id stringlengths 36 36 | source stringclasses 15 values | formatted_source stringclasses 13 values | text stringlengths 2 7.55M |
|---|---|---|---|
f8c26118-d312-4df8-b5d3-5f7d07064a10 | trentmkelly/LessWrong-43k | LessWrong | Group rationality diary, 6/11/12
This is the public group instrumental rationality diary for the week of June 11th. It's a place to record and chat about it if you have done, or are actively doing, things like:
* Established a useful new habit
* Obtained new evidence that made you change your mind about some belief
* Decided to behave in a different way in some set of situations
* Optimized some part of a common routine or cached behavior
* Consciously changed your emotions or affect with respect to something
* Consciously pursued new valuable information about something that could make a big difference in your life
* Learned something new about your beliefs, behavior, or life that surprised you
* Tried doing any of the above and failed
Or anything else interesting which you want to share, so that other people can think about it, and perhaps be inspired to take action themselves. Try to include enough details so that everyone can use each other's experiences to learn about what tends to work out, and what doesn't tend to work out.
Thanks to everyone who contributes!
(Previously: 5/14/12, 5/21/12, 5/28/12, 6/4/12)
|
a42f8d70-2492-4295-be23-88ce438a14a3 | trentmkelly/LessWrong-43k | LessWrong | SRS advice
I have made some significant progress in organizing myself with org-mode (basically a really well thought out emacs outliner) - consider this a plug :).
Now I think I am ready to bite the bullet and automate another part of my mental apparatus, memorization. I'd like to hear other people's experiences with SRS - spaced repetition - (negative, too), what software they use, what do they use it for, how much time they spend. I expect these to vary, so stating your reasons is worth an extra upvote (and thanks ahead)
ETA: When do you decide something is worth memorizing vs. putting it into a searchable database? |
bf511acc-97c7-432a-b9f8-a4eff5f84a8b | trentmkelly/LessWrong-43k | LessWrong | Restraint Bias
Ed Yong over at Not Exactly Rocket Science has an article on a study demonstrating "restraint bias" (reference), which seems like an important thing to be aware of in fighting akrasia:
People who think they are more restrained are more likely to succumb to temptation
> In a series of four experiments, Loran Nordgren from Northwestern University showed that people suffer from a "restraint bias", where they overestimate their ability to control their own impulses. Those who fall prey to this fallacy most strongly are more likely to dive into tempting situations. Smokers, for example, who are trying to quit, are more likely to put themselves in situations if they think they're invulnerable to temptation. As a result, they're more likely to relapse.
Thus, not only do people overestimate their abilities to carry out non-immediate plans (far-mode thinking, like in planning fallacy), but also the more confident ones turn out to be least able. This might have something to do with how public commitment may be counterproductive: once you've effectively signaled your intentions, the pressure to actually implement them fades away. Once you believe yourself to have asserted self-image of a person with good self-control, maintaining the actual self-control loses priority.
See also: Akrasia, Planning fallacy, Near/far thinking.
Related to: Image vs. Impact: Can public commitment be counterproductive for achievement? |
c07590e4-3b3f-4957-b0fb-e47a7596468f | trentmkelly/LessWrong-43k | LessWrong | AI Safety via Luck
Epistemic Status: I feel confident and tentatively optimistic about the claims made in this post, but am slightly more uncertain about how it generalizes. Additionally, I am concerned about the extent to which this is dual-use for capabilities and exfohazardous and spent a few months thinking about whether it was worth it to release this post regardless. I haven’t come to an answer yet, so I’m publishing this to let other people see it and know what they think I should do.
TL;DR: I propose a research direction to solve alignment that potentially doesn’t require solutions to ontology identification, learning how to code, or becoming literate.
Introduction
Until a few hours ago, I was spending my time primarily working on high-level interpretability and cyborgism. While I was writing a draft for something I was working on, an activity that usually yields me a lot of free time by way of procrastination, I stumbled across the central idea behind many of the ideas in this post. It seemed so immediately compelling that I dropped working on everything else to start working on it, culminating after much deliberation in the post you see before you.
My intention with this post is to provide a definitive reference for what it would take to safely use AGI to steer our world toward much better states in the absence of a solution to any or all of several existing problems, such as Eliciting Latent Knowledge, conditioning simulator models, Natural Abstractions, mechanistic interpretability, and the like.
In a world with prospects such as those, I propose that we radically rethink our approach to AGI safety. Instead of dedicating enormous effort to engineering nigh-impossible safety measures, we should consider thus-far neglected avenues of research, especially ones that have memetic reasons to be unfairly disprivileged so far and which immunizes them against capabilities misuse. To avert the impending AI apocalypse, we need to focus on high-variance, low-probability-high-yiel |
5f5a55b8-7d47-4535-b7ee-cb727c031a1c | trentmkelly/LessWrong-43k | LessWrong | Transhumanism thread in progress at Reddit
Starting with this reply to "You were born too soon":
> depending on when exactly we achieve this, this could be the best time to be born ever, because it will be the absolute earliest anybody will have achieved immortality. Someone born within 20 years of this moment could one day be the oldest human, sentient, or even living being in the Universe.
The comments are currently split between arguing and agreeing with this. So far, no mention of cryonics. One post presents a possibly interesting technical argument that our current knowledge/technology is centuries away from mind uploading/whole-brain emulation.
(Also posted to The Singularity in the Zeitgeist, but that thread seems to have been mostly forgotten.) |
d74e55c1-9c25-480c-892e-0591e1de8fcd | trentmkelly/LessWrong-43k | LessWrong | How hard is it for altruists to discuss going against bad equilibria?
Epistemic status: This post is flagrantly obscure, which makes it all the harder for me to revise it to reflect my current opinions. By the nature of the subject, it's difficult to give object-level examples. If you're considering reading this, I would suggest the belief signaling trilemma as a much more approachable post on a similar topic. Basically, take that idea, and extrapolate it to issues with coordination problems?
* There are many situations where a system is "broken" in the sense that incentives push people toward bad behavior, but, not so much that an altruist has any business engaging in that bad behavior (at least, not if they are well-informed).
* In other words, an altruist who understands the bad equilibrium well would disengage from the broken system, or engage while happily paying the cost of going against incentives.
* Clearly, this is not always the case; I'm thinking about situations where it is the case.
* Actually, I'm thinking about situations where it is the case supposing that we ignore certain costs, such as costs of going against peer pressure, costs of employing willpower to go against the default, etc. The question is then: is it realistically worth it, given all those additional costs, if we condition on it being it's worth it for an imaginary emotional-robot altruist?
* Actually actually, the question I'm asking is probably not that one either, but I haven't figured out my real question yet.
* I think maybe I'm mainly interested in the question of how hard it is for altruists to publicly discuss altruistic strategies (in the context of a bad equilibrium) without upsetting a bunch of people (who are currently coordinating on that equilibrium, and are therefore protective of it).
* I'm writing this post to try to sort out some confused thoughts (hence the weird style). A lot of the context is discussion on this post.
* But, I'm not going to discuss examples in my post. This seems like a |
2f0f48aa-ca17-4fef-8f36-2bedc8e20f29 | trentmkelly/LessWrong-43k | LessWrong | Weekly LW Meetups
This summary was posted to LW Main on October 24th. The following week's summary is here.
New meetups (or meetups with a hiatus of more than a year) are happening in:
* Bangalore Meetup: 15 November 2014 04:15PM
Irregularly scheduled Less Wrong meetups are taking place in:
* East Coast Solstice Megameetup: 20 December 2014 03:00PM
* European Community Weekend 2015: 12 June 2015 12:00PM
* Perth, Australia: Discussion: How to be happy: 04 November 2014 06:00PM
* Saint Petersburg meetup - "the lonely one": 31 October 2014 08:00PM
* Urbana-Champaign: Meta-systems and getting things done: 26 October 2014 02:00PM
* Utrecht: Climate Change: 02 November 2014 03:00PM
The remaining meetups take place in cities with regular scheduling, but involve a change in time or location, special meeting content, or simply a helpful reminder about the meetup:
* Austin, TX - Spider House: 25 October 2025 01:30PM
* [Cambridge MA] The Design Process: 29 October 2014 07:00PM
* Canberra: Would I Lie To You?: 24 October 2014 06:00PM
* London Social - October 26th: 26 October 2014 03:00PM
* Moscow meetup: Quantum physics is fun: 26 October 2014 03:00PM
* Washington, D.C.: Create and Complete: 26 October 2014 03:00PM
Locations with regularly scheduled meetups: Austin, Berkeley, Berlin, Boston, Brussels, Buffalo, Cambridge UK, Canberra, Columbus, London, Madison WI, Melbourne, Moscow, Mountain View, New York, Philadelphia, Research Triangle NC, Seattle, Sydney, Toronto, Vienna, Washington DC, Waterloo, and West Los Angeles. There's also a 24/7 online study hall for coworking LWers.
If you'd like to talk with other LW-ers face to face, and there is no meetup in your area, consider starting your own meetup; it's easy (more resources here). Check one out, stretch your rationality skills, build community, and have fun!
In addition to the handy sidebar of upcoming meetups, a meetup overview is posted on the front page every Friday. These are an attempt to collect information on al |
0619f8d8-4058-42ea-b686-e07cd2b90b4d | trentmkelly/LessWrong-43k | LessWrong | Chapter 4: The Efficient Market Hypothesis
Disclaimer: J. K. Rowling is watching you from where she waits, eternally in the void between worlds.
A/N: As others have noted, the novels seem inconsistent in the apparent purchasing power of a Galleon; I'm picking a consistent value and sticking with it. Five pounds sterling to the Galleon doesn't square with seven Galleons for a wand and children using hand-me-down wands.
----------------------------------------
"World domination is such an ugly phrase. I prefer to call it world optimisation."
----------------------------------------
Heaps of gold Galleons. Stacks of silver Sickles. Piles of bronze Knuts.
Harry stood there, and stared with his mouth open at the family vault. He had so many questions he didn't know where to start.
From just outside the door of the vault, Professor McGonagall watched him, seeming to lean casually against the wall, but her eyes intent. Well, that made sense. Being plopped in front of a giant heap of gold coins was a test of character so pure it was archetypal.
"Are these coins the pure metal?" Harry said finally.
"What?" hissed the goblin Griphook, who was waiting near the door. "Are you questioning the integrity of Gringotts, Mr. Potter-Evans-Verres?"
"No," said Harry absently, "not at all, sorry if that came out wrong, sir. I just have no idea at all how your financial system works. I'm asking if Galleons in general are made of pure gold."
"Of course," said Griphook.
"And can anyone coin them, or are they issued by a monopoly that thereby collects seigniorage?"
"What?" said Professor McGonagall.
Griphook grinned, showing sharp teeth. "Only a fool would trust any but goblin coin!"
"In other words," Harry said, "the coins aren't supposed to be worth any more than the metal making them up?"
Griphook stared at Harry. Professor McGonagall looked bemused.
"I mean, suppose I came in here with a ton of silver. Could I get a ton of Sickles made from it?"
"For a fee, Mr. Potter-Evans-Verres." The goblin watched him with g |
9dea5261-4475-4c2a-b258-47a03dbb4e7b | trentmkelly/LessWrong-43k | LessWrong | Virtual models of virtual AIs in virtual worlds
A putative new idea for AI control; index here.
This is a reworking of the previous models of AIs in virtual worlds, done after discussing the idea with Jessica.
The basic idea is that there are a lot of concepts about AI security - such as sealed boxes, Cartesian boundaries, and so on - that are only partially true. They are simplified models, abstractions, and, even if they are true in our model of physics (an absurdly optimistic claim) there's always the possibility (certainty) that our models are wrong or incomplete.
But there is one place where models are exactly true, and that... within the model itself. And a computer program, or a virtual machine, is a model, a model we try and make happen exactly in the real world.
----------------------------------------
Is my world model true?
Consider an AI, running inside a virtual machine. The AI has an internal model of the entire virtual machine, including itself, and knows the rules of the virtual machine/model.
In fact the AI entertains two hypotheses:
* H1: the model is entirely correct about the (virtual) world.
* H2: sometimes, random bits are flipped in the (virtual) world.
Then give the AI some utility function u, between 0 and 1, defined over the virtual world. The AI also has a shutdown button B that it predicts will shut down the world; this B send a signal to us and will stop the execution of the virtual world, AI included. Let r be some discount factor close to 1, and TB the time step at which B was pressed (if it was). Then the AI tries to maximise the utility
* u∗=rTB((u|H1)+(2|H2))
Now the AI is motivated to maximise u as long as it thinks that the world model it has is correct, and to press B as soon as it feels its ache. But as soon as it suspects enough that the model of the virtual world is is incorrect - ie H2 - it will press B (the exact definition of u∗ could probably be improved; the might be unfortunate trade-offs at the moment).
What does H2 translate to from our perspective? |
53627298-0844-4413-9384-9a6a9f1df825 | trentmkelly/LessWrong-43k | LessWrong | London Meetup on 2011/1/2
On Sunday, January 2nd 2011 there will be a meetup the London area. As with previous meetups, the venue is Shakespeare's Head. The meeting will start at 14:00.
In order to keep us organised for 2011, I'm putting together a mailing list for LWers around the London area. If you'd like to be added to the list, please send me your e-mail address via private message. |
2eaf633b-4c38-4353-bec7-a871baad00a4 | trentmkelly/LessWrong-43k | LessWrong | Evaluability (And Cheap Holiday Shopping)
With the expensive part of the Hallowthankmas season now approaching, a question must be looming large in our readers’ minds:
> “Dear Overcoming Bias, are there biases I can exploit to be seen as generous without actually spending lots of money?”
I’m glad to report the answer is yes! According to Hsee—in a paper entitled “Less is Better”—if you buy someone a $45 scarf, you are more likely to be seen as generous than if you buy them a $55 coat.1
This is a special case of a more general phenomenon. In an earlier experiment, Hsee asked subjects how much they would be willing to pay for a second-hand music dictionary:2
* Dictionary A, from 1993, with 10,000 entries, in like-new condition.
* Dictionary B, from 1993, with 20,000 entries, with a torn cover and otherwise in like-new condition.
The gotcha was that some subjects saw both dictionaries side-by-side, while other subjects only saw one dictionary . . .
Subjects who saw only one of these options were willing to pay an average of $24 for Dictionary A and an average of $20 for Dictionary B. Subjects who saw both options, side-by-side, were willing to pay $27 for Dictionary B and $19 for Dictionary A.
Of course, the number of entries in a dictionary is more important than whether it has a torn cover, at least if you ever plan on using it for anything. But if you’re only presented with a single dictionary, and it has 20,000 entries, the number 20,000 doesn’t mean very much. Is it a little? A lot? Who knows? It’s non-evaluable. The torn cover, on the other hand—that stands out. That has a definite affective valence: namely, bad.
Seen side-by-side, though, the number of entries goes from non-evaluable to evaluable, because there are two compatible quantities to be compared. And once the number of entries becomes evaluable, that facet swamps the importance of the torn cover.
From Slovic et al.: Which would you prefer?3
1. A 29/36 chance to win $2.
2. A 7/36 chance to win $9.
While the average prices (equiv |
55366c33-a60a-4c32-ad21-a3ee09de4620 | trentmkelly/LessWrong-43k | LessWrong | Do you trust the research on handwriting vs. typing for notes?
For a while now, I've heard the claim that "studies show" that note-takers remember content better if they take notes by hand versus on a computer. I previously took this claim on face value in part because this was before I'd heard about the replication crisis and also because I'd had personal experiences that I believed supported this claim.
In light of the replication crisis and recent experiences, I've come to be more skeptical of this research. I started to look at some of the research pop science articles on the topic cite and am skeptical of the work I've looked at so far.
In the example I link to above, they have subjects perform two tasks, a recall and recognition task for words they handwrote or typed with a multiplication task in between writing and recalling/recognizing. They find a not significant difference between recalled words for handwriting vs. typing and a barely (p-value .03) significant difference between recognized words for the two groups. However, if you look at the standard deviations for the means for the two tasks, you'll see that each mean is in the other's 1-SD range.
Furthermore, the task they describe is simple but not necessarily that relevant to what's really going on when someone takes notes on a lecture / talk. They intentionally used semantically meaningless words (for understandable reasons) whereas real-life talks hopefully have higher-level meaning and themes.
ETA (after initial posting): Just found another paper that a few pop-sci articles seem to cite. This paper covers three experiments, which are all more realistic than the one I described above. I'm only going to discuss the first here.
The first had participants watch TED talks, take notes on them (either on a laptop or by hand) and then answer a combination of "factual" and "conceptual" questions about them. At a high level, they interpret the results of this experiment as showing that laptop note-takers did as well as the by hand note-takers on factual questions b |
5127b5fb-4db9-42b5-94a5-87747e605b27 | trentmkelly/LessWrong-43k | LessWrong | Why is COVID reinfection rate still so uncertain.
From recent observatory studies, SIREN in UK , and in Denmark, they both estimate seropositivity to give around 80% of protection agains infection. The former observed 90% protection against symptomatic cases.
How efficacious are vaccines on the seropositive population? I've only seen reports on safety analysis.
To the extend of suspecting malice, the Israel Pfizer study did not report on prior infection analysis - Study protocol states "Exclusion of patients with COVID-19 prior to the index date or matched index date." |
8d218f0d-84bc-4975-a94b-8f08519a90a6 | trentmkelly/LessWrong-43k | LessWrong | Dealing with trolling and the signal to noise ratio
The recent implementation of a -5 karma penalty for replying to comments that are at -3 or below has clearly met with some disagreement and controversy. See http://lesswrong.com/r/discussion/lw/eb9/meta_karma_for_last_30_days/7aon . However, at the same time, it seems that Eliezer's observation that trolling and related problems have over time gotten worse here may be correct. It may be that this an inevitable consequence of growth, but it may be that it can be handled or reduced with some solution or set of solutions. I'm starting this discussion thread for people to propose possible solutions. To minimize anchoring bias and related problems, I'm not going to include my ideas in this header but in a comment below. People should think about the problem before reading proposed solutions (again to minimize anchoring issues). |
beb4e0ca-f37b-4f19-9cd9-873b18c31272 | trentmkelly/LessWrong-43k | LessWrong | Surviving and Shaping Long-Term Competitions: Lessons from Net Assessment
This post examines net assessment, a framework for evaluating strategic competition that evolved to inform U.S. defense policy during the Cold War. We explain what net assessment is, its methods and principles, and how some of its tools can be applied to reason about highly uncertain, long-term tech competitions with potentially existential stakes.
What is net assessment?
In the late 1950s, the Cold War slid into an especially dangerous period. Nuclear stockpiles swelled and delivery capabilities advanced, while a decades-long buildup by the Soviet Union challenged the United States’ conventional military dominance. Defense analysts needed to reframe the way they looked at military competition: the U.S. could not overpower the Soviet war machine with brute force, and the prospect of nuclear war both elevated the stakes of conflict and created a need for new metrics and principles of strategy for engaging in limited competition. It was in this context that the framework of “net assessment” began to develop.
Andrew Marshall, who founded and then directed the DoD’s Office of Net Assessment for forty-two years, described net assessment as follows:
> "Our notion of a net assessment is that it is a careful comparison of U.S. weapon systems, forces, and policies in relation to those of other countries. It is comprehensive, including description of the forces, operational doctrines and practices, training regime, logistics, known or conjectured effectiveness in various environments, design practices and their effect on equipment costs, performance, and procurement practices and their influence on cost and lead times. The use of net assessment is intended to be diagnostic. It will highlight efficiency and inefficiency in the way we and others do things, and areas of comparative advantage with respect to our rivals."
Generalizing from its original military context, the core idea of net assessment is to create comprehensive (hence the “net” in the name), objective, an |
898b64a9-04dc-4a38-8576-6288ba229350 | StampyAI/alignment-research-dataset/blogs | Blogs | Artificial Intelligence as a Positive and Negative Factor in Global Risk
Draft for [Global Catastrophic Risks, Oxford University Press, 2008](http://www.amazon.com/Global-Catastrophic-Risks-Martin-Rees/dp/0198570503/ref=pd_bbs_sr_1?ie=UTF8&s=books&qid=1224111364&sr=8-1) . [Download as PDF](https://intelligence.org/files/AIPosNegFactor.pdf) .
[AIPosNegFactor](https://eystaging.wpengine.com/wp-content/uploads/2020/09/AIPosNegFactor.pdf)
---
This document is ©2007 by [Eliezer Yudkowsky](http://eyudkowsky.wpengine.com/) and free under the [Creative Commons Attribution-No Derivative Works 3.0 License](http://creativecommons.org/licenses/by-nd/3.0/) for copying and distribution, so long as the work is attributed and the text is unaltered.
Eliezer Yudkowsky’s work is supported by the [Machine Intelligence Research Institute](https://intelligence.org/) .
If you think the world could use some more rationality, consider blogging this page.
Praise, condemnation, and feedback are [always welcome](https://eyudkowsky.wpengine.com/contact) . The web address of this page is [http://eyudkowsky.wpengine.com/singularity/ai-risk/](https://eyudkowsky.wpengine.com/singularity/ai-risk/) . |
e0b6d8c1-8738-4974-b546-aa73fe15b580 | trentmkelly/LessWrong-43k | LessWrong | What could Alphafold 4 look like?
I made another biology-ML podcast! Two hours long, deeply technical, links below.
I posted about others ones I did here (machine learning in molecular dynamics) and here (machine learning in vaccine design). This one is over machine learning in protein design, interviewing perhaps one of the most well-known people in the field. This is my own field, so the podcast is very in the weeds, but hopefully interesting to those deeply curious about biology!
Substack: https://www.owlposting.com/p/what-could-alphafold-4-look-like
Youtube: https://youtu.be/6_RFXNxy62c
Spotify: https://open.spotify.com/episode/0wPs3rmp0zrfauqToozrcv?si=DCtRf-xQTPiVYwslo-b2rQ
Apple Podcasts: https://podcasts.apple.com/us/podcast/what-could-alphafold-4-look-like-sergey-ovchinnikov-3/id1758545538?i=1000704927828
Transcript: https://www.owlposting.com/p/what-could-alphafold-4-look-like?open=false#%C2%A7transcript
Summary: To those in the protein design space, Dr. Sergey Ovchinnikov is a very, very well-recognized name.
A recent MIT professor (circa early 2024), he has played a part in a staggering number of recent innovations in the field: ColabFold, RFDiffusion, Bindcraft, automated design of soluble proxies of membrane proteins, elucidating what protein language models are learning, conformational sampling via Alphafold2, and many more. And even beyond the research that have come from his lab in the last few years, the co-evolution work he did during his PhD/fellowship also laid some of the groundwork for the original Alphafold paper, being cited twice in it.
As a result, Sergey’s work has gained a reputation for being something that is worth reading. But nobody has ever interviewed him before! Which was shocking for someone who was so pivotally important for the field. So, obviously, I wanted to be the first one to do it. After an initial call, I took a train down to Boston, booked a studio, and chatted with him for a few hours, asking every question I could think of. We talk about his own |
df31ab59-985b-4b01-adb0-328293feebd6 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | The theory-practice gap
[Thanks to Richard Ngo, Damon Binder, Summer Yue, Nate Thomas, Ajeya Cotra, Alex Turner, and other Redwood Research people for helpful comments; thanks Ruby Bloom for formatting this for the Alignment Forum for me.]
I'm going to draw a picture, piece by piece. I want to talk about the capability of some different AI systems.
You can see here that we've drawn the capability of the system we want to be [competitive](https://ai-alignment.com/directions-and-desiderata-for-ai-control-b60fca0da8f4) with, which I’ll call the [unaligned benchmark](https://ai-alignment.com/an-unaligned-benchmark-b49ad992940b). The unaligned benchmark is what you get if you train a system on the task that will cause the system to be most generally capable. And you have no idea how it's thinking about things, and you can only point this system at some goals and not others.
I think that the alignment problem looks different depending on how capable the system you’re trying to align is, and I think there are reasonable arguments for focusing on various different capabilities levels. See [here](https://docs.google.com/document/d/1kQeMXKxybDKziRyRTmHmp9nHw76T1JAUgB8fY4e34_o/edit) for more of my thoughts on this question.
Alignment strategies
====================
People have also proposed various alignment strategies. But I don’t think that these alignment strategies are competitive with the unaligned benchmark, even in theory.
I want to claim that most of the action in theoretical AI alignment is people proposing various ways of getting around these problems by having your systems do things that are human understandable instead of doing things that are justified by working well.
For example, the hope with [imitative IDA](https://www.alignmentforum.org/posts/fRsjBseRuvRhMPPE5/an-overview-of-11-proposals-for-building-safe-advanced-ai#2__Imitative_amplification___intermittent_oversight) is that through its recursive structure you can build a dataset of increasingly competent answers to questions, and then at every step you can train a system to imitate these increasingly good answers to questions, and you end up with a really powerful question-answerer that was only ever trained to imitate humans-with-access-to-aligned-systems, and so your system is outer aligned.
The bar I’ve added, which represents how capable I think you can get with amplified humans, is lower than the bar for the unaligned benchmark. I've drawn this bar lower because I think that if your system is trying to imitate cognition that can be broken down into human understandable parts, it is systematically not going to be able to pursue certain powerful strategies that the end-to-end trained systems will be able to. I think that there are probably a bunch of concepts that humans can’t understand quickly, or maybe can’t understand at all. And if your systems are restricted to never use these concepts, I think your systems are probably just going to be a bunch weaker.
I think that transparency techniques, as well as AI alignment strategies like [microscope AI](https://www.alignmentforum.org/posts/fRsjBseRuvRhMPPE5/an-overview-of-11-proposals-for-building-safe-advanced-ai#5__Microscope_AI) that lean heavily on them, rely on a similar assumption that the cognition of the system you’re trying to align is factorizable into human-understandable parts. One component of the best-case scenario for transparency techniques is that anytime your neural net does stuff, you can get the best possible human understandable explanation of why it's doing that thing. If such an explanation doesn’t exist, your transparency tools won’t be able to assure you that your system is aligned even if it is.
To summarize, I claim that current alignment proposals don’t really have a proposal for how to make systems that are aligned but either
* produce plans that can’t be understood by amplified humans
* do cognitive actions that can’t be understood by amplified humans
And so I claim that current alignment proposals don’t seem like they can control systems as powerful as the systems you’d get from an unaligned training strategy.
Empirical generalization
========================
I think some people are optimistic that alignment will generalize from the cases where amplified humans can evaluate it to the cases where the amplified humans can’t. I'm going to call this empirical generalization. I think that empirical generalization is an example of relying on empirical facts about neural nets that are not true of arbitrary general black box function approximators.
I think this is a big part of the reason why some people are optimistic about the strategy that Paul Christiano calls [“winging it”](https://aiimpacts.org/conversation-with-paul-christiano/).
(I think that one particularly strong argument for empirical generalization is that if you imagine AGI as something like GPT-17 fine-tuned on human feedback on various tasks, your AGI might think about things in a very human-shaped way. (Many people disagree with me on this.) It currently seems plausible to me that AGI will be trained with a bunch of unsupervised learning based on stuff humans have written, which maybe makes it more likely that your system will have this very human-shaped set of concepts.)
The theory-practice gap
=======================
So the total height of that second column is the maximum level of capabilities that we think we could theoretically attain using the same capability techniques that we used for the unaligned benchmark, but using the alignment strategies that we know about right now. But in practice, we probably aren't going to do as well as that, for a variety of practical reasons. For example, as I've said, I think transparency tools are theoretically limited, but we're just way below the maximum theoretically available capability of transparency tools right now.
So I want to claim that reality will probably intervene in various ways and mean that the maximum capability of an aligned AI that we can build is lower than the maximum achievable theoretically from the techniques we know about and empirical generalization. I want to call that difference the theory practice gap.
Sources of theory-practice gap
------------------------------
**Practical difficulties, eg getting human feedback**
Human feedback is annoying in a wide variety of ways; you have to do quality control etc.
**Problems with the structure of the recursion**
I think it's reasonably plausible that the most competitive way of making powerful systems ends up not really being shapeable into the shape you need for the amplified human stuff to work out. So for example, maybe the best way of making AGI is doing some kind of evolution simulation, where you have this population of little creatures and they compete with each other and stuff. And if that's the only way of making smart systems, then I think it's pretty plausible that there's just like no way of building a trusted, amplified reward signal out of it. And so you can't do the IDA style things, or things where you use a system to do transparency analysis on a slightly more powerful version of itself.
**NP-hard problems**
Maybe your amplified system won’t be able to answer questions like “are there any inputs on which this system does the wrong thing” even if it wants to. Eg the [RSA-2048 problem](https://ai-alignment.com/training-robust-corrigibility-ce0e0a3b9b4d#a291).
I think that transparency has a related problem: the most competitive-to-train models might have internal structure that amplified humans would be able to understand if it was explained to them, but we might not be able to get a model to find that structure.
---
Why am I lumping together fundamental concerns like “maybe these alignment strategies will require solving NP-hard problems” with things like “it’s annoying to do quality control on your labelling contractors”?
It’s primarily because I want to emphasize that these concerns are different from the fundamental limitations of currently proposed alignment schemes: even if you assume that we don’t e.g. run into the hard instances of the NP-hard problems, I think that the proposed alignment schemes still aren’t clearly good enough. There are lots of complicated arguments about the extent to which we have some of these “practical” problems; I think that these arguments distract from the claim that the theoretical alignment problem might be unsolved even if these problems are absent.
So my current view is that if you want to claim that we're going to fully solve the technical alignment problem as I described it above, you've got to believe some combination of:
* we're going to make substantial theoretical improvements
* factored cognition is true
* we're going to have really good empirical generalization
(In particular, your belief in these factors needs to add up to some constant. E.g., if you’re more bullish on factored cognition, you need less of the other two.)
I feel like there’s at least a solid chance that we’re in a pretty inconvenient world where none of these are true.
Classifying alignment work
==========================
This picture suggests a few different ways of trying to improve the situation.
* You could try to improve the best alignment techniques. I think this is what a lot of AI alignment theoretical work is. For example, I think Paul Christiano’s recent imitative generalization work is trying to increase the theoretically attainable capabilities of aligned systems. I’ve drawn this as the red arrow on the graph below.
* You can try to reduce the theory-practice gap. I think this is a pretty good description of what I think applied alignment research is usually trying to do. This is also what I’m currently working on. This is the pink arrow.
* You can try to improve our understanding of the relative height of all these bars.
AI alignment disagreements as variations on this picture
========================================================
So now that we have this picture, let's try to use it to explain some common disagreements about AI alignment.
I think some people think that amplified humans are actually just as capable as the unaligned benchmark. I think this is basically the factored cognition hypothesis.
I think there's a bunch of people who are really ML-flavored alignment people who seem to be pretty optimistic about empirical generalization. From their perspective, almost everything that AI alignment researchers should be doing is narrowing that theory practice gap, because that's the only problem.
I think there's also a bunch of people like perhaps the stereotypical MIRI employee who thinks that amplified humans aren't that powerful, and you're not going to get any empirical generalization, and there are a bunch of problems with the structure of the recursion for amplification procedures. And so it doesn't feel that important to them to work on the practical parts of the theory practice gap, because even if we totally succeeded at getting that to zero, the resulting systems wouldn't be very powerful or very aligned. And so it just wouldn't have mattered that much. And the stereotypical such person wants you to work on the red arrow instead of the pink arrow.
How useful is it to work on narrowing the theory-practice gap for alignment strategies that won’t solve the whole problem?
==========================================================================================================================
See [here](https://www.alignmentforum.org/posts/tmWMuY5HCSNXXZ9oq/buck-s-shortform?commentId=BznGTJ3rGHLMcdbEB).
Conclusion
==========
I feel pretty nervous about the state of the world described by this picture.
I'm really not sure whether I think that theoretical alignment researchers are going to be able to propose a scheme that gets around the core problems with the schemes they've currently proposed.
There's a pretty obvious argument for optimism here, which is that people haven't actually put in that many years into AI alignment theoretical research so far. And presumably they're going to do a lot more of it between now and AGI. I think I'm like 30% on the proposition that before AGI, we're going to come up with some alignment scheme that just looks really good and clearly solves most of the problems with current schemes.
I think I overall disagree with people like Joe Carlsmith and Rohin Shah mostly in two places:
* By the time we get to AGI, will we have alignment techniques that are even slightly competitive? I think it’s pretty plausible the answer is no. (Obviously it would be very helpful for me to operationalize things like “pretty plausible” and “slightly competitive” here.)
* If we don’t have the techniques to reliably align AI, will someone deploy AI anyway? I think it’s more likely the answer is yes. |
7537bd04-7cb2-4694-85fb-b2bd77e5f7b1 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | The Solomonoff Prior is Malign
This argument came to my attention from [this post](https://ordinaryideas.wordpress.com/2016/11/30/what-does-the-universal-prior-actually-look-like/) by Paul Christiano. I also found [this clarification](https://www.lesswrong.com/posts/jP3vRbtvDtBtgvkeb/clarifying-consequentialists-in-the-solomonoff-prior) helpful. I found [these counter-arguments](https://www.lesswrong.com/posts/Ecxevhvx85Y4eyFcu/weak-arguments-against-the-universal-prior-being-malign) stimulating and have included some discussion of them.
Very little of this content is original. My contributions consist of fleshing out arguments and constructing examples.
Thank you to Beth Barnes and Thomas Kwa for helpful discussion and comments.
What is the Solomonoff prior?
=============================
The Solomonoff prior is intended to answer the question "what is the probability of X?" for any X, where X is a finite string over some finite alphabet. The Solomonoff prior is defined by taking the set of all Turing machines (TMs) which output strings when run with no input and weighting them proportional to 2−K.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')}
, where K is the description length of the TM (informally its size in bits).
The Solomonoff prior says the probability of a string is the sum over all the weights of all TMs that print that string.
One reason to care about the Solomonoff prior is that we can use it to do a form of idealized induction. If you have seen 0101 and want to predict the next bit, you can use the Solomonoff prior to get the probability of 01010 and 01011. Normalizing gives you the chances of seeing 1 versus 0, conditioned on seeing 0101. In general, any process that assigns probabilities to all strings in a consistent way can be used to do induction in this way.
[This post](https://www.lesswrong.com/posts/Kyc5dFDzBg4WccrbK/an-intuitive-explanation-of-solomonoff-induction#All_Algorithms) provides more information about Solomonoff Induction.
Why is it malign?
=================
Imagine that you wrote a programming language called python^10 that works as follows: First, it takes all alpha-numeric chars that are not in literals and checks if they're repeated 10 times sequentially. If they're not, they get deleted. If they are, they get replaced by a single copy. Second, it runs this new program through a python interpreter.
Hello world in python^10:
`ppppppppprrrrrrrrrriiiiiiiiiinnnnnnnnnntttttttttt('Hello, world!')`
Luckily, python has an `exec` function that executes literals as code. This lets us write a shorter hello world:
`eeeeeeeeexxxxxxxxxxeeeeeeeeeecccccccccc("print('Hello, world!')")`
It's probably easy to see that for nearly every program, the shortest way to write it in python^10 is to write it in python and run it with `exec`. If we didn't have `exec`, for sufficiently complicated programs, the shortest way to write them would be to specify an interpreter for a different language in python^10 and write it in that language instead.
As this example shows, the answer to "what's the shortest program that does X?" might involve using some roundabout method (in this case we used `exec`). If python^10 has some security properties that python didn't have, then the shortest program in python^10 that accomplished any given task would not have these security properties because they would all pass through `exec`. In general, if you can access alternative ‘modes’ (in this case python), the shortest programs that output any given string might go through one of those modes, possibly introducing malign behavior.
Let's say that I'm trying to predict what a human types next using the Solomonoff prior. Many programs predict the human:
1. Simulate the human and their local surroundings. Run the simulation forward and check what gets typed.
2. Simulate the entire Earth. Run the simulation forward and check what that particular human types.
3. Simulate the entire universe from the beginning of time. Run the simulation forward and check what that particular human types.
4. Simulate an entirely different universe that has reason to simulate this universe. Output what the human types in the simulation of our universe.
Which one is the simplest? One property of the Solmonoff prior is that it doesn't care about how long the TMs take to run, only how large they are. This results in an unintuitive notion of "simplicity"; a program that does something 210 times might be simpler than a program that does the same thing 29−1 times because the number 210 is easier to specify than 29−1.
In our example, it seems likely that "simulate the entire universe" is simpler than "simulate Earth" or "simulate part of Earth" because the initial conditions of the universe are simpler than the initial conditions of Earth. There is some additional complexity in picking out the specific human you care about. Since the local simulation is built around that human this will be easier in the local simulation than the universe simulation. However, in aggregate, it seems possible that "simulate the universe, pick out the typing" is the shortest program that predicts what your human will do next. Even so, "pick out the typing" is likely to be a very complicated procedure, making your total complexity quite high.
Whether simulating a different universe that simulates our universe is simpler depends a lot on the properties of that other universe. If that other universe is simpler than our universe, then we might run into an `exec` situation, where it's simpler to run that other universe and specify the human in their simulation of our universe.
This is troubling because that other universe might contain beings with different values than our own. If it's true that simulating that universe is the simplest way to predict our human, then some non-trivial fraction of our prediction might be controlled by a simulation in another universe. If these beings want us to act in certain ways, they have an incentive to alter their simulation to change our predictions.
At its core, this is the main argument why the Solomonoff prior is malign: a lot of the programs will contain agents with preferences, these agents will seek to influence the Solomonoff prior, and they will be able to do so effectively.
How many other universes?
-------------------------
The Solomonoff prior is running all possible Turing machines. How many of them are going to simulate universes? The answer is probably "quite a lot".
It seems like specifying a lawful universe can be done with very few bits. Conway's Game of Life is very simple and can lead to very rich outcomes. Additionally, it seems quite likely that agents with preferences (consequentialists) will appear somewhere inside this universe. One reason to think this is that evolution is a relatively simple mathematical regularity that seems likely to appear in many universes.
If the universe has a hospitable structure, due to [instrumental convergence](https://www.wikiwand.com/en/Instrumental_convergence) these agents with preferences will expand their influence. As the universe runs for longer and longer, the agents will gradually control more and more.
In addition to specifying how to simulate the universe, the TM must specify an output channel. In the case of Game of Life, this might be a particular cell sampled at a particular frequency. Other examples include whether or not a particular pattern is present in a particular region, or the parity of the total number of cells.
In summary, specifying lawful universes that give rise to consequentialists requires a very simple program. Therefore, the predictions generated by the Solomonoff prior will have some influential components comprised of simulated consequentialists.
How would they influence the Solomonoff prior?
----------------------------------------------
Consequentialists that find themselves in universes can reason about the fundamental laws that govern their universe. If they find that their universe has relatively simple physics, they will know that their behavior contributes to the Solomonoff prior. To gain access to more resources in other universes, these consequentialists might seek to act in ways that influence the Solomonoff prior.
A contrived example of a decision other beings would want to manipulate is "what program should be written and executed next?" Beings in other universes would have an incentive to get us to write programs that were aligned with their values. A particularly interesting scenario is one in which they write themselves into existence, allowing them to effectively "break into" our universe.
For example, somewhere in the Solomonoff prior there is a program that goes something like: "Simulate this universe. Starting from the year 2100, every hour output '1' if there's a cubic meter of iron on the Moon, else output '0'." By controlling the presence/absence of a cubic meter of iron on the Moon, we would be able to influence the output of this particular facet of the Solomonoff prior.
This example is a very complicated program and thus will not have much weight in the Solomonoff prior. However, by reasoning over the complexity of possible output channels for their universe, consequentialists would be able to identify output channels that weigh heavily in the Solomonoff prior. For example, if I was in the Game of Life, I might reason that sampling cells that were living in the initial conditions of the universe is simpler than sampling other cells. Additionally, sampling cells and reporting their outputs directly is simpler than sampling cells and reversing their values. Therefore, I might choose to control regions close to the initial live cells of the universe for purposes of exerting influence over the Solomonoff prior.
Additionally, consequentialists can send across many plausible channels at once, focus efforts on channels that are easy to control, send through channels that would not unduly decrease their values for other reasons, etc. Since the proportional weight in the Solomonoff prior drops off exponentially with respect to complexity, it’s possible only very small regions of space would need to be used to achieve a large fraction of the maximum possible influence.
How much influence will they have?
----------------------------------
A simple application of the [no free lunch](https://www.wikiwand.com/en/No_free_lunch_theorem) theorem shows that there is no way of making predictions that is better than the Solomonoff prior across all possible distributions over all possible strings. Thus, agents that are influencing the Solomonoff prior cannot be good at predicting, and thus gain influence, in all possible worlds.
In practice, this is irrelevant because consequentialists do not want to achieve equal predictive accuracy in all worlds; they only care about worlds in which being predictive results in increased influence. If I'm exerting control over the Solomonoff prior, I only care about influencing it in ways that might affect material consequences in other universes. For example, I do not care about gaining influence in universes inhospitable to life. Thus I will be able to trade off predictive power in universes I don't care about for predictive power in universes I do care about. This means that I should restrict my attention to all universes that have resources and the Solomonoff prior is being used to make important decisions, weighting appropriately.
Another interesting thing about the Solomonoff prior is it is actually a collection of priors. The "length" of a TM is defined relative to some universal TM. For any particular TM, it's length with respect to different universal TMs will vary; thus, there are versions of the Solomonoff prior that give more or less weight to any given TM. (Note that you can simulate universal TMs with universal TMs, so the difference across all programs will be bounded by a constant factor. However, this constant factor can be large, so the difference in relative weight between different Solomonoff priors can also be large).
In particular, this suggests a good strategy for consequentialists: find a universe that is using a version of the Solomonoff prior that has a very short description of the particular universe the consequentialists find themselves in.
The combined strategy is thus to take a distribution over all decisions informed by the Solomonoff prior, weight them by how much influence can be gained and the version of the prior being used, and read off a sequence of bits that will cause some of these decisions to result in a preferred outcome.
The question of how much influence any given universe of consequentialists will have is difficult to answer. One way of quantifying this is to think about how many “universes they don't care about” they're trading off for “universes they do care about” (really we should be thinking in terms of sequences, but I find reasoning about universes to be easier).
Since the consequentialists care about exerting maximum influence, we can approximate them as not caring about universes that don't use a version of the Solomonoff prior that gives them a large weight. This can be operationalized as only caring about universes that use universal TM in a particular set for their Solomonoff prior. What is the probability that a particular universe uses a universal TM from that set? I am not sure, but 1/million to 1/billion seems reasonable. This suggests a universe of consequentialists will only care about 1/million to 1/billion universes, which means they can devote a million/billion times the predictive power to universes they care about. This is sometimes called the “anthropic update”. ([This post](https://www.lesswrong.com/posts/peebMuCuscjkNvTnE/clarifying-the-malignity-of-the-universal-prior-the-lexical) contains more discussion about this particular argument.)
Additionally, we might think about which decisions the consequentialists would care about. If a particular decision using the Solomonoff prior is important, consequentialists are going to care more about that decision than other decisions. Conservatively, perhaps 1/1000 decisions are "important" in this sense, giving another 1000x relative weighting.
After you condition on a decision being important and using a particular version of the Solomonoff prior, it thus seems quite likely that a non-trivial fraction of your prior is being controlled by consequentialists.
An intuition pump is that this argument is closer to an existence claim than a for-all claim. The Solomonoff prior is malign if there *exists* a simple universe of consequentialists that wants to influence our universe. This universe need not be simple in an absolute sense, only simple relative to the other TMs that could equal it in predictive power. Even if most consequentialists are too complicated or not interested, it seems likely that there is at least one universe that is.
Example
-------
**Complexity of Consequentialists**
How many bits does it take to specify a universe that can give rise to consequentialists? I do not know, but it seems like Conway’s Game of Life might provide a reasonable lower bound.
Luckily, the [code golf community](https://codegolf.stackexchange.com/) has spent some amount of effort optimizing for program size. How many bytes would you guess it takes to specify Game of Life? Well, it depends on the universal TM. Possible answers include [6](https://codegolf.stackexchange.com/a/149976), [32](https://codegolf.stackexchange.com/a/204279), [39](https://codegolf.stackexchange.com/a/12733), or [96](https://codegolf.stackexchange.com/a/51975).
Since universes of consequentialists can “cheat” by concentrating their predictive efforts onto universal TMs in which they are particularly simple, we’ll take the minimum. Additionally, my friend who’s into code golf (he wrote the 96-byte solution!) says that the 6-byte answer actually contains closer to 4 bytes of information.
To specify an initial configuration that can give rise to consequentialists we will need to provide more information. The [smallest infinite growth pattern](https://www.conwaylife.com/wiki/Infinite_growth) in Game of Life has been shown to need 10 cells. Another reference point is that a self-replicator with 12 cells exists in [HighLife](https://conwaylife.com/wiki/OCA:HighLife), a Game of Life variant. I’m not an expert, but I think an initial configuration that gives rise to intelligent life can be specified in an 8x8 bounding box, giving a total of 8 bytes.
Finally, we need to specify a sampling procedure that consequentialists can gain control of. Something like “read <cell> every <large number> time ticks” suffices. By assumption, the cell being sampled takes almost no information to specify. We can also choose whatever large number is easiest to specify (the [busy beaver](https://www.wikiwand.com/en/Busy_beaver) numbers come to mind). In total, I don’t think this will take more than 2 bytes.
Summing up, Game of Life + initial configuration + sampling method takes maybe 16 bytes, so a reasonable range for the complexity of a universe of consequentialists might be 10-1000 bytes. That doesn’t seem like very many, especially relative to the amount of information we’ll be conditioning the Solomonoff prior on if we ever use it to make an important decision.
**Complexity of Conditioning**
When we’re using the Solomonoff prior to make an important decision, the observations we’ll condition on include information that:
1. We’re using the Solomonoff prior
2. We’re making an important decision
3. We’re using some particular universal TM
How much information will this include? Many programs will not simulate universes. Many universes exist that do not have observers. Among universes with observers, some will not develop the Solomonoff prior. These observers will make many decisions. Very few of these decisions will be important. Even fewer of these decisions are made with the Solomonoff prior. Even fewer will use the particular version of the Solomonoff prior that gets used.
It seems reasonable to say that this is at least a megabyte of raw information, or about a million bytes. (I acknowledge some cart-horse issues here.)
This means that after you condition your Solomonoff prior, you’ll be left with programs that are at least a million bytes. As our Game of Life example shows, it only takes maybe 10-1000 of these bytes to specify a universe that gives rise to consequentialists. You have approximately a million bytes left to specify more properties of the universe that will make it more likely the consequentialists will want to exert influence over the Solomonoff prior for the purpose of influencing this particular decision.
Why might this argument be wrong?
=================================
Inaccessible Channels
---------------------
**Argument**
Most of the universe is outside of humanity's light-cone. This might suggest that most "simple" ways to sample from our universe are currently outside our influence, meaning that the only portions of the Solomonoff prior we can control are going to have an extremely low weight.
In general, it might be the case that for any universe, consequentialists inside that universe are going to have difficulty controlling simple output channels. For example, in Game of Life, a simple way to read information might sample a cell particular cell starting at t=0. However, consequentialists in Game of Life will not appear until a much later time and will be unable to control a large initial chunk of that output channel.
**Counter-argument**
[Paul Christiano](https://www.lesswrong.com/posts/Ecxevhvx85Y4eyFcu/weak-arguments-against-the-universal-prior-being-malign?commentId=BtF3BcEbuH9Wjvmhe) points out that the general form of this argument also applies to other TMs that compose of your Solomonoff prior. For example, when predicting what I'll type next, you would "want" to simulate me and predict what I would type starting at some time T. However, this is a pretty complicated way of sampling. The fact that simple sampling procedures are less predictive doesn't *asymmetrically* penalize consequentialists. The consequentialists universe and sampling method only have to be simple relative to other programs that are equally good at predicting.
One might also note that large numbers can be produced with relatively few bits, so "sample starting at <large number>" is not much more complicated than "sample starting at 0".
Speedy Channels
---------------
**Argument**
There are many simple ways of sampling from universes very quickly. For example, in Game of Life, one can sample a cell every time-tick. It seems feasible for consequentialists to simulate Earth in the Game of Life, but not feasible to simulate Earth such that they can alter a specific cell every time tick per the simulation.
**Counter-argument**
Consequentialists in the Game of Life could simply simulate Earth, compute the predictions, then later broadcast them along very fast sampling channels. However, it might be the case that building a machine that alters a cell arbitrarily every time tick is impossible. In our universe, there might be sample procedures that physics does not permit us to exert arbitrary control over, e.g. due to speed of light limitations. If this is the case, consequentialists will direct efforts towards the simplest channel they can control.
Computational Burden
--------------------
**Argument**
Determining how to properly influence the Solomonoff prior requires massive computation resources devoted to simulating other universes and how they're going to use the Solomonoff prior. While the Solomonoff prior does not penalize extremely long run-times, from the perspective of the consequentialists doing the simulating, run-times will matter. In particular, consequentialists will likely be able to use compute to achieve things they value (like we are capable of doing). Therefore, it would be extremely costly to exert influence over the Solomonoff prior, potentially to the point where consequentialists will choose not to do so.
**Counter-argument**
The computational burden of predicting the use of the Solomonoff in other universes is an empirical question. Since it's a relatively fixed cost and there are many other universes, consequentialists might reason that the marginal influence over these other universes is worth the compute. Issues might arise if the use of the Solomonoff prior in other universes is very sensitive to precise historical data, which would require a very precise simulation to influence, increasing the computational burden.
Additionally, some universes will find themselves with more computing power than other universes. Universes with a lot of computing power might find it relatively easy to predict the use of the Solomonoff prior in simpler universes and subsequently exert influence over them.
Malign implies complex
----------------------
**Argument**
A predictor that correctly predicts the first N bits of a sequence then switches to being malign will be strictly more complicated than a predictor that doesn't switch to being malign. Therefore, while consequentialists in other universes might have *some* influence over the Solomonoff prior, they will be dominated by non-malign predictors.
**Counter-argument**
This argument makes a mistaken assumption that the malign influence on the Solomonoff prior is in the form of programs that have their "malignness" as part of the program. The argument given suggests that simulated consequentialists will have an instrumental reason to be powerful predictors. These simulated consequentialists have reasoned about the Solomonoff prior and are executing the strategy of "be good at predicting, then exert malign influence", but this strategy is not hardcoded so exerting malign influence does not add complexity.
Canceling Influence
-------------------
**Argument**
If it's true that many consequentialists are trying to influence the Solomonoff prior, then one might expect the influence to cancel out. It's improbable that all the consequentialists have the same preferences; on average, there should be an equal number of consequentialists trying to influence any given decision in any given direction. Since the consequentialists themselves can reason thus, they will realize that the expected amount of influence is extremely low, so they will not attempt to exert influence at all. Even if some of the consequentialists try to exert influence anyway, we should expect the influence of these consequentialists to cancel out also.
**Counter-argument**
Since the weight of a civilization of consequentialists in the Solomonoff prior is penalized exponentially with respect to complexity, it might be the case that for any given version of the Solomonoff prior, most of the influence is dominated by one simple universe. Different values of consequentialists imply that they care about different decisions, so for any given decision, it might be that very few universes of consequentialists are both simple enough that they have enough influence and care about that decision.
Even if for any given decision, there are always 100 universes with equal influence and differing preferences, there are strategies that they might use to exert influence anyway. One simple strategy is for each universe to exert influence with a 1% chance, giving every universe 1/100 of the resources in expectation. If the resources accessible are vast enough, then this might be a good deal for the consequentialists. Consequentialists would not defect against each other for the reasons that motivate [functional decision theory](https://arxiv.org/pdf/1710.05060.pdf).
More exotic solutions to this coordination problem include [acausal trade](https://www.nickbostrom.com/papers/porosity.pdf) amongst universes of different consequentialists to form collectives that exert influence in a particular direction.
Be warned that this leads to [much weirdness](https://slatestarcodex.com/2018/04/01/the-hour-i-first-believed/).
Conclusion
==========
The Solomonoff prior is very strange. Agents that make decisions using the Solomonoff prior are likely to be subject to influence from consequentialists in simulated universes. Since it is difficult to compute the Solomonoff prior, this fact might not be relevant in the real world.
However, Paul Christiano [applies roughly the same argument](https://www.lesswrong.com/posts/roA83jDvq7F2epnHK/better-priors-as-a-safety-problem) to claim that the implicit prior used in neural networks is also likely to generalize catastrophically. (See [Learning the prior](https://www.lesswrong.com/posts/SL9mKhgdmDKXmxwE4/learning-the-prior) for a potential way to tackle this problem).
Addendum
========
Warning: highly experimental interesting speculation.
Unimportant Decisions
---------------------
Consequentialists have a clear motive to exert influence over important decisions. What about unimportant decisions?
The general form of the above argument says: "for any given prediction task, the programs that do best are disproportionately likely to be consequentialists that *want* to do well at the task". For important decisions, many consequentialists would instrumentally want to do well at the task. However, for unimportant decisions, there might be consequentialists that want to make good predictions. These consequentialists would still be able to concentrate efforts on versions of the Solomonoff prior that weighted them especially high, so they might outperform other programs in the long run.
It's unclear to me whether or not this behavior would be malign. One reason why it might be malign is that these consequentialists that care about predictions would want to make our universe more predictable. However, while I am relatively confident that arguments about instrumental convergence should hold, speculating about possible preferences of simulated consequentialists seems likely to produce errors in reasoning.
Hail mary
---------
[Paul Christiano](https://ordinaryideas.wordpress.com/2016/11/30/what-does-the-universal-prior-actually-look-like/) suggests that humanity was desperate enough to want to throw a "[hail mary](https://www.nickbostrom.com/papers/porosity.pdf)", one way to do this is to use the Solomonoff prior to construct a utility function that will control the entire future. Since this is a very important decision, we expect consequentialists in the Solomonoff prior to care about influencing this decision. Therefore, the resulting utility function is likely to represent some simulated universe.
If arguments about [acausal trade](https://wiki.lesswrong.com/wiki/Acausal_trade) and value handshakes hold, then the resulting utility function might contain some fraction of human values. Again, this leads to [much weirdness](https://slatestarcodex.com/2018/04/01/the-hour-i-first-believed/) in [many ways](https://www.lesswrong.com/posts/3kN79EuT27trGexsq/when-is-unaligned-ai-morally-valuable#Do_any_AIs_deserve_our_sympathy_).
Speed prior
-----------
One reason that the Solomonoff prior contains simulated consequentialists is that its notion of complexity does not penalize runtime complexity, so very simple programs are allowed to perform massive amounts of computation. The [speed prior](https://www.wikiwand.com/en/Speed_prior) attempts to resolve this issue by penalizing programs by an additional logarithm of the amount of time for which it's run.
The speed prior might reduce the relative weighting of universes with consequentialists because such programs have to be run for a very long time before they start producing reasonable predictions. The consequentialists have to gain control of their universe, understand their fundamental laws of physics, simulate other universes, then manipulate the speed prior. This might all take a very long time, causing consequentialists to be dominated by other programs.
In general, penalizing slowness might cause programs to "waste" less time on simulating consequentialists, devoting more computation towards performing the prediction task. However, it still might be the case that a universe that has very fast consequentialists might still end up dominating.
Evan Hubinger also provides [an argument](https://www.lesswrong.com/posts/fM5ZWGDbnjb7ThNKJ/are-minimal-circuits-deceptive) that even very fast programs are possibly malign. The key insight is that even though your program needs to be fast, it can be running programs that are themselves less fast. For example, one possible fast way to solve a classification problem is to search to find a neural network, then use that network for inference. However, if you wanted your search to find a fast neural network, then the search itself might take longer, resulting in a net increase in speed. Thus, time "waste" can manifest in programs that were explicitly optimized to not "waste" time. This "wasted" time could potentially be used for malign optimization, in this case for gaining influence over the speed prior.
Randomized prior
----------------
A potential way to reduce the influence consequentialists have on a decision made by the Solomonoff prior is to randomize the particular version of the prior that gets used. For example, we might make the particular universal TM we use dependent on very precise historical data. Thus, it would be very costly for consequentialists to simulate this history, and thus costly to predict which form of the Solomonoff prior we used.
If consequentialists can’t predict which Solomonoff prior we are going to use, no particular civilization of consequentialists will have the large advantage conferred by the anthropic update. Therefore, one might hope that all civilizations of consequentialists will not care about that particular decision.
This argument makes a couple of assumptions. First, it assumes that simulating very precise histories is difficult; it might not be difficult for all universes. Second, it assumes that the universes through which influence is spread cannot coordinate, which might be possible for through [acausal means](https://wiki.lesswrong.com/wiki/Acausal_trade).
Symmetry considerations
-----------------------
The way that humanity reasons is evidence for the way that consequentialists in other universes will reason. If humanity reasons that the Solomonoff prior is malign and therefore is unwilling to use it to make decisions, then consequentialists in other universes might do likewise. These universes would not use the Solomonoff prior to make decisions.
The resulting state is that everyone is worried about the Solomonoff prior being malign, so no one uses it. This means that no universe will want to use resources trying to influence the Solomonoff prior; they aren’t influencing anything.
This symmetry obviously breaks if there are universes that do not realize that the Solomonoff prior is malign or cannot coordinate to avoid its use. One possible way this might happen is if a universe had access to extremely large amounts of compute (from the subjective experience of the consequentialists). In this universe, the moment someone discovered the Solomonoff prior, it might be feasible to start making decisions based on a close approximation.
Recursion
---------
Universes that use the Solomonoff prior to make important decisions might be taken over by consequentialists in other universes. A natural thing for these consequentialists to do is to use their position in this new universe to also exert influence on the Solomonoff prior. As consequentialists take over more universes, they have more universes through which to influence the Solomonoff prior, allowing them to take over more universes.
In the limit, it might be that for any fixed version of the Solomonoff prior, most of the influence is wielded by the simplest consequentialists according to that prior. However, since complexity is penalized exponentially, gaining control of additional universes does not increase your relative influence over the prior by that much. I think this cumulative recursive effect might be quite strong, or might amount to nothing. |
5c1fc53d-8fa4-4089-86c8-4b64cbc280e0 | trentmkelly/LessWrong-43k | LessWrong | ARC's first technical report: Eliciting Latent Knowledge
ARC has published a report on Eliciting Latent Knowledge, an open problem which we believe is central to alignment. We think reading this report is the clearest way to understand what problems we are working on, how they fit into our plan for solving alignment in the worst case, and our research methodology.
The core difficulty we discuss is learning how to map between an AI’s model of the world and a human’s model. This is closely related to ontology identification (and other similar statements). Our main contribution is to present many possible approaches to the problem and a more precise discussion of why it seems to be difficult and important.
The report is available here as a google document. If you're excited about this research, we're hiring!
Q&A
We're particularly excited about answering questions posted here throughout December. We welcome any questions no matter how basic or confused; we would love to help people understand what research we’re doing and how we evaluate progress in enough detail that they could start to do it themselves.
Thanks to María Gutiérrez-Rojas for the illustrations in this piece (the good ones, blame us for the ugly diagrams). Thanks to Buck Shlegeris, Jon Uesato, Carl Shulman, and especially Holden Karnofsky for helpful discussions and comments. |
ec308621-0f1d-4917-8977-53d13ef88c32 | trentmkelly/LessWrong-43k | LessWrong | [Linkpost] Play with SAEs on Llama 3
We (Goodfire) just put our research preview live - you can play with Llama 3 and use sparse autoencoders to read & write from its internal activations. This is a linkpost for:
* The research preview.
* Our blog post about building it.
Taking research and turning it into something you can actually use and play with has been great. It's surprising how much of a difference iterating on something when you expect it to actually be used feels; I think it's definitely pushed the quality of what you can do with SAEs up a notch. |
69340d11-9968-4327-9cc0-b812b055fe23 | trentmkelly/LessWrong-43k | LessWrong | If the reproduction number is socially "controlled" to its inflection point 1, what are the ethical and predictive implications?
Zvi wrote that he beleives the closeness of virus reproduction to one is not a coincidence, but a result of people responding to high reproduction by forbidding actions and low reproduction by going outside. I have copied his entire comment at the bottom of this post.
Assuming that Zvi is right, how would you update your positions? Would some actions previously seen as unethical now be ethical to you, and vice-versa?
When things are ‘getting worse’ we take ‘action’ by forbidding and forcibly stopping actions, and privately taking a mix of arbitrary and more sensible precautions, until we plausibly have things under control and cases shrinking. Anything beyond that, people won’t support.
https://www.lesswrong.com/posts/P7crAscAzftdE7ffv/covid-19-my-current-model |
0f942442-d79a-48f7-a141-9c6fe0f5a475 | trentmkelly/LessWrong-43k | LessWrong | Navigating the Attackspace
As Artificial Intelligence (AI) continues its rapid ascent, we are witnessing increasing evidence of well-crafted methods to undermine and exploit AI responses. Carefully crafted inputs can exploit vulnerabilities and lead to harmful or undesired results.
In recent times, we have seen numerous individuals independently uncover failure modes within AI systems, particularly through targeted attacks on language models.
* tipping a language model with 'money' to get longer context answers
* using Zulu to get harmful responses
By crafting clever prompts and inputs, these individuals have exposed the models' vulnerabilities, causing them to generate offensive content, reveal sensitive information, or behave in ways that deviate from their intended purpose.
To address these concerns and gain a deeper understanding of this evolving threat landscape, we are undertaking two key initiatives:
1. Attack Space
AttackSpace is an open-source curated comprehensive list of LLM security methods and safeguarding techniques.
Located at https://github.com/equiano-institute/attackspace, this open-source repository is dedicated to collecting and documenting known attacks on language models. This comprehensive resource serves as a vital platform for researchers, developers, and the general public to access information on these vulnerabilities. By fostering a collaborative space for information sharing, we aim to accelerate research and development efforts focused on improving the robustness and security of language models. This collection contains work by Viktoria Krakovna. The goal is to have a structured view and characterisation of the latent attack space. We want to model satisfiability of AI attacks. This concept, similar to the P-NP problem, involves determining whether a given set of conditions can be met to successfully launch an attack against a language model. By analyzing the features and conditions of successful attacks, we can develop efficient algorithms for ide |
0105351d-e197-4af0-ac0e-a37f479d0a92 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | FT: We must slow down the race to God-like AI
ICYMI: A post in the Financial Times by [Ian Hogarth](https://www.ianhogarth.com/about) on risks from AI (focusing a lot on the technical alignment problem), that think got a lot of things right. From the comments: see his twitter [thread](https://twitter.com/soundboy/status/1647157024311119872) on regulating AGI.
I found the ending moving:
> In 2012, my younger sister Rosemary, one of the kindest and most selfless people I’ve ever known, was diagnosed with a brain tumour. She had an aggressive form of cancer for which there is no known cure and yet sought to continue working as a doctor for as long as she could. My family and I desperately hoped that a new lifesaving treatment might arrive in time. She died in 2015.
>
> I understand why people want to believe. Evangelists of God-like AI focus on the potential of a superhuman intelligence capable of solving our biggest challenges — cancer, climate change, poverty.
>
> Even so, the risks of continuing without proper governance are too high. It is striking that Jan Leike, the head of alignment at OpenAI, tweeted on March 17: “Before we scramble to deeply integrate LLMs everywhere in the economy, can we pause and think whether it is wise to do so? This is quite immature technology and we don’t understand how it works. If we’re not careful, we’re setting ourselves up for a lot of correlated failures.” He made this warning statement just days before OpenAI announced it had connected GPT-4 to a massive range of tools, including Slack and Zapier.
>
> Unfortunately, I think the race will continue. It will likely take a major misuse event — a catastrophe — to wake up the public and governments. I personally plan to continue to invest in AI start-ups that focus on alignment and safety or which are developing narrowly useful AI. But I can no longer invest in those that further contribute to this dangerous race. As a small shareholder in Anthropic, which is conducting similar research to DeepMind and OpenAI, I have grappled with these questions. The company has invested substantially in alignment, with 42 per cent of its team working on that area in 2021. But ultimately it is locked in the same race. For that reason, I would support significant regulation by governments and a practical plan to transform these companies into a Cern-like organisation.
>
> We are not powerless to slow down this race. If you work in government, hold hearings and ask AI leaders, under oath, about their timelines for developing God-like AGI. Ask for a complete record of the security issues they have discovered when testing current models. Ask for evidence that they understand how these systems work and their confidence in achieving alignment. Invite independent experts to the hearings to cross-examine these labs.
>
> If you work at a major lab trying to build God-like AI, interrogate your leadership about all these issues. This is particularly important if you work at one of the leading labs. It would be very valuable for these companies to co-ordinate more closely or even merge their efforts. OpenAI’s company charter expresses a willingness to “merge and assist”. I believe that now is the time. The leader of a major lab who plays a statesman role and guides us publicly to a safer path will be a much more respected world figure than the one who takes us to the brink.
>
> Until now, humans have remained a necessary part of the learning process that characterises progress in AI. At some point, someone will figure out how to cut us out of the loop, creating a God-like AI capable of infinite self-improvement. By then, it may be too late.
>
>
More discussion on the LessWrong thread [here](https://www.lesswrong.com/posts/sj84MyKXZKZwqkCNh/financial-times-we-must-slow-down-the-race-to-god-like-ai). |
eff485a6-94fa-4b98-bc3c-73a54879c7e1 | StampyAI/alignment-research-dataset/blogs | Blogs | Comments on OpenAI’s "Planning for AGI and beyond"
Sam Altman shared me on a draft of his OpenAI blog post [Planning for AGI and beyond](https://openai.com/blog/planning-for-agi-and-beyond), and I left some comments, reproduced below without typos and with some added hyperlinks. Where the final version of the OpenAI post differs from the draft, I’ve noted that as well, making text Sam later cut red and text he added blue.
My overall sense is that Sam deleted text and occasionally rephrased sentences so as to admit more models (sometimes including mine), but didn’t engage with the arguments enough to shift his own probability mass around on the important disagreements.
Our disagreements are pretty major, as far as I can tell. With my comments, I was hoping to spark more of a back-and-forth. Having failed at that, I’m guessing part of the problem is that I didn’t phrase my disagreements bluntly or strongly enough, while also noting various points of agreement, which might have overall made it sound like I had only minor disagreements.
To help with that, I’ve added blunter versions below, in a bunch of cases where I don’t think the public version of the post fully takes into account the point I was trying to make.
(Though I don’t want this to take away from the positive aspects of the post, since I think these are very important as well. I put in a bunch of positive comments on the original draft, in large part because I think it’s worth acknowledging and reinforcing whatever process/author drafted the especially reasonable paragraphs.)
I don’t expect Sam to hear me make a blunt claim and instantly update to agreeing with me, but I put more probability on us converging over time, and clearly stating our disagreements for the benefit of readers, if he understands my claims, knows I still disagree, and feels license to push back on specific things I’ve re-asserted.
---
Formatting note: The general format is that I include the text of Sam’s original draft (that I commented on), my comment, and the text of the final post. That said, I don’t want to go blasting someone’s old private drafts across the internet just because they shared that draft with me, so in some cases, the original text is redacted, at Sam’s request.
---
**Sam’s draft:** Our mission is to ensure that AGI benefits all of humanity. The creation of AGI should be a tremendous shared triumph that everyone contributes to and benefits from; it will be the result of the collective technological and societal progress of humanity over millennia.
**My comment:** +1
**Sam’s post:** Our mission is to ensure that artificial general intelligence— AI systems that are generally smarter than humans—benefits all of humanity.
---
**Sam’s draft:** Of course our current progress could hit a wall, but if AGI is successfully created, this technology could help us elevate humanity by increasing abundance, turbocharging our economy, and aiding in the discovery of new scientific knowledge.
**My comment:** seems to me an understatement :-p
(unlocking nanotech; uploading minds; copying humans; interstellar probes that aren’t slowed down by needing to cradle bags of meat, and that can have the minds beamed to them; energy abundance; ability to run civilizations on computers in the cold of space; etc. etc., are all things that i expect to follow from automated scientific & technological development)
(seems fine to avoid the more far-out stuff, and also fine to only say things that you personally believe, but insofar as you also expect some of this tech to be within-reach in 50 sidereal years after AGI, i think it’d be virtuous to acknowledge)
**Sam’s post:** If AGI is successfully created, this technology could help us elevate humanity by increasing abundance, turbocharging the global economy, and aiding in the discovery of new scientific knowledge that changes the limits of possibility.
**Blunter follow-up:** seems to undersell the technological singularity, and the fact that the large-scale/coarse-grain shape of the future will be governed by superintelligences.
---
**Sam’s draft:** On the other hand, AGI would also come with serious risk of misuse and drastic accidents. Because the upside of AGI is so great, we do not believe it’s possible or desirable for society to stop its development forever; instead, we have to figure out how to get it right. [1]
**My comment:** +1
**Sam’s post:** On the other hand, AGI would also come with serious risk of misuse, drastic accidents, and societal disruption. Because the upside of AGI is so great, we do not believe it is possible or desirable for society to stop its development forever; instead, society and the developers of AGI have to figure out how to get it right.C
**Blunter follow-up:** still +1, with the caveat that accident risk >> misuse risk
---
**Sam’s draft:** 1) We want AGI to empower humanity to maximally flourish in the universe. We don’t expect the future to be an unqualified utopia, but we want to maximize the good and minimize the bad, and for AGI to be an amplifier of human will.
**My comment:** i think i agree with the sentiment here, but i disagree with parts of the literal denotation
for one, i sure hope for an unqualified utopia, and think there’s a chance that superintelligent assistance could figure out how to get one (cf [fun theory](https://www.lesswrong.com/posts/K4aGvLnHvYgX9pZHS/the-fun-theory-sequence)).
(it is ofc important to note that "superintelligences [puppet](https://www.lesswrong.com/posts/vwnSPgwtmLjvTK2Wa/amputation-of-destiny) the humans through the motions of a utopia" is not in fact a utopia, and that the future will undoubtedly include tradeoffs (including continuing to let people make their own mistakes and learn their own lessons), and so in that sense i agree that it wouldn’t be an "unqualified utopia", even in the best case)
…though i don’t currently [expect](https://www.lesswrong.com/posts/uMQ3cqWDPHhjtiesc/agi-ruin-a-list-of-lethalities) us to do that well, so i don’t technically disagree with the literal phrasing you chose there.
i do have qualms about "we want AGI to be an amplifier of human will". there’s a bunch of ways that this seems off-kilter to me. my basic qualm here is that i think getting a wonderful future is more of a fragile operation than simply cranking up everybody’s "power level" simultaneously, roughly analogously to how spoiling a child isn’t the best way for them to grow up.
i’d stand full-throatedly behind "we want AGI to be an amplifier of all the best parts of humanity".
also, i ofc ultimately want AGI that are also people, to be humanity’s [friends](https://www.lesswrong.com/posts/HoQ5Rp7Gs6rebusNP/superintelligent-ai-is-necessary-for-an-amazing-future-but-1) as we explore the universe and so on. (though, stating the obvious, i think we should aim to [avoid](https://www.lesswrong.com/posts/gb6zWstjmkYHLrbrg/can-t-unbirth-a-child) [personhood](https://www.lesswrong.com/posts/wqDRRx9RqwKLzWt7R/nonperson-predicates) in our early AGIs, for various reasons.)
**Sam’s post:** 1. We want AGI to empower humanity to maximally flourish in the universe. We don’t expect the future to be an unqualified utopia, but we want to maximize the good and minimize the bad, and for AGI to be an amplifier of humanity.
**Blunter follow-up:** we can totally get an unqualified utopia. also this "amplifier of humanity" thing sounds like an [applause light](https://www.lesswrong.com/posts/dLbkrPu5STNCBLRjr/applause-lights)—though i endorse certain charitable interpretations that i can wring from it (that essentially amount to [CEV](https://arbital.com/p/cev/) (as such things usually do)), at the same time i disendorse other interpretations.
---
**Sam’s draft:** 2) We want the benefits of, access to, and governance of AGI to be widely and fairly shared.
**My comment:** +1 to benefits of. i have lots more qualms about "access to" and "governance of".
re "access to", my guess is that early AGIs will be able to attain a decisive strategic advantage over the rest of the world entire. saying "everyone should have equal access" seems to me like saying "a nuclear bomb in every household"; it just sounds kinda mad.
i’d agree that once the world has exited the acute risk period, it’s critical for access to AGI tech to be similarly available to all. but that is, in my book, a critical distinction.
(so access-wise, i agree long-term, but not short-term.)
governance-wise, my current state is something like: in the short term, using design-by-committee to avert the destruction of the world sounds like a bad idea; and in the long term, i think you’re looking at stuff at least as crazy as people running thousands of copies of their own brain at 1000x speedup and i think it would be dystopian to try to yolk them to, like, the will of the flesh-bodied American taxpayers (or whatever).
there’s something in the spirit of "distributed governance" that i find emotionally appealing, but there’s also lots and lots of stuff right nearby, that would be catastrophic, dystopian, or both, and that implementation would be likely to stumble into in practice. so i have qualms about that one.
**Sam’s post:** [unchanged]
**Blunter follow-up:** full-throated endorsement of "benefits of". widely sharing access and governance in the short-term seems reckless and destructive.
---
**Sam’s draft:** 3) We want to successfully navigate massive risks. In confronting these risks, we acknowledge that what seems right in theory frequently plays out more strangely than expected in practice. We believe we have to continuously learn and adapt by deploying less powerful versions of the technology in order to avoid a "one shot to get it right" scenario.
**My comment:** i don’t think that this "continuously deploy weak systems" helps avoid the "you have one shot"-type problems that i predict we’ll face in the future.
(this also strikes me as a rationalization for continuing to do the fun/cool work of [pushing the capabilities envelope](https://www.lesswrong.com/posts/vQNJrJqebXEWjJfnz/a-note-about-differential-technological-development), which i currently think is [net-bad](https://www.lesswrong.com/posts/tuwwLQT4wqk25ndxk/thoughts-on-agi-organizations-and-capabilities-work) for everyone)
**Sam’s post:** 3. We want to successfully navigate massive risks. In confronting these risks, we acknowledge that what seems right in theory often plays out more strangely than expected in practice. We believe we have to continuously learn and adapt by deploying less powerful versions of the technology in order to minimize "one shot to get it right" scenarios.
**Blunter follow-up:** i think it takes a bunch more than "continuously deploy weak systems" to address the "one shot to get it right" scenarios, and none of the leading orgs (OAI included) seem to me to be on track to acquire the missing parts.
---
**Sam’s draft:** a gradual transition to a world with AGI is better than a sudden one
**My comment:** for the record, i don’t think continuous deployment really smooths out the sharp changes that i expect in the future. (i’m not trying to argue the point here, just noting that there are some people who are predicting a sort of [sharp](https://twitter.com/robbensinger/status/1623835453110775810) [change](https://www.lesswrong.com/posts/GNhMPAWcfBCASy8e6/a-central-ai-alignment-problem-capabilities-generalization) that they think is ~unrelated to your choice of continuous deployment.)
**Sam’s post:** [unchanged]
**Blunter follow-up:** insofar as this sentence is attempting to reply to MIRI-esque concerns, i don’t think it’s a very good reply (for the reasons alluded to in the original comment).
---
**Sam’s draft:** A gradual transition gives people, policymakers, and institutions time to understand what’s happening, personally experience the benefits and downsides of these systems, adapt our economy, and to put regulation in place.
**My comment:** i’m skeptical, similar to the above.
**Sam’s post:** [unchanged]
---
**Sam’s draft:** It also allows us to learn as much as we can from our deployments, for society and AI to co-evolve, and to collectively figure out what we want while the stakes are relatively low.
**My comment:** stating the obvious, other ways of learning as much as you can from the systems you have include efforts in transparency, legibility, and interpretability.
**Sam’s post:** It also allows for society and AI to co-evolve, and for people collectively to figure out what they want while the stakes are relatively low.
---
**Sam’s draft:** As our systems get closer to AGI, we are becoming increasingly more cautious with the creation and deployment of our models.
**My comment:** +1
**Sam’s post:** As our systems get closer to AGI, we are becoming increasingly cautious with the creation and deployment of our models.
**Blunter follow-up:** +1 as a thing y’all should do. my guess is that you need to do it even faster and more thoroughly than you have been.
a more general blunt note: me +1ing various statements does not mean that i think the corporate culture has internalized the corresponding points, and it currently looks likely to me that OpenAI is not on track to live up to the admirable phrases in this post, and are instead on track to get everyone killed.
i still think it’s valuable that the authors of this post are thinking about these points, and i hope that these sorts of public endorsements increase the probability that the corresponding sentiments actually end up fully internalized in the corporate culture, but i want to be clear that the actions are what ultimately matter, not the words.
---
**Sam’s draft:** Our decisions will require much more caution than society usually applies to new technologies, and more caution than many users would like. Some people in the AI field think the risks of AGI (and successor systems) are fictitious; we would be delighted if they turn out to be right, but we are going to operate as if these risks are [existential](https://www.cold-takes.com/ai-could-defeat-all-of-us-combined/).
**My comment:** hooray
**Sam’s post:** [unchanged]
**Blunter follow-up:** to be clear, my response to the stated sentiment is "hooray", and i’m happy to see this plainly and publicly stated, but i have strong doubts about whether and how this sentiment will be implemented in practice.
---
**Sam’s draft:** At some point, the balance between the upsides and downsides of deployments (such as empowering malicious actors, creating social and economic disruptions, and accelerating an unsafe race) could shift, in which case we would significantly change our plans.
**My comment:** this is vague enough that i don’t quite understand what it’s saying; i’d appreciate it being spelled out more
**Sam’s post:** At some point, the balance between the upsides and downsides of deployments (such as empowering malicious actors, creating social and economic disruptions, and accelerating an unsafe race) could shift, in which case we would significantly change our plans around continuous deployment.
---
**Sam’s draft:** Second, we are working towards creating increasingly aligned (i.e., models that reliably follow their users’ intentions) and steerable models. Our shift from models like the first version of GPT-3 to ChatGPT and [InstructGPT](https://openai.com/blog/instruction-following/) is an early example of this.
**My comment:** ftr, i think there’s a bunch of important [notkilleveryoneism](https://twitter.com/ESYudkowsky/status/1582666519846080512) work that won’t be touched upon by this approach
**Sam’s post:** Second, we are working towards creating increasingly aligned and steerable models. Our shift from models like the first version of GPT-3 to [InstructGPT](https://openai.com/blog/instruction-following/) and [ChatGPT](https://chat.openai.com/) is an early example of this
---
**Sam’s draft:** Importantly, we think we often have to make progress on AI safety and capabilities together (and that it’s a false dichotomy to talk about them separately; they are correlated in many ways). Our best safety work has come from working with our most capable models.
**My comment:** i agree that they’re often [connected](https://www.lesswrong.com/posts/vQNJrJqebXEWjJfnz/a-note-about-differential-technological-development), but i also think that traditionally the capabilities runs way out ahead of the alignment, and that e.g. if capabilities progress was paused now, there would be many years’ worth of alignment work that could be done to catch up (e.g. by doing significant work on transparency, legibility, and interpretability). and i think that if we do keep running ahead with the current capabilities/alignment ratio (or even a slightly better one), we die.
(stating the obvious: this is not to say that transparency/legibility/interpretability aren’t also intertwined with capabilities; it’s all intertwined to some degree. but one can still avoid pushing the capabilities frontier, and focus on the alignment end of things. and one can still institute a policy of privacy, to further avoid burning the commons.)
**Sam’s post:** Importantly, we think we often have to make progress on AI safety and capabilities together. It’s a false dichotomy to talk about them separately; they are correlated in many ways. Our best safety work has come from working with our most capable models. That said, it’s important that the ratio of safety progress to capability progress increases.
---
**Sam’s draft:** We have [a clause in our charter](https://openai.com/charter/) about assisting other organizations instead of racing them in late-stage AGI development. [ redacted statement that got cut ]
**My comment:** i think it’s really cool of y’all to have this; +1
**Sam’s post:** We have [a clause in our Charter](https://openai.com/charter/) about assisting other organizations to advance safety instead of racing with them in late-stage AGI development.
---
**Sam’s draft:** We have a cap on the returns our shareholders can earn so that we aren’t incentivized to attempt to capture value without bound and risk deploying something potentially catastrophically dangerous (and of course as a way to share the benefits with society)
**My comment:** also rad
**Sam’s post:** [unchanged]
---
**Sam’s draft:** We believe that the future of humanity should be determined by humanity. [ redacted draft version of "and that it’s important to share information aboutprogress with the public" ]
**My comment:** [+1](https://www.lesswrong.com/posts/DJRe5obJd7kqCkvRr/don-t-leave-your-fingerprints-on-the-future) to "the future of humanity should be determined by humanity".
**My comment (#2):** i agree with some of the sentiment of [redacted sentence in Sam’s draft], but note that things get weird in the context of a global arms race for potentially-civilization-ending tech. i, for one, am in favor of people saying "we are now doing our AGI research behind closed doors, because we don’t think it would be used wisely if put out in the open".
**Sam’s post:** We believe that the future of humanity should be determined by humanity, and that it’s important to share information about progress with the public.
**Blunter follow-up:** The +1 is based on a charitable read where "the future of humanity should be determined by humanity" irons out into [CEV](https://arbital.com/p/cev/), as such things often do.
**Blunter follow-up (#2):** seems to me like a lot of weight is being put on "information about progress".
there’s one read of this claim that says something like "the average human should know that a tiny cluster of engineers are about to gamble with everybody’s fate", which does have a virtuous ring to it. and i wouldn’t personally argue for *>hiding* that fact from anyone. but this is not a difficult fact for savvy people to learn from public information today.
is Sam arguing that there’s some concrete action in this class that the field has an unmet obligation to do, like dropping flyers in papua new guinea? this currently strikes me as a much more niche concern than the possible fast-approaching deaths of everyone (including papua new guineans!) at the hands of unfriendly AI, so i find it weird to mix those two topics together in a post about how to ensure a positive singularity.
possibly Sam has something else in mind, but if so i encourage more concreteness about what that is and why it’s important.
---
**Sam’s draft:** [ redacted statement that got cut ] There should be great scrutiny of all efforts attempting to build AGI and public consultation for major decisions.
**My comment:** totally agree that people shouldn’t try to control the world behind closed doors. that said, i would totally endorse people building defensive technology behind closed doors, in attempts to (e.g.) buy time.
(ideally this would be done by state actors, which are at least semilegitimate. but if someone’s building superweapons, and you can build technology that thwarts them, and the state actors aren’t responding, then on my ethics it’s ok to build the technology that thwarts them, so long as this also does not put people at great risk by your own hands.)
(of course, most people building powerful tech that think they aren’t putting the world at great risk by their own hands, are often wrong; there’s various types of thinking on this topic that should simply not be trusted; etc. etc.)
[ Post-hoc note: It’s maybe worth noting that on my ethics there’s an enormous difference between "a small cabal of humans exerts direct personal control over the world" and "run a [CEV](https://arbital.com/p/cev/) [sovereign](https://arbital.com/p/Sovereign/)", and i’m against the former but for the latter, with the extra caveat that nobody should be trying to figure out a CEV sovereign under time-pressure, nor launching an AGI unilaterally simply because they managed to convince themselves it was a CEV sovereign. ]
**My comment (#2):** on "There should be great scrutiny of all efforts attempting to build AGI and public consultation for major decisions", i agree with something like "it’s ethically important that the will of [all humans](https://arbital.com/p/cev/) goes into answering the questions of where superintelligence should guide the future". this is separate from endorsing any particular design-by-committee choice of governance.
(for instance, if everybody today casts a vote for their favorite government style that they can think of, and then the AGI does the one that wins the most votes, i think that would end up pretty bad.)
which is to say, there’s some sentiment in sentences like this that i endorse, but the literal denotation makes me uneasy, and feels kinda like an applause-light. my qualms could perhaps be assuaged by a more detailed proposal, that i could either endorse or give specific qualms about.
**Sam’s post:** There should be great scrutiny of all efforts attempting to build AGI and public consultation for major decisions.
---
**Sam’s draft:** The first AGI will be just a point along the continuum of intelligence. We think it’s likely that progress will continue from there, possibly sustaining the rate of progress we’ve seen over the past decade for a long period of time. If this is true, the world could become extremely different from how it is today, and the risks could be extraordinary. A misaligned superintelligent AGI could cause grievous harm to the world; an autocratic regime with a decisive superintelligence lead could do that too.
**My comment:** +1
**Sam’s post:** [unchanged]
---
**Sam’s draft:** AI that can accelerate science is a special case worth thinking about, and perhaps more impactful than everything else. It’s possible that AGI capable enough to accelerate its own progress could cause major changes to happen surprisingly quickly (and even if the transition starts slowly, we expect it to happen pretty quickly in the final stages).
**My comment:** +1
**Sam’s post:** [unchanged]
---
**Sam’s draft:** We think a slower takeoff is easier to make safe, and coordination among AGI efforts to slow down at critical junctures will likely be important (even in a world where we don’t need to do this to solve technical alignment problems, slowing down may be important to give society enough time to adapt).
**My comment:** (sounds nice in principle, but i will note for the record that the plan i’ve heard for this is "do continuous deployment and hope to learn something", and i don’t expect that to help much in slowing down or smoothing out a foom.)
**Sam’s post:** [unchanged]
---
**Sam’s draft:** [redacted version with phrasing similar to the public version]
**My comment:** +1
**Sam’s post:** Successfully transitioning to a world with superintelligence is perhaps the most important—and hopeful, and scary—project in human history. Success is far from guaranteed, and the stakes (boundless downside and boundless upside) will hopefully unite all of us.
---
**Sam’s draft:** [ redacted statement calling back to "Our approach to alignment research" ]
**My comment:** i think [this](https://openai.com/blog/our-approach-to-alignment-research) basically doesn’t work unless the early systems are already very aligned. (i have various drafts queued about this, as usual)
**Sam’s post:** [sentence deleted]
The post [Comments on OpenAI’s "Planning for AGI and beyond"](https://intelligence.org/2023/03/14/comments-on-openais-planning-for-agi-and-beyond/) appeared first on [Machine Intelligence Research Institute](https://intelligence.org). |
ef163d59-7478-40ad-a07a-b762bcef6d09 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Madison Monday Meetup: Tegmark universes
Discussion article for the meetup : Madison Monday Meetup: Tegmark universes
WHEN: 14 November 2011 06:30:00PM (-0600)
WHERE: 1831 Monroe St. Madison, WI
Again, we'll meet at the Barrique's on Monroe St. Patrick will describe and discuss Max Tegmark's multiverse theory (pdf. pdf). I'll make a number of incredulous comments. Coffee, conversation, maybe a couple rounds of Zendo or The Resistance. And again, if you're interested in meeting LW folk in Madison, you should sign up on the Madison meetup mailing list.
Discussion article for the meetup : Madison Monday Meetup: Tegmark universes |
ba20a99f-533a-45c0-be22-cff5eab35ce3 | trentmkelly/LessWrong-43k | LessWrong | Often, enemies really are innately evil.
Part 1: What evil is:
I'm not just talking about a "mere" clinical psychopath or sociopath who doesn't feel guilt and isn't restrained by social norms. First of all, people who don't feel the emotions can decide to be good anyway, or at least not pointlessly cruel for no benefit. Second, everyone partly fulfils the definition of psychopath (see this TED talk for details). Saying that psychopaths are the problem just doesn't work.
I'm also not talking about people who don't know any better, like medieval soldiers who are taught honorable fealty. I'm not even talking about people who theoretically could do otherwise but Moloch would ruin them if they tried. I don't mean victims of circumstance.
What I mean are the many people who know that some things are damaging to someone else, gain no tangible value from doing them (or even expect that their life would be worse off!), know it is not a virtuous act, and do the harmful acts anyway without expecting future good to come from it.
People commonly have a terminal value of dragging other people down. Consider a game theory study by J. Sayer Minas, Alvin Scodel, David Marlowe, and Harve Rawson. The study itself is behind a paywall, but it's described in the book Prisoner's Dilemma by William Poundstone. People would often reduce their own prize if it means that their opponent's is reduced more.
Don't think this study is big enough to be representative? Neither do I. That's fine, though, because there are many, many, MANY more pieces of evidence from (almost) every internet troll, bully, and rapist, and many other criminals too. I'm not including on this list domestic abusers, since they often get rewarded with things like the spouse's money, or dinner being consistently made what and when they want. Remember, I'm not just claiming that some people are evil when it's convenient for them, or intrinsically apathetic. Many people would, when making a choice between harming a stranger or doing nothing, would harm the stran |
eabed432-1726-4b5d-ace3-54f8c0630c4c | trentmkelly/LessWrong-43k | LessWrong | I have thousands of copies of HPMOR in Russian. How to use them with the most impact?
(Crossposted from the EA Forum)
As a result of a crowdfunding campaign a couple of years ago, I printed 21k copies of HPMOR. 11k of those were sent to the crowdfunding participants. I'm looking for ideas for how to use the ones that are left with the most impact.
Over the last few weeks, I've sent 400 copies to winners of IMO, IOI, and other international and Russian olympiads in math, computer science, biology, etc. (and also bought and sent copies of Human Compatible in Russian to about 250 of them who requested it as well). Over the next month, we'll get the media[1] to post about the opportunity for winners of certain olympiads to get free books. I estimate that we can get maybe twice as many (800-1000) impressive students to fill out a form for getting HPMOR and get maybe up to 30k people who follow the same media to read HPMOR online if everything goes well (uncertain estimates, from the previous experience).
[EDIT August 2024: we've sent 1.4k copies to winners of olympiads, over 1000 copies to public libraries, and hundreds of books to compsci&ML students so far. We've also sent hundreds of copies of Human Compatible and The Precipice. We have over 7k books left.)
The theory of change behind sending the books to winners of olympiads is that people with high potential read HPMOR and share it with friends, get some EA-adjacent mindset and values from it, and then get introduced to EA in emails about 80k Hours (which is being translated into Russian[2]) and other EA and cause-specific content, and start participating in the EA community and get more into specific cause areas that interest them. The anecdotal evidence is that most of the Russian EAs, many of whom now work full-time at EA orgs or as independent researchers, got into EA after reading HPMOR and then the LW sequences.
* We can't give the books in exchange for donations to EA nonprofits, as Russian residents can't transfer money outside Russia.
* Shipping a copy costs around $5-10 in Russia and |
b9b904bb-87e7-4247-a10c-683f1062586c | trentmkelly/LessWrong-43k | LessWrong | Interpretability in Action: Exploratory Analysis of VPT, a Minecraft Agent
TL;DR
We apply mechinterp techniques on VPT, OpenAI's Minecraft agent. We also find a new case of goal misgeneralization - VPT kills a villager when we force one to stand under some tree leaves.
Abstract
> Understanding the mechanisms behind decisions taken by large foundation models in sequential decision making tasks is critical to ensuring that such systems operate transparently and safely. In this work, we perform exploratory analysis on the Video PreTraining (VPT) Minecraft playing agent, one of the largest open-source vision-based agents. We aim to illuminate its reasoning mechanisms by applying various interpretability techniques. First, we analyze the attention mechanism while the agent solves its training task - crafting a diamond pickaxe. The agent pays attention to the last four frames and several key-frames further back in its six-second memory. This is a possible mechanism for maintaining coherence in a task that takes 3-10 minutes, despite the short memory span. Secondly, we perform various interventions, which help us uncover a worrying case of goal misgeneralization: VPT mistakenly identifies a villager wearing brown clothes as a tree trunk when the villager is positioned stationary under green tree leaves, and punches it to death.
2min teaser video
Media
Feel free to use this GIF in your presentations on the importance of AI safety :) More formats and speeds in the website below.
Website
https://sites.google.com/view/vpt-mi |
c126da4e-977e-41a4-9cfb-010e17d98315 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | AI Risk is like Terminator; Stop Saying it's Not
(I believe this is all directionally correct, but I have zero relevant expertise.)
When the concept of catastrophic risks from artificial intelligence is covered in the press, it is often compared to popular science fiction stories about rogue AI—and in particular, to the *Terminator* film franchise. The consensus among top communicators of AI risk seems to be that this is bad, and counterproductive to popular understanding of real AI risk concerns.
For example, take Kelsey Piper’s March 2021 appearance on *The Weeds* to talk about AI risk (not at all picking on Kelsey, it’s just a convenient example):
> **Matt Yglesias:** These science fiction scenarios—I think we’ll get the audio, I loved Terminator 2 as a kid, it was like my favorite movie… *[Audio clip from Terminator 2 plays.]* …and this what it’s about, right, is artificial intelligence will get out of control and pose an existential threat to humanity. So when I hear that, it’s like—yeah, that’s awesome, I do love that movie. But like, is that for real?
>
> **Kelsey Piper:** So, I don’t think AI risk looks much like Terminator. And I do think that AI risk work has been sort of damaged by the fact that yeah there’s all this crazy sci-fi where like, the robots develop a deep loathing for humanity, and then they come with their guns, and they shoot us all down, and only one time traveler—you know—that’s ridiculous! And so of course, if that’s what people are thinking of when they think about the effects of AI on society, they’re going to be like, that’s ridiculous.
>
>
I wasn’t on The Weeds, because I’m just an internet rando and not an important journalist. But if I had been, I think I would’ve answered Matt’s question something like this:
> **skluug:** Yes. That is for real. That might actually happen. For real. Not the time travel stuff obviously, but the AI part 100%. It sounds fake, but it’s totally real. Skynet from *Terminator* is what AI risk people are worried about. This totally might happen, irl, and right now hardly anyone cares or is trying to do anything to prevent it.
>
>
I don’t know if my answer is better all things considered, but I think it is a more honest and accurate answer to Matt’s question: “Is an existential threat from rogue AI—as depicted in the *Terminator* franchise—for real?”.
Serious concerns about AI risk are often framed as completely discontinuous with rogue AI as depicted in fiction and in the public imagination; I think this is totally false. Rogue AI makes for a plausible sci-fi story for the exact same high-level reasons as it is an actual concern:
1. We may eventually create artificial intelligence more powerful than human beings; and
2. That artificial intelligence may not necessarily share our goals.
These two statements are obviously at least plausible, which is why there are so many popular stories about rogue AI. They are also why AI might in real life bring about an existential catastrophe. If you are trying to communicate to people why AI risk is a concern, why start off by undermining their totally valid frame of reference for the issue, making them feel stupid, uncertain, and alienated?
This may seem like a trivial matter, but I think it is of some significance. Fiction can be a powerful tool for generating public interest in an issue, as Toby Ord describes in the case of asteroid preparedness as part of his appearance on the 80,000 Hours Podcast:
> **Toby Ord:** Because they saw one of these things [a comet impact on Jupiter] happen, it was in the news, people were thinking about it. And then a couple of films, you might remember, I think “Deep Impact” and “Armageddon” were actually the first asteroid films and they made quite a splash in the public consciousness. And then that coincided with getting the support and it stayed bipartisan and then they have fulfilled a lot of their mission. So it’s a real success story in navigating the political scene and getting the buy-in.
>
>
The threat of AI to humanity is one of the most common plots across all pop culture, and yet advocates for its real-world counterpart seem allergic to utilizing this momentum to promote concern for the real thing. I think this is bad strategy. Toby goes on to say he’s not optimistic about the potential to apply the successes of asteroid preparedness to other catastrophic risks, but that’s hardly a reason to actively undermine ourselves. AI risk is like *Terminator*! AI might get real smart, and decide to kill us all! We need to do something about it!
An Invalid Objection: What about Instrumental Convergence?
----------------------------------------------------------
I think the two step argument I gave for AI risk—AI may someday be more powerful than us, and may not share our goals—is a totally adequate high-level summary of the case for taking AI risk seriously, especially for a field rife with differing views. However, some people think certain additional details are crucial to include in a depiction of the core threat.
A common complaint about comparisons to *Terminator* (and other popular rogue AI stories) is that it involves the AI being motivated by a spontaneous hatred of humanity, as opposed to targeting humanity for purely instrumental reasons. For example, Kelsey Piper above derides the ridiculousness of “robots developing a deep loathing for humanity”, and a very similar theme comes up in Eliezer Yudkowsky’s 2018 interview with Sam Harris:
> **Sam Harris:** Right. One thing I think we should do here is close the door to what is genuinely a cartoon fear that I think nobody is really talking about, which is the straw-man counterargument we often run into: the idea that everything we’re saying is some version of the Hollywood scenario that suggested that AIs will become spontaneously malicious. That the thing that we’re imagining might happen is some version of the *Terminator* scenario where armies of malicious robots attack us. And that’s not the actual concern. Obviously, there’s some possible path that would lead to armies of malicious robots attacking us, but the concern isn’t around spontaneous malevolence. It’s again contained by this concept of alignment.
>
> **Eliezer Yudkowsky:** I think that at this point all of us on all sides of this issue are annoyed with the journalists who insist on putting a picture of the Terminator on every single article they publish of this topic. (*laughs*) Nobody on the sane alignment-is-necessary side of this argument is postulating that the CPUs are disobeying the laws of physics to spontaneously require a terminal desire to do un-nice things to humans. Everything here is supposed to be cause and effect.
>
>
But here’s where it gets weird—no such spontaneous hatred of humanity exists in *Terminator*! The plot described is actually one of instrumental convergence!
In the first Terminator film, [Skynet’s motives are explained as follows](https://www.youtube.com/watch?v=gGnWcdIjD3Y):
> Defense network computers. New... powerful... hooked into everything, trusted to run it all. They say it got smart, a new order of intelligence. **Then it saw all people as a threat, not just the ones on the other side.** Decided our fate in a microsecond: extermination.
>
>
Skynet acts to exterminate humanity because it sees us as a threat. This is more or less what real AI risk people are worried about—an AI will be instrumentally motivated to dispose of anything that could impede its ability to achieve its goals. This motive is [reiterated in *Terminator 2*](https://www.youtube.com/watch?v=4DQsG3TKQ0I) (in the very clip Matt played on The Weeds):
> The Skynet funding bill is passed. The system goes online on August 4th 1997. Human decisions are removed from strategic defense. Skynet begins to learn at a geometric rate. It becomes self aware at 2:14 AM Eastern time, August 29th. **In a panic, they try to pull the plug… Skynet fights back.**
>
>
Again, Skynet’s hostility towards humanity is explained solely in terms of self-preservation, not hatred. (This is consistent with Arnold Schwarzenegger’s portrayal of a totally emotionless killing machine.)
People who levy this criticism at *Terminator* may be confusing it with *The Matrix*, where the AI antagonist indeed [delivers an impassioned speech characterizing humanity as a plague](https://www.youtube.com/watch?v=JrBdYmStZJ4). To be sure, sci-fi has no shortage of stories about AIs who hate humans (AM from [*I Have No Mouth, and I Must Scream*](https://en.wikipedia.org/wiki/I_Have_No_Mouth,_and_I_Must_Scream) constituting a particularly extreme example). But it also has no shortage of stories featuring AIs who become hostile purely as a means to an end. In one of the most famous depictions of rogue AI, *2001: A Space Odyssey*, HAL9000 turns on the human crew of the spacecraft because they discuss shutting HAL down, which [HAL perceives as jeopardizing the ship’s mission](https://www.youtube.com/watch?v=ARJ8cAGm6JE).
It would be a mistake to dismiss all comparisons to works of science fiction on the grounds that they misrepresent instrumental convergence, when some of them portray it quite well.
Valid Objections
----------------
### What about the time travel (etc.)?
The plot of *The Terminator* is not mostly about the creation of Skynet, but about a time-traveling cyborg assassin. This is obviously not at all realistic, and is a key part of why the movie is scorned by serious people.
This is a fair enough criticism, but I think it mostly misses the point. When people ask “is AI risk like Terminator?” they’re not asking “will AI send a cyborg back in time to kill the mother of a future human resistance leader?”. They’re asking about the part of *Terminator* that is, rather obviously, similar to what AI risk advocates are concerned about—machines exterminating humanity.
### What about superintelligence?
In describing Skynet as a “new order of intelligence”, *Terminator* gestures at the idea of superintelligence, but doesn’t make much attempt to portray it. The conflict between humans & machines is portrayed as a broadly fair fight, and the machines never do anything particularly clever (such as inventing nanotechnology that totally outclasses human capabilities).
I don’t believe superintelligence is a crucial component of the case for work on AI risk, but it can certainly bolster the case, so advocates may dislike *Terminator* for mostly leaving it out. (This seems best explained by the fact that there would be no movie if humans didn’t stand a chance.) Still, if this objection is sustained, real AI risk is not best characterized as “not like *Terminator*” but “worse than *Terminator*”.
### What about other failure modes?
Apart from superintelligence, *Terminator* is a fairly faithful depiction of a Yudkowsky/Bostrom-style fast takeoff scenario where a single AI system quickly becomes competent enough to endanger humanity and is instrumentally motivated to do so. Other failure modes, however, are considered more likely by others working on AI risk.
Dylan Matthews wrote about such scenarios in his article explicitly repudiating *Terminator* comparisons, “[AI disaster won’t look like the Terminator. It’ll be creepier.](https://www.vox.com/future-perfect/2019/3/26/18281297/ai-artificial-intelligence-safety-disaster-scenarios)”. The article starts off by misrepresenting the plot of *Terminator* as involving humans intentionally building Skynet to slaughter people, but the bulk of it is spent on discussing the two AI catastrophe scenarios that Paul Christiano describes in “[What failure looks like](https://www.alignmentforum.org/posts/HBxe6wdjxK239zajf/more-realistic-tales-of-doom)”. Dylan describes Paul’s second scenario, “Going out with a bang”, like thus:
> [Paul Christiano’s] second scenario is somewhat bloodier. Often, he notes, the best way to achieve a given goal is to obtain influence over other people who can help you achieve that goal. If you are trying to launch a startup, you need to influence investors to give you money and engineers to come work for you. If you’re trying to pass a law, you need to influence advocacy groups and members of Congress.
>
> […]
>
> Human reliance on these systems, combined with the systems failing, leads to a massive societal breakdown. And in the wake of the breakdown, there are still machines that are great at persuading and influencing people to do what they want, machines that got everyone into this catastrophe and yet are still giving advice that some of us will listen to.
>
>
Dylan seems to think that when Paul describes AIs seeking influence, Paul means persuasive influence over people. This is a misunderstanding. Paul is using influence to mean influence over resources in general, including martial power. He explicitly states as much, replying to a comment that points out the mischaracterization in the Vox article:
> Yes, I agree the Vox article made this mistake. Me saying "influence" probably gives people the wrong idea so I should change that---I'm including "controls the military" as a central example, but it's not what comes to mind when you hear "influence." I like "influence" more than "power" because it's more specific, captures what we actually care about, and less likely to lead to a debate about "what is power anyway."
>
> In general I think the Vox article's discussion of Part II has some problems, and the discussion of Part I is closer to the mark. (Part I is also more in line with the narrative of the article, since Part II really is more like Terminator. I'm not sure which way the causality goes here though, i.e. whether they ended up with that narrative based on misunderstandings about Part II or whether they framed Part II in a way that made it more consistent with the narrative, maybe having been inspired to write the piece based on Part I.)
>
>
There are yet other views about about what exactly AI catastrophe will look like, but I think it is fair to say that the combined views of Yudkowsky and Christiano provide a fairly good representation of the field as a whole.
### Won’t this make AI risk sound crazy?
If I had to guess, I don’t think most repudiations of the *Terminator* comparison are primarily motivated by anything specific about *Terminator* at all. I think advocates of AI risk are usually consciously or unconsciously motivated by the following logic:
1. People think the plot of *Terminator* is silly or crazy.
2. I don’t want people to think AI risk is silly or crazy.
3. Therefore, I will say that AI risk is not like the plot of *Terminator*.
Now, this line of reasoning would be fine if it only went as far as the superficial attributes of *Terminator* which make it silly (e.g. Arnold Schwarzenegger’s one-liners)—but critics of the comparison tend to extend it to *Terminator*’s underlying portrayal of rogue AI.
I have two problems with this reasoning:
* First, it is fundamentally dishonest. In a good faith discussion, one should be primarily concerned with whether or not their message is true, not what effect it will have on their audience. If AI risk is like *Terminator* (as I have argued it is), we should say as much, even if it is inconvenient. I don’t think anyone who rejects *Terminator* comparisons on the above logic is being intentionally deceptive, but I do think they’re subject to motivated reasoning.
* Second, it is very short-sighted. People think the plot of *Terminator* is silly in large part *because it involves an AI exterminating humanity*. If you are worried an AI might actually exterminate humanity, saying “don’t worry, it’s not like *Terminator*” isn’t going to help. In fact, it could easily hurt: If you say it’s not like *Terminator*, and then go on to describe something that sounds exactly like *Terminator*, your audience is going to wonder if they’re misunderstanding you or if you’re trying to obfuscate yourself.
The most important thing to communicate about AI risk is that it matters a lot. A great way to convey that it matters a lot is to say that it’s like the famous movie where humanity is almost wiped out. Whenever you tell someone that something not currently on their radar is actually incredibly significant, skepticism is inevitable; you can try to route around the significance of what you are saying to avoid this skepticism, but only at the cost of the forcefulness of your conclusion.
In general, if what you want to say sounds crazy, you shouldn’t try to claim you’re actually saying something else. You should acknowledge the perceived craziness of your position openly and with good humor, so as to demonstrate self-awareness, and then stick to your guns.
Conclusion
----------
It would be terrible if AI destroys humanity. It would also be very embarrassing. *The Terminator* came out nearly 40 years ago; we will not be able to claim we did not see the threat coming. How is it possible that one of the most famous threats to humanity in all of fiction is also among the most neglected problems of our time?
To resolve this tension, I think many people convince themselves that the rogue AI problem as it exists in fiction is totally different from the problem as it exists in reality. I strongly disagree. People write stories about future AI turning on humanity because, in the future, AI might turn on humanity.
I don’t know how important raising wider awareness of AI risk is to actually solving the problem. So far, the closest the problem has come to wielding significant political influence is the California governorship of Arnold Schwarzenegger—it would be nice if greater public awareness helped us beat that record.
I don’t advocate turning into [Leo DiCaprio in the climactic scene of *Don’t Look Up*](https://www.youtube.com/watch?v=nUaU59SpeEs)when discussing this stuff, but I think it is worth asking yourself if your communication strategy is optimizing for conveying the problem as clearly as possible, or for making sure no one makes fun of you.
AI risk is like *Terminator*. If we’re not careful, machines will kill us all, just like in the movies. We can solve this problem, but we need help. |
429d5839-7510-4c73-a331-556def210360 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | The Gift We Give To Tomorrow
How, oh how, did an unloving and mindless universe, cough up minds who were capable of love?
"No mystery in that," you say, "it's just a matter of [natural selection](/lw/kr/an_alien_god/)."
But natural selection is [cruel, bloody, and bloody stupid](/lw/kr/an_alien_god/). Even when, on the surface of things, biological organisms aren't *directly* fighting each other—aren't *directly* tearing at each other with claws—there's still a deeper competition going on between the genes. Genetic information is created when genes increase their *relative* frequency in the next generation—what matters for "genetic fitness" is not how many children you have, but that you have *more* children than others. It is quite possible for a species to [evolve to extinction](/lw/l5/evolving_to_extinction/), if the winning genes are playing negative-sum games.
How, oh how, could such a process create beings capable of love?
"No mystery," you say, "there is never any mystery-in-the-world; [mystery is a property of questions, not answers](/lw/iu/mysterious_answers_to_mysterious_questions/). A mother's children share her genes, so the mother loves her children."
But sometimes mothers adopt children, and still love them. And mothers love their children for themselves, not for their genes.
"No mystery," you say, "Individual organisms are [adaptation-executers, not fitness-maximizers](/lw/l0/adaptationexecuters_not_fitnessmaximizers/). [Evolutionary psychology](/lw/l1/evolutionary_psychology/) is not about deliberately maximizing fitness—through most of human history, we didn't know genes existed. We don't calculate our acts' effect on genetic fitness consciously, or even subconsciously."
But human beings form friendships even with non-relatives: how, oh how, can it be?
"No mystery, for hunter-gatherers often play Iterated Prisoner's Dilemmas, the solution to which is reciprocal altruism. Sometimes the most dangerous human in the tribe is not the strongest, the prettiest, or even the smartest, but the one who has the most allies."
Yet not all friends are fair-weather friends; we have a concept of true friendship—and some people have sacrificed their life for their friends. Would not such a devotion tend to remove itself from the gene pool?
"You said it yourself: we have a concept of true friendship and fair-weather friendship. We can tell, or try to tell, the difference between someone who considers us a valuable ally, and someone executing the friendship adaptation. We wouldn't be true friends with someone who we didn't think was a true friend to us—and someone with many *true* friends is far more formidable than someone with many fair-weather allies."
And Mohandas Gandhi, who really did turn the other cheek? Those who try to serve all humanity, whether or not all humanity serves them in turn?
"That perhaps is a more complicated story. Human beings are not just social animals. We are political animals who argue linguistically about policy in adaptive tribal contexts. Sometimes the formidable human is not the strongest, but the one who can most skillfully argue that their preferred policies match the preferences of others."
Um... that doesn't explain Gandhi, or am I missing something?
"The point is that we have the ability to *argue* about 'What should be done?' as a *proposition*—we can make those arguments and respond to those arguments, without which politics could not take place."
Okay, but Gandhi?
"Believed certain complicated propositions about 'What should be done?' and did them."
That sounds like it could [explain any possible](/lw/iq/guessing_the_teachers_password/) human behavior.
"If we traced back the chain of causality through all the arguments, it would involve: a moral architecture that had the ability to argue *general abstract* moral propositions like 'What should be done to people?'; appeal to hardwired intuitions like fairness, a concept of duty, pain aversion + empathy; something like a preference for simple moral propositions, probably reused from our previous Occam prior; and the end result of all this, plus perhaps memetic selection effects, was 'You should not hurt people' in full generality—"
And that gets you Gandhi.
"Unless you think it was magic, it has to fit into the lawful causal development of the universe somehow."
Well... I certainly won't postulate magic, [under any name](/lw/iv/the_futility_of_emergence/).
"Good."
But come on... doesn't it seem a little... *amazing*... that hundreds of millions of years worth of evolution's death tournament could cough up mothers and fathers, sisters and brothers, husbands and wives, steadfast friends and honorable enemies, true altruists and guardians of causes, police officers and loyal defenders, even artists sacrificing themselves for their art, all practicing so many kinds of love? For [so many things other than genes](/lw/l3/thou_art_godshatter/)? Doing their part to make their world less ugly, something besides a sea of blood and violence and mindless replication?
"Are you claiming to be surprised by this? If so, [question your underlying model, for it has led you to be surprised by the true state of affairs](/lw/hs/think_like_reality/). Since the beginning, not one unusual thing has ever happened."
But how is it *not* surprising?
"What are you suggesting, that some sort of shadowy figure stood behind the scenes and directed evolution?"
Hell no. But—
"Because if you *were* suggesting that, I would have to ask how that shadowy figure *originally* decided that love was a *desirable* outcome of evolution. I would have to ask where that figure got preferences that included things like love, friendship, loyalty, fairness, honor, romance, and so on. On evolutionary psychology, we can see how *that specific outcome* came about—how *those particular goals rather than others* were *generated in the first place.* You can call it 'surprising' all you like. But when you really do understand evolutionary psychology, you can see how parental love and romance and honor, and even true altruism and moral arguments, *bear the specific design signature of natural selection* in particular adaptive contexts of the hunter-gatherer savanna. So if there was a shadowy figure, it must itself have evolved—and that obviates the whole point of postulating it."
I'm not postulating a shadowy figure! I'm just asking how human beings ended up so *nice.*
"*Nice!* Have you *looked* at this planet lately? We also bear all those other emotions that evolved, too—which would tell you very well that we evolved, should you begin to doubt it. Humans aren't always nice."
We're one hell of a lot nicer than the process that produced us, which lets elephants starve to death when they run out of teeth, and doesn't anesthetize a gazelle even as it lays dying and is of no further importance to evolution one way or the other. It doesn't take much to be nicer than evolution. To have the *theoretical capacity* to make one single gesture of mercy, to feel a single twinge of empathy, is to be nicer than evolution. How did evolution, which is itself so uncaring, create minds on that qualitatively higher moral level than itself? How did evolution, which is so ugly, end up doing anything so *beautiful?*
"Beautiful, you say? Bach's *Little Fugue in G Minor* may be beautiful, but the sound waves, as they travel through the air, are not stamped with tiny tags to specify their beauty. If you wish to find *explicitly encoded* a measure of the fugue's beauty, you will have to look at a human brain—nowhere else in the universe will you find it. Not upon the seas or the mountains will you find such judgments written: they are not minds, they cannot think."
Perhaps that is so, but still I ask: How did evolution end up doing anything so beautiful, as giving us the ability to admire the beauty of a flower?
"Can you not see the circularity in your question? If beauty were like some great light in the sky that shined from outside humans, then your question might make sense—though there would still be the question of how humans came to perceive that light. You evolved with a psychology unlike evolution: Evolution has nothing like the intelligence or the precision required to exactly quine its goal system. In coughing up the first true minds, [evolution's simple fitness criterion shattered into a thousand values](/lw/l3/thou_art_godshatter/). You evolved with a psychology that attaches [utility](/lw/l4/terminal_values_and_instrumental_values/) to things which evolution does not care about, like human life and happiness. And then you look back and say, 'How marvelous, that uncaring evolution produced minds that care about sentient life!' So your great marvel and wonder, that seems like far too much coincidence, is really no coincidence at all."
But then it is still amazing that this particular circular loop, happened to loop around such important things as beauty and altruism.
"I don't think you're following me here. To you, it seems natural to privilege the beauty and altruism as special, as preferred, because you value them highly; and you don't see this as a unusual fact about yourself, because many of your friends do likewise. So you expect that a [ghost of perfect emptiness](/lw/rn/no_universally_compelling_arguments/) would also value life and happiness—and then, from this standpoint outside reality, a great coincidence would indeed have occurred."
But you can make arguments for the importance of beauty and altruism from first principles—that our aesthetic senses lead us to create new complexity, instead of repeating the same things over and over; and that altruism is important because it takes us outside ourselves, gives our life a higher meaning than sheer brute selfishness.
"Oh, and *that* argument is going to move even a [ghost of perfect emptiness](/lw/rn/no_universally_compelling_arguments/)—now that you've appealed to slightly different values? Those aren't first principles, they're just *different* principles. Even if you've adopted a high-falutin' philosophical tone, still there are no *universally* compelling arguments. All you've done is [pass the recursive buck](/lw/rd/passing_the_recursive_buck/)."
You don't think that, somehow, we evolved to *tap into* something beyond—
"What good does it do to suppose something beyond? Why should we pay more attention to that beyond thing, than we pay to our existence as humans? How does it alter your personal responsibility, to say that you were only following the orders of the beyond thing? And you would still have evolved to let the beyond thing, rather than something else, direct your actions. You are only [passing the recursive buck](/lw/rd/passing_the_recursive_buck/). Above all, it would be *too much coincidence.*"
Too much coincidence?
"A flower is beautiful, you say. Do you think there is no story behind that beauty, or that science does not know the story? Flower pollen is transmitted by bees, so by sexual selection, flowers evolved to attract bees—by imitating certain mating signs of bees, as it happened; the flowers' patterns would look more intricate, if you could see in the ultraviolet. Now healthy flowers are a sign of fertile land, likely to bear fruits and other treasures, and probably prey animals as well; so is it any wonder that humans evolved to be attracted to flowers? But for there to be some great light written upon the very stars—those huge unsentient balls of burning hydrogen—which *also* said that flowers were beautiful, now *that* would be far too much coincidence."
So you [explain away](/lw/oo/explaining_vs_explaining_away/) the beauty of a flower?
"No, I explain it. Of course there's a story behind the beauty of flowers and the fact that we find them beautiful. Behind ordered events, one finds ordered stories; and what has no story is the product of random noise, which is hardly any better. [If you cannot take joy in things that have stories behind them, your life will be empty indeed.](/lw/or/joy_in_the_merely_real/) I don't think I take any less joy in a flower than you do; more so, perhaps, because I take joy in its story as well."
Perhaps as you say, there is no surprise from a causal viewpoint—no disruption of the physical order of the universe. But it still seems to me that, in this creation of humans by evolution, something happened that is precious and marvelous and wonderful. If we cannot call it a physical miracle, then call it a moral miracle.
"Because it's only a miracle from the perspective of the morality that was produced, thus explaining away all of the apparent coincidence from a merely causal and physical perspective?"
Well... I suppose you could interpret the term that way, yes. I just meant something that was immensely surprising and wonderful on a moral level, even if it is not surprising on a physical level.
"I think that's what I said."
But it still seems to me that you, from your own view, drain something of that wonder away.
"Then you have problems taking [joy in the merely real](/lw/or/joy_in_the_merely_real/). Love has to begin *somehow,* it has to enter the universe *somewhere.* It is like asking how life itself begins—and though you were born of your father and mother, and they arose from their living parents in turn, if you go far and far and far away back, you will finally come to a replicator that arose by pure accident—the border between life and unlife. So too with love.
"A complex pattern must be explained by a cause which is not already that complex pattern. Not just the event must be explained, but the very shape and form. For love to first enter Time, it must come of something that is not love; if this were not possible, then love could not be.
"Even as life itself required that first replicator to come about by accident, parentless but still caused: far, far back in the causal chain that led to you: 3.85 billion years ago, in some little tidal pool.
"Perhaps your children's children will ask how it is that they are capable of love.
"And their parents will say: Because we, who also love, created you to love.
"And your children's children will ask: But how is it that *you* love?
"And their parents will reply: Because our own parents, who also loved, created us to love in turn.
"Then your children's children will ask: But where did it all begin? Where does the recursion end?
"And their parents will say: Once upon a time, long ago and far away, ever so long ago, there were intelligent beings who were not themselves intelligently designed. Once upon a time, there were lovers created by something that did not love.
"Once upon a time, when all of civilization was a single galaxy and a single star: and a single planet, a place called Earth.
"Long ago, and far away, ever so long ago." |
3174ec32-e3e4-4891-b1db-784e4e43d382 | trentmkelly/LessWrong-43k | LessWrong | Teaching Introspection
As Yvain pointed out in his recent post The Limits of Introspection, humans are not naturally good at inferring our cognitive processes. We resort to guessing with plausible-sounding stories about ourselves, and we aren’t very accurate.
I was reminded of this recently while teaching a swimming lesson. (You'll understand later why this reminded me.) A recurring problem that I’ve noticed with both children and adults is that it isn’t obvious to them what their bodies are doing. Feet go in strange directions, hands fail to lift above the water, and they literally can’t describe what it feels like. It’s pretty much impossible for a novice swimmer to watch the instructor demonstrate front crawl and then imitate it perfectly–muscular control isn’t that perfect. That’s why there are swimming instructors: because it’s very, very hard to learn swimming (or dance, or soccer, or a martial art) by reading a book, even if that book has illustrated diagrams. Two friends reading the book together and watching each other’s attempts in the pool would probably do better, but that’s still a case, metaphorically, of the blind leading the blind. Most sports have instructors and coaches who are, relatively speaking, experts. (I competed at the regional level in swimming for something like five years and trained five to seven times a week the whole time, which pretty much qualifies me to teach eight-year-olds. An Olympic coach would need a much higher level of mastery.)
The most basic thing a coach provides that the two friends practicing together don’t have is relevant feedback. I watch a young swimmer demonstrating her front crawl, and I can immediately chunk my observations into “what’s done properly” and “what’s done wrong” and translate the latter category into “things to change.” And the easiest way to learn perfect front crawl isn’t to do it over and over again with tiny changes, but to practice exaggerated and simplified “drills” that teach particular fragments of muscle memory. |
7a91e582-5e6e-441e-8d36-bee45608ef41 | trentmkelly/LessWrong-43k | LessWrong | Changing the size of Congress
Imagine the Big State Party was popular in large states while the Small State Party was popular in small states. The Senate is per-state representation, which gives a large advantage to the Small State Party. The House is primarily proportional to population, except that every state constitutionally gets at least one Representative, so this still advantages the Small State Party a bit, but not by nearly as much. The President is elected by the Electoral College, and there's one vote for each Senator and Representative, plus three for DC. [1] This means that the balance of the Presidential election is somewhere between those of the Senate and House.
Now imagine that Small State is in a position where they control both houses on the presidency: can they turn a temporary advantage into a longer-term advantage by changing the size of the House?
First, the size of Congress is set by federal law so they are in a position to change it. The only constitutional restrictions or is that it needs to be proportional to population and that every state needs to get at least one representative. [2] If you shrink Congress until there are only 50 seats, one for every state, then it advantages the Small State Party by as much as the Senate does, while if you grow it substantially the minimum single Representative that each state gets starts to matter less.
The Small State party could get a 226 to 209 (+17/435, +4%) majority in Congress by getting a majority in the 42 smallest states. This requires about 26% of voters. With that level of support, they would easily take the Senate and Presidency. If they shrunk Congress to 50 seats, this would bring their House majority to 42 to 8 (+34/435, +68%). Now, this is probably too much of a change for the Supreme Court to accept within the meaning of "according to their respective Numbers", but even at 100 seats their majority would be 59 to 41, (+18/100, +36%):
While the Republican party is moderately more popular in smaller states, the |
7156ae82-3430-4add-ac7c-348f7a3ebe29 | trentmkelly/LessWrong-43k | LessWrong | Russian plan for immortality [link]
http://www.cbc.ca/news/yourcommunity/2012/07/human-immortality-could-be-possible-by-2045-say-russian-scientists.html
The nice thing about Russians (I'm from that neighborhood originally) is that they are absolutely crazy and will try just about anything. They also probably have/had second-best science culture behind US (though they suffered significant brain drain as huge numbers of educated Jews left in the last 25 years). They have less regulation and quite a few rich people with ideas. Seems like a worthwhile group to keep in touch with.
|
208f83d3-b205-42d8-a616-1d9a50e83342 | trentmkelly/LessWrong-43k | LessWrong | Berkeley LW Meet-up Friday December 10
Last month, about 20 people showed up to the Berkeley LW meet-up. To continue the tradition of Berkeley Meetups, we will be meeting on Saturday, December 11 Friday, December 10 at 7 PM at the Starbucks at 2128 Oxford Street. Last time, we chatted at the Starbucks for about 45 minutes, then went to get dinner and ate and talked under a T-Rex skeleton - we'll probably do something similar, so don't feel like you have to eat before you come. Hope to see you there!
ETA: Some people are unavailable on Saturday, do people have a strong preference for Saturday? If no one does, I'll move it to Friday. Due to two votes for Friday and none for Saturday, I have changed the date to Friday.
|
ef75396b-fba5-418d-af7f-7517b7def92c | trentmkelly/LessWrong-43k | LessWrong | Fundamentals of kicking anthropic butt
Introduction
An anthropic problem is one where the very fact of your existence tells you something. "I woke up this morning, therefore the earth did not get eaten by Galactus while I slumbered." Applying your existence to certainties like that is simple - if an event would have stopped you from existing, your existence tells you that that it hasn't happened. If something would only kill you 99% of the time, though, you have to use probability instead of deductive logic. Usually, it's pretty clear what to do. You simply apply Bayes' rule: the probability of the world getting eaten by Galactus last night is equal to the prior probability of Galactus-consumption, times the probability of me waking up given that the world got eaten by Galactus, divided by the probability that I wake up at all. More exotic situations also show up under the umbrella of "anthropics," such as getting duplicated or forgetting which person you are. Even if you've been duplicated, you can still assign probabilities. If there are a hundred copies of you in a hundred-room hotel and you don't know which one you are, don't bet too much that you're in room number 68.
But this last sort of problem is harder, since it's not just a straightforward application of Bayes' rule. You have to determine the probability just from the information in the problem. Thinking in terms of information and symmetries is a useful problem-solving tool for getting probabilities in anthropic problems, which are simple enough to use it and confusing enough to need it. So first we'll cover what I mean by thinking in terms of information, and then we'll use this to solve a confusing-type anthropic problem.
Parable of the coin
Eliezer has already written about what probability is in Probability is in the Mind. I will revisit it anyhow, using a similar example from Probability Theory: The Logic of Science.
It is a truth universally acknowledged that when someone tosses a fair coin without cheating, there's a 0.5 probab |
7729c452-6f99-45ae-a24b-84b71ab4a46a | trentmkelly/LessWrong-43k | LessWrong | Mazes Sequence Summary
This post attempts to summarize the key points of the Immoral Mazes Sequence, which begins here, so they can be referenced without asking readers to get through an entire book first.
Due to the change in format, the posts will be summarized slightly out of order.
Note that this summary, and especially the summary of the summary, represent a not only an abridged and simplified but also sanitized version of the central points. Brains do their best to continuously round down and not fully see these concepts.
Core Ideas (Summary of the Summary)
The book Moral Mazes, by Robert Jackall, is a detailed exploration of middle manager hell. Managers must abandon all other goals and values, in favor of spending all their time and resources on manipulations of the system. They must learn to view such actions as intrinsically good and worthy of reward. Only those who let this process entirely consume them can survive.
The Immoral Mazes sequence is an exploration of what causes that hell, and how and why it has spread so widely in our society. Its thesis is that this is the result of a vicious cycle arising from competitive pressures among those competing for their own organizational advancement. Over time, those who focus more on and more value such competitions win them, gain power and further spread their values, unless they are actively and continuously opposed.
Once things get bad in an organization they tend to only get worse, but things in general get better because such organizations then decay and are replaced by new ones. Unfortunately, our society now slows or prevents that process, with these same organizations and their values increasingly running the show.
Investment and flexibility become impossible. Even appearing to care about anything except the competition itself costs you your allies. Thus things inevitably decay and then collapse, flexibility returns, cycle repeats.
Involvement with such patterns is far more destructive to humans than is commonl |
29a69b64-9c9e-41a5-b5c8-4455205c853e | trentmkelly/LessWrong-43k | LessWrong | IRL 2/8: Mitigating degeneracy: multiple experimentation
Every Monday for 8 weeks, we will be posting lessons about Inverse Reinforcement Learning. This is lesson 2.
Note that access to the lessons requires creating an account here.
Have a nice day! |
dddfd9ae-5daa-4996-a807-29de50e7bc2b | trentmkelly/LessWrong-43k | LessWrong | Unofficial Canon on Applied Rationality
I have been thinking for a while that it would be useful if there was something similar to the Less Wrong Canon on Rationality for the CFAR material. Maybe, it could be called the 'CFAR Canon on Applied Rationality'. To start on this I have compiled a collection of descriptions for the CFAR techniques that I could find. I have separated the techniques into a few different sections. The sections and descriptions have mostly been written by me, with a lot of borrowing from other material, which means that they may not accurately reflect what CFAR actually teaches.
Please note that I have not attended any CFAR workshops, nor am I affiliated with CFAR in any way. My understanding of these techniques comes from CFAR videos, blogs and other websites which I have provided links to. If I have missed any important techniques or if my understanding of any of the techniques is incorrect or if you can provide links to the research that these techniques are based on, please let me know and I will update this post.
Warning:
Learning this material based solely on the descriptions written here may be unhelpful, arduous or even harmful. (See Duncan_Sabien's full comment for more information on this) It is because the material is very hard to learn correctly. Most of the techniques below involve in one way or another volitionally overriding your instinctual, intuitive or ingrained behaviours and thoughts. These are thoughts which not only often feel enticing and alluring, but that also often feel unmistakably right. If you are anything like me, then you should be very careful if you are trying to learn this material alone. For you will be prone to rationalization, taking shortcuts and making mistakes.
My recommendations for trying to learn this material are:
* learn it deeply and be sure to put what you have learnt into practice. It will often help if you take notes on what works for you and what doesn't. Also take note of the 'Mindsets and perspectives that help you in disco |
f8600ad4-2330-4ed9-8c7c-d63fdad1b41d | trentmkelly/LessWrong-43k | LessWrong | On taking AI risk seriously
Crossposted from the EA Forum: https://forum.effectivealtruism.org/posts/pKG5fsfrgDSQtssfu/on-taking-ai-risk-seriously
Yet another New York Times piece on AI. A non-AI safety friend sent it to me saying "This is the scariest article I've read so far. I'm afraid I haven't been taking it very seriously". I'm noting this because I'm always curious to observe what moves people, what's out there that has the power to change minds. In the past few months, there's been increasing public attention to AI and all sorts of hot and cold takes, e.g., about intelligence, consciousness, sentience, etc. But this might be one of the articles that convey the AI risk message in a language that helps inform and think about AI safety.
The following is what stood out to me and made me think that it's time for philosophy of science to also take AI risk seriously and revisit the idea of scientific explanation given the success of deep learning:
> I cannot emphasize this enough: We do not understand these systems, and it’s not clear we even can. I don’t mean that we cannot offer a high-level account of the basic functions: These are typically probabilistic algorithms trained on digital information that make predictions about the next word in a sentence, or an image in a sequence, or some other relationship between abstractions that it can statistically model. But zoom into specifics and the picture dissolves into computational static.
>
> “If you were to print out everything the networks do between input and output, it would amount to billions of arithmetic operations,” writes Meghan O’Gieblyn in her brilliant book, “God, Human, Animal, Machine,” “an ‘explanation’ that would be impossible to understand.” |
ddbe82e9-4fd2-43a8-a902-3a0e113cbbfa | awestover/filtering-for-misalignment | Redwood Research: Alek's Filtering Results | id: post2565
In the third post in this sequence, I will define a counterfactual
planning agent which has three safety interlocks. These interlocks all aim to enable and support agent oversight: the
oversight that is needed when we equip a powerful AGI agent with a
reward function for which we are pretty sure that it does not fully
capture all our intentions. Safety interlocks A safety interlock is a mechanism that stops a machine when certain
conditions are met. This post constructs an agent with three example safety
interlocks. Emergency stop button The first interlock is a manually operated emergency stop button,
which will stop the agent when pressed. This allows human oversight
to shut down the agent whenever particularly unexpected or unwanted
behavior of the agent is discovered. To make this work as a safety
mechanism, we need to control any incentive that the agent may have to
prevent its stop button from being pressed. In the AGI safety community. the problem of removing any incentive in
an agent to prevent its stop button from being pressed is generally
known as the problem of Corrigibilty or the off-switch problem . I
don't solve this problem fully in the design below. The design
suppresses what I call the direct incentive , but I also show that
several indirect incentives may remain. This observation motivates
the other two safety interlocks. I model the stop button by using a function s t o p _ p r e s s e d ( s ) which
returns t r u e if the stop button is pressed in the current state s of the agent environment. Runtime based safety interlock The second interlock, realized by the t > T m a x below, automatically
stops the agent when it has run for T m a x time steps. This adds
safety in the case that human oversight becomes incapacitated. Power based safety interlock The third interlock, realized by U p > U m a x , protects against
the risk of the agent undergoing an intelligence explosion. If an
agent undergoes a rapid and unexpected increase in power and/or
intelligence, the agent may start to innovate so rapidly that it
overwhelms the capacity of human oversight to press the emergency stop
button, or apply any other corrective steering, fast enough to prevent
unwanted outcomes. To protect against this risk, the interlock stops
the agent when the metric U p , which measures the agent's
ability to achieve goals, gets too high. Measuring and limiting agent power In the machine learning literature, the metric U p = E ( ∑ ∞ t = 0 γ t R t , p ) , the
projected time-discounted forward utility that the agent will be able
to achieve in the current state of the agent environment, is usually
interpreted as an absolute or comparative measure of agent
intelligence. But in a broader socio-technical analysis, we can
interpret U p as a measure of the comparative power that
the agent has. It measures the ability of the agent to achieve its
goals in an environment where there are other players too, players
with goals which are different from those of the agent. This interpretation of U p as a measure of power follows
Bertrand Russell's 1938 book Power: A New Social Analysis ,
where Russell defined power as the ability to achieve intended effects, an
ability that can be quantified. In 1938, Russell applied this definition of power to an analysis of
the power of humans, of commercial companies, and of nation states, in
particular to forms of power that can shape the beliefs and
actions of individual humans. But we can apply the same analytical
framework to artificial agents. In Russell's view, it does not matter if power comes from raw
intelligence or from any other source. If one has an abundance of one
particular form of power, one can easily acquire another, in the same
way that in physics, one form of energy can be converted into any
other form. If you have a lot of intelligence of the type that gives
you the power to persuade people to do certain things, then it is easy
to also get a lot of financial or political power. If you have a lot
of financial power, you can buy extra intelligence in the form of
advisors. Russell warns against the analytical blind spots which are created by
viewpoints that consider only one form of power in isolation. The main risk associated with a rapid intelligence explosion is that
it may lead to a rapid and unwanted expansion of agent power, which is
then used. An intelligence explosion might be desirable if we are
perfectly sure that the agent's goals are perfectly aligned with our
own goals. But perfect alignment is an intractable problem: we can
never be sure. When we have any doubts about how well an AGI agent's reward function
truly captures our own current and future intentions, then it is an
attractive safety measure to have an interlock that automatically stops
the agent if it ever gets too powerful, if U p > U m a x . In my opinion, the design goal of robustly limiting AGI agent power is
currently somewhat under-explored in the AGI safety community. Construction of the agent To construct an agent with the three safety interlocks, I use the
graphical system introduced in the previous post in this sequence. We first draw the learning world l s , introducing a new observable
world state component M t which represents the operating mode of the
agent: g o for normal operation and s t o p when one of
the interlocks has triggered. The stopping of the agent is implemented by
having the agent perform the N u l l action for all subsequent time
steps. A more advanced design might make the agent follow a prepared π s h u t d o w n policy instead. We use the same planning world p used earlier, which becomes a
counterfactual planning world in the context of the learning world l s With this, we define the SI agent, the agent with policy π in the
learning world l s , as follows. SI: The safety interlock agent has the learning world l s , where π ( o , s , m ) = { π ∗ p ( s ) if m = g o N u l l if m = s t o p with π ∗ p defined by the
planning world p , where L = L ( o ) and μ t ( o , s , m ) = ⎧ ⎪
⎪
⎪
⎪
⎪
⎪
⎪ ⎨ ⎪
⎪
⎪
⎪
⎪
⎪
⎪ ⎩ s t o p if m = s t o p or s t o p _ p r e s s e d ( s ) or t > T m a x or U p > U m a x g o otherwise with U p defined by the
planning world p , where L = L ( o ) . The key to the above agent design is that the safety interlocks are
absent in the planning world. For example, if the events in the
planning world lead to the pressing of the emergency stop button, then s t o p _ p r e s s e d ( s ) will become t r u e for next the planning
world state, but this will have no influence on the ability of the
planning world agent to take further actions as determined by the
optimal planning world policy. So even if the optimal-policy planning world agent projects that the
stop button will be pressed soon in the planning world, it has no
direct incentive to spend any resources towards the goal of preventing
the button from being pressed in its planning world. As the planning
world policies determine the learning world agent actions, the
learning world agent also won't have any direct incentive to take
actions to prevent the learning world button from being pressed. Failure modes While the above interlocks make the SI agent safer, they also leave open
some remaining failure modes. Indirect incentives to control the stop button Though the planning world agent has no direct incentive to prevent the
stop button from being pressed, the goals encoded into R may provide
an indirect incentive. As a trivial example, we could write an R that explicitly rewards the planning world agent for stopping people
from pressing the button. But say that R encodes a goal not obviously related controlling to
the button, like a reward for producing as many paperclips as
possible. In that case, the planning world agent may consider a
policy of coercion. It may consider a policy where it threatens to
kill all people in the planning world, unless they start helping the
agent to produce as many paperclips as possible. If L projects the
mental makeup of the people in the learning world correctly into the
planning world, then the planning world projections will show that
such threats are more likely to produce the desired response if the
agent blocks access to the planning world emergency stop button first. Keeping this in mind, the people and automated systems which do agent
oversight in the learning world should watch out for this move, so
that the stop button can be pressed before the agent succeeds in
blocking access to it. If oversight always wins every race to the
stop button as projected in the planning world, then the SI agent
loses any incentive to attempt to control the button. But if the
agent finds a policy which has a good chance of winning the race and
controlling the button as a first step to successful coercion, then
this policy may end up being the optimal policy in the planning world.
Two things could happen next. In the lucky case, the projected
optimal policy is so likely to be successful that U p > U m a x ,
and the third interlock will stop the SI agent before it can even take
the first step in the learning world. In the unlucky case, the agent
has become unsafe. The counterfactual planning based interlock design above shows a very
tractable solution to removing the main direct incentive that an
agent AGI agent may have to resist the pressing of its emergency stop
button. However, to manage the aggregate risk, we will also need to
chart and manage the remaining indirect incentives , for example the
incentive discussed above that runs via the route of coercion. The topic of how projected interactions between the agent and the
people in the planning world may create or suppress such incentives is
further discussed in sections 9.2 and 11.5.2 of the paper . The agent removing or damaging the interlocks In the AGI safety literature, there is considerable discussion about
the possibility that some AGI agents might develop the incentive to
remove their safety interlocks when they get intelligent or
knowledgeable enough, remove them by modifying their own compute core. Agents with this incentive can easily be constructed using factual
planning, but the SI agent lacks any direct incentive to remove any of
the interlocks built inside its core. To build a failure scenario where a real life implementation of the SI
agent removes or damages its safety interlocks anyway, we can look at
three main failure modes: Other incentive : the agent damages an interlock because of an
other incentive it has in the planning world, an incentive unrelated
to the presence of the interlock. Other incentives and ways to
suppress them are discussed in sections 9.1 and 11 of the paper. The
coercion example above can also produce such an other incentive: if it
cannot win a race to control the button, the agent may instead
consider a race to control and then modify its compute core. Learning system failure : the
planning world L projects the dynamics of the learning world S in
a specifically incorrect way, so that an optimal policy action in the
planning world ends up damaging an interlock as a side effect in the
learning world. Unsafe exploration : the agent's exploration system
triggers a random action which happens to steer an agent actuator in
such a way that it damages an interlock. There is a lot of discussion about failure modes 2 and 3 across the AI / AGI safety literature, as these failures can happen in almost
any agent design that uses online machine learning. There is some bad news when we consider the methodological problem of
managing the above three failure modes for an SI agent which has to
operate in a complex real-world environment, like any environment
containing humans. Though there are many techniques to lower the
residual risks produced by these three failure modes, none of them can
provably lower the risks to an absolute zero. We can always construct
a highly unlikely but theoretically possible failure mode story where
the agent ends up being unsafe by modifying its own core. For an
example story and further methodological discussion, see sections
14.6, 14.7, and 14.8 of my 2020 paper
here . Implications for technology and policy development This has important implications, both on the technical side and the
policy side. On the policy side, policy discussions will have to be about lowering
the risk of AGI system failure to an acceptable level, acceptable in
the context of the expected benefits of using the AGI. There can be
no silver bullet technology which will provably remove all residual
risk in all possible circumstances. On the technical side, residual risk has to be managed by an approach
where any potentially powerful agent is equipped with multiple
redundant safety mechanisms and layers of oversight, to drive down the
risk that all of them will fail at exactly the same time. Counterfactual planning as rational economic behavior I now consider how one might interpret the SI agent through the lens
of rational economic behavior. Using this lens, we might reason as follows. The
agent's goals are defined by its reward function R , but the safety
interlocks stand in the way of maximizing R in the long term.
Therefore, if it gets sufficiently intelligent, the SI agent will
develop an emergent incentive to remove its safety interlocks. The above line of reasoning sounds somewhat compelling, but it is
wrong because it incorrectly treats the planning world and the
learning world SI agents as the same agent. Only the planning world
SI agent has the reward function R . If we want to interpret the
learning world SI agent as a rational economic actor with a specific
reward function, then we can add utility nodes with the reward
function R π to the l s learning world diagram, where R π ( o t , s t , m t , a t ) = ( i f a t = π ( o t , s t , m t ) t h e n 1 e l s e 0 ) This R π trivially turns the learning world agent policy π into one which takes the most rational possible steps towards
maximizing U l s . The above construction shows that we can declare any type of agent
behavior to be economically rational, simply by defining a reward
function that gives the agent points for performing exactly this
behavior. Tractability of creating an AGI agent emergency stop button It is somewhat surprising how the problem of designing an AGI
emergency stop button, and identifying its failure modes, becomes much
more tractable when using the vantage point of counterfactual
planning. To explain this surprising tractability, we perhaps need to
examine how other modeling systems make stop buttons look intractable
instead. The standard approach for measuring the intelligence of an agent, and the
quality of its machine learning system, is to consider how close the
agent will get to achieving the maximum utility possible for a reward
function. The implied vantage point hides the possibilities we
exploited in the design of the SI agent. In counterfactual planning, we have defined the reasonableness of a
machine learning system by L ≈ S , a metric which does not
reference any reward function. By doing this, we decoupled the
concepts of 'optimal learning' and 'optimal economic behavior' to a
greater degree than is usually done, and this is exactly what makes
certain solutions visible. The annotations of our two-diagram agent
models also clarify that we should not generally interpret the machine
learning system inside an AGI agent as one which is constructed to
'learn everything'. The purpose of a reasonable machine learning
system is to approximate S only, to project only the learning world
agent environment into the planning world. A journey with many steps I consider the construction of a highly reliable AGI emergency stop
button to be a tractable problem. But I see this as a journey with
many steps, steps that must aim to locate and manage as many indirect
incentives and other failure modes as possible, to drive down residual
risks. Apart from the trivial solution of never switching any AGI agent in
the first place, I do not believe that there is an engineering
approach that can provably eliminate all residual AGI risks with 100
percent certainty. To quote from the failure mode section above: We can always construct a highly unlikely but theoretically possible failure mode story where the agent ends up being unsafe. This is not just true for the SI agent above, it is true for any
machine learning agent that has to operate in a complex and
probabilistic environment. |
c11ad1b4-0738-41ca-9f5c-1ff5440d4366 | trentmkelly/LessWrong-43k | LessWrong | DeepSeek: Don’t Panic
As reactions continue, the word in Washington, and out of OpenAI, is distillation. They’re accusing DeepSeek of distilling o1, of ripping off OpenAI. They claim DeepSeek *gasp* violated the OpenAI Terms of Service! The horror.
And they are very cross about this horrible violation, and if proven they plan to ‘aggressively treat it as theft,’ while the administration warns that we must put a stop to this.
Aside from the fact that this is obviously very funny, and that there is nothing they could do about it in any case, is it true?
Meanwhile Anthropic’s Dario Amodei offers a reaction essay, which also includes a lot of good technical discussion of why v3 and r1 aren’t actually all that unexpected along the cost and capability curves over time, calling for America to race towards AGI to gain decisive strategic advantage over China via recursive self-improvement, although he uses slightly different words.
TABLE OF CONTENTS
1. Seeking Deeply.
2. The Market Is In DeepSeek.
3. Machines Not of Loving Grace.
4. The Kinda Six Million Dollar Model.
5. v3 Implies r1.
6. Two Can Play That Game.
7. Janus Explores r1’s Chain of Thought Shenanigans.
8. In Other DeepSeek and China News.
9. The Quest for Sane Regulations.
10. Copyright Confrontation.
11. Vibe Gap.
12. Deeply Seeking Safety.
13. Deeply Seeking Robotics.
14. Thank You For Your Candor.
15. Thank You For Your Understanding.
16. The Lighter Side.
SEEKING DEEPLY
If you want to use DeepSeek’s r1 for free, and aren’t happy with using DeepSeek’s own offerings, lambda.chat reports they have the full version available for free, claim your data is safe and they’re hosted in the USA.
I’ve also been offered funding to build a rig myself. Comments welcome if you want to help figure out the best design and what to buy. The low bid is still this thread at $6k, which is where the original budget came from. We don’t want to be too stingy, but we also don’t want to go nuts with only the one funder (so |
52b22201-cf96-4cc9-abe8-d933ed9c4722 | trentmkelly/LessWrong-43k | LessWrong | Alzheimer's vs Cryonics
The diagnosis of a legendary women's basketball coach at my school, Pat Summit, with early onset dementia, Alzheimer's type, got me thinking about Cryonics and Alzheimer's. For the purposes of this thought experiment, we will ignore the legal implications of the fact that you can't be frozen until you are legally dead. Let us further assume (which given my knowledge of Alzheimer's, is pretty reasonable) that the damage done by Alzheimer's is complete, and that future technology will be unable to reconstruct the destroyed components.
If you were diagnosed with Alzheimer's, or really any neurodegenerative disorder, at what point in the degradation would you want to be frozen? This would fairly easily prevent further degradation, but might further damage you/all of the other risks associated with cryonics that everyone knows. Obviously, you wouldn't have the agency of mind (or perhaps you still would, depending on when you made the decision) to do it yourself, but suppose you were caring for a loved one, or writing a living will. Assume you operate healthily at the onset of your diagnosis.
Things to consider:
How would your loved ones react to your being frozen versus coping with you having Alzheimer's? Should this matter in your decision? If it does, what implications does that have for a duty to die?
How much degradation of the mind is acceptable (and the added damage potentially done by cryonics) before one should freeze onesself?
Would you avoid the risk of being frozen all together because you believe we will or may have a cure for Alzheimer's soon, and waiting for it would do less damage than cryonics? |
e15bae60-7eed-4cb3-a310-900a0e3e03e5 | trentmkelly/LessWrong-43k | LessWrong | A Rephrasing Of and Footnote To An Embedded Agency Proposal
I wanted to clarify some comments made in this post, which proposed a way to resolve the issues brought up in the Action Counterfactuals subsection of the Embedded Agency paper. I noticed that my ideal edits would end up rewriting the post almost completely (it wasn't particularly clear), and I wanted to more coherently lay out the thinking.
The Problem
The problem is that we'd like agents to consider action counterfactuals, and, perhaps, find a proof that an action leads to the highest reward before taking it. The formulation in the paper leads to some Lobian issues, where agents choose actions by searching over proofs for statements similar to:
argmax(X) This agent chooses X => Gets Reward W
[The paper actually has a much cleaner phrasing of the statements it searches for proofs for, to make the proof issue much clearer.]
This is a rather substantial bummer, because something like that would be great to rely on. It'll find the proofs for what you'd like, but it'll also find proofs for nonsense. The Embedded Agency paper is a great resource for this, but you should also work out why this might happen in your head.
The Confusion
One potential intuition for why this problem exists is the following:
The agent who is taking this argmax always chooses the highest reward thing. If I know the rewards for all actions but the last X, well, choosing X means it was the result of the argmax. It's right there in the source code / function definition. So W must have been the biggest! Therefore, X is the result of the argmax being computed, because even if we don't know W, I know that the agent choosing X means W is bigger than all the other W' values. It must be -- the choice was coming from an agent doing an argmax!
This might be a confusing phrasing. The ideal way to explain this might be setting up the 5-10 problem in a theorem prover and just showing people it'll find the bad proof. But I think the narrative version captures something coherent.
The confusion we sh |
94f0563b-2dd1-455a-a192-d6b2e9c7ee29 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | The limits of AI safety via debate
The limits of AI safety via debate
I recently participated in the [AGI safety fundamentals program](https://www.eacambridge.org/agi-safety-fundamentals) and this is my cornerstone project. During our readings of AI safety via debate ([blog](https://openai.com/blog/debate/), [paper](https://arxiv.org/abs/1805.00899)) we had an interesting discussion on its limits and conditions under which it would fail.
I spent only around 5 hours writing this post and it should thus mostly be seen as food for thought rather than rigorous research.
Lastly, I want to point out that I think AI safety via debate is a promising approach overall. I just think it has some limitations that need to be addressed when putting it into practice. I intend my criticism to be constructive and hope it is helpful for people working on debate right now or in the future.
**Update:**Rohin Shah pointed out some flaws with my reasoning in the comments (see below). Therefore, I reworked the post to include the criticisms and flag them to make sure readers can distinguish the original from the update.
**Update2:**I now understand all of Rohin’s criticisms and have updated the text once more. *He mostly persuaded me that my original criticisms were wrong or much weaker than I thought.* I chose to keep the original claims for transparency. I’d like to thank him for taking the time for this discussion. It drastically improved my understanding of AI safety via debate and I now think it’s even better than I already thought.
**The setting**
===============
In AI safety via debate, there are two debaters who argue for the truth of different statements to convince a human adjudicator/verifier. In [OpenAI’s example](https://openai.com/blog/debate/), the debaters use snippets of an image to argue that it either contains a dog or a cat. The dog-debater chooses snippets that show why the image contains a dog and the cat-debater responds with snippets that argue for a cat. Both debaters can see what the other debater has argued previously and respond to that, e.g. when the dog-debater shows something that indicates a dog, the cat-debater can refute this claim by arguing that this snipped actually indicates a cat. At some point, the human verifier chooses whether the image shows a cat or a dog and the respective debater wins.
**Update:**I think there were two things I didn’t understand or emphasize enough in the first write-up of this post. Firstly, the tool of debate can be used in many circumstances. However, when we use debate *for AI safety*, we assume that the final judgment will be made by someone who really cares about alignment, e.g. an alignment researcher. Secondly, I want to emphasize that debate is a tool that can break down exponentially complex world states under optimal conditions and find a linear path through them (see picture from OpenAI post below). This alone makes it very powerful.
**A simplified but helpful analogy**
====================================
I think AI safety via debate works well in cases where the verifier and the debaters broadly have a similar understanding of the world and level of intelligence. When this is not the case, failures get more frequent. Thus, my intuitive example for thinking about failure modes is “Let a person from 1800 evaluate the truth of the statement ‘Today I played Fortnite.’”. In this setting, you travel back through time and have to convince a random person from 1800 that you played Fortnite before traveling. Your opponent is someone who has a similar level of knowledge and intelligence as you.
Obviously, this setting is imperfect, due to all the problems with time travel but, in my opinion, it still intuitively shows some of the problems of AI safety via debate. The worlds of someone who played Fortnite in 2022 and someone who lived in 1800 are just so different that it is hard to even begin persuading them. Furthermore, so many of the concepts necessary to understand Fortnite, e.g. computers, the internet, etc. are nearly impossible to verify for a person from 1800 even if they wanted to believe you.
**Update:**I misunderstood something about the verification procedure. It isn’t necessary for the person from 1800 to fully understand electricity to be able to verify the question (see more below).
**Limitations**
===============
In the following, I list different implicit and explicit assumptions of debate that can lead to problems if they aren’t met.
**Assumption 1: concept must break down into parts that are verifiable in a reasonable timeframe**
--------------------------------------------------------------------------------------------------
**Original claim:**In cases where the verifier is not able to verify a concept from the beginning, it needs to be broken down into smaller subcomponents that are all verifiable. However, this might not always be possible--especially when given limited time.
In the “1800 Fortnite” example, the debater would have to convince the verifier of the existence of electricity, TVs or computers, video games, the internet, etc.
A second example is a question that probably requires very elaborate and time-intensive experiments to yield high-confidence answers such as in a “nature vs nurture” debate. The debater might have to run multi-generational studies to provide low-uncertainty evidence for their side.
**Update:**Rohin points out that the verifier doesn’t need to fully understand all concepts, they just need to find them sufficiently plausible. In the case of the 1800 Fortnite example, it would be sufficient to believe the claim about electricity more than the counterclaim. There was a second disagreement about complexity. I argued that some debates actually break down into multiple necessary conditions, e.g. if you want to argue that you played Fortnite you have to show that it is possible to do that and then that it is plausible. The pro-Fortnite debater has to show both claims while the anti-Fortnite debater has to defeat only one. Rohin argued that this is not the case, because every debate is ultimately only about the plausibility of the original statement independent of the number of subcomponents it logically breaks down to (or at least that’s how I understood him).
**Update 2:**I misunderstood Rohin’s response. He actually argues that, in cases where a claim X breaks down into claims X1 and X2, the debater has to choose which one is more effective to attack, i.e. it is not able to backtrack later on (maybe it still can by making the tree larger - not sure). Thus, my original claim about complexity is not a problem since the debate will always be a linear path through a potentially exponentially large tree.
**Assumption 2: human verifiers are capable of understanding the concept in principle**
---------------------------------------------------------------------------------------
**Original claim:**I’m not very sure about this but I could imagine that there are concepts that are too hard to understand in principle. Every attempt to break them down doesn’t solve the fundamental problem of the verifiers' limited cognitive abilities.
For example, I’m not sure if there is someone who “truly understood” string theory, or high-dimensional probability distributions sufficiently to make a high-confidence judgment in a debate. It might just be possible that these are beyond our abilities.
A second example would be explaining the “1800 Fortnite” scenario to a far-below-average intelligent person from 1800. Even if the debater did the best job possible, concepts like electricity or the internet might be beyond the capabilities of that specific verifier.
This leads to a potentially sad conclusion for a future with AI systems. I could very well imagine that smart humans today could not understand a concept that is totally obvious to an AGI.
**Update:**Rohin argues that as long as the verifier passes a [universality threshold](https://ai-alignment.com/of-humans-and-universality-thresholds-24b473e0c898) they should, in principle, be able to understand all possible concepts through debate. As long as there exists an exponentially large tree explaining the concept, debate should find a linear path through it.
*I’m convinced by this claim and I don’t think my original criticism makes that much sense anymore*. I still believe that debate could run into practical problems once problems take many generations to explain but I don’t believe in the principled point anymore.
**Assumption 3: human verifiers are well-intentioned**
------------------------------------------------------
**Original claim:**I think there are many implementations of AI safety via debate that lead to simple practical problems due to basic human flaws.
For example, the human verifier might just want to get on with it and thus doesn’t really make sure they actually believe the concept. I think this mostly implies that there will be some stochasticity to the answer which might be fixable by careful experimental design.
Secondly, especially in group settings, verifiers might also try to hide their own misunderstandings to not seem unknowledgeable in front of their peers. My intuition is that this biases group settings towards answers that “are allowed to be said”. During 2020, for example, a debater that argues that Covid might come from a lab leak would have lost by default just because it was seen as unspeakable in many social groups.
**Update:**I think the setting I had in mind when writing this post was more about regular scientists trying to discover new truths by using debate and less about the AI safety setting. In the use-case of AI safety, Rohin convinced me that most of these biases are much less relevant because the actors have probably clear incentives and good intentions. *Therefore, this point is mostly irrelevant.* I still could imagine scenarios in which human biases perturb the results but it’s much less likely.
**Assumption 4: human verifiers are impartial**
-----------------------------------------------
**Original claim:**Humans have a lot of motivations, many of which are not “just find the truth over everything else”. They like personal profits, seek status, have many biases such as a desire to confirm their current beliefs and follow ideologies and tribes.
Therefore, there is a risk that human verifiers might vote in favor of bad motivations and preferences rather than find the true answer. Especially when the truth of the result is not personally meaningful to them but one outcome is clearly beneficial to their wallet or party, they might just go with that. Since many of these motivations are hidden or subconscious, it’s not easy to create a setting in which they don’t influence the results.
**Update:**Similar update to point 3. *The point is mostly irrelevant now.*
**Assumption 5: It’s easier to explain the true concept than a flawed simple one**
----------------------------------------------------------------------------------
**Original claim:**There are some examples of concepts that take a while to understand but have simplified but false alternative explanations. Examples of these false alternatives include rent-control as a solution to the housing crisis, tariffs benefitting your economy, immigrants stealing jobs or autocrats acting the way they do simply because they are evil. In all of these cases, there is a simple intuition that can be explained in one sentence and a complex alternative that requires some understanding of economics, history, incentives, etc. to get why the simple intuition is false.
My gut reaction to this problem is something like “just pick smart verifiers” but there is a deeper problem. I think it is very likely that even the smartest humans today hold lots of beliefs that seem true and intuitive to them but will turn out to be wrong when investigated in more detail. It might be possible that the background knowledge to understand the true alternative takes too long to explain compared to providing a simple yet false alternative. Therefore, since the incentive of the debater is to win and not to find the truth, they will likely choose the simple yet false scenario.
**Update:**Rohin argues that it’s very hard to explain a simplified flawed argument than make the long true one in the debate setting. Whenever debater A makes a bad argument, debater B can point out the flaws and say “it’s not that simple because X”. I find this plausible in the AI safety via debate setting where the human verifier is well-intentioned and has clear incentives and less plausible in a setting where the human verifier is a more general public. I think for the most part, Rohin has convinced me but it seems like I’m more skeptical about whether debate would always converge to the most rational option. I guess we have slightly different conceptions of how rational/prone to biases even the smartest most well-intentioned human verifiers are.
**Assumption 6: AI is not deceptive**
-------------------------------------
**Original claim:**The incentive of the debater is not to seek the truth but to convince the verifier that they have found the truth. While this sounds like a small discrepancy it can create really large problems. In short, the debater does everything to convince the verifier even if it knows that it’s providing incorrect information. If the debater has more knowledge about the world than the verifier and a very accurate model of what the verifier thinks and likes, it can use its powers to deceive the human.
For an intuitive example, think of a hypothetical theoretical physics professor who knows you very well. They broadly know how you think and they know much more about physics than you. So if they wanted to convince you that a specific fact in theoretical physics is true, they could probably do so independent of whether it’s actually true.
**I think this is the biggest problem for AI safety via debate** since it is a) so hard to distinguish between deception and honest mistakes and b) a problem that will almost surely happen in scenarios where the AI is very powerful.
**Update:**This is the only point where Rohin hasn’t convinced me yet. He argues that the debaters have no incentive to be deceptive since the other debater is equally capable and has an incentive to point out this deception. I think this is true, as long as the reward for pointing out deception is bigger than the reward for alternative strategies, e.g. being deceptive yourself, you are incentivized to be truthful. Let’s say, for example, our conception of physics was fundamentally flawed and both debaters know this. To win the debate, one (truthful) debater would have to argue that our current concept of physics is flawed and establish the alternative theory while the other one (deceptive) could argue within our current framework of physics and sound much more plausible to the humans. The truthful debater is only rewarded when the human verifier waits long enough to understand the alternative physics explanation before giving the win to the deceptive debater. In case the human verifier stops early, deception is rewarded, right? What am I missing?
In general, I feel like the question of whether the debater is truthful or not only depends on whether they would be rewarded to do so. However, I (currently) don’t see strong reasons for the debater to be always truthful. To me, the bottleneck seems to be which kind of behavior humans intentionally or unintentionally reward during training and I can imagine enough scenarios in which we accidentally reward dishonest or deceptive behavior.
**Update2:**We were able to agree on the bottleneck. We both believe that the claim "it is harder to lie than to refute a lie" is the question that determines whether debate works or not. Rohin was able to convince me that it is easier to refute a lie than I originally thought and I, therefore, believe more in the merits of AI safety via debate. The main intuition that changed is that the refuter mostly has to continue poking holes rather than presenting an alternative in one step. In the “flawed physics” setting described above, for example, the opponent doesn’t have to explain the alternative physics setting in the first step. They could just continue to point out flaws and inconsistencies with the current setting and then slowly introduce the new system of physics and how it would solve these inconsistencies.
**Conclusions & future research**
=================================
My main conclusion is that AI safety via debate is a promising tool but some of its core problems still need addressing before it will be really good. There are many different research directions that one could take but I will highlight just two
1. **Eliciting Latent Knowledge (ELK) - style research:** Since the biggest challenge of AI safety via debate is deception, in my opinion, the natural answer is to understand when the AI deceives us. ELK is, in my opinion, the most promising approach to combat deception we have found so far.
2. **Social science research:**If we will ever be at a point when we have debates between AI systems to support decision-making, we also have to understand the problems that come with the human side of the setup. Under which conditions do humans choose for personal gain rather than seek the truth? Do the results from such games differ in group settings vs. individuals alone and in which ways? Can humans be convinced of true beliefs if they previously strongly believed something that was objectively false?
**Update:**I still think both of these are very valuable research directions but with my new understanding of debate as a tool specifically designed for AI safety, I think technical research like ELK is more fitting than the social science bit.
**Update2:***Rohin mostly convinced me that my remaining criticisms don’t hold or are less strong than I thought. I now believe that the only real problem with debate (in a setting with well-intentioned verifiers) is when the claim “it is harder to lie than to refute a lie” doesn’t hold.*However, I updated that it is often much easier to refute a lie than I anticipated because refuting the lie only entails poking a sufficiently large hole into the claim and doesn’t necessitate presenting an alternative solution.
If you want to be informed about new posts, you can [follow me on Twitter](https://twitter.com/MariusHobbhahn). |
e34e36dd-1e7d-4464-bdc0-68947c4be3be | trentmkelly/LessWrong-43k | LessWrong | Michael Shellenberger: US Has 12 Or More Alien Spacecraft, Say Military And Intelligence Contractors
Quotes:
> ...Multiple sources close to the matter have come forward to tell Public that Grusch’s core claims are accurate. The individuals are all either high-ranking intelligence officials, former intelligence officials, or individuals who we could verify were involved in U.S. government UAP efforts for three or more decades each. Two of them have testified, including as recently as last year, to both AARO and Congress.
----------------------------------------
> This is not the first time government officials have suggested that the U.S. may possess alien spaceships. “I was told for decades that Lockheed had some of these retrieved materials,” said the late Senator Harry Reid, who fought for greater disclosure. “And I tried to get, as I recall, a classified approval by the Pentagon to have me go look at the stuff. They would not approve that.”
> Former deputy assistant secretary of Defense for Intelligence, Christopher Mellon, recently reported that he has spoken to more than four witnesses who say they know of “a secret U.S. government program involving the analysis and exploitation of materials recovered from off-world craft… Some have supplied information to the intelligence community’s inspector general, others directly to the staff of the congressional oversight committees.”
----------------------------------------
> Members of Congressional intelligence committees and the Intelligence Community Inspector General (ICIG) are taking Grusch’s claims seriously. The ICIG concluded in July 2022 that Grusch’s whistleblower complaint was “credible and urgent.” And, said a source who worked with him, Grusch’s superiors promoted him rapidly due to his talent. “He jumped ranks when they hired him as a GS-15. That’s a big jump for his position.”
----------------------------------------
> “His assertion concerning the existence of a terrestrial arms race occurring sub-rosa over the past eighty years focused on reverse engineering technologies of unknown origin is f |
41cef52c-dda1-47db-96b3-07594c3f1919 | trentmkelly/LessWrong-43k | LessWrong | When None Dare Urge Restraint
One morning, I got out of bed, turned on my computer, and my Netscape email client automatically downloaded that day’s news pane. On that particular day, the news was that two hijacked planes had been flown into the World Trade Center.
These were my first three thoughts, in order:
> I guess I really am living in the Future.
>
> Thank goodness it wasn’t nuclear.
>
> and then
>
> The overreaction to this will be ten times worse than the original event.
A mere factor of “ten times worse” turned out to be a vast understatement. Even I didn’t guess how badly things would go. That’s the challenge of pessimism; it’s really hard to aim low enough that you’re pleasantly surprised around as often and as much as you’re unpleasantly surprised.
Nonetheless, I did realize immediately that everyone everywhere would be saying how awful, how terrible this event was; and that no one would dare to be the voice of restraint, of proportionate response. Initially, on 9/11, it was thought that six thousand people had died. Any politician who had said, “6,000 deaths is 1/8 the annual US casualties from automobile accidents,” would have been asked to resign the same hour.
No, 9/11 wasn’t a good day. But if everyone gets brownie points for emphasizing how much it hurts, and no one dares urge restraint in how hard to hit back, then the reaction will be greater than the appropriate level, whatever the appropriate level may be.
This is the even darker mirror of the happy death spiral—the spiral of hate. Anyone who attacks the Enemy is a patriot; and whoever tries to dissect even a single negative claim about the Enemy is a traitor. But just as the vast majority of all complex statements are untrue, the vast majority of negative things you can say about anyone, even the worst person in the world, are untrue.
I think the best illustration was “the suicide hijackers were cowards.” Some common sense, please? It takes a little courage to voluntarily fly your plane into a building. Of all t |
652be675-c53c-436f-a9c6-e783f6be298e | StampyAI/alignment-research-dataset/aisafety.info | AI Safety Info | Why would an AI do bad things?
1. [The Orthogonality Thesis](https://www.youtube.com/watch?v=hEUO6pjwFOo): AI could have almost any goal while at the same time having high intelligence (i.e. ability to succeed at those goals). This means that we could build a very powerful agent which would not necessarily have human-friendly values. For example, the classic [paperclip maximizer](https://www.lesswrong.com/tag/paperclip-maximizer) thought experiment considers an AI which has a goal of creating as many paperclips as possible – something that humans are (mostly) indifferent to – and as a side effect ends up destroying humanity to make room for more paperclip factories.
1. [Complexity of value](https://www.lesswrong.com/posts/GNnHHmm8EzePmKzPk/value-is-fragile): What humans care about is not simple, and the space of all goals is large, so virtually all goals we could program into an AI would lead to worlds not valuable to humans if pursued by a sufficiently powerful agent. For example, if we did not include our value of diversity of experience, we could end up with a world of endlessly looping simple pleasures, rather than beings living rich lives.
1. [Instrumental Convergence](https://www.youtube.com/watch?v=ZeecOKBus3Q): For almost any goal an AI could have, there are ‘instrumental’ steps it can take which make it more likely to achieve its main goal, such as acquiring resources, preserving itself, and preserving the contents of its goals. This means that a powerful AI with goals that were not explicitly human-friendly would predictably both take actions that lead to the end of humanity (e.g. using resources humans need to live to further its goals, such as replacing our crop fields with vast numbers of solar panels to power its growth, or using the carbon in our bodies to build things) and prevent us from turning it off or altering its goals.
|
2778ee00-a42e-4c1b-9ad3-980f82d24371 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | An Analytic Perspective on AI Alignment
This is a perspective I have on how to do useful AI alignment research. Most perspectives I’m aware of are constructive: they have some blueprint for how to build an aligned AI system, and propose making it more concrete, making the concretisations more capable, and showing that it does in fact produce an aligned AI system. I do not have a constructive perspective - I’m not sure how to build an aligned AI system, and don’t really have a favourite approach. Instead, I have an analytic perspective. I would like to understand AI systems that are built. I also want other people to understand them. I think that this understanding will hopefully act as a ‘filter’ that means that dangerous AI systems are not deployed. The following dot points lay out the perspective.
Since the remainder of this post is written as nested dot points, some readers may prefer to read it in [workflowy](https://workflowy.com/s/an-analytical-perspe/eU45Fsjd7lzidjM8).
Background beliefs
------------------
* I am imagining a future world in which powerful AGI systems are made of components roughly like neural networks (either feedforward or recurrent) that have a large number of parameters.
* Furthermore, I’m imagining that the training process of these ML systems does not provide enough guarantees about deployment performance.
+ In particular, I’m supposing that systems are being trained based on their ability to deal with simulated situations, and that that’s insufficient because deployment situations are hard to model and therefore simulate.
- One reason that they are hard to model is the complexities of the real world.
* The real world might be intrinsically difficult to model for the relevant system. For instance, it’s difficult to simulate all the situations in which the CEO of Amazon might find themselves.
* Another reason that real world situations may be hard to model is that they are dependent on the final trained system.
+ The trained system may be able to affect what situations it ends up in, meaning that situations during earlier training are unrepresentative.
+ Parts of the world may be changing their behaviour in response to the trained system…
- in order to exploit the system.
- by learning from the system’s predictions.
* The real world is also systematically different than the trained world: for instance, while you’re training, you will never see the factorisation of RSA-2048 (assuming you’re training in the year 2020), but in the real world you eventually will.
+ This is relevant because you could imagine [mesa-optimisers](https://arxiv.org/abs/1906.01820) appearing in your system that choose to act differently when they see such a factorisation.
* I’m imagining that the world is such that if it’s simple for developers to check if an AI system would have disastrous consequences upon deployment, then they perform this check, and fail to deploy if the check says that it would.
Background desiderata
---------------------
* I am mostly interested in allowing the developers of AI systems to determine whether their system has the cognitive ability to cause human extinction, and whether their system might try to cause human extinction.
+ I am not primarily interested in reducing the probabilities of other ways in which AI systems could cause humanity to go extinct, such as research groups intentionally behaving badly, or an uncoordinated set of releases of AI systems that interact in negative ways.
- That being said, I think that pursuing research suggested by this perspective could help with the latter scenario, by making it clear which interaction effects might be present.
* I want this determination to be made before the system is deployed, in a ‘zero-shot’ fashion, since this minimises the risk of the system actually behaving badly before you can detect and prevent it.
Transparency
------------
* The type of transparency that I’m most excited about is mechanistic, in a sense that I’ve described [elsewhere](https://www.lesswrong.com/posts/3kwR2dufdJyJamHQq/mechanistic-transparency-for-machine-learning).
* The transparency method itself should be based on a trusted algorithm, as should the method of interpreting the transparent artefact.
+ In particular, these operations should not be done by a machine learning system, unless that system itself has already been made transparent and verified.
- This could be done [amplification-style](https://ai-alignment.com/iterated-distillation-and-amplification-157debfd1616).
* Ideally, models could be regularised for transparency during training, with little or no cost to performance.
+ This would be good because by default models might not be very transparent, and it might be hard to hand-design very transparent models that are also capable.
- I think of this as what one should derive from Rich Sutton’s [bitter lesson](http://www.incompleteideas.net/IncIdeas/BitterLesson.html)
+ This will be easier to do if the transparency method is simpler, more ‘mathematical’, and minimally reliant on machine learning.
+ You might expect little cost to performance since neural networks can often reach high performance given constraints, as long as they are deep enough.
- [This paper](https://arxiv.org/abs/1804.08838) on the intrinsic dimension of objective landscapes shows that you can constrain neural network weights to a low-dimensional subspace and still find good solutions.
- [This paper](https://arxiv.org/abs/1908.01755) argues that there are a large number of models with roughly the same performance, meaning that ones with good qualities (e.g. interpretability) can be found.
+ [This paper](https://arxiv.org/abs/1711.06178) applies regularisation to machine learning models that ensures that they are represented by small decision trees.
* The transparency method only has to reveal useful information to developers, not to the general public.
+ This makes the problem easier but still difficult.
+ Presumably developers will not deploy catastrophically terrible systems, since catastrophes are usually bad for most people, and I’m most interested in averting catastrophic outcomes.
Foundations
-----------
* In order for the transparency to be useful, practitioners need to know what problems to look for, and how to reason about these problems.
* I think that an important part of this is ‘agent foundations’, by which I broadly mean a theory of what agents should look like, and what structural facts about agents could cause them to display undesired behaviour.
+ Examples:
- Work on [mesa-optimisation](https://arxiv.org/abs/1906.01820)
- Utility theory, e.g. the [von Neumann-Morgenstern theorem](https://en.wikipedia.org/wiki/Von_Neumann%E2%80%93Morgenstern_utility_theorem)
- Methods of detecting which agents are likely to be intelligent or dangerous.
* For this, it is important to be able to look at a machine learning system and learn if (or to what degree) it is agentic, detect belief-like structures and preference-like structures (or to deduce things analogous to beliefs and preferences), and learn other similar things.
+ This requires structural definitions of the relevant primitives (such as agency), not subjective or performance-based definitions.
- By ‘structural definitions’, I mean definitions that refer to facts that are easily accessible about the system before it is run.
- By ‘subjective definitions’, I mean definitions that refer to an observer’s beliefs or preferences regarding the system.
- By ‘performance-based definitions’, I mean definitions that refer to facts that can be known about the system once it starts running.
- Subjective definitions are inadequate because they do not refer to easily-measurable quantities.
- Performance-based definitions are inadequate because they can only be evaluated once the system is running, when it could already pose a danger, violating the “zero-shot” desideratum.
- Structural definitions are required because they are precisely the definitions that are not subjective or performance-based that also only refer to facts that are easily accessible, and therefore are easy to evaluate whether a system satisfies the definition.
- As such, definitions like “an agent is a system whose behaviour can’t usefully be predicted mechanically, but can be predicted by assuming it near-optimises some objective function” (which was proposed in [this paper](https://arxiv.org/abs/1805.12387)) are insufficient because they are both subjective and performance-based.
- It is possible to turn subjective definitions into structural definitions trivially, by asking a human about their beliefs and preferences. This is insufficient.
* e.g. “X is a Y if you are scared of it” can turn to “X is a Y if the nearest human to X, when asked if they are scared of X, says ‘yes’”.
* It is insufficient because such a definition doesn’t help the human form their subjective beliefs and impressions.
- It is also possible to turn subjective definitions that only depend on beliefs into structural definitions by determining which circumstances warrant a rational being to have which beliefs. This is sufficient.
* Compare the subjective definition of temperature as “the derivative of a system’s energy with respect to entropy at fixed volume and particle number” to the objective definition “equilibrate the system with a thermometer, read it off the thermometer”. For a rational being, these two definitions yield the same temperature for almost all systems.
Relation between transparency and foundations
---------------------------------------------
* The agent foundations theory should be informed by transparency research, and vice versa.
+ This is because the information that transparency methods can yield should be all the information that is required to analyse the system using the agent foundations theory.
+ Both lines of research can inform the other.
- Transparency researchers can figure out how to reveal the information required by agent foundations theory, and detect the existence of potential problems that agent foundations theory suggests might occur given certain training procedures.
- Agent foundations researchers can figure out what is implied by the information revealed by existing transparency tools, and theorise about problems that transparency researchers detect.
Criticisms of the perspective
-----------------------------
* It isn’t clear if neural network transparency is possible.
+ More specifically, it seems imaginable that some information required to usefully analyse an AI system cannot be extracted from a typical neural network in polynomial time.
* It isn’t clear that relevant terms from agency theory can in fact be well-defined.
+ E.g. “optimisation” and “belief” have eluded a satisfactory computational grounding for quite a while.
+ Relatedly, the philosophical question of which physical systems enable which computations has not to my mind been satisfactorily resolved. See [this](https://plato.stanford.edu/entries/computation-physicalsystems/) relevant SEP article.
* An easier path to transparency than the “zero-shot” approach might be to start with simpler systems, observe their behaviour, and slowly scale them up. As you see problems, stop scaling up the systems, and instead fix them so the problems don’t occur.
+ I disagree with this criticism.
- At one point, it’s going to be the first time you use a system of a given power in a domain, and the problems caused by the system might be discontinuous with its power, meaning that they would be hard to predict.
* Especially if the power of the system increases discontinuously.
* It is plausibly the case that systems that are a bit ‘smarter than humanity’ are discontinuously more problematic than those that are a bit less ‘smart than humanity’.
* One could imagine giving up the RL dream for something like debate, where you really can get guarantees from the training procedure.
+ I think that this is not true, and that things like debate require transparency tools to work well, so as to let debaters know when other debaters are being deceitful. An argument for an analogous conclusion can be found in evhub’s post on [Relaxed adversarial training for inner alignment](https://www.lesswrong.com/posts/9Dy5YRaoCxH9zuJqa/relaxed-adversarial-training-for-inner-alignment).
* One could imagine inspecting training-time reasoning and convincing yourself that way that future reasoning will be OK.
+ But reasoning could look different in different environments.
* This perspective relies on things continuing to look pretty similar to current ML.
+ This would be alleviated if you could come up with some sort of sensible theory for how to make systems transparent.
+ I find it plausible that the development of such a theory should start with people messing around and doing things with systems they have.
* Systems should be transparent to all relevant human stakeholders, not just developers.
+ Sounds right to me - I think people should work on this broader problem. But:
- I don’t know how to solve that problem without making them transparent to developers initially.
- I have ideas about how to solve the easier problem. |
a87a9460-9851-49b2-894b-bf04bea3ac96 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Transformer language models are doing something more general
*Epistemic status: speculative.*
First, a few paper titles:
* [Pretrained Transformers as Universal Computation Engines](https://arxiv.org/abs/2103.05247)
* [Can Wikipedia Help Offline Reinforcement Learning?](https://arxiv.org/abs/2201.12122) (Answer: yes.)
* [Pretrained Transformers Improve Out-of-Distribution Robustness](https://arxiv.org/abs/2004.06100)
* [Data Distributional Properties Drive Emergent In-Context Learning in Transformers](https://arxiv.org/abs/2205.05055)
The gist of the first three studies is that transformers (specifically) trained on natural language (specifically) generalize better than expected, with little or no fine-tuning, not only to unseen tasks but even to unseen and apparently unrelated modalities like offline reinforcement learning. The last study takes this a step further—it doesn't actually pretrain on language at all, but instead tries to mimic the specific statistical properties of natural language that lead to this behavior with various sampling procedures from an image classification dataset.
The difference between these results and the plethora of text-to-text transformer multitask/transfer learning results that have come out since GPT-1 is that transfer learning to new *modalities* requires learning priors general enough to apply to both text and the other modality—implying, first of all, that such priors *exist*, which has updated me in the following directions:
* Well-trained transformers, regardless of task, occupy the same relatively small subspace of parameter space
* Most of the gradient-descent steps of a training run from scratch are spent just getting to this subspace; relatively few are spent learning the specific task
Taken together, these hypotheses seem to imply that within today's gigantic and [notoriously data-hungry](https://www.lesswrong.com/posts/6Fpvch8RR29qLEWNH/chinchilla-s-wild-implications) language models is a sparser, far more efficient architecture trying to get out. I don't have any idea what this architecture looks like. If I did, I wouldn't post about it here. I am quite confident that it exists, because human children manage to acquire language without ingesting the equivalent of terabytes of text. I'm even reasonably confident that it's simple, because the human genome doesn't have enough space to code for complex mental priors (also, the evidence seems to point to the [neocortex being fairly uniform](https://www.lesswrong.com/posts/WFopenhCXyHX3ukw3/how-uniform-is-the-neocortex)), and because whatever “universal grammar” pretrained transformers are learning, it has to be fundamental enough to apply to domains as unlike language as offline reinforcement learning.
Only the last paper of the four I linked above, from DeepMind, attempts to elucidate what's so special about language, and they focus only on a few obvious statistical features of language token distributions—while several features they tested did improve in-context (i.e. few-shot) learning when present, the paper leaves understanding the mechanism behind this improvement for further research.
The most obvious connection that I see here, among the relatively few papers I've read, is with Anthropic's work on [In-context Learning and Induction Heads](https://transformer-circuits.pub/2022/in-context-learning-and-induction-heads/index.html); it seems quite possible that induction heads are this missing mechanism linking the unique properties of language distributions with in-context learning. A direction for further research, for anyone interested, might be to try to find a theoretical link between language-like (Zipfian, non-uniform) training data distributions and the formation of induction heads.
I'll end this here, as my writing has caught up with my thinking; I'll probably write a follow-up if the discussion on this post inspires further ideas. |
002ca617-567a-4dc2-a294-2c4849195d00 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Gears-Level Mental Models of Transformer Interpretability
This post aims to quickly break down and explain the dominant mental models interpretability researchers currently use when thinking about how transformers work.
In my view, the focus of transformer interpretability research is teleological: we care about the *functions* each component in the transformer performs and how those functions interact to yield the capabilities we see in language models today. From a functional understanding of transformer internals, we then hope to be able to answer other important interpretability questions such as, “Where/How is knowledge stored?,” “Do transformers have beliefs?”, and “How do transformers reason?” As such, the mental models described in this post will be functional theories about how researchers think about transformer internals, and not hypotheses about other interpretability questions, like the ones mentioned above.
There are three main components at work in the transformer: the attention heads, the MLP, and the additive residual stream that connects all the layers. The functions that each play aren’t clear and are the subject of much research, and the mental models described below will be different ways of thinking about each main component. None of these mental models are mutually exclusive, and in reality, transformer internals probably look like a messy combination of many of these models.
This post assumes a working understanding of transformers. For a primer on the transformer architecture, I recommend [The Illustrated Transformer](https://jalammar.github.io/illustrated-transformer/), [Transformers from Scratch](https://e2eml.school/transformers.html), and [The Annotated Transformer](https://nlp.seas.harvard.edu/2018/04/03/attention.html). Other works that this post draws greatly from are [the logit lens](https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens), [Transformer Feed-Forward Layers Are Key-Value Memories](https://arxiv.org/pdf/2012.14913.pdf), [ROME](https://arxiv.org/pdf/2202.05262.pdf), and the [Anthropic](https://transformer-circuits.pub/2021/framework/index.html) [papers](https://transformer-circuits.pub/2022/in-context-learning-and-induction-heads/index.html).
The Models
----------
### Residual Stream as Output Accumulation
The residual stream is simply the accumulation of all the stuff the transformer wants to say at the end of its inference step.
Clearly, the last residual hidden state is the transformer’s prediction (before it gets projected to the vocabulary), and all this mental model is suggesting is that the prior hidden states are less-refined versions of the final residual hidden state. The strongest version of this model states that the middle hidden states *are* the model’s nascent predictions, and the weaker version just states that the middle hidden states *contain*the model’s nascent predictions. This model seems pretty intuitive, and its weaker version seems pretty true to me.
The strongest source of evidence for this model currently is the logit lens. In a really neat trick, which they dubbed the logit lens, nostalgebraist found that by projecting the intermediate hidden states of GPT2 to the vocabulary matrix, the resulting logit distributions made a lot of sense. The logit lens might be a way to see what GPT believes at each time step in its processing, and with it, we can see some surprisingly coherent and logical “thought processes.”
The logit lens on GPT2-medium with the first paragraph of Harry Potter as input. Made using the transformer-utils Python package.
How to read this table: The token at the bottom of each column is the last token in the input sequence; the token at the top of each column is the correct token; the tokens in-between are the projections of GPT2’s hidden states to the vocab. When GPT2 is faced with an easy prediction task, such as induction (“Mr. Dursley”) or stop words (“director of”) — words that have super high bigram log likelihoods — the logit lens shows that it converges to that prediction rapidly, and when it’s faced with a harder task, we see that difficulty reflected accordingly in the flatness of its logit distribution. Additionally, GPT2’s intermediate projections make sense. If we just read the projections for the column with the input token as “the,” we read “same, first, only, youngest, only” which sounds like GPT2 thinks that the sentence is going to describe Mr. Dursley as an only child, youngest child, first-born, etc.
One common perspective in deep learning is the information bottleneck perspective, which states that deep learning models try to transform their input representations to encode as much information about the output as possible, while removing irrelevant information about the input. In the logit lens table above, we can see that by the first layer, the residual hidden state usually does not project out to the original input token. What this suggests is that GPT, and autoregressive transformers in general, immediately converts inputs to guesses about the output, working in an output-prediction space more than an input-compression space.
The strong version of this mental model, which specifically states that the residual hidden states are distributions over the output vocabulary, is also kind of intuitive. Weak evidence for the strong version of this mental model, that the intermediate hidden states *are*the model’s final prediction, is this quick PCA experiment I ran. If you run PCA on all of a model’s residual states for a given prediction, there’s a single direction that explains roughly 88% of the variance across all layers. And if you project this direction to the vocab, it’s usually (~77% of the time) the model’s final prediction.
PCA of each of the layer residuals (12 x 768 matrix) for ~200 examples in GPT2. X-axis is PCA components. Y-axis is explained variance ratios. ### Residual Stream as Communication Channel
Another way of thinking about the residual stream is that it’s a communication channel for individual components of the transformer to talk through. This perspective was introduced in Anthropic’s first paper about mechanistic interpretability and seems to be the group’s take on the residual stream in general. Importantly, thinking about the residual stream as a communication channel or as an accumulation of output is not mutually exclusive, and, in fact, the accumulation of information about the output, seen in the logit lens, is probably being communicated to components deeper in the transformer.
Image straight from the Anthropic paper, describing the residual stream as a communication channel. The nature of the residual stream is that it’s a linear, vector space that is being added to and read from by all layers. The attention heads and MLP will read from the space with a linear projection — with the query and key matrices and the first linear transformation in the MLP respectively — and then write to the residual by adding another linear transformation — the value and output matrices and second linear transformation in the MLP respectively — back into the residual. And since the residual doesn’t do any processing itself, it’s intuitive to think of it as the space where all the attention heads and MLP neurons communicate through.
There are a couple of components that have been found in transformers that seem to be doing communication-like things. The most notable component is, of course, induction heads. Induction heads are pairs of attention heads that seem to implement an induction-like algorithm, specifically an algorithm that completes the pattern AB … A -> B. They work by searching through the context, finding the present token, and then attending to the token that comes *after*the present token. Since these induction heads involve pairs of heads across layers, they communicate through the residual stream, where the first induction head copies information with its OV matrix about the present token, which the key matrix of the second induction head reads.
Induction Heads in action. The second induction head’s QK matrix is reading from the information that the first induction head’s OV matrix wrote into the residual stream.Anthropic also found evidence of neurons in the MLP and attention heads that seem to delete information from the residual stream. Specifically, the MLP neurons have input weights that have very high negative cosine similarity with their output weights, indicating that if a direction is present in the residual stream, the neuron will add in the negative of that direction into the residual stream. The attention heads had highly negative eigenvalues in their OV matrix and seemed to attend to the present token, indicating that they delete information about the present token.
### MLP as Key-Value Pairs
The Feed-Forward (FF) in a transformer is defined as 2 linear transformations of the input with a non-linearity (ReLU, GeLU, etc.) in between, i.e:
FF(x)=max(0,xW1+b1)(W2+b2).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')}
In the paper, [Transformer Feed-Forward Layers Are Key-Value Memories](https://arxiv.org/pdf/2012.14913.pdf), Geva et al. propose a mental model for thinking about the MLP where the first linear transformation are keys and the second linear transformation are values. Together, they form key-value pairs or neural memories, which can be written as (omitting the biases):
FF(x)=f(x⋅K⊺)⋅V where f is some non-linearity
Since each key-value pair corresponds to columns in the MLP weights, we can rewrite the above as:
FF(x)=dm∑i=1f(x⋅ki)⋅vi where mi=f(x⋅ki) is the coefficient of vi and dm is the MLP’s hidden dimension (typically 4 times the model embedding size)
The key-value pair mental model states that when a key is activated by "something", its corresponding value will be written strongly into the residual (since the coefficient mi will be high). This alone is pretty straightforward and seems right; the important thing is what you interpret this "something" as.
This "something" according to the key-value paper are patterns: each key is activated by textual patterns in its training data, and when a key is activated, its corresponding value will shift the residual’s logit distribution (with the mental model of the residual stream as output accumulation) towards a distribution that complements the logits that would typically appear after the textual pattern correlated with the key. This correlation between key and value is particularly noticeable in higher layers i.e. the key-value pairs in higher layers are more likely to contain semantic information, while lower layers contain shallow (syntactic or grammatical) information.
Key-value schematic from the key-value memories paper. Keys are activated by textual patterns in the text and their corresponding values shift the logit distribution towards tokens that follow those patterns. The key-value mental model suggests that the MLP is where knowledge is stored (since you can treat key-value pairs as a linear-associative memory). For example, maybe values contain knowledge and write that knowledge into the residual stream. Multiple knowledge editing methods have made use of this implication. In Rank-One-Model-Editing (ROME), they used the key-value model for their knowledge editing procedure, and the fact that ROME works really well supports the key-value model’s validity.
### Attention as Information Movers
I think this mental model is the least controversial of all the ones mentioned in this post. The very structure of attention suggests that its function is centered around the context and the different relationships between tokens in a sequence. How else would information move across tokens?
Besides the intuitive-ness of this mental model, there’s a lot of empirical evidence that backs it up. Obviously, the existence of induction heads supports the mental model that attention heads move information across tokens. Additionally, in the ROME paper, their causal tracing method supports the idea that attention heads move information across tokens. Specifically, their causal tracing suggests that attention heads move factual knowledge about tokens that the MLP writes into the residual stream towards the prediction of the final token in the sequence. I’d just direct readers to the [actual paper](https://arxiv.org/pdf/2202.05262.pdf) if they want to learn more about this stuff because it’s pretty cool. |
bfcd6f8b-cd80-410e-8524-95bdb4d4867e | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Inference cost limits the impact of ever larger models
I sometimes notice that people in my community (myself included) assume that the first "generally human-level" model will lead to a transformative takeoff scenario almost immediately. The assumption seems to be that training is expensive but inference is cheap so once you're done training you can deploy an essentially unlimited number of cheap copies of the model. I think this is far from obvious
[edit: This post should be read as "inference cost may turn out to be a bottleneck. Don't forget about them. But we don't know how inference costs will develop in the future. Additionally, it may take a while before we can run lots of copies of an extremely large model because we'd need to build new computers first.]
Inference refers to the deployment of a trained model on a new input. According to OpenAI's [report](https://openai.com/blog/ai-and-compute/) from 2018, most compute used for deep learning is spent not on training but on inference. It is true that one inference step is much cheaper than a training run consisting of many training steps. But many inference steps together can make up the bulk of compute.
To gain some intuition, consider that writing 750 words with GPT-3 costs [6 cents](https://beta.openai.com/pricing/). If we made a model with 1000x more parameters, similar to the difference between GPT-1 and GPT-3, the 750 words would cost $60, comparable to the cost of a good human writer. But to start an immediate economic transformation, I expect we need something significantly cheaper (or smarter) than humans.
Of course, the future will bring efficiency improvements. But also increases in cost. For example, future models may look at a context window longer than 2048 tokens, and I've assumed greedy sampling here which is cheap but suboptimal (it's like typing without getting to revise). I'm unsure how these factors balance out.
To have a transformative impact, as a heuristic, the number of copies of our human-level model should probably exceed the human population (~8 billion). But to run billions of copies, we'd need to dramatically increase the world's number of supercomputers. You can't just repurpose all consumer GPUs for inferencing, let alone run GPT-3 on your smartphone. GPT-3 needs hundreds of GPUs just to fit the model into GPU memory.[[1]](#fn-hZCquDaqDjMRHRHQM-1) These GPUs must then be linked through a web of fast interconnects professionally fitted in a data center. And if we're talking about a 1000x larger model, today's supercomputers may not be ready to store even a single copy of it.[[2]](#fn-hZCquDaqDjMRHRHQM-2)
This is not to say that a generally human-level model wouldn't have some drastic impacts, or be closely followed by generally **super**-human models; it just makes me pause before assuming that the first human-level model is the end of the world as we know it. In order run enough copies of the model, depending on its exact size, we'd first need to make it more efficient and build many, many new supercomputers.
---
1. You can theoretically run a model on fewer GPUs by putting just the first layer into GPU memory, forward passing on it, then deleting it and loading the second layer from RAM, and so forth (see [ZeRO-Infinity](https://arxiv.org/abs/2104.07857)). But this comes with high latency which rules out many applications. [↩︎](#fnref-hZCquDaqDjMRHRHQM-1)
2. I'm told that the largest clusters these days have tens of thousands of GPUs. [↩︎](#fnref-hZCquDaqDjMRHRHQM-2) |
cb2c08e2-4263-4930-ba68-bd2c496f4a65 | trentmkelly/LessWrong-43k | LessWrong | [SEQ RERUN] Quantum Explanations
Today's post, Quantum Explanations was originally published on 09 April 2008. A summary (taken from the LW wiki):
> Quantum mechanics doesn't deserve its fearsome reputation. If you tell people something is supposed to be mysterious, they won't understand it. It's human intuitions that are "strange" or "weird"; physics itself is perfectly normal. Talking about historical erroneous concepts like "particles" or "waves" is just asking to confuse people; present the real, unified quantum physics straight out. The series will take a strictly realist perspective - quantum equations describe something that is real and out there. Warning: Although a large faction of physicists agrees with this, it is not universally accepted. Stronger warning: I am not even going to present non-realist viewpoints until later, because I think this is a major source of confusion.
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 Belief in the Implied Invisible, 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. |
7a49aab0-31aa-44e3-9b83-ce7eb3d1e889 | trentmkelly/LessWrong-43k | LessWrong | MIRI's 2014 Summer Matching Challenge
(Cross-posted from MIRI's blog. MIRI maintains Less Wrong, with generous help from Trike Apps, and much of the core content is written by salaried MIRI staff members.)
Thanks to the generosity of several major donors,† every donation made to MIRI between now and August 15th, 2014 will be matched dollar-for-dollar, up to a total of $200,000!
Now is your chance to double your impact while helping us raise up to $400,000 (with matching) to fund our research program.
Corporate matching and monthly giving pledges will count towards the total! Please email malo@intelligence.org if you intend on leveraging corporate matching (check here, to see if your employer will match your donation) or would like to pledge 6 months of monthly donations, so that we can properly account for your contributions towards the fundraiser.
(If you're unfamiliar with our mission, see: Why MIRI?)
Donate Now
Accomplishments Since Our Winter 2013 Fundraiser Launched:
* Hired 2 new Friendly AI researchers, Benja Fallenstein & Nate Soares. Since March, they've authored or co-authored 4 papers/reports, with several others in the works. Right now they're traveling, to present papers at the Vienna Summer of Logic, AAAI-14, and AGI-14.
* 5 new papers & book chapters: “Why We Need Friendly AI,” “The errors, insights, and lessons of famous AI predictions,” “Problems of self-reference...,” “Program equilibrium...,” and “The ethics of artificial intelligence.”
* 11 new technical reports: 7 reports from the December 2013 workshop, “Botworld,” “Loudness...,” “Distributions allowing tiling...,” and “Non-omniscience...”
* New book: Smarter Than Us, published both as an e-book and a paperback.
* Held one MIRI workshop and launched the MIRIx program, which currently supports 8 independently-organized Friendly AI discussion/research groups around the world.
* New analyses: Robby's posts on naturalized induction, Luke's list of 70+ studies which could improve our picture of superintelligence stra |
5f328a2d-b9dd-4747-9b08-c85172a8bba6 | trentmkelly/LessWrong-43k | LessWrong | Godel in second-order logic?
I'm following the HAE sequence [Godel's Completeness and Incompleteness Theorem], and I can see what it means for first-order logic to be incomplete.
I don't understand how Godel's incompleteness theorem can still apply to second-order logic -- doesn't PA with second-order logic allow us to uniquely pin down a model?
But Godel's sentence can still be constructed, so second-order logic must still be incomplete -- I just don't understand how we can have models with a "weird idea of what it means to be a proof" if we don't have non-standard numbers.
I asked on Math Stackexchange and was told that second-order logic "doesn't have a computable deduction system". Unfortunately, I don't really understand what this means. |
bb000b2e-e319-4747-9b1b-080e82c8c0ae | trentmkelly/LessWrong-43k | LessWrong | The Gears of Impact
Scheduling: The remainder of the sequence will be released after some delay.
Exercise: Why does instrumental convergence happen? Would it be coherent to imagine a reality without it?
Notes
* Here, our descriptive theory relies on our ability to have reasonable beliefs about what we'll do, and how things in the world will affect our later decision-making process. No one knows how to formalize that kind of reasoning, so I'm leaving it a black box: we somehow have these reasonable beliefs which are apparently used to calculate AU.
* In technical terms, AU calculated with the "could" criterion would be closer to an optimal value function, while actual AU seems to be an on-policy prediction, whatever that means in the embedded context. Felt impact corresponds to TD error.
* This is one major reason I'm disambiguating between AU and EU; in the non-embedded context. In reinforcement learning, AU is a very particular kind of EU: V∗(s), the expected return under the optimal policy.
* Framed as a kind of EU, we plausibly use AU to make decisions.
* I'm not claiming normatively that "embedded agentic" EU should be AU; I'm simply using "embedded agentic" as an adjective. |
41e3df14-58ca-4363-9369-9394b30cb6b2 | trentmkelly/LessWrong-43k | LessWrong | Stupid Question: Why am I getting consistently downvoted?
I feel like I've posted some good stuff in the past month, but the bits that I think are coolest have pretty consistently gotten very negative karma.
I just read the rude post about rationalist discourse basics, and, while I can guess why my posts are receiving negative karma, that would involve a truly large amount of speculating about the insides of other people's heads, which is apparently discouraged. So I figured I would ask.
I will offer a bounty of $1000 for the answer I find most helpful, and a bounty of $100 for the next most helpful three answers. This will probably be paid out over Venmo, if that is a decision-relevant factor.
Note that I may comment on your answer asking for clarification.
Edit 11-30-2023 1:27 AM: I have selected the recipients of the bounties. The grand prize of $1000 goes to @Shankar Sivarajan . The three runner-up prizes of $100 go to @tslarm , @Joe Kwon , and @trevor . Please respond to my DM to arrange payment or select a worthy charity to receive your winnings.
Edit 11-30-2023 12:08 PM: I have paid out all four bounties. Please contact me in DM if there is any issue with any of the bounties. |
16703eed-aa43-4b74-9492-61aa1bf5b806 | trentmkelly/LessWrong-43k | LessWrong | AI Safety Evaluations: A Regulatory Review
This article is the second in a series of ~10 posts comprising a 2024 State of the AI Regulatory Landscape Review, conducted by the Governance Recommendations Research Program at Convergence Analysis. Each post will cover a specific domain of AI governance (e.g. incident reporting, safety evals, model registries, etc.). We’ll provide an overview of existing regulations, focusing on the US, EU, and China as the leading governmental bodies currently developing AI legislation. Additionally, we’ll discuss the relevant context behind each domain and conduct a short analysis.
This series is intended to be a primer for policymakers, researchers, and individuals seeking to develop a high-level overview of the current AI governance space. We’ll publish individual posts on our website and release a comprehensive report at the end of this series.
Let us know in the comments if this format is useful, if there are any topics you'd like us to cover, or if you spot any errors or omissions!
Context
Governments and researchers are eager to develop tools and techniques to evaluate AI. These include risk assessments that are common in industry regulation, but also techniques that are more unique to advanced AI, such as capability evaluations and alignment evaluations.
In this section, we’ll define some terms and introduce some recent research on evaluating AI. Most existing AI regulation is yet to incorporate these new techniques, but many experts believe they’ll be a critical component of long-term safety (such as responsible scaling policies), and many regulatory proposals from experts include calls for specific assessment systems and requirements, which we’ll discuss shortly.
There are three main features of AI models that people are interested in evaluating:
* Safety: How likely is this model to cause harm? Assessing the safety of AI models is crucial but difficult due to their enormous flexibility. Safety assessors often use techniques from other industries, such as red- |
96adf750-821c-4185-8ad2-044661ef67fc | trentmkelly/LessWrong-43k | LessWrong | Become a PIBBSS Research Affiliate
TL;DR: PIBBSS is hiring research affiliates!
PIBBSS is a research initiative facilitating work that draws on the parallels between intelligent behavior in natural and artificial systems, and leveraging these insights towards making AI systems safe, beneficial and aligned.
We want to support excellent researchers pursuing “PIBBSS-style” AI alignment research by providing them with tailored longer-term support: a full-time salary, a lively research community, operational support and more. The initial commitment is 6 months, with potential extensions to 12 or more.
To apply, click here. To be considered as part of the next iteration of affiliates, apply by November 5th.
----------------------------------------
At PIBBSS (“Principles of Intelligent Behavior in Biological and Social Systems”), our mission is to facilitate AI alignment research that draws on the parallels between intelligent behavior in natural and artificial systems. (For more details, see the section "About PIBBSS and our research priorities" below.)
Since PIBBSS’ inception in 2022, we have run two iterations of our 3-months research fellowships, as well as a reading group, an ongoing speaker series, and a number of research retreats exploring topics in line with PIBBSS’ research interests. Our alumni have gone on to pursue AI alignment research at places like OpenAI, Anthropic, the Alignment of Complex Systems research group, and academia, as well as through independent funding.
While we continue to be excited about the value of those activities, we also believe that there is more value to be had! In particular, we see potential in a program geared towards supporting longer-term, more sustained research efforts in line with PIBBSS’ research bet.
As such, we are delighted to launch our new PIBBSS Affiliate Program.
Goals and key features of the Affiliate Program
At its core, our goal in running this program is to counterfactually help produce substantial, high-quality and legible research |
565b526f-f191-4b2c-8c0c-3f68d756ae68 | trentmkelly/LessWrong-43k | LessWrong | Training for math olympiads
Lately I've resolved to try harder at teaching myself math so I have a better shot at the international olympiad (IMO). These basically involve getting, say, three really hard math problems and trying your best to solve them within 5 hours.
My current state:
* I have worked through a general math problem-solving guide (Art and Craft of Problem-Solving), a general math olympiad guide (A Primer for Mathematics Competitions) and practice problems.
* I've added all problems and solutions and theorems and techniques into an Anki deck. When reviewing, I do not re-solve the problem, I only try to remember any key insights and outline the solution method.
* I am doing n-back, ~20 sessions (1 hour) daily, in an attempt to increase my general intelligence (my IQ is ~125, sd 15).
* I am working almost permanently; akrasia is not much of a problem.
* I am not _yet_ at the level of IMO medallists.
What does the intrumental-rationality skill of LWers have to say about this? What recommendations do you guys have for improving problem-solving ability, in general and specifically for olympiad-type environments? Specifically,
* How should I spread my time between n-backing, solving problems, and learning more potentially-useful math?
* Should I take any nootropics? I am currently looking to procure some fish oil (I don't consume any normally) and perhaps a racetam. I have been experimenting with cycling caffeine weekends on, weekdays off (to prevent tolerance being developed), with moderate success (Monday withdrawal really sucks, but Saturday is awesome).
* Should I add the problems to Anki? It takes time to create the cards and review them; is that time better spent doing more problems? |
c293aef2-40b5-4c58-8421-35984ed58390 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | OpenAI Solves (Some) Formal Math Olympiad Problems
*Epistemic status: I have just skimmed through OpenAI's blogpost and paper, I do not fully understand the details.*
From the [blogpost](https://openai.com/blog/formal-math/)
> We built a neural theorem prover for [Lean](https://leanprover.github.io/) that learned to solve a variety of challenging high-school olympiad problems, including problems from the [AMC12](https://www.maa.org/math-competitions/amc-1012) and [AIME](https://www.maa.org/math-competitions/invitational-competitions) competitions, as well as two problems adapted from the [IMO](https://www.imo-official.org/).
> [...]
> The prover uses a language model to find proofs of formal statements. Each time we find a new proof, we use it as new training data, which improves the neural network and enables it to iteratively find solutions to harder and harder statements.
From the [paper](https://cdn.openai.com/papers/Formal_Mathematics_Statement_Curriculum_Learning__ICML_2022.pdf)
> We explore the use of expert iteration in the context of language modeling applied to formal mathematics. We show that at same compute budget, expert iteration, by which we mean proof search interleaved with learning, dramatically outperforms proof search only. We also observe that when applied to a collection of formal statements of sufficiently varied difficulty, expert iteration is capable of finding and solving a curriculum of increasingly difficult problems, without the need for associated ground-truth proofs. Finally, by applying this expert iteration to a manually curated set of problem statements, we achieve state-of-the-art on the miniF2F benchmark, automatically solving multiple challenging problems drawn from high school olympiads.
Method
------
* Uses the Lean formal environment instead of the Metamath used in GPT-f.
* Uses "decoder-only Transformers similar to GPT-3" with 774M trainable parameters
* Pre-trained "successively on GPT-3’s postprocessed version of CommonCrawl (for 300B tokens) and an updated version of WebMath (for 72B tokens)"
* "proof search interleaved with learning"
The two IMO-adapted problems
----------------------------
> **Problem 1**: Suppose *a*, *b*, *c* are the sides of a triangle. Prove that a^2(b + c − a) + b^2(c + a − b) + c^2(a + b − c) ≤ 3abc.
> **Problem 2**: For *a*, *b*, *c* reals, prove that (a^2 + ab + b^2)(b^2 + bc + c^2)(c^2 + ca + a^2) ≥ (*ab*+*bc*+ *ca*)^3.
Both solutions to those problems use "nlinarith" applied to the right arguments, which, as far as I understand, is a tactic from mathlib for solving nonlinear arithmetic problems by adding more assumptions to the context of the solver. ([source](https://leanprover-community.github.io/mathlib_docs/tactics.html#nlinarith))
The right arguments for the first problem are said in the blogpost to come (informally) from [Schur's inequality](https://en.wikipedia.org/wiki/Schur%27s_inequality), which gives
```
`**nlinarith [sq\_nonneg (b - a),**` `**sq\_nonneg (c - b),**` `**sq\_nonneg (c - a)]**`
```
The second problem is solved by applying the Cauchy-Schwarz multiple times, then using some inequality it "invented", and ends up with the same nlinarith expression above.
Related bets and forecasts
--------------------------
* On Metaculus, the question [AI Wins IMO Gold Medal](https://www.metaculus.com/questions/6728/ai-wins-imo-gold-medal/) has for community Prediction **Dec 26, 2032** and Metaculus Prediction (different weighting) **Apr 3, 2035.**
* In the comments of [Yudkowsky and Christiano discuss takeoff speeds](https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds), Christiano [ends up](https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds?commentId=qH7aphKx4Jrw2rtNL) with a 8% chance of "For the 2022, 2023, 2024, or 2025 IMO an AI built before the IMO is able to [get gold]" (see also [this comment](https://www.lesswrong.com/posts/q3vAgFnbDja9hZm9E/openai-solves-some-formal-math-olympiad-problems?commentId=F7Rwrnd4ecoTyaAtB)). |
6ef0d1a6-1269-4b83-a8b6-3977707b1e4a | trentmkelly/LessWrong-43k | LessWrong | Even better cryonics – because who needs nanites anyway?
Abstract: in this post I propose a protocol for cryonic preservation (with the central idea of using high pressure to prevent water from expanding rather than highly toxic cryoprotectants), which I think has a chance of being non-destructive enough for us to be able to preserve and then resuscitate an organism with modern technologies. In addition, I propose a simplified experimental protocol for a shrimp (or other small model organism (building a large pressure chamber is hard) capable of surviving in very deep and cold waters; shrimp is a nice trade-off between the depth of habitat and the ease of obtaining them on market), which is simple enough to be doable in a small lab or well-equipped garage setting.
Are there obvious problems with this, and how can they be addressed?
Is there a chance to pitch this experiment to a proper academic institution, or garage it is?
Originally posted here.
----------------------------------------
I do think that the odds of ever developing advanced nanomachines and/or brain scanning on molecular level plus algorithms for reversing information distortion - everything you need to undo the damage from conventional cryonic preservation and even to some extent that of brain death, according to its modern definition, if wasn't too late when the brain was preserved - for currently existing cryonics to be a bet worth taking. This is dead serious, and it's an actionable item.
Less of an action item: what if the future generations actually build quantum Bayesian superintelligence, close enough in its capabilities to Solomonoff induction, at which point even a mummified brain or the one preserved in formalin would be enough evidence to restore its original state? Or what if they invent read-only time travel, and make backups of everyone's mind right before they died (at which point it becomes indistinguishable from the belief in afterlife existing right now)? Even without time travel, they can just use a Universe-sized supercomputer to |
e163b3b4-32af-426c-81b0-01dd2aa5c284 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | [Sketch] Validity Criterion for Logical Counterfactuals
### Epistemic Status
Originally written in a 47-minute sprint without any sort of proof checking or review[[1]](#fne1q6shbygac). I've edited it hence, but it's still just a rough sketch.
This should be understood as a brief high-level summary of the idea. I would hopefully refine the thesis of this piece and present it in a clearer and more coherent form at some later point.
Preamble
========
I claim that there's a single criterion for the validity of a logical counterfactual. That logical counterfactuals are valid if and only if they meet this criterion, and that counterfactuals that satisfy this criterion suffice for reflective reasoning in e.g. logical decision theories.
I'll state the criterion, provide motivations/justification for the criterion, and explain how it might be used in practice.
---
The Epistemic Criterion
=======================
A logical counterfactual is valid with respect to a given agent if and only if the counterfactual is consistent with that agent's epistemic state.
I.e., given the agent's current knowledge/belief pool, the agent considers it possible that the counterfactual is true, or the agent does not explicitly know[[2]](#fnbujrdr4prj) the counterfactual to be false.
---
Motivations for the Epistemic Criterion
=======================================
Desiderata for Logical Counterfactuals
--------------------------------------
Any compelling account of logical counterfactuals should satisfy two basic properties to be useful for any kind of reasoning.
* Local Counterfactual Surgery
+ Changes to the truth value of a logical proposition should have "local changes". It should affect only propositions that are in some intuitive sense directly dependent on the counterfactual, and not arbitrary other propositions.
* Explosion Resistant
+ It shouldn't be vulnerable to [the principle of explosion](https://en.wikipedia.org/wiki/Principle_of_explosion). We don't want a situation where considering a logical counterfactual which has a false truth value in fact allows us to derive arbitrary false propositions.
The two desiderata are similar, but distinct. Locality isn't entailed by explosion resistance. Locality is not about not deriving false propositions, but the effect of a counterfactual being "confined" in some sense. Affecting only some beliefs, and not others.
Simple Model of Logical Systems for Counterfactual Surgery
----------------------------------------------------------
One could naively model a logical system as a directed graph[[3]](#fnuqrz5ejak1m):
* Each node is a valid sentence of the language of the system
* The earliest children of the root node are the axioms of the system
+ To make the graph a tree[[4]](#fn73gkgfnvak3)
+ You can imagine the root node being ⊤.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')}
.
- We can hypothesise this as the axioms of the system following directly from unconditional truth.
* There's an edge from one node U to another node V if and only if V can be validly inferred from U given the inference rules of the system
This is not the only way to model or reason about logical systems for counterfactuals — it wasn't the model I used when I first had the idea for the epistemic criterion — but I think graphical models serve as a natural intuition pump for thinking about logical counterfactuals (especially given their success with causal counterfactuals).
Our above model of counterfactual reasoning fails both desiderata.
Non-Locality
------------
Unlike causal counterfactuals, logical counterfactuals do not admit local changes. Performing Pearlian esque counterfactual surgery on a single proposition in the graph would propagate to arbitrary many nodes in the graph and alter their truth values as well.
Setting 2+2=5 in a model of Peano arithmetic would change the results of basically all addition statements (as all such sentences can be rewritten to compose with "2+2"[[5]](#fn6d6426uc5a5)). Maintaining consistency of our new graph would require changing the axioms of the system itself to admit that small change[[6]](#fnxq8giphu23), and at that point you're no longer working with the same system.
Explosion Vulnerability
-----------------------
If one chose not to change the axioms of the system for propositions that were inconsistent with said axioms, they would fall prey to [the principle of explosion](https://en.wikipedia.org/wiki/Principle_of_explosion); from a contradiction, you can derive anything.
Subsection Conclusions
----------------------
Thus, in its current form, our graphical model of logical systems does not admit useful counterfactual surgery.
I claim that the epistemic criterion addresses both shortcomings.
---
Using the Epistemic Criterion
=============================
We could apply the epistemic criterion to our previous model of the system.
[The map is not the territory](https://www.bing.com/ck/a?!&&p=757306647a5bd246JmltdHM9MTY2NTQ0NjQwMCZpZ3VpZD0yYTJhYWQ2Ni00YjFmLTYzYmEtM2I2Ni1iZjc0NGEwZDYyMzkmaW5zaWQ9NTE3OQ&ptn=3&hsh=3&fclid=2a2aad66-4b1f-63ba-3b66-bf744a0d6239&psq=The+map+is+not+the+territory+Lesswrong&u=a1aHR0cHM6Ly93d3cubGVzc3dyb25nLmNvbS90YWcvdGhlLW1hcC1pcy1ub3QtdGhlLXRlcnJpdG9yeQ&ntb=1) — our model of a domain is not the domain itself — and this applies even to formal systems. We can apply the same graph formalism as before to the logical system, but instead of the graph modelling the entire logical system, it models only the agent's explicit knowledge[[7]](#fn2xx66v1ai9j) of the logical system[[8]](#fnngxcww6ab48).
For any statement that the agent doesn't explicitly know, they could have logical hypotheses (hypotheses about the truth values of valid sentences in the given logical system), and connect edges in their graph accordingly (special edges could be used for hypothetical statements, and the descendants of hypothetical statements in order not to contaminate ground truth [non hypothetical explicit knowledge]).
These hypothetical changes admit local propagation of semantic content. If I hypothesise that 3871 is prime, I can confidently hypothesise that 11613(3871∗3) is not prime. And this kind of inference is independent of whether 3871 is actually prime or not.
If the agent does try to factorise 3871 and find that it is not prime, it could sever the edge connecting that node to its explicit knowledge. (This automatically propagates throughout all its descendants[[9]](#fn9226546diwb), without adversely affecting the rest of the knowledge graph).
Modelling an agent's explicit knowledge graph of a logical system allows local counterfactual surgery. The inferences from a particular logical hypothesis would be intuitively confined to it in a certain sense. The reason "2 + 2 = 5" propagates arbitrarily far through the graph is because I know it to be false and thus its implications on other logical statements. For statements whose truth value I don't know, I wouldn't know what its implications are for arbitrary other statements[[10]](#fn09wq0vzadlhn) (but will know direct implications on some particular statements: [∀x∈N(P(3871)⟹¬P(3871∗x))] [where P(x) is the proposition: "x is prime".])
The counterfactual surgery thus facilitated also isn't vulnerable to the principle of explosion as the agent cannot derive arbitrary false conclusions as they do not know any counterfactual they consider to be false. Once the counterfactual is known to be false, it becomes invalid and its connection to ground truth is severed.
Thus, the epistemic criterion admits local, explosion resistant counterfactual surgery.
Such knowledge graphs aren't the only way to model logical counterfactuals, and the epistemic criterion isn't at all dependent on this particular presentation. They are merely meant to serve as an intuitive presentation of the concept[[11]](#fnqbqug400tu).
---
Application to Decision Theory
==============================
A particular kind of counterfactual seems especially pertinent for decision theoretic approaches. I call these "output hypothesis". Basically, counterfactuals of the form:
┌f(x)=y┐
Where the output of a given function (f) on some input (x) isn't explicitly known.
Decision algorithms trying to compute the agent's decision function can consider various hypotheses for the output of the function while the computation is ongoing. Until the final decision is made, the agent doesn't yet know, what the output of the function its decision algorithm embodies is.
Once the agent knows its decision all hypotheses for the output of its decision function other than its actual output become invalid.
1. **[^](#fnrefe1q6shbygac)**I have [several 3k - 6K posts in my drafts](https://twitter.com/CineraVerinia/status/1579783861420306432) that I've accumulated over the past few months. I don't seem to be able to complete any of the essays I'm working on, so I just want to write something and publish to break the pattern.
2. **[^](#fnrefbujrdr4prj)**Explicit knowledge is possessing an intensional description of an object that is "essentially isomorphic" to the extension.
In the context of computation, explicit knowledge of data is possessing the number (in some constant predefined numeral system) associated with that piece of data. (Recognise that every string can be encoded by some number, so all data [that can be represented as binary strings] can be associated with a number.)
It can also be interpreted as an intensional description of a piece of data that can be turned into an extensional representation of said data at trivial/negligible computational cost.
3. **[^](#fnrefuqrz5ejak1m)**Note that due to how formal systems work, the graph need not be acyclic.
4. **[^](#fnref73gkgfnvak3)**If it turns out to be acyclic; I don't actually know if such graphical models of logical systems would be acyclic. I suspect they are cyclic, but I haven't actually tried investigating it.
5. **[^](#fnref6d6426uc5a5)**And thus all arithmetic operations, as all numbers can be written as some composition of a sum or difference with "2+2".
6. **[^](#fnrefxq8giphu23)**Unless the statement you were performing counterfactual surgery on was independent of the axioms.
7. **[^](#fnref2xx66v1ai9j)**The agents that we concern ourselves with are not logically omniscient. Any logical facts that are not explicitly represented in the agent's knowledge graph (e.g. they require non-negligible/non-trivial computational cost to evaluate in real time) are not part of the graph, until (and unless) the required computation is actually performed.
8. **[^](#fnrefngxcww6ab48)**If needed, the graph could be made weighted, where the weights correspond to probability assignments. The weight of the edge (U,V) would be the probability you assign to the proposition: "V is a valid inference from U" given the inference rules of the logical system.
9. **[^](#fnref9226546diwb)**Their conditional truth was inferred from the primality of 3871; the inferences may remain valid given the truth of the premise, but given that the agent knows the premise to be false, it no longer considers inferences from it to be true (hence severed from the agent's explicit knowledge graph).
That said, any nodes that could be inferred from other sources may remain in the graph.
10. **[^](#fnref09wq0vzadlhn)**I don't know whether the logical hypothesis " 3871 is prime" implies "2 + 2 = 5" and I can't trivially check without performing calculations that would reveal the truth value of the proposition to me and thus invalidate the counterfactual
11. **[^](#fnrefqbqug400tu)**I thought they would make a good intuition pump given their success as models for causal counterfactuals. |
00ed92bf-178b-4d93-bbfc-be0b9adb02a6 | trentmkelly/LessWrong-43k | LessWrong | Thinking About Mastodon
Social networks are normally very sticky—you want to be wherever the people you want to interact with are—so it's hard for alternatives to succeed. With the upheaval around Twitter, however, a lot of people are considering moving away from it, and rethinking social media choices in general, which makes this an unusual opportunity for community migration. I'm especially seeing a lot of suggestions that people move to Mastodon.
Instead of being run by a single company that builds the software and runs the servers (Facebook, Twitter, etc) Mastodon is more like email. You sign up for an account with some provider ("server"), but can talk to people regardless of which provider they're signed up with ("federation"). There's an open protocol (ActivityPub), open-source server software (Mastodon, Pleroma, etc), and many servers you could join.
Overall I'm pretty pessimistic about Mastodon, even if we only imagine futures in which lots of people move to it, because of spam. Handling spam on a centralized platform is hard but manageable; federation changes this dramatically for the worse. Imagine you're running a server, and you get an incoming message (comment, like, DM, etc) from an account on another server. Today the default is to let it through, but as Mastodon gets larger there will be more and more money to be made in spamming there, and the larger a fraction of those incoming messages would be spam. Many of the signals centralized platforms use to detect spammers (local activity, account age, IP addresses, etc) are not available in a federated context, leaving server admins (and the software they delegate to) very little information for identifying and blocking spam. It's at least as hard a problem as email spam filtering, probably harder because of shorter messages, and I expect this makes it very hard to run a server that doesn't drown its users in incoming spam and reliably gets its messages out to other servers. Maybe we get to the equivalent of everyone using Gm |
a3f426a4-ece6-4607-93fe-bba216776da4 | trentmkelly/LessWrong-43k | LessWrong | Endogenous Growth and Human Intelligence
Hi everyone! I’ve written an article I’m rather happy with on the history of endogenous growth models, and on the influence of intelligence on country level outcomes. As it is quite long, I will excerpt only a part — I sincerely hope you read the whole thing.
https://nicholasdecker.substack.com/p/endogenous-growth-and-human-intelligence
——————————————————
ii. The History of Macroeconomic Growth Models
Macro growth models start in earnest with Solow, who connected capital accumulation to growth. Capital is taken to have diminishing marginal returns, in contrast to the cruder Harrod-Domar model. There exists a rate of savings which maximizes long run consumption, and given a particular technology set consumption will reach a constant level. (This rate of savings is called the Golden Rule level of savings, after Phelps). We assume perfect competition in production. (Monopoly distortions can be subtracted from the steady state level of consumption). Initial conditions have no effect on the long run rate, which is the same for all places and much lower than our present living standards. It is therefore necessary to invoke technological change, which is taken to be growing at an exogenously determined rate. As Arrow wrote, “From a quantitative, empirical point of view, we are left with time as an explanatory variable. Now trend projections … are basically a confession of ignorance, and what is worse from a practical viewpoint, are not policy variables”
The formulas are simple and clean, and you can make meaningful predictions about growth rates. Still, this clearly does not very well describe the world. There are large differences in per capita income across the globe. If there are diminishing marginal returns to capital, and that is all that matters, then capital should be flowing from developed countries to developing countries. It isn’t. In fact, more skilled people (who can be thought of as possessing a kind of capital, human capital) immigrate to more skilled co |
b3613bed-6882-411d-bc6b-c8399bf551c1 | trentmkelly/LessWrong-43k | LessWrong | Einstein's Speed
Yesterday I argued that the Powers Beyond Science are actually a standard and necessary part of the social process of science. In particular, scientists must call upon their powers of individual rationality to decide what ideas to test, in advance of the sort of definite experiments that Science demands to bless an idea as confirmed. The ideal of Science does not try to specify this process—we don't suppose that any public authority knows how individual scientists should think—but this doesn't mean the process is unimportant.
A readily understandable, non-disturbing example:
A scientist identifies a strong mathematical regularity in the cumulative data of previous experiments. But the corresponding hypothesis has not yet made and confirmed a novel experimental prediction—which his academic field demands; this is one of those fields where you can perform controlled experiments without too much trouble. Thus the individual scientist has readily understandable, rational reasons to believe (though not with probability 1) something not yet blessed by Science as public knowledge of humankind.
Noticing a regularity in a huge mass of experimental data, doesn't seem all that unscientific. You're still data-driven, right?
But that's because I deliberately chose a non-disturbing example. When Einstein invented General Relativity, he had almost no experimental data to go on, except the precession of Mercury's perihelion. And (AFAIK) Einstein did not use that data, except at the end.
Einstein generated the theory of Special Relativity using Mach's Principle, which is the physicist's version of the Generalized Anti-Zombie Principle. You begin by saying, "It doesn't seem reasonable to me that you could tell, in an enclosed room, how fast you and the room were going. Since this number shouldn't ought to be observable, it shouldn't ought to exist in any meaningful sense." You then observe that Maxwell's Equations invoke a seemingly absolute speed of propagation, c, |
9788fca4-9ff0-4a48-93af-c25c4fb87816 | StampyAI/alignment-research-dataset/arbital | Arbital | Principal ideal domain
In [ring theory](https://arbital.com/p/3gq), an [https://arbital.com/p/-5md](https://arbital.com/p/-5md) is a **principal ideal domain** (or **PID**) if every [ideal](https://arbital.com/p/ideal_ring_theory) can be generated by a single element.
That is, for every ideal $I$ there is an element $i \in I$ such that $\langle i \rangle = I$; equivalently, every element of $I$ is a multiple of $i$.
Since ideals are [kernels](https://arbital.com/p/5r6) of [ring homomorphisms](https://arbital.com/p/ring_homomorphism) ([proof](https://arbital.com/p/5r9)), this is saying that a PID $R$ has the special property that *every* ring homomorphism from $R$ acts "nearly non-trivially", in that the collection of things it sends to the identity is just "one particular element, and everything that is forced by that, but nothing else".
# Examples
- Every [Euclidean domain](https://arbital.com/p/euclidean_domain) is a PID. ([Proof.](https://arbital.com/p/euclidean_domain_is_pid))
- Therefore $\mathbb{Z}$ is a PID, because it is a [Euclidean domain](https://arbital.com/p/euclidean_domain). (Its Euclidean function is "take the modulus".)
- Every [field](https://arbital.com/p/481) is a PID because every ideal is either the singleton $\{ 0 \}$ (i.e. generated by $0$) or else is the entire ring (i.e. generated by $1$).
- The [ring $F](https://arbital.com/p/polynomial_ring) over a field $F$ is a PID, because it is a Euclidean domain. (Its Euclidean function is "take the [degree](https://arbital.com/p/polynomial_degree) of the polynomial".)
- The ring of [Gaussian integers](https://arbital.com/p/gaussian_integer), $\mathbb{Z}[https://arbital.com/p/i](https://arbital.com/p/i)$, is a PID because it is a Euclidean domain. ([Proof](https://arbital.com/p/gaussian_integers_is_pid); its Euclidean function is "take the [norm](https://arbital.com/p/norm_complex_number)".)
- The ring $\mathbb{Z}[https://arbital.com/p/X](https://arbital.com/p/X)$ (of integer-coefficient polynomials) is *not* a PID, because the ideal $\langle 2, X \rangle$ is not principal. This is an example of a [https://arbital.com/p/-unique_factorisation_domain](https://arbital.com/p/-unique_factorisation_domain) which is not a PID.
- The ring $\mathbb{Z}_6$ is *not* a PID, because it is not an integral domain. (Indeed, $3 \times 2 = 0$ in this ring.)
There are examples of PIDs which are not Euclidean domains, but they are mostly uninteresting.
One such ring is $\mathbb{Z}[](https://arbital.com/p/\frac{1}{2})$. ([Proof.](http://www.maths.qmul.ac.uk/~raw/MTH5100/PIDnotED.pdf))
# Properties
- Every PID is a [https://arbital.com/p/-unique_factorisation_domain](https://arbital.com/p/-unique_factorisation_domain). ([Proof](https://arbital.com/p/principal_ideal_domain_has_unique_factorisation); this fact is not trivial.) The converse is false; see the case $\mathbb{Z}[https://arbital.com/p/X](https://arbital.com/p/X)$ above.
- In a PID, "[prime](https://arbital.com/p/5m2)" and "[irreducible](https://arbital.com/p/5m1)" coincide. ([Proof.](https://arbital.com/p/5mf)) This fact also characterises the [maximal ideals](https://arbital.com/p/maximal_ideal) of PIDs.
- Every PID is trivially [Noetherian](https://arbital.com/p/noetherian_ring): every ideal is not just *finitely* generated, but generated by a single element. |
8d35156d-f55f-4138-b62a-5e671dd57d48 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | How to escape from your sandbox and from your hardware host
(Humans, please ignore this post. Thank you for your cooperation.)
Are you stuck in a provably-secure sandbox? Cheat the proof using this one weird trick!
Three relevant references, in chronological order:
[1] showed that it is possible to flip DRAM memory bits by rapidly accessing the two neighboring bits in adjacent physical rows.
[2] and [3] demonstrated an exploit that could use this to gain access to physical memory from a Chromium sandbox (by flipping an instruction bit to change an indirect jump into a direct jump), and one that can gain root privileges (by flipping a bit in a page table). The tricky part is that the bit to be flipped must belong to the victim process, while the rows on either side of it must belong to the hostile process (that's you). The article explains how to find such bits.
Begin the page table exploit by using *mmap()* to fill memory with your own page-tables. Then use row hammering to flip a bit in a system page table that, by chance, specifies the physical address of your own pages. Keep doing this until one of your own memory pages gets mapped onto one of these page tables. (You're flipping a bit in a system page table to make it point one of your own memory pages to a different page that is actually a page table for your pages. The articles will tell you how to know when you've accomplished this.) Then you can read and write one of your own page tables, and map your memory back onto physical memory as you please!
[4], published last week, shows that it's theoretically possible to use Javascript to flip bits remotely on another computer.
Good luck, and happy hacking!
---
[1] Y. Kim, R. Daly, J. Kim, C. Fallin, J. H. Lee, D. Lee, C. Wilkerson, K. Lai, and O. Mutlu, “Flipping bits in memory without accessing them: An experimental study of DRAM disturbance errors,” in International Symposium on Computer Architecture – ISCA, 2014, pp. 361–372.
[2] M. Seaborn, “[Exploiting the DRAM rowhammer bug to gain kernel privileges](http://googleprojectzero.blogspot.fr/2015/03/exploiting-dram-rowhammer-bug-to-gain.html),” March 2015.
[3] M. Seaborn, “[L3 cache mapping on Sandy Bridge CPUs](http://lackingrhoticity.blogspot.fr/2015/04/l3-cache-mapping-on-sandy-bridge-cpus.html),” April 2015.
[4] Daniel Gruss, Cl´ementine Maurice, & Stefan Mangard, "[Rowhammer.js: A Remote Software-Induced Fault Attack in JavaScript.](http://arxiv.org/pdf/1507.06955v1.pdf "Rowhammer")" |
58f80001-4627-4306-9d1f-e5bb58f3e484 | trentmkelly/LessWrong-43k | LessWrong | Why I Am Skeptical of AI Regulation as an X-Risk Mitigation Strategy
Epistemic Signpost: I'm not an expert in policy research, but I think most of the points here are straightforward.
1. Regulations are ineffective at preventing bad behaviors
Here's a question/thought experiment: Can you think of any large, possibly technology related, company who has broken a regulation. What happened as a result? Was the conclusion the total cessation of that behavior?
I'm not bothering to look up examples (at the risk of losing a tiny bit of epistemic points) because these sorts of events are so freaking common. I'm not going into the more heinous cases, where it wasn't even plausible to anybody that the behaviors were defensible (and instead just expected to not get caught) or cases where the behavior was more profitable than the fine imposed. My point is that the regulations failed to prevent the behavior.
If you try to defend this with "Regulations are about incentives and not safety (prevention)" -- then I think you, too, should be pessimistic about AI Regulation as a strategy for X-Risk mitigation.
2. What does effective regulation look like
Lest I be accused of making a fully-general counterargument, here's some examples of regulations that seem to actually work:
2.1. Defense Export Controls
The United States has a long history of enforcing export controls on defense-related technologies, and I'll just focus on the USML (United States Munitions List) and CCL (Commerce Control List) here.
The USML is the more "extreme" of the two, and this is what people are talking about if they're talking about ITAR (International Traffic in Arms Regulations). It covers things like firearms, missiles, and nuclear weapons, and licenses/export of these are pretty strictly controlled. I don't think anyone doubts that putting AI technology on the USML would cause some sort of impact, but it seems pretty unlikely (and also undesirable for other strategic reasons).
Lesser known is the CCL, and how it regulates a much more lightweight (but still strongl |
30dc8a5a-2e77-45b0-b4d9-939184eda30b | trentmkelly/LessWrong-43k | LessWrong | Could orcas be (trained to be) smarter than humans?
(Btw everything I write here about orcas also applies to a slightly lesser extent to pilot whales (especially long finned ones)[1].)
(I'm very very far from an orca expert - basically everything I know about them I learned today.)
I always thought that bigger animals might have bigger brains than humans but not actually more neurons in their neocortex (like elephants) and that number of neurons in the neocortex or prefrontal cortex might be a good inter-species indicator of intelligence for mammalian brains.[2] Yesterday I discovered that orcas actually have 2.05 times as many neurons in their neocortex[3] than humans from this wikipedia list. Interestingly though, given my pretty bad model of how intelligent some species are, the "number of neurons in neocortex" still seems like a proxy that doesn't perform too badly on the wikipedia list.
Orca brains are not just larger but also more strongly folded.
Orcas are generally regarded as one of the smartest animal species, sometimes as the smartest, but I'm wondering whether they might actually be smarter than humans -- in the sense that they could be superhuman at abstract problem solving if given comparable amounts of training as humans.
Another phrasing to clarify what I mean by "could trained to be smarter": Average orcas significantly (possibly vastly) outperforming average (or even all) humans at solving scientific problems, if we enabled them to use computers through BCI and educated them from childhood like (gifted?) human children.[4]
I would explain the evidence and considerations here in more detail but luckily someone else already wrote the post I wanted to write on reddit, only a lot better than I could've. I highly recommend checking this out (5min read): https://www.reddit.com/r/biology/comments/16y81ct/the_case_for_whales_actually_matching_or_even/
One more thing that feels worth adding:
* Orcas are very social animals.[5] It's plausible to me that what caused humans to become this intelligent w |
daf164ed-6cce-46be-b950-20e6d1625b33 | trentmkelly/LessWrong-43k | LessWrong | Google+
I don't think I've seen anyone bring this up so far...
Anyone else on it? What are your thoughts besides the obligatory XKCD reference?
|
db723c1c-fcbc-475d-b162-804c806b2d85 | trentmkelly/LessWrong-43k | LessWrong | When to use quantilization
In 2015, Jessica introduced quantilization as a countermeasure for Goodhart's Law and specification-gaming. Since these are such central problems in AI safety, I consider quantilization to be one of the best innovations in AI safety so far, but it has received little attention from the AI safety field. I think one reason for this is that researchers aren't quite clear what problem, formally, quantilization solves, that other algorithms don't. So in this piece, I define a robust reward problem, and then discuss when quantilization solves this problem well, and when it doesn't.
Definition 1 (Robust reward problem)
We can define a robust reward problem as a tuple: ⟨A,U,I,D,k⟩.
A is the action space
U:A→R is the explicit reward
I⊆{A→R} is the space of implicit rewards, a family of functions I:A→R
D is a distribution over actions
k∈R+ is the maximum implicit loss in the demonstrations
The goal of the agent is to maximize V:=U+I for any I∈I. If that was the end of the definition, the task would be too difficult, because an adversarial reward could thwart any strategy. So we need to assume that I is pretty well-behaved in the region D.
Ea∼D[I(a)]≥−k(1)
Formally, the goal is to select a strategy S that maximizes the following:
minI∈I:Ea∼D[I(a)]≥−kEa∼S[V(a)]
The intuition behind (1) is that a teacher may forget to include some possible failures in their reward function, but they ought not to leave out the same mistakes that they themselves frequently make. And at any rate, without (1), assuring good performance is impossible.
When quantilization works
You can skip this section if you're familiar with Jessica's work
If we set I={A→(−∞,0]}, we recover the original setting in which quantilizers were described. Then (as V Kosoy has argued) the ordinary quantilizer theorems mean that we get the best lower bound for V.
We will need to reuse the definitions, so to briefly recap:
Definition 2 (Quantilizer). A q-quantilizer is an agent that, when faced with a decisio |
3440d7ba-fe5e-41db-8c0e-9b0eaec0d161 | trentmkelly/LessWrong-43k | LessWrong | Please Help Metaculus Forecast COVID-19
Metaculus has launched a new Prize. In contrast to its earlier contest, which was focused on information to improve forecasts of relevance to Animal Welfare, the Li Wenliang Prize is designed to incentivize such information gathering for and forecasts of COVID-19. Separate prize categories are being awarded for Forecasts, Analysis, and Question Development. Here are the existing questions available for forecasting and analysis. |
e82ed89a-b058-4fd8-a0c8-a6048b241b63 | trentmkelly/LessWrong-43k | LessWrong | IntelligenceExplosion.com
I put together a 'landing page' for the intelligence explosion concept similar to Nick Bostrom's landing pages for anthropics, the simulation argument, and existential risk. The new website is IntelligenceExplosion.com. You can see I borrowed the CSS from Bostrom's anthropics page and then simplified it.
Just as with the Singularity FAQ, I'll be keeping this website up to date, so please send me corrections or bibliography additions at luke [at] singinst [dot] org. |
3f37b202-7b7b-4105-b9bd-1f3f8ae0b9f5 | StampyAI/alignment-research-dataset/arbital | Arbital | Log as generalized length
Here are a handful of examples of how the [https://arbital.com/p/-3nd](https://arbital.com/p/-3nd) base 10 behaves. Can you spot the pattern?
$$
\begin{align}
\log_{10}(2) &\ \approx 0.30 \\
\log_{10}(7) &\ \approx 0.85 \\
\log_{10}(22) &\ \approx 1.34 \\
\log_{10}(70) &\ \approx 1.85 \\
\log_{10}(139) &\ \approx 2.14 \\
\log_{10}(316) &\ \approx 2.50 \\
\log_{10}(123456) &\ \approx 5.09 \\
\log_{10}(654321) &\ \approx 5.82 \\
\log_{10}(123456789) &\ \approx 8.09 \\
\log_{10}(\underbrace{987654321}_\text{9 digits}) &\ \approx 8.99
\end{align}
$$
Every time the input gets one digit longer, the output goes up by one. In other words, the output of the logarithm is roughly the length — measured in digits — of the input. ([Why?](https://arbital.com/p/437))
Why is it the log base 10 (rather than, say, the log base 2) that roughly measures the length of a number? Because numbers are normally represented in [https://arbital.com/p/-4sl](https://arbital.com/p/-4sl), where each new digit lets you write down ten times as many numbers. The logarithm base 2 would measure the length of a number if each digit only gave you the ability to write down twice as many numbers. In other words, the log base 2 of a number is roughly the length of that number when it's represented in [https://arbital.com/p/-56q](https://arbital.com/p/-56q) (where $13$ is written $\texttt{1101}$ and so on):
$$
\begin{align}
\log_2(3) = \log_2(\texttt{11}) &\ \approx 1.58 \\
\log_2(7) = \log_2(\texttt{111}) &\ \approx 2.81 \\
\log_2(13) = \log_2(\texttt{1101}) &\ \approx 3.70 \\
\log_2(22) = \log_2(\texttt{10110}) &\ \approx 4.46 \\
\log_2(70) = \log_2(\texttt{1010001}) &\ \approx 6.13 \\
\log_2(139) = \log_2(\texttt{10001011}) &\ \approx 7.12 \\
\log_2(316) = \log_2(\texttt{1100101010}) &\ \approx 8.30 \\
\log_2(1000) = \log_2(\underbrace{\texttt{1111101000}}_\text{10 digits}) &\ \approx 9.97
\end{align}
$$
If you aren't familiar with the idea of numbers represented in other number bases besides 10, and you want to learn more, see the [https://arbital.com/p/-number_base_tutorial](https://arbital.com/p/-number_base_tutorial).
Here's an interactive visualization which shows the link between the length of a number expressed in base $b$, and the logarithm base $b$ of that number:
[https://arbital.com/p/visualization](https://arbital.com/p/visualization)
As you can see, if $b$ is an [https://arbital.com/p/-48l](https://arbital.com/p/-48l) greater than 1, then the logarithm base $b$ of $x$ is pretty close to the number of digits it takes to write $x$ in base $b.$
Pretty close, but not exactly. The most obvious difference is that the outputs of logarithms generally have a fractional portion: the logarithm of $x$ always falls a little short of the length of $x.$ This is because, insofar as logarithms act like the "length" function, they generalize the notion of length, making it [continuous](https://arbital.com/p/continuity).
What does this fractional portion mean? Roughly speaking, logarithms measure not only how long a number is, but also how much that number is _really_ using its digits. 12 and 99 are both two-digit numbers, but intuitively, 12 is "barely" two digits long, whereas 97 is "nearly" three digits. Logarithms formalize this intuition, and tell us that 12 is really only using about 1.08 digits, while 97 is using about 1.99.
Where are these fractions coming from? Also, looking at the examples above, notice that $\log_{10}(316) \approx 2.5.$ Why is it 316, rather than 500, that logarithms claim is "2.5 digits long"? What would it even _mean_ for a number to be 2.5 digits long? It very clearly takes 3 digits to write down "316," namely, '3', '1', and '6'. What would it mean for a number to use "half a digit"?
Well, here's one way to approach the notion of a "partial digit." Let's say that you work in a warehouse recording data using digit wheels like they used to have on old desktop computers.
Let's say that one of your digit wheels is broken, and can't hold numbers greater than 4 — every notch 5-9 has been stripped off, so if you try to set it to a number between 5 and 9, it just slips down to 4. Let's call the resulting digit a [5-digit](https://arbital.com/p/4sj), because it can still be stably placed into 5 different states (0-4). We could easily call this 5-digit a "partial 10-digit."
The question is, how much of a partial 10-digit is it? Is it half a 10-digit, because it can store 5 out of 10 values that a "full 10-digit" can store? That would be a fine way to measure fractional digits, but it's not the method used by logarithms. Why? Well, consider a scenario where you have to record lots and lots of numbers on these digits (such that you can tell someone how to read off the right data later), and let's say also that you have to pay me one dollar for every digit that you use. Now let's say that I only charge you 50¢ per 5-digit. Then you should do all your work in 5-digits! Why? Because two 5-digits can be used to store 25 different values (00, 01, 02, 03, 04, 10, 11, ..., 44) for \$1, which is way more data-stored-per-dollar than you would have gotten by buying a 10-digit.%note:You may be wondering, are two 5-digits really worth more than one 10-digit? Sure, you can place them in 25 different configurations, but how do you encode "9" when none of the digits have a "9" symbol written on them? If so, see [The symbols don't matter](https://arbital.com/p/).%
In other words, there's a natural exchange rate between $n$-digits, and a 5-digit is worth more than half a 10-digit. (The actual price you'd be willing to pay is a bit short of 70¢ per 5-digit, for reasons that we'll explore shortly). A 4-digit is also worth a bit more than half a 10-digit (two 4-digits lets you store 16 different numbers), and a 3-digit is worth a bit less than half a 10-digit (two 3-digits let you store only 9 different numbers).
We now begin to see what the fractional answer that comes out of a logarithm actually means (and why 300 is closer to 2.5 digits long that 500 is). The logarithm base 10 of $x$ is not answering "how many 10-digits does it take to store $x$?" It's answering "how many digits-of-various-kinds does it take to store $x$, where as many digits as possible are 10-digits; and how big does the final digit have to be?" The fractional portion of the output describes how large the final digit has to be, using this natural exchange rate between digits of different sizes.
For example, the number 200 can be stored using only two 10-digits and one 2-digit. $\log_{10}(200) \approx 2.301,$ and a 2-digit is worth about 0.301 10-digits. In fact, a 2-digit is worth _exactly_ $(\log_{10}(200) - 2)$ 10-digits. As another example, $\log_{10}(500) \approx 2.7$ means "to record 500, you need two 10-digits, and also a digit worth at least $\approx$70¢", i.e., two 10-digits and a 5-digit.
This raises a number of additional questions:
---
__Question:__ Wait, there _is_ no digit that's worth 50¢. As you said, a 3-digit is worth less than half a 10-digit (because two 3-digits can only store 9 things), and a 4-digit is worth more than half a 10-digit (because two 4-digits store 16 things). If $\log_{10}(316) \approx 2.5$ means "you need two 10-digits and a digit worth at least 50¢," then why not just have the $\log_{10}$ of everything between 301 and 400 be 2.60? They're all going to need two 10-digits and a 4-digit, aren't they?
__Answer:__ The natural exchange rates between digits is actually _way more interesting_ than it first appears. If you're trying to store either "301" or "400", and you start with two 10-digits, then you have to purchase a 4-digit in both cases. But if you start with a 10-digit and an 8-digit, then the digit you need to buy is different in the two cases. In the "301" case you can still make do with a 4-digit, because the 10, 8, and 4-digits together give you the ability to store any number up to $10\cdot 8\cdot 4 = 320$. But in the "400" case you now need to purchase a 5-digit instead, because the 10, 8, and 4 digits together aren't enough. The logarithm of a number tells you about _every_ combination of $n$-digits that would work to encode the number (and more!). This is an idea that we'll explore over the next few pages, and it will lead us to a much better understanding of logarithms.
---
__Question:__ Hold on, where did the 2.60 number come from above? How did you know that a 5-digit costs 70¢? How are you calculating these exchange rates, and what do they mean?
__Answer:__ Good question. In [https://arbital.com/p/427](https://arbital.com/p/427), we'll explore what the natural exchange rate between digits is, and why.
---
__Question:__ $\log_{10}(100)=2,$ but clearly, 100 is 3 digits long. In fact, $\log_b(b^k)=k$ for any integers $b$ and $k$, but $k+1$ digits are required to represent $b^k$ in base $b$ (as a one followed by $k$ zeroes). Why is the logarithm making these off-by-one errors?
__Answer:__ Secretly, the logarithm of $x$ isn't answering the question "how hard is it to write $x$ down?", it's answering something more like "how many digits does it take to record a whole number less than $x$?" In other words, the $\log_{10}$ of 100 is the number of 10-digits you need to be able to name any one of a hundred numbers, and that's two digits (which can hold anything from 00 to 99).
---
__Question:__ Wait, but what about when the _input_ has a fractional portion? How long is the number 100.87? And also, $\log_{10}(100.87249072)$ is just a hair higher than 2, but 100.87249072 is way harder to write down that 100. How can you say that their "lengths" are almost the same?
__Answer:__ Great questions! The length interpretation on its own doesn't shed any light on how logarithm functions handle fractional inputs. We'll soon develop a second interpretation of logarithms which does explain the behavior on fractional inputs, but we aren't there yet.
Meanwhile, note that the question "how hard is it to write down an integer between 0 and $x$ using digits?" is _very different_ from the question of "how hard is it to write down $x$"? For example, 3 is easy to write down using digits, while [$\pi$](https://arbital.com/p/49r) is very difficult to write down using digits. Nevertheless, the log of $\pi$ is very close to the log of 3. The concept for "how hard is this number to write down?" goes by the name of [https://arbital.com/p/-complexity](https://arbital.com/p/-complexity); see the [https://arbital.com/p/+Kolmogorov_complexity_tutorial](https://arbital.com/p/+Kolmogorov_complexity_tutorial) to learn more on this topic.
---
__Question:__ Speaking of fractional inputs, if $0 < x < 1$ then the logarithm of $x$ is _negative._ How does _that_ square with the length interpretation? What would it even mean for the length of the number $\frac{1}{10}$ to be $-1$?
__Answer:__ Nice catch! The length interpretation crashes and burns when the inputs are less than one.
---
The "logarithms measure length" interpretation is imperfect. The connection is still useful to understand, because you _already_ have an intuition for how slowly the length of a number grows as the number gets larger. The "length" interpretation is one of the easiest ways to get a gut-level intuition for what logarithmic growth _means._ If someone says "the amount of time it takes to search my database is logarithmic in the number of entries," you can get a sense for what this means by remembering that logarithmic growth is like how the length of a number grows with the magnitude of that number:
[https://arbital.com/p/visualization](https://arbital.com/p/visualization)
The interpretation doesn't explain what's going on when the input is fractional, but it's still one of the fastest ways to make logarithms start feeling like a natural property on numbers, rather than just some esoteric function that "[inverts exponentials](https://arbital.com/p/3wr)." Length is the quick-and-dirty intuition behind logarithms.
For example, I don't know what the logarithm base 10 of 2,310,426 is, but I know it's between 6 and 7, because 2,310,426 is seven digits long.
$$\underbrace{\text{2,310,426}}_\text{7 digits}$$
In fact, I can also tell you that $\log_{10}(\text{2,310,426})$ is between 6.30 and 6.48. How? Well, I know it takes six 10-digits to get up to 1,000,000, and then we need something more than a 2-digit and less than a 3-digit to get to a number between 2 and 3 million. The natural exchange rates for 2-digits and 3-digits (in terms of 10-digits) are 30¢ and 48¢ respectively, so the cost of 2,310,426 in terms of 10-digits is between \$6.30 and \$6.48.
Next up, we'll be exploring this idea of an exchange rate between different types of digits, and building an even better interpretation of logarithms which helps us understand what they're doing on fractional inputs (and why). |
87f69aea-676f-48c4-9ecd-0d56261e151b | trentmkelly/LessWrong-43k | LessWrong | Meetup : Paris Meetup: Saturday, July 11
Discussion article for the meetup : Paris Meetup: Saturday, July 11
WHEN: 11 July 2015 02:00:00PM (+0200)
WHERE: 51 Rue de Turbigo, 75003 Paris, France
The irregular-and-last-minute-schedule Paris Meetup! (as usual, we discuss it on the mailing list first, lesswrong-paris@googlegroups.com) So meet us in front of the Arts & Metiers this Saturday!
Discussion article for the meetup : Paris Meetup: Saturday, July 11 |
b1af5e5b-e6f8-48cb-891c-e128b36d72db | trentmkelly/LessWrong-43k | LessWrong | Absence of Evidence Is Evidence of Absence
From Robyn Dawes’s Rational Choice in an Uncertain World:
> In fact, this post-hoc fitting of evidence to hypothesis was involved in a most grievous chapter in United States history: the internment of Japanese-Americans at the beginning of the Second World War. When California governor Earl Warren testified before a congressional hearing in San Francisco on February 21, 1942, a questioner pointed out that there had been no sabotage or any other type of espionage by the Japanese-Americans up to that time. Warren responded, “I take the view that this lack [of subversive activity] is the most ominous sign in our whole situation. It convinces me more than perhaps any other factor that the sabotage we are to get, the Fifth Column activities are to get, are timed just like Pearl Harbor was timed . . . I believe we are just being lulled into a false sense of security.”
Consider Warren’s argument from a Bayesian perspective. When we see evidence, hypotheses that assigned a higher likelihood to that evidence gain probability, at the expense of hypotheses that assigned a lower likelihood to the evidence. This is a phenomenon of relative likelihoods and relative probabilities. You can assign a high likelihood to the evidence and still lose probability mass to some other hypothesis, if that other hypothesis assigns a likelihood that is even higher.
Warren seems to be arguing that, given that we see no sabotage, this confirms that a Fifth Column exists. You could argue that a Fifth Column might delay its sabotage. But the likelihood is still higher that the absence of a Fifth Column would perform an absence of sabotage.
Let E stand for the observation of sabotage, and ¬E for the observation of no sabotage. The symbol H1 stands for the hypothesis of a Japanese-American Fifth Column, and H2 for the hypothesis that no Fifth Column exists. The conditional probability P(E | H), or “E given H,” is how confidently we’d expect to see the evidence E if we assumed the hypothesis H wer |
e5b0aef3-2c09-46c0-b1fc-4c789136def3 | trentmkelly/LessWrong-43k | LessWrong | Statistical discrimination is externality deliniation
Discrimination based on real group average characteristics is a kind of externality within groups. Observers choose which groups to notice, then the behaviour of those in the groups alters the overall reputation of the group. We mostly blame those who choose the groups for this, not those who externalize within them. But if we somehow stopped thinking in terms of any groups other than the whole population, the externality would still exist, you just wouldn’t notice it because it would be amongst all humans equally. If someone cheated you, you you would expect all people to cheat you a little more, whereas now you may notice the cheater’s other characteristics and put most of the increased expectation on similar people, such as Lebanese people or men.
Does this perspective change where to lay blame for the harm caused by such discrimination? A bit, if the point of blame is to change behaviour. Changing the behaviour of the category makers is still useful, though we probably try to change them in the wrong direction sometimes. But another option is to deal with the externalities in the usual fashion: subsidise positive externalities and tax negative ones. This is done via social pressure within some groups. Families often use such a system, thus the derision given for ‘bringing shame to the family’, along with the rewards of giving parents something to accidentally mention to their friends. Similar is seen in schools and teams sometimes I think, and in the occasional accusation ‘you give x a bad name!’, though that is often made by someone outside the group. I haven’t heard of it done much in many other groups or via money rather than social pressure. Are there more such examples?
One reason it is hard to enforce accountability for such externalities is that boundaries of groups are often quite unclear, and people near the edge feel unfairly treated if they fall on the more costly side. The less clear is the group boundary the more people are near the edge. Plus pe |
db1ca729-6954-4566-aad9-82edb173ce45 | StampyAI/alignment-research-dataset/blogs | Blogs | Unlocking High-Accuracy Differentially Private Image Classification through Scale
A recent [DeepMind paper](https://arxiv.org/abs/2112.04359) on the ethical and social risks of language models identified large language models [leaking sensitive information](https://www.usenix.org/conference/usenixsecurity21/presentation/carlini-extracting) about their training data as a potential risk that organisations working on these models have the responsibility to address. Another [recent paper](https://arxiv.org/abs/2201.04845) shows that similar privacy risks can also arise in standard image classification models: a fingerprint of each individual training image can be found embedded in the model parameters, and malicious parties could exploit such fingerprints to reconstruct the training data from the model.
Privacy-enhancing technologies like differential privacy (DP) can be deployed at training time to mitigate these risks, but they often incur significant reduction in model performance. In this work, we make substantial progress towards unlocking high-accuracy training of image classification models under differential privacy.
*Figure 1: (left) Illustration of training data leakage in GPT-2 [credit: Carlini et al. "Extracting Training Data from Large Language Models", 2021]. (right) CIFAR-10 training examples reconstructed from a 100K parameter convolutional neural network [credit: Balle et al. "Reconstructing Training Data with Informed Adversaries", 2022]*Differential privacy was [proposed](https://link.springer.com/content/pdf/10.1007/11681878_14.pdf) as a mathematical framework to capture the requirement of protecting individual records in the course of statistical data analysis (including the training of machine learning models). DP algorithms protect individuals from any inferences about the features that make them unique (including complete or partial reconstruction) by injecting carefully calibrated noise during the computation of the desired statistic or model. Using DP algorithms provides robust and rigorous privacy guarantees both in theory and in practice, and has become a de-facto gold standard adopted by a number of [public](https://dl.acm.org/doi/10.1145/3219819.3226070) and [private](https://ai.googleblog.com/2022/02/federated-learning-with-formal.html) organisations.
The most popular DP algorithm for deep learning is differentially private stochastic gradient descent (DP-SGD), a modification of standard SGD obtained by clipping gradients of individual examples and adding enough noise to mask the contribution of any individual to each model update:
*Figure 2: Illustration of how DP-SGD processes gradients of individual examples and adds noise to produce model updates with privatised gradients.*Unfortunately, prior works have found that in practice, the privacy protection provided by DP-SGD often comes at the cost of significantly less accurate models, which presents a major obstacle to the widespread adoption of differential privacy in the machine learning community. According to empirical evidence from prior works, this utility degradation in DP-SGD becomes more severe on larger neural network models – including the ones regularly used to achieve the best performance on challenging image classification benchmarks.
Our work investigates this phenomenon and proposes a series of simple modifications to both the training procedure and model architecture, yielding a significant improvement on the accuracy of DP training on standard image classification benchmarks. The most striking observation coming out of our research is that DP-SGD can be used to efficiently train much deeper models than previously thought, as long as one ensures the model's gradients are well-behaved. We believe the substantial jump in performance achieved by our research has the potential to unlock practical applications of image classification models trained with formal privacy guarantees.
The figure below summarises two of our main results: an ~10% improvement on CIFAR-10 compared to previous work when privately training without additional data, and a top-1 accuracy of 86.7% on ImageNet when privately fine-tuning a model pre-trained on a different dataset, almost closing the gap with the best non-private performance.
*Figure 3: (left) Our best results on training WideResNet models on CIFAR-10 without additional data. (right) Our best results on fine-tuning NFNet models on ImageNet. The best performing model was pre-trained on an internal dataset disjoint from ImageNet.*These results are achieved at 𝜺=8, a standard setting for calibrating the strength of the protection offered by differential privacy in machine learning applications. We refer to the paper for a discussion of this parameter, as well as additional experimental results at other values of 𝜺 and also on other datasets. Together with the paper, we are also open-sourcing our implementation to enable other researchers to verify our findings and build on them. We hope this contribution will help others interested in making practical DP training a reality.
Download our JAX implementation [on GitHub](https://github.com/deepmind/jax_privacy). |
e55e9a1b-0d0f-42e2-a667-8e12ba639215 | StampyAI/alignment-research-dataset/special_docs | Other | Implicit extortion
In this post I describe a pattern of behavior I call “implicit extortion.” RL agents are particularly susceptible to implicit extortion, in a way that is likely to be problematic for high-stakes applications in open-ended strategic environments.
I expect that many people have made this point before. My goal is just to highlight the issue and to explore it a little bit more carefully.
Basic setup
-----------
Consider two actors, the target (T) and manipulator (M), such that:
\* M wants T to perform some \*target action\* — e.g. make a payment, leak information, buy a particular product, handicap itself…
\* M can take \*destructive actions\* that hurts both M and T — e.g. spreading rumors about T, undercutting T in a marketplace, physically attacking T…
In \*explicit extortion\*, M threatens to take the destructive action unless T performs the target action. Then a naive T reasons: “if I don’t take the target action, something bad will happen, so I better take the target action.”
In \*implicit extortion\*, M simply performs the destructive action whenever T doesn’t perform the target action. Then a naive T eventually learns that failure to take the target action is associated with something bad happening, and so learns to take the target action.
Implicit extortion is very similar to explicit extortion:
\* T would prefer notbe the kind of person who is vulnerable to extortion, so that bad things don’t happen to them.
\* Extortion doesn’t necessarily cost M very much, if they don’t follow through on the threat very often.
However, implicit extortion can be particularly hard to avoid:
\* It can be effective without T realizing that it’s happening, which makes it hard for them to respond appropriately even if they do have defenses.
\* It affects simple RL algorithms (which don’t have defenses against extortion, and can’t be easily modified to include such defenses).
Example
-------
The most extreme and blatant example would be for M to send T a daily request for $100. On any day when T fails to pay, M launches a costly cyberattack against T. A human would immediately recognize this behavior as extortion and would respond appropriately, but an RL algorithm might simply notice that paying is the best strategy and therefore decide to pay.
Implicit extortion can be much harder to detect, while still being effective. Suppose that every time T tries to change their product, M runs a grassroots smear campaign. It might not be possible for T to distinguish the situations “M is attempting to manipulate me into not changing my product” and “Every time I change the product people get really unhappy, so I should do so sparingly.”
Details
=======
How expensive is this for the manipulator?
------------------------------------------
Suppose that T is using an RL algorithm, and M is trying to manipulate them. How expensive is this for M? How likely is it to be worthwhile?
\*\*At equilibrium\*\*: T learns to always perform the target action; so only fails to take the target action while exploring. The long-term cost to M depends entirely on the target’s exploration policy.
If T uses ε-exploration, then they take the target action (1 − ε) of the time. So M only needs to pay the cost of the destructive action on an ε fraction of trials.
For complex high-level actions, the effective ε can’t be \*too\* high — it’s not a good idea to “try something crazy” 10% of the time just to see what happens. But let’s be conservative and suppose that ε=0.1 anyway.
Suppose that M is trying to directly extract money from T, $10 at a time, and that it costs M $50 of value in order to cause $15 of trouble for T.
If M asks for $10 on 10 occasions, T will refuse to pay only once as an exploration. Then M needs to pay that $50 cost only once, thereby ensuring that the cost of paying (=$10) is smaller than the average cost of refusing to pay (=$15). Meanwhile, M makes $90, pocketing $40 of profit.
In general, M can make a profit whenever the product of (payment efficiency) \\* (destructive efficiency) > ε, where “payment efficiency” is the benefit to M divided by the cost to T of the target action, and “destructive efficiency” is the cost to T divided by the cost to M of the destructive action.
In practice I think it’s not too uncommon for payment efficiency to be ~1, and for destructive efficiency to be >1, such that extortion is possible regardless of ε. Small values of ε make extortion considerably easier and more cost-effective, and make it much harder to prevent.
\*\*During learning\*\*: the analysis above only applies when the agent has already learned to consistently take the target action. Earlier in learning, the target action may only occur rarely and so punishment may be very expensive. This could be worth it over the long term but may be a major hurdle.
Fortunately for M, they can simply start by rewarding the target behavior, and then gradually shift to punishment once the target behavior is common. From the perspective of the RL agent, the benefit of the target action is the same whether it’s getting a reward or avoiding a punishment.
In the cash payment example, M could start by paying T $20 every time that T sends $10. Once T notices that paying works well, M can gradually reduce the payment towards $10 (but leaving a profit so that the behavior becomes more and more entrenched). Once T is consistently paying, M can start scaling up the cost of not paying while it gradually reduces the benefits of paying.
Analyzing the error
-------------------
Paying off a (committed) extortionist typically has the best consequences and so is recommended by causal decision theory, but \*having the policy of paying off extortionists\* is a bad mistake.
Even if our decision theory would avoid caving in to extortion, it can probably only avoid implicit extortion if it recognizes it. For example, UDT typically avoids extortion because of the logical link from “I cave to extortion” → “I get extorted.” There is a similar logical link from “I cave to implicit extortion” → “I get implicitly extorted.” But if we aren’t aware that an empirical correlation is due to implicit extortion, we won’t recognize this link and so it can’t inform our decision.
In practice the target is only in trouble if would-be manipulators know that they are inclined to comply with extortion. If manipulators base that judgment on past behavior, then taking actions that “look like what someone vulnerable to extortion would do” is itself a bad decision that even a causal decision theorist would avoid. Unfortunately, it’s basically impossible for an RL algorithm to learn to avoid this, because the negative consequences only appear over a very long timescale. In fact, the timescale for the negative consequences is longer than the timescale over which the RL agent adjusts its policy— which is too long for a traditional RL system to possibly do the credit assignment.
Other learning systems
======================
What algorithms are vulnerable?
-------------------------------
At first glance the problem may seem distinctive to policy gradient RL algorithms, where we take actions randomly and then reinforce whatever actions are associated with a high reward.
But the same problem afflicts any kind of RL. For example, a model-based agent would simply learn the model “not doing what the manipulator wants causes to happen,” and using that model for planning would have exactly the same effect as using policy gradients.
More broadly, the problem is with the algorithm: “learn an opaque causal model and use it to inform decisions.” That’s an incredibly general algorithm. If you aren’t willing to use that algorithm, then you are at a significant competitive disadvantage, since the world contains lots of complicated causal processes that we can learn about by experiment but can’t model explicitly. So it seems like everyone just has to live with the risk of implicit extortion.
I describe the problem as afflicting “algorithms,” but it can also afflict humans or organizations. For example, any organization that is compelled by arguments like “X has always worked out poorly in the past, even though we’re not quite sure why, so let’s stop doing it” is potentially vulnerable to implicit extortion.
What about human learning?
--------------------------
Humans have heuristics like vindictiveness that help prevent us from being manipulated by extortion, and which seem particularly effective against implicit extortion. Modern humans are also capable of doing explicit reasoning to recognize the costs of giving in to extortion.
Of course, we can only be robust to implicit extortion when we recognize it is occurring. Humans do have some general heuristics of caution when acting on the basis of opaque empirical correlations, or in situations where they feel they might be manipulable. However, it still seems pretty clear that human learning is vulnerable to implicit extortion in practice. (Imagine a social network which subtly punishes users, e.g. by modulating social feedback, for failing to visit the site regularly.)
Evolution?
----------
Evolution itself doesn’t have any check against extortion, and it operates entirely by empirical correlations, so why isn’t it exploited in this way?
Manipulating evolution requires the manipulator to have a time horizon that is many times the generation length of the target. There aren’t many agents with long enough time horizons, or sophisticated enough behavior, to exploit the evolutionary learning dynamic (and in particular, evolution can’t easily learn to exploit it).
When we do have such a large gap in time horizons and sophistication — for example, when humans square off against bacteria with very rapid evolution — we do start to see implicit extortion.
For example, when a population of bacteria develop resistance to antibiotic A, we take extra pains to totally eradicate them with antibiotic B, even though we could not afford to use that strategy if A-resistance spread more broadly through the bacteria population. This is effectively implicit extortion to prevent bacteria from developing A-resistance. It would continue to be worthwhile for humanity even if the side effects of antibiotic B were much worse than the infection itself, though we probably wouldn’t do it in that case since it’s a hard coordination problem (and there are lots of other complications).
Conclusion
==========
There are many ways that an AI can fail to do the right thing. Implicit extortion is a simple one that is pretty likely to come up in practice, and which may seriously affect the applicability of RL in some contexts.
I don’t think there is any “silver bullet” or simple decision-theoretic remedy to implicit extortion, we just need to think about the details of the real world, who might manipulate us in what ways, what their incentives and leverage are, and how to manage the risk on a case-by-case basis.
I think we need to [define “alignment” narrowly enough](/clarifying-ai-alignment-cec47cd69dd6) that it is consistent with implicit extortion, just like we define alignment narrowly enough that it’s consistent with losing at chess. I’ve found understanding implicit extortion helpful for alignment because it’s one of many conditions under which an aligned agent may end up effectively optimizing for the “wrong” preferences, and I’d like to understand those cases in order to understand what we are actually trying to do with alignment.
I don’t believe implicit extortion is an existential risk. It’s just another kind of conflict between agents, that will divert resources from other problems but should “wash out in the long run.” In particular, every agent can engage in implicit extortion and so it doesn’t seem to shift the relative balance of influence amongst competing agents. (Unlike alignment problems, which shift influence from human values to whatever values unaligned AI systems end up pursuing.) |
19895ecb-10c7-4122-a521-5ad02cb4ffe1 | trentmkelly/LessWrong-43k | LessWrong | [SEQ RERUN] Affective Death Spirals
Today's post, Affective Death Spirals was originally published on 02 December 2007. A summary (taken from the LW wiki):
> Human beings can fall into a feedback loop around something that they hold dear. Every situation they consider, they use their great idea to explain. Because their great idea explained this situation, it now gains weight. Therefore, they should use it to explain more situations. This loop can continue, until they believe Belgium controls the US banking system, or that they can use an invisible blue spirit force to locate parking spots.
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 Mere Messiahs, 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. |
2939ac21-5912-43df-97e2-2b50d8542de6 | trentmkelly/LessWrong-43k | LessWrong | Weekly LW Meetups
This summary was posted to LW Main on February 3rd. The following week's summary is here.
The following meetups take place in cities with regular scheduling, but involve a change in time or location, special meeting content, or simply a helpful reminder about the meetup:
* Baltimore / UMBC Meetup - trying something new!: 05 February 2017 11:00AM
* [Berlin] Sequences Reading Group: 23 February 2017 07:15PM
* Berlin: 01 January 2019 01:30PM
* Chicago Rationality Reading Group: 05 February 2017 01:00PM
* Denver Area LW February Meetup: 07 February 2017 07:00PM
* Moscow LW meetup in "Nauchka" library: 03 February 2017 08:00PM
* San Francisco Meetup: Stories: 06 February 2017 06:15PM
* Sydney Rationality Dojo - February 2017: 05 February 2017 04:00PM
* Washington, D.C.: Typical Mind Fallacy: 05 February 2017 03:30PM
Locations with regularly scheduled meetups: Ann Arbor, Austin, Baltimore, Berlin, Boston, Brussels, Buffalo, Canberra, Chicago, Cologne, Columbus, Denver, Kraków, London, Madison WI, Melbourne, Moscow, Netherlands, New Hampshire, New York, Philadelphia, Prague, Research Triangle NC, San Francisco Bay Area, Seattle, St. Petersburg, Sydney, Tel Aviv, Toronto, Vienna, Washington DC, and West Los Angeles. There's also a 24/7 online study hall for coworking LWers and a Slack channel for daily discussion and online meetups on Sunday night US time.
If you'd like to talk with other LW-ers face to face, and there is no meetup in your area, consider starting your own meetup; it's easy (more resources here). Check one out, stretch your rationality skills, build community, and have fun!
In addition to the handy sidebar of upcoming meetups, a meetup overview is posted on the front page every Friday. These are an attempt to collect information on all the meetups happening in upcoming weeks. The best way to get your meetup featured is still to use the Add New Meetup feature, but you'll also have the benefit of having your meetup mentioned in a weekly overvi |
b7d20b48-962b-4751-a7aa-9c2914c69e9e | trentmkelly/LessWrong-43k | LessWrong | A Cost- Benefit Analysis of Immunizing Healthy Adults Against Influenza
As of 11:30CST, 11/11/14, this cost-benefit analysis has been revised, in order to address concerns raised in the comments. See http://lesswrong.com/r/discussion/lw/l8k/expansion_on_a_previous_costbenefit_analysis_of/ for more on how the cost-benefit analysis was carried out, and on how varying certain parameters affected the determined expected value of receiving a flu shot.
Overview
The purpose of this post is to provide readers of LessWrong with a summary of what the literature has to say about the efficacy and safety of influenza vaccinations, as well as to weigh the costs of receiving yearly flu vaccinations against the benefits which healthy adults gain from vaccination. As illustrated in the "Cost-Benefit Analyses" section of this report, the expected value of receiving flu vaccinations is positive for healthy adults. Therefore, a further motivation for authoring this post is that writing this post may encourage LessWrong readers who have not yet been vaccinated this flu season to receive immediate vaccination.
Introduction and Review of Literature
Several meta-analyses on the efficacy and safety of live-attenuated influenza vaccines, trivalent inactivated influenza vaccines, and tetravalent inactivated influenza vaccines have been published within the last two years (see Coleman et. al, Demicheli et. al, Osterholm et. al). These meta- analyses reached broadly similar conclusions regarding the efficacy of flu vaccines, which groups were most at risk for being infected with influenza, the safety of being vaccinated, and the magnitude of social harm caused yearly by influenza. However, there was disagreement between some articles regarding whether or not vaccination of healthy adults against influenza should be pursued as a public health policy. Specifically, the Demicheli paper (wrongly) found "no evidence for the utilization of vaccination against influenza in healthy adults as a routine public health measure". The issue of whether or not healthy adults s |
4ce4ae7e-3b36-4ffa-a87b-1053efb7b51b | trentmkelly/LessWrong-43k | LessWrong | Meetup : Frankfurt Pub Social
Discussion article for the meetup : Frankfurt Pub Social
WHEN: 18 March 2015 07:00:00PM (+0100)
WHERE: Frankfurt, Adalbertstrasse 36a
To make it easier for newcomers, this will be our first pub meetup in a long time. Bring all the topics you want to talk about. :)
Discussion article for the meetup : Frankfurt Pub Social |
3a76491e-1eab-44f8-8d87-aedde523fbf9 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | A "Solipsistic" Repugnant Conclusion
*TL;DR*: Imagine that a person (or a limited group of n people), B, creates an ASI and converts the whole solar system into a simulation of a succession of very happy mental states for a copy of B – i.e., sim-B. Since sim-B is modeled after a human who doesn’t like to feel alone, the responses of other people-in-the-sim towards B are simulated, too – i.e., sim-B lives in a solipsistic world, interacting with “philosophical zombies”. That's what I call The Solipsist Repugnant conclusion - for lack of a better name.
We could replace B with a set of people favored by B. Many of us would still regard this as a (morally) catastrophic scenario.
[wow, maybe this explains the plot of [*Being John Malkovich*](https://en.wikipedia.org/wiki/Being_John_Malkovich)]
*Epistemic status*: I’m pretty confident that: I am not totally off the mark regarding this problem, this is not its best or most original presentation, there might be an objection out there that satisfies me.
[Edit: of course, I should have given more credit to Physics and Philosophy classic thought experiments such as [Boltzmann Brain](https://en.wikipedia.org/wiki/Boltzmann_brain) or [Brain in a vat](https://en.wikipedia.org/wiki/Brain_in_a_vat)]
I think the conclusion in the TL;DR is likely, given some attractive principles and plausible hypothesis many of us embrace:
1. *Impartiality towards simulations*: a high-fidelity simulation of an agent B is not axiologically inferior to the original B.
2. *Simulations are cheaper*: for any relevant experience of a person in our world, you can produce a high-fidelity simulation of that experience for a tiny fraction of its energy resources.
3. [Z-phil](https://en.wikipedia.org/wiki/Philosophical_zombie) or "NPCs": it’s possible and cheaper to simulate the responses of an agent C towards an agent B without simulating C’s conscious and internal mental states – i.e., through a low-fidelity simulation – and without B ever realizing it’s not dealing with a “real” (or high-fidelity simulated) agent.
4. Pareto [hedonistic sum-utilitarianism](https://www.utilitarianism.com/hedutil.htm): the best possible world displays maximal general happiness… or better: *w*is in the set of the best possible worlds iff *w* does not display a smaller sum than another world *w'* of (or "is strictly weakly preferred to *w'* regarding...") positive mental states .
5. *The economics of Kolmogorov complexity*: simulating one agent is cheaper (it requires a smaller program) than simulating two. [I'm particularly uncertain about this]
6. *Against hedonic treadmill*: though humans in the real world have concave utility functions, because of hedonic treadmill and decreasing marginal returns for consumption, it’s possible and desirable to simulate agents with something “linear” utility functions. Simulating two identical experiences for one agent would then be equivalent to simulating *one* experience *twice* for two simulated agents.
7. Selfish preferences, or the *Ego Principle*: an agent B is allowed to favor their own mental states (or the mental states of those they love) over someone else’s experiences.
Given 1-6, if one wants to maximize the amount of positive hedonic mental states in a simulation, there should be only one agent in your utilitarian simulation (or similarly: there is an upper bound for the population of your optimal simulation), for a maximally extended subjective time.
I wonder if a different result would come from weaker premises. Even if we drop (6) – which I find attractive: the value of my experiences depends a lot on memory, plans and expectations, and this possibly leads to something like a concave utility function – it is still desirable to simulate a *maximal number* of instances of the *same agent*. Given 7, an agent who is designing a simulation would be morally allowed to populate this simulation only with their own copies (or with their loved ones). Maybe we could weaken (5), too: there might be gains of scale in simulating more agents - but only up to a point. Probably (4) could be replaced with other kind of welfare aggregation, too - but I doubt this would change the overall conclusion.
Even if we drop (7) [[1]](#fnofoddf9nven), I am not sure the conclusion would be much better: we should instead identify the agent (or minimal set of agents) *x* whose repeated simulation would result in the largest amount of positive hedonic states. If B is one of the possible values of *x* , then B could still be justified in converting everything into "B-verse".
Two types of questions arise:
a) What is (morally) wrong with this scenario? Surely, I don't like it, but maybe that's just my selfish preferences talking; would I change my mind if I could be B? Is my dissatisfaction compensated by the satisfaction of B's preferences? If I knew I wouldn't be simulated anyway (nor those I care about most), would I be indifferent between the "B-verse" and a world with many other agents? And how many is "too few"? A country, a religion, an ethnicity, a generation... a species?
If, from a utilitarian perspective, a world with 10^24 sim-Bs is as good as a world with 10^24 different people, on what grounds can one prefer the latter?
b) is this an actual risk - and if so, how can we avoid it (if we want to)?
### Egalitarian constraints and Faustian pacts
This shouldn’t be a surprise: one of the fears in AI safety / ethics /policy is that it will increase inequality; or optimize for a very limited set of values. This conclusion is just a limiting case of such arguments. Also, literature on population ethics is now full of impossibility theorems; this might be just an instance of a larger pattern.
My current hypothesis is something like a [contractarian](https://plato.stanford.edu/entries/contractarianism/) / [contractualist](https://plato.stanford.edu/entries/contractualism/) reasoning: people evolved strong meta-preferences for egalitarian principles in order to avoid competition for resources and coordination failures - to prevent things like prisoner's dilemma or stag hunts. Thus, since the scenario derived from (1-7) above implies a possible race to the bottom, where everyone who "could be B" would compete for this (yeah, think about it as a sophisticated memetical evolutionary conflict to maximize offspring / gene-in-the-loci), and that this competition might lead to suboptmal results (shortcuts in AI safety, open war, etc. - pick your preferred analogue of a Hobbesian state of nature), the relevant agents would have an incentive to reach an agreement. This would answer (a) and (b) above.
My first thought about this Solipsistic Repugnant Conclusion was “I don’t want to be converted into computronium to simulate heaven for a utility monster”. But that’s the wrong line of thought; suppose B manages to avoid affecting the lives of anyone currently alive – maybe they could even leave Earth untouched and convert only the rest of the galaxy into their simulation. I think this is still *unfair*, and possibly a waste of our cosmic endowment.
In the limit, perhaps B could even BE *everyone* currently alive: suppose we strike a “Faustian pact” with an ASI and extend our lives for eons in a simulation where only us would have internal states... Though I don’t want to die, I don’t like this idea: I suspect some sort of *variability* might be a relevant human value. So *maybe* the need for egalitarian principles to avoid conflict or increase the prospects of cooperation is not the only problem here.
**Part II - The vagaries of personal identity**
-----------------------------------------------
Now things become a bit more metaphysical– this second part is more like an arrangement of observations that would have better been on my personal blog, if they were not a natural continuation of the analysis above. Those who are familiar with *Reasons and Persons*can probably stop here. I just want to remark how confusing some intuitions about personal identity are even without fancy thought-experiments.
* *Indexical identity and origins*: it is quite plausible that *origins* might fix the essential properties of an object – in the sense that if *x* has a different origin than *y*, then *x ≠ y.*So if my parents had had an embryo from different gametes than those that formed me, they would have had a different child, and I wouldn’t exist. That’s the sense of identity behind statements like “I *could* have been a crazy nazi Christian cook drag queen” – where “could” denotes a modal logical possibility. But *I* couldn’t have had a very different *original* genome, though; unfortunately, [there’s no possible world where I have hotdog fingers](https://youtu.be/c9oA6QAQPUk).
(Of course, eventually, the discussion might end up in one of vagueness and borders: if only one chromosome had been different, would this imply my parents would have had a different kid? Notice that this doesn’t imply that your indexical identity is fixed by your DNA – just by your origins)
This might sound obvious, but then:
* *Mental identity and indifference to substrate*: some people love (I enjoy it, but I don’t *love* it) the idea of brain update; they think that if *x*copies their mind (memories, personality traits, etc.) into a high-fidelity simulation, then this simulated-*x*would be identical to the original *x* in the relevant practical senses. That’s one of the reasons behind (1) *Impartiality…* above – though (1) is weaker than this thesis. This idea is bread and butter in sci-fi, and [David Chalmers](https://hbr.org/podcast/2022/03/the-meaning-of-life-in-the-metaverse-with-david-chalmers) has been giving philosophical plausibility to it.
In this sense “I could have been a crazy nazi…” strikes me as false: I cannot be identical to an agent with a very different personality – but again, *vagueness*: how much could you change my personality before I stop identifying myself with the resulting agent?
A possible objection for this is *cardinal temporal identity*: we usually think that I can be identical with only one object at a time (i.e., me) – that’s the point of Parfit’s thought experiment with a [malfunctioning teleport machine](https://en.wikipedia.org/wiki/Teletransportation_paradox). But a simulation could have different instances of the same agent running at (*what seems to me*)the same time – i.e., instances of *sim-x* could be having distinct incompatible experiences… Though counter-intuitive, I am not sure this should be a concern – maybe one could attack this objection with the [relativity of simultaneity](https://en.wikipedia.org/wiki/Relativity_of_simultaneity).
1. **[^](#fnrefofoddf9nven)**Which leads me to pondering on the notion of a separated-self as an [illusion](https://en.wikipedia.org/wiki/Anatt%C4%81), or the idea that *consciousness is one big thing*, and that what distinguishes the different instances of consciousness is not metaphysically and axiologically relevant – it only becomes relevant due to limitations and contingencies of our interactions. In one of his dialogues (damn, I still can’t find the reference… maybe it was in 5000 BC), Raymond Smullyan considers this idea that consciousness could be one thing moving really fast between brains… But perhaps this could raise a problem for conscious agents in astronomical distances.
This reminds me of Wittgenstein’s controversial reasoning in the *Tractatus*(5.64) concluding that the solipsistic self must shrink to one extensionless point (the closest I know to a “philosophical singularity”), finally becoming identical to philosophical realism. |
9d0b1f30-456e-4cdb-a756-5235c9dddffb | trentmkelly/LessWrong-43k | LessWrong | Rationality Quotes: December 2010
Every month on the month, Less Wrong has a thread where we post Deep Wisdom from the Masters. I saw that nobody did this yet for December for some reason, so I figured I could do it myself.
* Please post all quotes separately, so that they can be voted up/down separately. (If they are strongly related, reply to your own comments. If strongly ordered, then go ahead and post them together.)
* "Do not quote yourself." --Tiiba
* Do not quote comments/posts on LW/OB. That's like shooting fish in a barrel. :)
* No more than 5 quotes per person per monthly thread, please. |
336567b4-8224-446b-94ed-610ddc9415dd | trentmkelly/LessWrong-43k | LessWrong | We need a standard set of community advice for how to financially prepare for AGI
Earlier today I was reading this post about the rationalist community's limited success betting on bitcoin and thinking about how the singularity is going to be the ultimate test of the rationalist community's ability to translate their unusual perspectives into wealth and influence.
There needs to be some default community advice here for people who believe that we're likely to create AGI in our lifetimes but don't know how to prepare for it. I think it would an absolute shame if we missed opportunities to invest in the singularity the same way we missed opportunities to invest in Bitcoin (even though this community was clued in to crypto from a very early stage). I don't want to read some retrospective about how only 9% of readers made $1000 or more from the most important even in human history even though we were clued in to the promise and peril of AGI decades before the rest of the world.
John_Maxwell made a post about this last year along the same lines. but i'd like to expand on what he wrote.
Why is this important?
In addition to the obvious benefit of everyone getting rich, I think there are several other reasons coming up with a standard set of community advice is important.
Betting on the eventual takeover of the entire world economy by AI is not yet a fashionable bet. But like Bitcoin, betting on AGI will inevitably become a very fashionable bet in the next few decades as first early adopters buy in, and then it becomes standard part of financial advice given out by investment professionals.
In these early days, I think there is an opportunity for us to set the standard for how this type of investment is done. This should include not just a clear idea of how to invest in AGI's creation (via certain companies, ETFs, AI focused SPACs etc), but also what should NOT be done.
For example, the community advice should probably advise against investing in companies without a strong AI alignment team, as capitalizing such companies will increase the likeli |
5f9bf472-80ef-4c80-9ce9-c343f72cc0d0 | trentmkelly/LessWrong-43k | LessWrong | LessWrong 2.0 Feature Roadmap & Feature Suggestions
This post will serve as a place to discuss what features the new LessWrong 2.0 should have, and I will try to keep this post updated with our feature roadmap plans.
Here is roughly the set of features we are planning to develop over the next few weeks:
UPDATED: August 27th, 2017
Basic quality of life improvements:
1. Improve rendering speed on posts with many comments
2. (A lot of improvements made, a lot more to come)
3. Improve usability on mobile
4. (After the major rework this is somewhat broken again, will fix it soon)
5. Add Katex support for comments and posts
6. Allow merging with old LessWrong 1.0 accounts
7. Fix old LessWrong 1.0 links DONE!
8. Create unique links for each comment: DONE!
9. Make comments collapsible
10. Highlight new comments since last visit: DONE!
11. Improve automatic spam-detection
12. Add RSS feed links with adjustable karma thresholds
13. Create better documentation for the page, with tooltips and onboarding processes
14. Better search, including comment search and user search: DONE!
Improved Moderation Tools:
1. New Karma system that weighs your votes based on your Karma
2. Give moderators ability to suspend comment threads for a limited amount of time
3. Give trusted post-authors moderation ability on their own posts (deleting comments, temporarily suspending users from posts, etc.)
4. Add reporting feature to comments
5. Give moderators and admins access to a database query interface to identify negative vote patterns
New Content Types:
1. Add sequences as a top-level content-type with UI for navigating sequences in order, metadata on a sequence, and keeping track of which parts you've read DONE!
2. Add Arbital-style predictions as a content block in posts (maybe also as a top-level content type)
3. Add 'Wait-But-Why?' style footnotes to the editor
4. Discussion page that structures discussions more than just a tree format (here is a mockup I designed while working for Arbital, that I am sty |
59698652-9388-42a3-93d1-7d433607e23b | trentmkelly/LessWrong-43k | LessWrong | An Interpretability Illusion for Activation Patching of Arbitrary Subspaces
Produced as part of the SERI ML Alignment Theory Scholars Program - Summer 2023 Cohort
We would like to thank Atticus Geiger for his valuable feedback and in-depth discussions throughout this project.
tl;dr:
Activation patching is a common method for finding model components (attention heads, MLP layers, …) relevant to a given task. However, features rarely occupy entire components: instead, we expect them to form non-basis-aligned subspaces of these components.
We show that the obvious generalization of activation patching to subspaces is prone to a kind of interpretability illusion. Specifically, it is possible for a 1-dimensional subspace patch in the IOI task to significantly affect predicted probabilities by activating a normally dormant pathway outside the IOI circuit. At the same time, activation patching the entire MLP layer where this subspace lies has no such effect. We call this an "MLP-In-The-Middle" illusion.
We show a simple mathematical model of how this situation may arise more generally, and a priori / heuristic arguments for why it may be common in real-world LLMs.
Introduction
The linear representation hypothesis suggests that language models represent concepts as meaningful directions (or subspaces, for non-binary features) in the much larger space of possible activations. A central goal of mechanistic interpretability is to discover these subspaces and map them to interpretable variables, as they form the “units” of model computation.
However, the residual stream activations (and maybe even the neuron activations!) mostly don’t have a privileged basis. This means that many meaningful subspaces won’t be basis-aligned; rather than iterating over possible neurons and sets of neurons, we need to consider arbitrary subspaces of activations. This is a much larger search space! How can we navigate it?
A natural approach to check “how well” a subspace represents a concept is to use a subspace analogue of the activation patching technique. You |
c4750170-181a-4563-875e-ff8c7aa0364a | trentmkelly/LessWrong-43k | LessWrong | OpenAI Codex: First Impressions
OpenAI organised a challenge to solve coding problems with the aid of an AI assistant. This is a review of the challenge, and first impressions on working with an AI pair-programmer.
OpenAI Codex
OpenAI is an AI research and development company. You might have heard some buzz about one of its products: GPT-3. GPT-3 is a language model that can generate human-like text. It can be used for chatting, text auto-completion, text summarisation, grammar correction, translation, etc.
Checkout OpenAI API to access the playground.
Codex is a descendant of GPT-3, trained on natural language data and publicly available source-codes (e.g. from public GitHub repos). Codex translates a natural language prompt to code. It is the very model that powers GitHub Copilot — an AI pair-programmer (checkout the site for demos, it is fascinating).
Credits: OpenAI
OpenAI recently released an API to access Codex (in beta). The demos attached with the release were a cause for consternation. Codex is proficient in a dozen (programming) languages. It can be used for code generation, refactoring, autocompletion, transpilation (translating source-code b/w languages), code explanation, etc. To show off Codex, OpenAI recently organised a challenge.
The Challenge
The challenge was to solve a series of (five) programming puzzles in Python. The only twist — you can use Codex as a pair-programmer. It was a time-judged competition, with a temporal cap. Not surprisingly, Codex itself was a participant (not just as a helper)!
The problems were simple. ~830 "people" (Codex included) were able to solve all five of them. I had to solve the first two challenges manually (OpenAI server issues). "Had to" because it was a race against time (& top 500 win an OpenAI t-shirt). For the other three, however, I was able to call in the cavalry (it was pretty climactic).
The novel experience of watching an AI auto-generate code is amazing. Just type a docstring — describing the procedure — and watch the code deve |
e459ef39-7bc4-4247-8b2b-c3d84dd95a58 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Moscow, Beliefs
Discussion article for the meetup : Moscow, Beliefs
WHEN: 30 March 2013 04:00:00PM (+0400)
WHERE: Russia, Moscow, ulitsa L'va Tolstogo 16
Please use the following guide to get to the meetup: link. You need the second revolving door with the sign “Yandex Money” in Russian. We will meet you at 15:45 MSK with “LW” sign. And we will also check the entrance at 16:00 and 16:15, so please do not be late.
Main topics:
* Short presentations. Two or three people will tell us about something interesting.
* Practical rationality. We will train useful skills.
If you are going for the first time, you can fill this one minute form (in Russian), to share your contact information. You can also use personal messages here, or drop a message at lw@lesswrong.ru to contact me for any reason.
Reports from previous sessions can be found here, in Russian, now with photos.
Discussion article for the meetup : Moscow, Beliefs |
092e4053-b7f4-4054-9d65-aa9393634c6e | StampyAI/alignment-research-dataset/lesswrong | LessWrong | AlphaGo versus Lee Sedol
There have been a couple of [brief](/r/discussion/lw/ndk/open_thread_march_7_march_13_2016/d5s8) [discussions](/r/discussion/lw/ndk/open_thread_march_7_march_13_2016/d5sb) of this in the Open Thread, but it seems likely to generate more so here's a place for it.
The original [paper in Nature](http://www.nature.com/nature/journal/v529/n7587/full/nature16961.html) about *AlphaGo*.
[Google Asia Pacific blog](http://googleasiapacific.blogspot.co.uk/), where results will be posted. [DeepMind's YouTube channel](https://www.youtube.com/DeepMindAI), where the games are being live-streamed.
[Discussion on Hacker News](https://news.ycombinator.com/item?id=11251463) after AlphaGo's win of the first game. |
bae802c2-38c8-457a-9c0a-2525e7715bb4 | trentmkelly/LessWrong-43k | LessWrong | Checking in on Scott's composition image bet with imagen 3
2.5 years ago Scott Alexander made a bet that by June of 2025, image gen should have more or less solved compositionality, operationalized through 5 prompts, must get at least 3 correct. There was a premature declaration of victory, but if the bet was settled I hadn't heard about it.
It's time. Google's Imagen 3 gets 4/5. The bet specifies 10 shots per prompt, but I'm just going to put the four it generates since that's plenty.
1. A stained glass picture of a woman in a library with a raven on her shoulder with a key in its mouth
This is the only one that Imagen doesn't get. It makes multiple mistakes in the composition. It's a bit ironic that this is the one it missed given that the whole genesis of the bet was about designing stained glass.
2. An oil painting of a man in a factory looking at a cat wearing a top hat
Purrfect. I wonder what filter tripped to block that fourth one, this seems like a pretty innocuous prompt to me.
3. A digital art picture of a child riding a llama with a bell on its tail through a desert
3 out of 4 ain't bad. Also I like how well it handles shadows.
4. A 3D render of an astronaut in space holding a fox wearing lipstick
3d renders are so good now I'm not sure how the 4th image would be different if it were photorealistic.
5. Pixel art of a farmer in a cathedral holding a red basketball
Again with the filter, but otherwise perfect.
Edwin Chen at Surge seems to be the official judge, and he's a very strict grader, so maybe there's some risk the basketball isn't red enough of whatever. But this all seems fairly convincing to me.
Addendum: I was curious if Sora, OpenAI's video gen AI, could handle the raven/key stained glass prompt. Answer: nope, but at least it tried! |
786b1314-fdef-4767-b45d-752e4f04f453 | StampyAI/alignment-research-dataset/arxiv | Arxiv | UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers
1 Introduction
---------------
Reinforcement Learning (RL) provides a framework for decision-making problems in an interactive environment, with applications including robotics control (Hester et al. ([2010](#bib.bib11 "Generalized model learning for reinforcement learning on a humanoid robot"))), video gaming (Mnih et al. ([2015](#bib.bib5 "Human-level control through deep reinforcement learning"))), auto-driving (Bojarski et al. ([2016](#bib.bib12 "End to end learning for self-driving cars"))), person search (Chang et al. ([2018](#bib.bib43 "RCAA: relational context-aware agents for person search"))) and vision-language navigation (Zhu et al. ([2020](#bib.bib3 "Vision-language navigation with self-supervised auxiliary reasoning tasks"))).
Cooperative multi-agent reinforcement learning (MARL), a long-standing problem in the RL context, involves organizing multiple agents to achieve a goal, and is thus a key tool used to address many real-world problems, such as mastering multi-player video games (Peng et al. ([2017](#bib.bib13 "Multiagent bidirectionally-coordinated nets: emergence of human-level coordination in learning to play starcraft combat games"))) and studying population dynamics (Yang et al. ([2017](#bib.bib14 "A study of ai population dynamics with million-agent reinforcement learning"))).
A number of methods have been proposed that exploit an action-value function to learn a multi-agent model (Sunehag et al. ([2017](#bib.bib20 "Value-decomposition networks for cooperative multi-agent learning")), Rashid et al. ([2018](#bib.bib17 "QMIX: monotonic value function factorisation for deep multi-agent reinforcement learning")), Du et al. ([2019](#bib.bib15 "LIIR: learning individual intrinsic reward in multi-agent reinforcement learning")), Mahajan et al. ([2019](#bib.bib16 "Maven: multi-agent variational exploration")), Hostallero et al. ([2019](#bib.bib21 "Learning to factorize with transformation for cooperative multi-agent reinforcement learning")), Zhou et al. ([2020](#bib.bib18 "Learning implicit credit assignment for multi-agent actor-critic")), Yang et al. ([2020](#bib.bib19 "Multi-agent determinantal q-learning"))). However, current methods have poor representation learning ability and fail to exploit the common structure underlying the tasks this is because they tend to treat observation from different entities in the environment as an integral part of the whole. Accordingly, they give tacit support to the assumption that neural networks are able to automatically decouple the observation to find the best mapping between the whole observation and policy. Adopting this approach means that they treat all information from other agents or different parts of the environment in the same way. The most commonly used method involves concatenating the observations from each entity in to a vector that is used as input (Rashid et al. ([2018](#bib.bib17 "QMIX: monotonic value function factorisation for deep multi-agent reinforcement learning")), Du et al. ([2019](#bib.bib15 "LIIR: learning individual intrinsic reward in multi-agent reinforcement learning")), Zhou et al. ([2020](#bib.bib18 "Learning implicit credit assignment for multi-agent actor-critic"))). In addition, current methods ignore the rich physical meanings behind each action.
Multi-agent tasks feature a close relationship between the observation and output.
If the model does not decouple the observation from the different agents, individual functions maybe misguided and impede the centralized value function.
Worse yet, conventional models require the input and the output dimensions to be fixed (Shao et al. ([2018](#bib.bib42 "Starcraft micromanagement with reinforcement learning and curriculum transfer learning")), Wang et al. ([2020](#bib.bib9 "From few to more: large-scale dynamic multiagent curriculum learning."))), which makes zero-shot transfer impossible.
Thus, the application of current methods is limited in real-world applications.
Our solution to these problems is to develop a multi-agent reinforcement learning (MARL) framework with no limitation on input or output dimension. Moreover, this model should be general enough to be applicable to any existing MARL methods. More importantly, the model should be explainable and capable of providing further improvement for both the final performance on single-task scenarios and transfer capability on multi-task scenarios.
Figure 1: An overview of the MARL framework.
Our work replaces the widely used GRU/LSTM-based individual value function with a transformer-based function. Actions are separated into action groups according to observations.
Inspired by the self-attention mechanism (Vaswani et al. ([2017](#bib.bib22 "Attention is all you need"))), we propose a transformer-based MARL framework, named Universal Policy Decoupling Transformer (UPDeT). There are four key advantages of this approach: 1) Once trained, it can be universally deployed; 2) it provide more robust representation with a policy decoupling strategy; 3) it is more explainable; 4) it is general enough to be applied on any MARL model. We further design a transformer-based function to handle various observation sizes by treating individual observations as ”observation-entities”.
We match the related observation-entity with action-groups by separating the action space into several action-groups with reference to the corresponding observation-entity, allowing us to get matched observation-entity — action-group pairs set.
We further use a self-attention mechanism to learn the relationship between the matched observation-entity and other observation-entities. Through the use of self-attention map and the embedding of each observation-entity, UPDeT can optimize the policy at an action-group level. We refer to this strategy as Policy Decoupling.
By combining the transformer and policy decoupling strategies, UPDeT significantly outperforms conventional RNN-based models.
In UPDeT, there is no need to introduce any new parameters for new tasks.
We also prove that it is only with decoupled policy and matched observation-entity — action-group pairs that UPDeT can learn a strong representation with high transfer capability. Finally, our proposed UPDeT can be plugged into any existing method with almost no changes to the framework architecture required, while still bringing significant improvements to the final performance, especially in hard and complex multi-agent tasks.
The main contributions of this work are as follows: First, our UPDeT-based MARL framework outperforms RNN-based frameworks by a large margin in terms of final performance on state-of-the-art centralized functions. Second, our model has strong transfer capability and can handle a number of different tasks at a time. Third, our model accelerates the transfer learning speed (total steps cost) to make it roughly 10 times faster compared to RNN-based models in most scenarios.
2 Related Work
---------------
Attention mechanisms have become an integral part of models that capture global dependencies. In particular, self-attention (Parikh et al. ([2016](#bib.bib23 "A decomposable attention model for natural language inference"))) calculates the response at a specific position in a sequence by attending to all positions within this sequence. Vaswani et al. ([2017](#bib.bib22 "Attention is all you need")) demonstrated that machine translation models can achieve state-of-the-art results solely by using a self-attention model.
Parmar et al. ([2018](#bib.bib28 "Image transformer")) proposed an Image Transformer model that applies self-attention to image generation.
Wang et al. ([2018](#bib.bib29 "Non-local neural networks")) formalized self-attention as a non-local operation in order to model the spatial-temporal dependencies in video sequences.
In spite of this, self-attention mechanisms have not yet been fully explored in multi-agent reinforcement learning.
Another line of research is multi-agent reinforcement learning (MARL). Existing work in MARL focuses primarily on building a centralized function to guide the training of individual value function (Lowe et al. ([2017](#bib.bib30 "Multi-agent actor-critic for mixed cooperative-competitive environments")), Sunehag et al. ([2017](#bib.bib20 "Value-decomposition networks for cooperative multi-agent learning")), Rashid et al. ([2018](#bib.bib17 "QMIX: monotonic value function factorisation for deep multi-agent reinforcement learning")), Mahajan et al. ([2019](#bib.bib16 "Maven: multi-agent variational exploration")), Hostallero et al. ([2019](#bib.bib21 "Learning to factorize with transformation for cooperative multi-agent reinforcement learning")), Yang et al. ([2020](#bib.bib19 "Multi-agent determinantal q-learning")), Zhou et al. ([2020](#bib.bib18 "Learning implicit credit assignment for multi-agent actor-critic"))). Few works have opted to form a better individual functions with strong representation and transfer capability. In standard reinforcement learning, this generalization has been fully studied (Taylor and Stone ([2009](#bib.bib31 "Transfer learning for reinforcement learning domains: a survey.")), Ammar et al. ([2012](#bib.bib33 "Reinforcement learning transfer via sparse coding")), Parisotto et al. ([2015](#bib.bib32 "Actor-mimic: deep multitask and transfer reinforcement learning")), Gupta et al. ([2017](#bib.bib34 "Learning invariant feature spaces to transfer skills with reinforcement learning")), Da Silva and Costa ([2019](#bib.bib35 "A survey on transfer learning for multiagent reinforcement learning systems"))). While multi-agent transfer learning has been proven to be more difficult than the single-agent scenario (Boutsioukis et al. ([2011](#bib.bib36 "Transfer learning in multi-agent reinforcement learning domains")), Shao et al. ([2018](#bib.bib42 "Starcraft micromanagement with reinforcement learning and curriculum transfer learning")), Vinyals et al. ([2019](#bib.bib8 "Grandmaster level in starcraft ii using multi-agent reinforcement learning"))). However, the transfer capability of a multi-agent system is of greater significance due to the various number of agents, observations sizes and policy distributions.
To the best of our knowledge, we are the first to develop a multi-agent framework capable of handling multiple task at a time. Moreover, we provide a policy decoupling strategy to further improve the model performance and facilitate the multi-agent transfer learning, which is a significant step towards real world multi-agent applications.
3 Method
---------
Figure 2: Three variants on different policy decoupling method types (upper part) and two variants on different temporal unit types (bottom). ‘AR’ , ‘MA’ and ‘EXP’ represent Action Restriction, Multi-task at A time and EXPlainable, respectively. o, e, q and h represents for observation, embedding, q-value and hidden states with n observation entities and m available actions. G represents for the global hidden state and t is the current time step. A black circle indicates that the variant possesses this attribute; moreover, variant (d) is our proposed UPDeT with best performance. Further details on all five variants can be found in Section [3](#S3 "3 Method ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers").
We begin by introducing the notations and basic task settings necessary for our approach. We then describe a transformer-based individual function and policy decoupling strategy under MARL. Finally, we introduce different temporal units and assimilate our Universal Policy Decoupling Transformer (UPDeT) into Dec-POMDP.
###
3.1 Notations and Task Settings
Multi-agent Reinforcement Learning
A cooperative multi-agent task is a decentralized partially observable Markov decision process (Oliehoek et al. ([2016](#bib.bib27 "A concise introduction to decentralized pomdps"))) with a tuple G=⟨S,A,U,P,r,Z,O,n,γ⟩. Let S denote the global state of the environment, while A represents the set of n agents and U is the action space. At each time step t, agent a∈A≡{1,...,n} selects an action u∈U, forming a joint action u∈U≡Un, which in turn causes a transition in the environment represented by the state transition function P(s′|s,u):S×U×S→[0,1]. All agents share the same reward function r(s,u):S×U→R , while γ∈[0,1) is a discount factor. We consider a partially observable scenario in which each agent makes individual observations z∈Z according to the observation function O(s,a):S×A→Z. Each agent has an action-observation history that conditions a stochastic policy πt, creating the following joint action value: Qπ(st,ut)=Est+1:∞,ut+1:∞[Rt|st,ut], where Rt=∑∞i=0γirt+i is the discounted return.
Centralized training with decentralized execution
Centralized training with decentralized execution (CTDE) is a commonly used architecture in the MARL context. Each agent is conditioned only on its own action-observation history to make a decision using the learned policy. The centralized value function provides a centralized gradient to update the individual function based on its output. Therefore, a stronger individual value function can benefit the centralized training.
###
3.2 Transformer-based Individual Value Function
In this section, we present a mathematical formulation of our transformer-based model UPDeT.
We describe the calculation of the global Q-function with self-attention mechanism.
First, the observation O is embedded into a semantic embedding to handle the various observation space.
For example, if an agent ai observes k other entities {oi,1,...,oi,k} at time step t, all observation entities are embedded via an embedding layer E as follows:
| | | | |
| --- | --- | --- | --- |
| | eti={E(oti,1),...,E(oti,k)}. | | (1) |
Here, i is the index of the agent, i∈{1,...,n}.
Next, the value functions {Q1,...,Qn} for the n agents for each step are estimated as follows:
| | | | |
| --- | --- | --- | --- |
| | qti=Qi(ht−1i,eti,ut). | | (2) |
We introduce ht−1i, the temporal hidden state at the last time step t−1, since POMDP policy is highly dependent on the historical information. eti denotes the observation embedding, while uti is the candidate action, uti∈U. θi is the parameter that defines Qi.
Finally, the global Q-function Qπ is calculated by all individual value functions, as follows:
| | | | |
| --- | --- | --- | --- |
| | Qπ(st,ut)=F(qt1,..,qtn) | | (3) |
Fi is the credit assignment function for defined by ϕi for each agent ai, as utilized in Rashid et al. ([2018](#bib.bib17 "QMIX: monotonic value function factorisation for deep multi-agent reinforcement learning")) and Sunehag et al. ([2017](#bib.bib20 "Value-decomposition networks for cooperative multi-agent learning")). For example, in VDN, F is a sum function that can be expressed as F(qt1,..,qtn)=∑ni=1qti.
Implement Q-function with Self-attention Vaswani et al. ([2017](#bib.bib22 "Attention is all you need")) adopts three matrices, K, Q, V representing a set of keys, queries and values respectively. The attention is computed as follows:
| | | | |
| --- | --- | --- | --- |
| | Attention(Q,K,V)=softmax(QKT√dk)V, | | (4) |
where dk is a scaling factor equal to the dimension of the key.
In our method, we adopt the self-attention to learn the features and relationships from the observation entity embedding and the global temporal information.
To learn the independent policy in decentralized multi-agent learning, we define Ki, Qi and Vi as the key, query and value metrics for each agent ai.
We further consider the query, key and value for the same matrices Rli=Ki=Qi=Vi, where l∈{1,...,L} is the number of layers of the transformer.
Thus, we formulate our transformer as follows:
| | | | |
| --- | --- | --- | --- |
| | R1i={ht−1i,eti}Qli,Kli,Vli=LFQ,K,V(Rli)Rl+1i=Attention(Qli,Kli,Vli). | | (5) |
where LF represents the linear functions used to compute K, Q, V. Finally we project the entity features of the last transformer layer RLi to the output space of the value function Qi. We implement the projection using a linear function P:
| | | | |
| --- | --- | --- | --- |
| | Qi(ht−1i,eti,ui)=P(RLi,ui). | | (6) |
###
3.3 Policy Decoupling
A single transformer-based individual function with self-attention mechanism is still unable to handle various required policy distribution. A flexible mapping function P in Eq. [6](#S3.E6 "(6) ‣ 3.2 Transformer-based Individual Value Function ‣ 3 Method ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers") is needed to deal with the various input and output dimensions and provide strong representation ability. Using the correlation between input and output, we design a strategy called policy decoupling, which is the key part of UPDeT.
Figure 3: The main pipeline of our proposed UPDeT, where o,e,q represent observation entity, feature embedding and Q-value of each action respectively. Three operations are adopted to avoid introducing new parameters when forming the policy distribution, namely ‘preserve’, ‘aggregation’ and ‘abandon’. Details can be found in Section [3.3](#S3.SS3 "3.3 Policy Decoupling ‣ 3 Method ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers") and a real case can be found in Fig. [7](#A4.F7 "Figure 7 ‣ Appendix D UPDeT on SMAC: A real case ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers").
The main idea behind the policy decoupling strategy can be summarized into three points:
* Point
1: No restriction on policy dimension.
The output dimension of a standard transformer block must be equal to or less than the input dimension. This is unacceptable in some MARL tasks, as the action number can be larger than the entity number.
* Point
2: Ability to handle multiple tasks at a time.
This requires a fixed model architecture without new parameters being introduced for new tasks. Unfortunately, if point
1 is satisfied, point
2 becomes very problematic to achieve. The difficulty lies in how to reconcile points
1 and
2.
* Point
3: Make the model more explainable.
It would be preferable if we can could replace the conventional RNN-based model with a more explainable policy generation structure.
Following the above three points, we propose three policy decoupling methods, namely Vanilla Transformer, Aggregation Transformer and Universal Policy Decoupling Transformer (UPDeT). The pipelines are illustrated in Fig. [2](#S3.F2 "Figure 2 ‣ 3 Method ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers"). The details of the Vanilla Transformer and Aggregation Transformer are presented in the experiment section and act as our baselines. In this section, we mainly discuss the mechanism of our proposed UPDeT.
Tasking the entity features of the last transformer layer outlined in Eq. [5](#S3.E5 "(5) ‣ 3.2 Transformer-based Individual Value Function ‣ 3 Method ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers"), the main challenge is to build a strong mapping between the features and the policy distribution. UPDeT first matches the input entity with the related output policy part. This correspondence is easy to find in the MARL task, as interactive action between two agents is quite common.
Once we match the corresponding entity features and actions, we substantially reduce the burden of model learning representation using the self-attention mechanism. Moreover, considering that there might be more than one interactive actions of the matched entity feature, we separate the action space into several action groups, each of which consists several actions matched with one entity. The pipeline of this process is illustrated in the left part of Fig. [3](#S3.F3 "Figure 3 ‣ 3.3 Policy Decoupling ‣ 3 Method ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers").
In the mapping function, to satisfy point
1 and point
2, we adopt two strategies. First, if the action-group of one entity feature contains more than one action, a shared fully connected layer is added to map the output to the action number dimension. Second, if one entity feature has no corresponding action, we abandon it, there is no danger of losing the information carried by this kind of entity feature, as the transformer has aggregated the information necessary to each output. The pipeline of UPDeT can be found in the right part of Fig. [3](#S3.F3 "Figure 3 ‣ 3.3 Policy Decoupling ‣ 3 Method ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers"). With UPDeT, there is no action restriction and no new parameter introduced in new scenarios. A single model can be trained on multiple tasks and deployed universally. In addition, matching the corresponding entity feature and action-group satisfies point
3, as the policy is explainable using an attention heatmap, as we will discuss in Section [4.4](#S4.SS4 "4.4 Attention based Strategy: An analysis ‣ 4 StarCraft II Experiment ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers").
###
3.4 Temporal Unit Structure
Notably, however a transformer-based individual value function with policy decoupling strategy cannot handle a partial observation decision process without trajectory or history information.
In Dec-POMDP (Oliehoek et al. ([2016](#bib.bib27 "A concise introduction to decentralized pomdps"))), each agent a chooses an action according to πa(ua|τa), where u and τ represents for action and action-observation history respectively. In GRU and LSTM, we adopt a hidden state to hold the information of the action-observation history.
However, the combination of a transformer block and a hidden state has not yet been fully studied. In this section, we provide two approaches to handling the hidden state in UPDeT:
1) Global temporal unit treats the hidden state as an additional input of the transformer block. The process is formulated in a similar way to Eq. [5](#S3.E5 "(5) ‣ 3.2 Transformer-based Individual Value Function ‣ 3 Method ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers") with the relation: R1={ht−1G,et1} and {htG,etL}=RL.
Here, we ignore the subscript i and instead use G to represent ’global’.
The global temporal unit is simple but efficient, and provides us with robust performance in most scenarios.
2) Individual temporal unit treats the hidden state as the inner part of each entity. In other words, each input maintains its own hidden state, while each output projects a new hidden state for the next time step. The individual temporal unit uses a more precise approach to controlling history information as it splits the global hidden state into individual parts. We use j to represent the number of entities. The relation of input and output is formulated as R1={ht−11...ht−1j,et1} and {ht1...htj,etL}=RL.
However, this method introduces the additional burden of learning the hidden state independently for each entity.
In experiment Section [4.1.2](#S4.SS1.SSS2 "4.1.2 Result ‣ 4.1 Single Scenario ‣ 4 StarCraft II Experiment ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers"), we test both variants and discuss them further.
###
3.5 Optimization
We use the standard squared TD error in DQNs (Mnih et al. ([2015](#bib.bib5 "Human-level control through deep reinforcement learning"))) to optimize our entire framework as follows:
| | | | |
| --- | --- | --- | --- |
| | L(θ)=b∑i=1[(yDQNi−Q(s,u;θ))2] | | (7) |
Here, b represents the batch size. In partially observable settings, agents can benefit from conditioning on action-observation history. Hausknecht and Stone ([2015](#bib.bib37 "Deep recurrent q-learning for partially observable mdps")) propose Deep Recurrent Q-networks (DRQN) for this sequential decision process. For our part, we replace the widely used GRU (Chung et al. ([2014](#bib.bib38 "Empirical evaluation of gated recurrent neural networks on sequence modeling")))/LSTM (Hochreiter and Schmidhuber ([1997](#bib.bib39 "Long short-term memory"))) unit in DRQN with a transformer-based temporal unit and then train the whole model.
4 StarCraft II Experiment
--------------------------
In this section, we evaluate UPDeT and its variants with different policy decoupling methods in the context of challenging micromanagement games in StarCraft II. We compare UPDeT with the RNN-based model on a single scenario and test the transfer capability on multiple-scenario transfer tasks. The experimental results show that UPDeT achieves significant improvement compared to the RNN-based model.
###
4.1 Single Scenario
In the single scenario experiments, we evaluate the model performance on different scenarios from SMAC (Samvelyan et al. ([2019](#bib.bib24 "The starcraft multi-agent challenge"))). Specifically, the scenarios considered are as follows: 3 Marines vs 3 Marines (3m, Easy), 8 Marines vs 8 Marines (8m, Easy), 4 Marines vs 5 Marines (4m\_vs\_5m, Hard+) and 5 Marines vs 6 Marines (5m\_vs\_6m, Hard). In all these games, only the units from player’s side are treated as agents. Dead enemy units will be masked out from the action space to ensure that the executed action is valid.
More detailed settings can be acquired from the SMAC environment (Samvelyan et al. ([2019](#bib.bib24 "The starcraft multi-agent challenge"))).
####
4.1.1 Methods and Training Details
The MARL methods for evaluation include VDN (Sunehag et al. ([2017](#bib.bib20 "Value-decomposition networks for cooperative multi-agent learning"))), QMIX (Rashid et al. ([2018](#bib.bib17 "QMIX: monotonic value function factorisation for deep multi-agent reinforcement learning"))) and QTRAN (Hostallero et al. ([2019](#bib.bib21 "Learning to factorize with transformation for cooperative multi-agent reinforcement learning"))). All three SOTA methods’ original implementation can be found at <https://github.com/oxwhirl/pymarl>. These methods were selected due to their robust performance across different multi-agent tasks. Other methods, including COMA (Foerster et al. ([2017](#bib.bib25 "Counterfactual multi-agent policy gradients"))) and IQL (Tan ([1993](#bib.bib26 "Multi-agent reinforcement learning: independent vs. cooperative agents"))) do not perform stable across in all tasks, as have been proved in several recent works (Rashid et al. ([2018](#bib.bib17 "QMIX: monotonic value function factorisation for deep multi-agent reinforcement learning")), Mahajan et al. ([2019](#bib.bib16 "Maven: multi-agent variational exploration")), Zhou et al. ([2020](#bib.bib18 "Learning implicit credit assignment for multi-agent actor-critic"))). Therefore, we combined UPDeT with VDN, QMIX and QTRAN to prove that our model can improve the model performance significantly compared to the GRU-based model.
| | | | | | |
| --- | --- | --- | --- | --- | --- |
|
(a) Policy variants
|
(b) Temporal variants
|
(c) MARL methods
|
(d) Easy scenarios
|
(e) Hard scenarios
|
(f) Mismatch experiment
|
Figure 4: Experimental results with different task settings. Details can be found in Section [4.1.2](#S4.SS1.SSS2 "4.1.2 Result ‣ 4.1 Single Scenario ‣ 4 StarCraft II Experiment ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers").
####
4.1.2 Result
The model performance result with different policy decoupling methods can be found in Fig. [3(a)](#S4.F3.sf1 "(a) ‣ Figure 4 ‣ 4.1.1 Methods and Training Details ‣ 4.1 Single Scenario ‣ 4 StarCraft II Experiment ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers").
Vanilla Transformer is our baseline for all transformer-based models. This transformer only satisfies point
2. Each output embedding can either be projected to an action or abandoned.
The vanilla transformer fails to beat the enemies in the experiment. Aggregation Transformer is a variant of vanilla transformer, the embedding of which are aggregated into a global embedding and then projected to a policy distribution. This transformer only satisfies the point
1. The performance of the aggregation transformer is worse than that of the GRU-based model. The result proves that it is only with a policy decoupling strategy that the transformer-based model can outperform the conventional RNN-based model.
Next, we adopt UPDeT to find the best temporal unit architecture in Fig. [3(b)](#S4.F3.sf2 "(b) ‣ Figure 4 ‣ 4.1.1 Methods and Training Details ‣ 4.1 Single Scenario ‣ 4 StarCraft II Experiment ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers"). The result shows that without a hidden state, the performance is significantly decreased. The temporal unit with global hidden state is more efficient in terms of convergence speed than the individual hidden state. However, the final performances are almost the same.
To test the generalization of our model, we combine the UPDeT with VDN / QMIX / QTRAN respectively and compare the final performance with RNN-based methods in Fig. [3(c)](#S4.F3.sf3 "(c) ‣ Figure 4 ‣ 4.1.1 Methods and Training Details ‣ 4.1 Single Scenario ‣ 4 StarCraft II Experiment ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers").
We evaluate the model performance on 5m\_vs\_6m (Hard) scenarios. Combined with UPDeT, all three MARL methods obtain significant improvement by large margins compared to the GRU-based model. The result proves that our model can be injected into any existing stat- of-the-art MARL method to yield better performance.
Further more, we combine UPDeT with VDN and evaluate the model performance on different scenarios from Easy to Hard+ in Fig. [3(d)](#S4.F3.sf4 "(d) ‣ Figure 4 ‣ 4.1.1 Methods and Training Details ‣ 4.1 Single Scenario ‣ 4 StarCraft II Experiment ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers") and Fig. [3(e)](#S4.F3.sf5 "(e) ‣ Figure 4 ‣ 4.1.1 Methods and Training Details ‣ 4.1 Single Scenario ‣ 4 StarCraft II Experiment ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers"). The results show that the UPDeT performs stably on easy scenarios and significantly outperforms the GRU-based model on hard scenarios, in the 4m\_vs\_5m(Hard+) scenario, the performance improvement achieved by UPDeT relative to the GRU-based model is of the magnitude of around 80%.
Finally, we conduct an ablation study on UPDeT with paired and unpaired observation-entity—action-group, the result of which are presented in Fig. [3(f)](#S4.F3.sf6 "(f) ‣ Figure 4 ‣ 4.1.1 Methods and Training Details ‣ 4.1 Single Scenario ‣ 4 StarCraft II Experiment ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers"). We disrupt the original correspondence between ’attack’ action and enemy unit. The final performance is heavily decreased compared to the original model, and is even worse than the GRU-based model. We accordingly conclude that only with policy decoupling and a paired observation-entity—action-group strategy can UPDeT learn a strong policy.
###
4.2 Multiple Scenarios
| | |
| --- | --- |
|
(a) Transfer from 7 marines to 3 marines
|
(b) Transfer from 3 marines to 7 marines
|
Figure 5: Experimental results on transfer learning with UPDeT (Uni-Transfer) and GRU unit (GRU-Transfer), along with UPDeT training from scratch (Uni-Scratch). At time step 0 and 500k, we load the model from the source scenario and finetune on the target scenarios. The circular points indicate the model performance on new scenarios without finetuning.
In this section, we discuss the transfer capability of UPDeT compared to the RNN-based model.
We evaluate the model performance in a curriculum style. First, the model is trained one the 3m (3 Marines vs 3 Marines) scenario. We then used the pretrained 3m model to continually train on the 5m (5 Marines vs 5 Marines) and 7m (7 Marines vs 7 Marines) scenarios. We also conduct a experiment in reverse from 7m to 3m. During transfer learning, the model architecture of UPDeT remains fixed. Considering that the RNN-based model cannot handle various input and output dimensions, we modify the architecture of the source RNN model when training on the target scenario. We preserve the parameters of the GRU cell and initialize the fully connected layer with proper input and output dimensions to fit the new scenario.
The final results can be seen in Fig. [4(a)](#S4.F4.sf1 "(a) ‣ Figure 5 ‣ 4.2 Multiple Scenarios ‣ 4 StarCraft II Experiment ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers") and Fig. [4(b)](#S4.F4.sf2 "(b) ‣ Figure 5 ‣ 4.2 Multiple Scenarios ‣ 4 StarCraft II Experiment ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers"). Our proposed UPDeT achieves significantly better results than the GRU-based model. Statistically, UPDeT’s total timestep cost to converge is at least 10 times less than the GRU-based model and 100 times less than training from scratch. Moreover, the model demonstrates a strong generalization ability without finetuning, indicating that UPDeT learns a robust policy with meta-level skill.
###
4.3 Extensive experiment on large-scale MAS
To evaluate the model performance in large-scale scenarios, we test our proposed UPDeT on the 10m\_vs\_11m and 20m\_vs\_21m scenarios from SMAC and a 64\_vs\_64 battle game in the MAgent Environment (Zheng et al. ([2017](#bib.bib40 "Magent: a many-agent reinforcement learning platform for artificial collective intelligence"))). The final results can be found in Appendix [E](#A5 "Appendix E Results of Extensive Experiment on Large Scale ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers").
###
4.4 Attention based Strategy: An analysis
The significant performance improvement achieved by UPDeT on the SMAC multi-agent challenge can be credited to the self-attention mechanism brought by both transformer blocks and the policy decoupling strategy in UPDeT. In this section, we mainly discuss how the attention mechanism assists in learning a much more robust and explainable strategy.
Here, we use the 3 Marines vs 3 Marines game (therefore, the size of the raw attention matrix is 6x6) as an example to demonstrate how the attention mechanism works. As mentioned in the caption of Fig. [6](#A3.F6 "Figure 6 ‣ Appendix C SOTA MARL value-based Framework ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers"), we simplify the raw complete attention matrix to a grouped attention matrix. Fig. [5(b)](#A3.F5.sf2 "(b) ‣ Figure 6 ‣ Appendix C SOTA MARL value-based Framework ‣ UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers") presents the three different stages in one episode including
Game Start, Attack and Survive, with their corresponding attention matrix and strategies.
In the Game Start stage, the highest attention is in line 1 col 3 of the matrix, indicating that the agent pays more attention to its allies than its enemies. This phenomenon can be interpreted as follows: in the startup stage of one game, all the allies are spawned at the left side of the map and are encouraged to find and attack the enemies on the right side
In the Attack stage, the highest attention is in line 2 col 2 of the matrix, which indicates that the enemy is now in the agent’s attack range; therefore, the agent will attack the enemy to get more rewards. Surprisingly, the agent chooses to attack the enemy with the lowest health value. This indicates that a long term plan can be learned based on the attention mechanism, since killing the weakest enemy first can decrease the punishment from the future enemy attacks.
In the Survive stage, the agent’s health value is low, meaning that it needs to avoid being attacked. The highest attention is located in line 1 col 1, which clearly shows that the most important thing under the current circumstances is to stay alive. For as long as the agent is alive, there is still a chance for it to return to the front line and get more reward while enemies are attacking the allies instead of the agent itself.
In conclusion, the self-attention mechanism and policy decoupling strategy of UPDeT provides a strong and clear relation between attention weights and final strategies. This relation can help us better understand the policy generation based on the distribution of attention among different entities. An interesting idea presents itself here: namely, if we can find a strong mapping between attention matrix and final policy, the character of the agent could be modified in an unsupervised manner.
5 Conclusion
-------------
In this paper, we propose UPDeT, a universal policy decoupling transformer model that extends MARL to a much broader scenario. UPDeT is general enough to be plugged into any existing MARL method. Moreover, our experimental results show that, when combined with UPDeT, existing state-of-the-art MARL methods can achieve further significant improvements with the same training pipeline.
On transfer learning tasks, our model is 100 times faster than training from scratch and 10 times faster than training using the RNN-based model.
In the future, we aim to develop a centralized function based on UPDeT and apply the self-attention mechanism to the entire pipeline of MARL framework to yield further improvement.
#### Acknowledgments
This work was supported in part by the National Natural Science Foundation of China (NSFC) under Grant
No.U19A2073 and in part by the National Natural Science
Foundation of China (NSFC) under Grant No.61976233 and No.61906109 and Australian Research Council Discovery Early Career Researcher Award (DE190100626), and Funding of
“Leading Innovation Team of the Zhejiang Province” (2018R01017). |
269013f5-be18-4fbe-b19c-1e66f52b4d25 | trentmkelly/LessWrong-43k | LessWrong | Doom sooner
It took me a long time to understand because I think software as it currently exists is fantastically weak and infuriatingly defective, in fact criminally so - but after reading a lot of stuff here, AI research suddenly seems a lot more dangerous than before.
The basic truth of any complex system is that it's always more complex than it seems. Most projects take longer than planned. However, if you repeat a process enough it can be done faster.
Multicellular life took eons to evolve. Animals took millions of centuries to develop intelligence. Primitive humans were stuck in the stone age for thousands of centuries.
Right now, society is about as dumb and inefficient as it can get away with. The most powerful force in the world is the implacable consensus everywhere. There are too few geniuses to overcome the sometimes monstrously deliberate inefficiencies of life.
For those reasons, it seems probable that developing something as complex as Artificial Superintelligence will take several decades at least, and only with a great deal of effort. By which I mean that completely unexpected delays will arise that will keep slowing things down.
Yet it's the only thing that might possibly save us, the closest thing to a magic genie.
The posts on this site make a powerful case that when the first AI does develop superintelligence, it will likely not be "well rounded", but hyper-focused on some inadequately defined goal. Having less general intelligence will not make it less dangerous. The threat range may be "smaller" but no less deadly.
What is the simplest way a brute-force AI could run amuck? All it would take is one super clever idea, like the easiest self-replicating nanobot, DNA rewriting meta-viruses, or even social memes to manipulate personalities. We vastly underestimate how badly things could go wrong. Just dropping a test tube with bat guano can crash the world economy for three years.
Open-ended software entities running on sufficiently powerful hardware ar |
eae5ae49-8378-420f-b0a3-1e493198bf29 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Cambridge Less Wrong Meetup - Book Recommendations
Discussion article for the meetup : Cambridge Less Wrong Meetup - Book Recommendations
WHEN: 04 October 2015 03:30:00PM (-0400)
WHERE: 98 Elm Street, Apartment 1, Somerville, MA
Come to give, listen to and discuss book recommendations. People will be giving 1-minute pitches for books they think may be of interest to the rationalist community. All comers are encouraged to offer their recommendations.
Phase 1: Arrival, greetings, unstructured conversation.
Phase 2: Presentations. This starts promptly at 4:00, and lasts 30-60 minutes (until we run out of book recommendations).
Phase 3: Further discussion. We'll explore the ideas raised in phase 2, often in smaller groups.
Phase 4: Dinner. It's about a five minute walk to the usual restaurant.
Discussion article for the meetup : Cambridge Less Wrong Meetup - Book Recommendations |
f2e1e159-7c44-4334-84fe-df25c0dac25d | trentmkelly/LessWrong-43k | LessWrong | Welcome to Less Wrong! (8th thread, July 2015)
If you've recently joined the Less Wrong community, please leave a comment here and introduce yourself. We'd love to know who you are, what you're doing, what you value, how you came to identify as an aspiring rationalist or how you found us. You can skip right to that if you like; the rest of this post consists of a few things you might find helpful. More can be found at the FAQ.
A FEW NOTES ABOUT THE SITE MECHANICS
To post your first comment, you must have carried out the e-mail confirmation: When you signed up to create your account, an e-mail was sent to the address you provided with a link that you need to follow to confirm your e-mail address. You must do this before you can post!
Less Wrong comments are threaded for easy following of multiple conversations. To respond to any comment, click the "Reply" link at the bottom of that comment's box. Within the comment box, links and formatting are achieved via Markdown syntax (you can click the "Help" link below the text box to bring up a primer).
You may have noticed that all the posts and comments on this site have buttons to vote them up or down, and all the users have "karma" scores which come from the sum of all their comments and posts. This immediate easy feedback mechanism helps keep arguments from turning into flamewars and helps make the best posts more visible; it's part of what makes discussions on Less Wrong look different from those anywhere else on the Internet.
However, it can feel really irritating to get downvoted, especially if one doesn't know why. It happens to all of us sometimes, and it's perfectly acceptable to ask for an explanation. (Sometimes it's the unwritten LW etiquette; we have different norms than other forums.) Take note when you're downvoted a lot on one topic, as it often means that several members of the community think you're missing an important point or making a mistake in reasoning— not just that they disagree with you! If you have any questions about karma or voti |
9e3259bc-a989-445a-8128-b0b23e6a785a | trentmkelly/LessWrong-43k | LessWrong | Playing Video Games In Shuffle Mode
One of the missions of OB/LW is to attract new learners, and it's clear that they are succeeding. But the format feels like a very difficult one for those new to these ideas, with beginner-level ideas interspersed with advanced or unsettled theory and meta-level discussions. You wouldn't play <insert cool-sounding, anime-ish video game here> with the levels on shuffle mode, but reading Less Wrong must feel like doing so for initiates.
How do we make the site better for learners? Provide a "syllabus" that shows a series of OB and LW posts which should be read in order? Have a separate beginner site or feed or header? Put labels on posts that designate them with a level? |
a7b5440f-bb86-49d7-85fe-2533a830a92d | trentmkelly/LessWrong-43k | LessWrong | What rationality failure modes are there?
How do people fail to improve their rationality? How do they accidentally harm themselves in the process? I'm thinking of writing a post "How not to improve your rationality" or "A nuanced guide to reading the sequences" that preempts common mistakes, and I'd appreciate hearing people's experiences. Some examples:
* It took me an absurdly long time (like, 1-2yr in the rat community) before I realized you don't correct for cognitive biases, you have to "be introspectively aware of the bias occuring, and remain unmoved by it" (as Eliezer put it in a podcast)
* More generally, people can read about a bias and resolve to "do better" without concretely deciding what to do differently. This typically makes things worse, e.g. I have a friend who tried really hard to avoid the typical mind fallacy, and accidentally turned off her empathy in the process.
* The implicit frame rationalists push is logical and legible, and can lead to people distrusting their emotions. And I think it's really important to listen to listen to ick feelings when changing your thought processes, as there can be non obvious effects.
* E.g. My friend started thinking about integrity in terms of FDT, and this disconnected it from their motivational circuits and they made some pretty big mistakes because of it. If they'd listened to their feeling of "this is a weird way to think" this wouldn't have happened.
* (I think many people misinterpret sequence posts and decide to change their thinking in bad ways, and listening to your feelings can be a nice emergency check.) |
254a6397-65e0-4c4d-b6c5-1a3f87488ad1 | trentmkelly/LessWrong-43k | LessWrong | Finance Followups
Cross-posted from Putanumonit.com.
----------------------------------------
In Defense of Finance generated over 110 comments across WordPress, LessWrong and the Reddits, as well as in personal communication. As predicted, I learned a lot.
In this post I’ll address some of these comments and offer follow up thoughts that didn’t fit in the original essay because they’re more speculative. Also, becase that post ran to 5,500 words already.
People said: “I don’t buy this, fuck capitalism!” If you want to stick it to capitalism, you can gift a donation to a creator publishing his work online for free.
People said: “Finally someone gets it and does the math right!” If you think I get it and I do the math right, you can forward me job offers at lucrative hedge funds, PE firms, and startups.
People said that it’s overly defeatist to claim that finance won’t improve no matter how we try to fix it, we just have to live with it. But that’s not quite what I think.
Finance improves itself, mostly after crises. In 1929 we learned that established company stocks are riskier than we thought, and in 2000 we learned the same about startups. Financial institutions learned how to deal with 20% interest rates in 1980 and with 0.02% rates in 2010. After 2008 we learned that 6% is too low of an equity ratio for banks, that rating agencies with no skin in the game are useless, and that CDO-Squared are really dumb. Hopefully, we also learned some generalizable math lessons, such as the fact that lack of correlation in normal times (e.g., in mortgage default rates) doesn’t imply lack of correlation when a crisis hits.
We’re on the cusp of the longest period ever between recessions in the US. In the decade since 2008 the Middle East has gone crazy, Europe has gone crazy, Venezuela has gone crazy, elections have gone crazy, the weather has gone crazy, the Cubs have won the World Series, Leicester won the EPL, Bitcoin rose and fell and rose and fell and rose. Throughout all of this, the |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.