but this is just wrapping almost all of the difficulty into defining the utility function, which would have to do induction on transparent hypotheses, recognize valuable things, do decision theory, and translate action evaluations into a reward function (HT Soares). The right way to deal with this may be to build a reasoning system to be understandable from the ground up, where each piece has, in some appropriate sense, a clear function with a comprehensible meaning in the context of the rest of the system; or to bypass the problem entirely with high-level learning of human-like behaviors, goals, world models, etc. (e.g. [apprenticeship learning](https://medium.com/ai-control/elaborations-of-apprenticeship-learning-eb93a53ae3ca#.ib19tgosh)). But we may want to start with a reasoning system designed with other desiderata in mind, for example to satisfy guarantees like limited optimization power or robustness under changes. Then later if we want to be able to identify valuable referents of hypotheses, it would be nice to understand the “parts” of different hypotheses. In particular, it would be nice to know when two parts of two different hypotheses “do the same thing”. #### Rice’s theorem: mostly false (HT Eisenstat) Rice’s theorem—you can’t determine any extensional input-output behavior of programs in general—stops us from saying in full generality when two programs do the same thing (in terms of I/O). But, there may be important senses in which this is mostly false: it may be possible to mostly discern the behavior of most programs. This is one way to view the logical uncertainty program. #### Parts of hypotheses I suggest that two “parts” of two hypotheses are “doing the same work” when they are logically informative about each other. For example, a (spacetime chunk containing a) chair in a physical model of the world, is informative about an abstract wire-frame simulation of the same chair, at least for certain questions such as “what happens if I tip the chair by τ/6 radians?”. This should hold even if both hypotheses are fully specified and deterministic, as long as the predictor is not powerful enough to just simulate each one directly rather than guessing. Unfortunately, I’m pretty sure that there’s no good way to, in general, decompose programs neatly so that different parts do separate things, and so that anything the program is modeling, is modeled in some particular part. Now I’ll sketch one notion of part that I think doesn’t work, in case it helps anyone think of something better. Say that all hypotheses are represented as cellular automaton computations, e.g. as Turing machines. Consider the computation trace of a given hypothesis X. Then define a part of X to be a small, cheaply computable subtrace of the trace of X, where by subtrace I just mean a subset of the bits in the trace. Cheaply computable is meant to capture the notion of something that is easily recognizable in the world X while you are looking at it. This can be viewed as a small collection of observed nodes in the causal graph corresponding to the cellular automaton implementing X. Then we say that a part X′ of a hypothesis X is informative about a computation Y′, to the extent that P(Y′\|K∪{X′=¯x} changes from P(Y′\|K). If Y′ is also a subtrace of some hypothesis Y, this gives a notion of how much two parts of two hypotheses are doing the same work. #### Potential uses for identifying parts Finding valuable parts of hypotheses. Say we have some computation X that implements something relevant to computing values of possible futures. This X doesn’t necessarily have to be intrinsically valuable by itself, like a simulation of a happy human, since it could be something like a strong but non-conscious Go program, and humans may have values about what sort of thing their opponent is. (It’s not actually obvious to me that I have any non-extensional values over anything that isn’t on the “inside” of a conscious mind, i.e. any non-conscious Go player with the same input-output behavior across counterfactuals is just as sweet.) In any case, we might be able to identify instances of X in a given model by searching for parts of the model that are informative about X and vice versa. Finding such computations X seems hard. Ontological crises. Say we have an AI system with a running model of the world, and value bindings into parts of that model. Then some considerations such as new data arise, and the AI system switches its main model to an entirely new sort of model. For example, the agent might switch from thinking the world runs on particles and classical mechanics, to thinking the world runs on quantum mechanics. See [“Ontological crises in artificial agents’ value systems”](http://arxiv.org/pdf/1105.3821v1.pdf). Then we want the agent to attach values to parts of the new model that correspond to values attached to parts of the old model. This might be accomplished by matching up parts of models as gestured at above. E.g. a happy human embedded in classical physics or a happy human embedded in quantum mechanics may be “made of different stuff”, but still do very similar computations and thereby be logically informative about each other. Plan-pivotal parts of a hypothesis class. Even more speculatively, it could be possible to identify “pivotal parts” of hypotheses that lead to a decision or a belief. That is, if we want to understand why an AI made a decision or came to a conclusion, it could help to look at a single class of corresponding parts across many hypotheses, and see how much the predicted behavior of those parts “influenced” the decision, possibly again using logical informativeness. ### Not-obviously-this-person predicate We might hope to prevent an agent from running computations that are conscious and possibly suffering, by installing an overseer to veto any computations that are logically informative about a reference human simulation H. This works just as in ontology identification, but it is distinct because here we are trying to avoid moral patients actually occurring inside the agent’s hypotheses, rather than trying to locate references to moral patients. Problems with logical information --------------------------------- ### Empirical bits can be logical bits If the environment contains computers running computations that are informative about Y, then empirical observations can be relevant to predicting Y. So a good theory of “total information” should call on a general predictor rather than just a logical predictor. These might be the same thing, e.g. if observations are phrased as logical statements about the state of the agent’s sensors. ### Strong dependence on a theory of logical uncertainty Many computational questions are entangled with many other computational questions [citation needed], so using logical information to understand the structure of computations depends on a good quantitative theory of logical uncertainty. Then we can speak meaningfully of “how much” one computation is informative about another. Indeed, this notion of logical information may depend in undesirable ways on the free parameter of a “good” predictor, including the choices of computational resources available to the predictor. This is akin to the free choice of a UTM for Solomonoff induction; different predictor predictors may use, in different ways, the results of X to guess the results of Y, and so could make different judgments about how informative X is about Y. (As in the case of Solomonoff induction, there is some hope that this would wash out in any cases big enough to be of interest.) For example, if the predictor is clever enough or computationally powerful enough to predict Y, it will think nothing is at all informative about Y, because all the conditional distributions will just be the same point distribution on the actual outcome of Y. This may not capture what we cared about. For example, if Y implements versions of a conscious computation X, we want to detect this; but the predictor tells us nothing about which things Y implements. More abstractly, logical informativeness formulated in terms of some predictor is relying on ignorance to detect logical dependencies; this is not necessarily a problem, but seems to demand a canonical idea of prediction under logical uncertainty. ### Dependence on irrelevant knowledge We want to get notions of how informative X is about Y, given a knowledge base K. In particular, P needs to have enough time to consider what is even in K. But this means that we need to allow P more time to think about Y as K gets larger, even if X and Y are fixed. This is problematic because at some point P will simply be able to compute Y, and therefore be useless for saying whether X is informative about Y. (There may be some clever way to preprocess K to avoid this problem, but it does seem like a serious problem.) ### Pointers to things are hard to recognize Say we want to find all the chairs in some large hypothesis X. We consider some subtrace X′, which does in fact point to a chair in X. But, this pointer may be long (as a program), and so it might take a lot of resources for the predictor to think about X′ enough to recognize that X′ points to a chair. If the chair is another fixed computation, then for most implementations X′ of chairs in other hypotheses, it is easier for P to just compute any question about the chair, than to recognize X′ as a chair and use its behavior as information. This seems like a serious problem. I don’t think P would have a similar problem identifying parts of hypotheses that are informative or not in a way that depends on the results of difficult computations happening inside a hypothesis. For example, say X runs a very complicated computation, and then depending on its output, either halts or implements a chair. We might worry that P can’t tell if the second stage is useful or not for modeling a chair, because it can’t compute the complicated thing. But P can still tell that this is the setup, so if it is given the results and they start to look like a chair is being modeled, P will know that it can use the rest of the bits as information. Also note that locating parts of large hypotheses requires a search over vast numbers of possible parts, which seems extremely expensive. ### Subtraces can sneak in information The programs serving as pointers to subtraces of X could just think about Y, and then find bits in the trace of X to encode information about Y, mostly ignoring whatever X is doing. This is supposed to be avoided by simply adding the value of certain bits of X’s trace to K (rather than adding the result of the computation that looks at the trace). But this assumes that the predictor somehow updates on its observations of logical facts without updating on the fact that it was shown those facts, which is probably not even a desirable property of a good predictor. ### Problems with the non-person predicate Recall the proposal to prevent mindcrime by vetoing computations that are informative about a human. Agent: Okay to run computation X? Overseer: Hold on, let me make sure it is safe. Overseer: \commits a whole lot of mindcrime\ Overseer: Um, yep, that is definitely mindcrime, no you may not run X. Agent: Whew, glad I checked. Overseer: \grimaces inwardly\ Even if the predictor can successfully detect potential mindcrime, it may itself commit mindcrime, especially while thinking about computations that include mindcrime. This might be partially sidestepped by not computing the answer to X and just adding possible outputs of X to K, but the resulting counterfactuals might not make sense. More fundamentally, the overseer blacklists computations that are definitely bad because they are informative about H, rather than whitelisting computations that are definitely safe. There may be many situations where an agent commits mindcrime without modeling any existing human, or commit mindcrime on very nonhuman but nevertheless morally valuable minds. ### Values over counterfactual behavior We may have values that depend on the counterfactual behavior of certain parts of the environment (“If I hadn’t tried so hard to beat the video game, the win screen wouldn’t be displayed.”). In this context, we might try saying that a part of X implements Y if adding any possible results of that subtrace to K would make P more informed about Y. But there’s no reason to expect that this sort of counterlogical will be reasonable. For example, instead of updating Y on counterfactual results of X, the predictor might just stop thinking that X has anything to do with Y (because X is apparently doing some weird thing). ### Identifying things, not human things At best, this ontology identification scheme still has nothing to do with locating things that humans value. That would require something like writing down computations that exemplify different kinds of value-relevant parts of the world, which seems like most of the problem.
835c5fd6-b532-45dc-9d13-c11a94b7c9c7	StampyAI/alignment-research-dataset/arbital	Arbital	Interest in mathematical foundations in Bayesianism [https://arbital.com/p/multiple-choice](https://arbital.com/p/multiple-choice)

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.