text stringlengths 300 320k | source stringlengths 52 154 |
|---|---|
# How bad a future do ML researchers expect?
_Katja Grace, 8 March 2023_
In our [survey last year](https://wiki.aiimpacts.org/doku.php?id=ai_timelines:predictions_of_human-level_ai_timelines:ai_timeline_surveys:2022_expert_survey_on_progress_in_ai), we asked publishing machine learning researchers how they would divi... | https://www.lesswrong.com/posts/Zq2HaihaDy7sSarMz/how-bad-a-future-do-ml-researchers-expect |
# Speed running everyone through the bad alignment bingo. $5k bounty for a LW conversational agent
There's a wave of people, of various degrees of knowledge and influence, currently waking up to the ideas of AI existential risk. They seem to be literally going through every box of the bad alignement bingo card takes. ... | https://www.lesswrong.com/posts/SLRLuiuDykfTdmesK/speed-running-everyone-through-the-bad-alignment-bingo-usd5k |
# Paper Summary: The Effectiveness of AI Existential Risk Communication to the American and Dutch Public
*This is a summary of the following paper by Alexia Georgiadis (Existential Risk Observatory): *[https://existentialriskobservatory.org/papers\_and\_reports/The\_Effectiveness\_of\_AI\_Existential\_Risk\_Communicat... | https://www.lesswrong.com/posts/werC3aynFD92PEAh9/paper-summary-the-effectiveness-of-ai-existential-risk |
# AI #3
Welcome to week three of the AI era. Another long week and another giant post. I _intend_ to take a nice four-day vacation in Mexico starting today, during which I won’t do any writing. I’m sure the pace of things will slow down real soon now. I mean, they have to. Don’t they?
If not, I’m going to have to lea... | https://www.lesswrong.com/posts/AgaBzvuBJg2evEjqh/ai-3 |
# Just Imitate Humans?
Do people think we could make a singleton (or achieve [global coordination and preventative policing](https://nickbostrom.com/papers/vulnerable.pdf)) just by imitating human policies on computers? If so, this seems pretty safe to me.
Some reasons for optimism: 1) these could be run much faster ... | https://www.lesswrong.com/posts/LTFaD96D9kWuTibWr/just-imitate-humans |
# Impact Measure Testing with Honey Pots and Myopia
Suppose we have an impact measure that we think might work. That is, it might tame a misaligned agent. There isn't an obvious way to test whether it works: if we just try it out, then if it's ineffective, that's an existential loss. This is a proposal for how to test... | https://www.lesswrong.com/posts/2R8XqCt6f6Ss2TzwM/impact-measure-testing-with-honey-pots-and-myopia |
# Utility uncertainty vs. expected information gain
It is a relatively intuitive thought that if a Bayesian agent is uncertain about its utility function, it will act more conservatively until it has a better handle on what its true utility function is.
This might be deeply flawed in a way that I'm not aware of, but ... | https://www.lesswrong.com/posts/Pkr97mB9Y4rkx5DdZ/utility-uncertainty-vs-expected-information-gain |
# The Translucent Thoughts Hypotheses and Their Implications
*Epistemic status: Uncertain about the validity of the claims I’m making here, and looking for feedback about the research directions I’m suggesting.*
*Thanks to Marius Hobbhahn, Johannes Treutlein, Siméon Campos, and Jean-Stanislas Denain for helpful feedb... | https://www.lesswrong.com/posts/r3xwHzMmMf25peeHE/the-translucent-thoughts-hypotheses-and-their-implications |
# Some ML-Related Math I Now Understand Better
Here are some simple Math facts rarely taught in ML & Math lectures:
* SVD is decomposing a matrix into a sum of simple read-and-write operations
* There is exponentially much room for close vectors in high dimensional space
* Layer Normalization is a projection
S... | https://www.lesswrong.com/posts/Tu8DYNCg63F4HYbXE/some-ml-related-math-i-now-understand-better |
# Anthropic's Core Views on AI Safety
> We founded Anthropic because we believe the impact of AI might be comparable to that of the industrial and scientific revolutions, but we aren’t confident it will go well. And we also believe this level of impact could start to arrive soon – perhaps in the coming decade.
>
> Th... | https://www.lesswrong.com/posts/xhKr5KtvdJRssMeJ3/anthropic-s-core-views-on-ai-safety |
# Why do we assume there is a "real" shoggoth behind the LLM? Why not masks all the way down?
In recent discourse, Large Language Models (LLMs) are often depicted as presenting a human face over a vast alien intelligence (the [shoggoth](https://en.wikipedia.org/wiki/Shoggoth)), as in this popular image or this Eliezer... | https://www.lesswrong.com/posts/bYzkipnDqzMgBaLr8/why-do-we-assume-there-is-a-real-shoggoth-behind-the-llm-why |
# Anthropic: Core Views on AI Safety: When, Why, What, and How
We founded Anthropic because we believe the impact of AI might be comparable to that of the industrial and scientific revolutions, but we aren’t confident it will go well. And we also believe this level of impact could start to arrive soon – perhaps in the... | https://www.lesswrong.com/posts/MLDXcEBJ2jX7BcSmN/anthropic-core-views-on-ai-safety-when-why-what-and-how |
# What Jason has been reading, March 2023
A new monthly feature, let me know what you think.
**Books**
---------
**Matt Ridley, **[***How Innovation Works***](https://www.amazon.com/How-Innovation-Works-Flourishes-Freedom-ebook/dp/B07WSBV7YZ) **(2020).** About halfway through, lots of interesting case studies, very ... | https://www.lesswrong.com/posts/6dJiPs4qCP4TfJrBw/what-jason-has-been-reading-march-2023 |
# Questions about Conjecure's CoEm proposal
Conjecture recently released an [AI safety proposal](https://www.lesswrong.com/posts/ngEvKav9w57XrGQnb/cognitive-emulation-a-naive-ai-safety-proposal). The three of us spent a few hours discussing the proposal and identifying questions that we have. (First, we each re-read t... | https://www.lesswrong.com/posts/LanufchfpiTiDe2NF/questions-about-conjecure-s-coem-proposal |
# What's Not Our Problem
Cross-posted, as always, [from Putanumonit](https://putanumonit.com/2023/03/06/whats-not-our-problem/).
Putanumonit started in 2015, and like every rationalist blogger at the time my main inspirations were [Scot... | https://www.lesswrong.com/posts/QJr9K4Ssch3v48a2Y/what-s-not-our-problem |
# Why Not Just Outsource Alignment Research To An AI?
Warmup: The Expert
------------------
If you haven’t seen “The Expert” before, I recommend it as a warmup for this post:
> The Client: “We need you to draw seven red lines, all strictly perpendicular. Some with green ink, some with transparent. Can you do that?”
... | https://www.lesswrong.com/posts/3gAccKDW6nRKFumpP/why-not-just-outsource-alignment-research-to-an-ai |
# Japan AI Alignment Conference
[Conjecture](https://www.conjecture.dev/) and [ARAYA](https://www.araya.org/en/) are hosting and organizing the first Japan AI Alignment Conference. The conference will take place in Tokyo, Japan on March 11 and 12. Details about the event can be found [here](https://jac2023.ai/). This ... | https://www.lesswrong.com/posts/tAQRxccEDYZY5vxvy/japan-ai-alignment-conference |
# Reflections On The Feasibility Of Scalable-Oversight
*This post was inspired by discussions on the feasibility of scaling RLHF for aligning AGI during EAG 2023 in Oakland. The first draft was finished during the Apart Research Thinkaton the following week. Thanks to Lee Sharkey, Daniel Ziegler, Tom Henighan, Stephen... | https://www.lesswrong.com/posts/8yimdZcEWSKkutHhZ/reflections-on-the-feasibility-of-scalable-oversight |
# Thoughts on the OpenAI alignment plan: will AI research assistants be net-positive for AI existential risk?
*Note: I really appreciate the work that the OpenAI alignment team put into their alignment plan writeup and related posts, especially Jan Leike, the leader of that team. I believe open discussions about align... | https://www.lesswrong.com/posts/6RC3BNopCtzKaTeR6/thoughts-on-the-openai-alignment-plan-will-ai-research |
# Long-term memory for LLM via self-replicating prompt
*Infohazard assessment: the idea is so obvious that it is either well-known or will be found soon, and it is better to known about it advance and prepare *
### *LLM needs long-term memory to be effective agent*
In a science fiction story, there was a character w... | https://www.lesswrong.com/posts/vyE5eonRR9EW7B2ha/long-term-memory-for-llm-via-self-replicating-prompt |
# Dice Decision Making
*Epistemic status: I have been using dice on and off in my decision-making processes for a few years and am confident it has been a net positive for me (85%). I believe that experimenting with dice decisions would be beneficial for most people who struggle with decision-making (60%). I have a ve... | https://www.lesswrong.com/posts/PhHFoJwaFLfMfQsHF/dice-decision-making |
# Meetup Tip: Nametags
Summary
-------
Nametags are nice to have at meetups. When I was planning my second big meetup, I typed “nametag” into Amazon and got [these](https://www.amazon.com/dp/B098FGKMLS). They’ve worked great but whatever your local stationary store or online shopping service suggests is probably just... | https://www.lesswrong.com/posts/DdGCgpdmtKDaL5jT2/meetup-tip-nametags |
# Rationalism and social rationalism
*As with all my blog posts, these are only very preliminary thoughts. I worry perhaps their scope gives them a sense of self-seriousness I don’t intend. This piece is about floating ideas that might be worth discussing, rather than strong claims to truth.*
Cate Hall tweeted:
“In ... | https://www.lesswrong.com/posts/Q8CkyrvibYkvwajoB/rationalism-and-social-rationalism |
# Will the Bitcoin fee market actually work?
Based on other posts I've seen on here, many of you are Bitcoiners. While I clearly see the value proposition, I have a few questions about the protocol that haven't been answered, and I figure the lesswrong audience is my best change.
**1.) Will the fee market actually wo... | https://www.lesswrong.com/posts/AeATMCFaotHfGiyaK/will-the-bitcoin-fee-market-actually-work |
# The Power of Intelligence - The Animation
This video is an animation of [The Power of Intelligence](https://www.lesswrong.com/posts/aiQabnugDhcrFtr9n/the-power-of-intelligence), by [@Eliezer Yudkowsky](https://www.lesswrong.com/users/eliezer_yudkowsky?mention=user)
The [Sorting Pebbles Into Correct Heaps](https://... | https://www.lesswrong.com/posts/9JhhZNeykXvv3WsaR/the-power-of-intelligence-the-animation |
# Counterarguments to Core AI X-Risk Stories?
I've added a [tag](https://www.lesswrong.com/tag/object-level-ai-risk-skepticim) for object-level AI risk skepticism arguments. I've included [my own post about deceptive alignment](https://www.lesswrong.com/posts/RTkatYxJWvXR4Qbyd/deceptive-alignment-is-less-than-1-likely... | https://www.lesswrong.com/posts/K5TFTLzobmxfQvvfp/counterarguments-to-core-ai-x-risk-stories |
# Against Complete Blackout Curtains For Sleep
*Epistemic status: moderately confident that something like this is directionally true, but good data or arguments could change my mind. Note the lack of citations — this is based on memory.*
Most rationalists I meet believe the following:
1. Sleep is very important
2.... | https://www.lesswrong.com/posts/rQuuvLvudiKqmGRQ7/against-complete-blackout-curtains-for-sleep |
# Agents synchronization
There is a line in alignment-related thinking, of looking for ways that agents will tend to be similar. An early example is convergent instrumental goals, and a later is natural abstractions. Those two ideas share an important attribute - trying to think on something "mental" (values, abstract... | https://www.lesswrong.com/posts/sAMGhYoGLLNNHN2r7/agents-synchronization |
# How to Support Someone Who is Struggling
*\[Crossposted from* [*my blog*](https://bordersofpsychology.substack.com/)*\]*
There’s no shortage of pain, tragedy and loss in the world. [And if you’re anything like me](https://bordersofpsychology.substack.com/p/dedication), you don’t always know how to be helpful when a... | https://www.lesswrong.com/posts/ZdWfD9dhAvNgFs6Dh/how-to-support-someone-who-is-struggling-1 |
# Understanding and controlling a maze-solving policy network
Previously: [Predictions for shard theory mechanistic interpretability results](https://www.lesswrong.com/posts/JusJcepE2qohiC3hm/predictions-for-shard-theory-mechanistic-interpretability)
 hypotheses about *computations* (i.e. computable[^1] logical/analytic facts). This is achieved via a novel (AFAIK) branch of automata theory, which generalizes both ordinary and t... | https://www.lesswrong.com/posts/jxE47v3aezEYaydg9/compositional-language-for-hypotheses-about-computations |
# [Linkpost] Scott Alexander reacts to OpenAI's latest post
Scott Alexander recently wrote a post about OpenAI's [Planning for AGI and beyond](https://openai.com/blog/planning-for-agi-and-beyond). I found it thoughtful, and I think others here might want to read or discuss it.
Some highlights:
**ExxonMobil analogy*... | https://www.lesswrong.com/posts/nALdMXkxkLzysKtzC/linkpost-scott-alexander-reacts-to-openai-s-latest-post |
# Parasitic Language Games: maintaining ambiguity to hide conflict while burning the commons
> “They are playing a game. They are playing at not playing a game. If I show them I see they are, I shall break the rules and they will punish me. I must play their game, of not seeing I see the game”
> \- R. D. Laing
"**I... | https://www.lesswrong.com/posts/3JzndpGm4ZgQ4GT3S/parasitic-language-games-maintaining-ambiguity-to-hide |
# "You'll Never Persuade People Like That"
Sometimes, when someone is arguing for some proposition, their interlocutor will reply that the speaker's choice of arguments or tone wouldn't be effective at persuading some third party.
This would seem to be an odd change of topic. If I was arguing for this-and-such propos... | https://www.lesswrong.com/posts/67NrgoFKCWmnG3afd/you-ll-never-persuade-people-like-that |
# "Liquidity" vs "solvency" in bank runs (and some notes on Silicon Valley Bank)
_Epistemic status: Reference post, then some evidenced speculation about emerging current events (as of 2023-03-12 morning)._
## A "liquidity" crisis
There's one kind of "bank run" where the story, in stylized terms, starts like this:
... | https://www.lesswrong.com/posts/HqzpJGXRjdrbimwRn/liquidity-vs-solvency-in-bank-runs-and-some-notes-on-silicon |
# What problems do African-Americans face? An initial investigation using Standpoint Epistemology and Surveys
*This post is also available at* [*my substack*](https://tailcalled.substack.com/p/what-problems-do-african-americans)*.*
This post started from a bit of a weird place. I was in a Discord chatroom, and someon... | https://www.lesswrong.com/posts/SELnnesv6Fz2WqiMs/what-problems-do-african-americans-face-an-initial |
# Paper Replication Walkthrough: Reverse-Engineering Modular Addition
I'm excited about trying different formats for mechanistic interpretability education! I've [made a video walkthrough](https://neelnanda.io/modular-addition-walkthrough) where we replicate my paper, [Progress Measures for Grokking via Mechanistic In... | https://www.lesswrong.com/posts/6Ghvdb2iwLAyGT6A3/paper-replication-walkthrough-reverse-engineering-modular |
# Creating a Discord server for Mechanistic Interpretability Projects
*TL;DR: I think there likely are a lot of people who want to work on mechanistic interpretability projects but couldn’t get into AI safety camp, so I created a *[*Discord server*](https://discord.gg/cMr5YqbU4y) *for people to organize themselves int... | https://www.lesswrong.com/posts/gGysHedD8aHKueb35/creating-a-discord-server-for-mechanistic-interpretability |
# Are there cognitive realms?
*[Metadata: crossposted from [https://tsvibt.blogspot.com/2022/11/are-there-cognitive-realms.html](https://tsvibt.blogspot.com/2022/11/are-there-cognitive-realms.html). First completed November 16, 2022. This essay is more like research notes than exposition, so context may be missing, th... | https://www.lesswrong.com/posts/NvwjExA7FcPDoo3L7/are-there-cognitive-realms |
# Musicians and Mouths
Most musicians massively underuse their mouths, and the ones that use their mouths underuse their fingers. This is very weird and I don't understand it.
The most common instruments I see are probably piano, guitar, drums, and violin. Add in mandolin, bass, accordion, cello, viola, and banjo, an... | https://www.lesswrong.com/posts/3gW4MxKdzXfXN737f/musicians-and-mouths |
# An AI risk argument that resonates with NYTimes readers
Ezra Klein of NYT put out [a surprisingly sympathetic post on AI risk](https://www.nytimes.com/2023/03/12/opinion/chatbots-artificial-intelligence-future-weirdness.html) in the Sunday edition. It even [quotes Paul Christiano](https://www.nytimes.com/2023/03/12/... | https://www.lesswrong.com/posts/FzhedhEFAcKJZkgJS/an-ai-risk-argument-that-resonates-with-nytimes-readers |
# Yudkowsky on AGI risk on the Bankless podcast
Eliezer gave a very frank overview of his take on AI two weeks ago on the cryptocurrency show Bankless:
I've posted a transcript of the show and a [**follow-up Q&A**](https://twitter.com/i/spaces/1PlJQpZogzVGE) below.
Andrea_Miotti, remember, and vonk previously poste... | https://www.lesswrong.com/posts/e4pYaNt89mottpkWZ/yudkowsky-on-agi-risk-on-the-bankless-podcast |
# Gold, Silver, Red: A color scheme for understanding people
A few months ago, I built a predictive algorithm to determine the probability of romantic relationships between individuals in my periphery.
The algorithm works like this:
-Personal Factors. How likely are you as an individual to find yourself in a new rel... | https://www.lesswrong.com/posts/ifGzvYvaWj7dHz4w2/gold-silver-red-a-color-scheme-for-understanding-people |
# Nose / throat treatments for respiratory infections
After hearing various people interested in nasal sprays for preventing covid or the common cold, I did a shallow dive on what might be worth trying in this area.
Overall, there are several treatments that seem safe and cheap and some of them are probably effective... | https://www.lesswrong.com/posts/cA24NRxMDDdWxPJuR/nose-throat-treatments-for-respiratory-infections |
# What is calibration?
A forecaster is well-calibrated if, for every $p\in[0,1]$, of the propositions that they assign probability approximately $p$ to, the fraction of them that are true is approximately $p$. However, there is no natural probability distribution over propositions, so this notion is not well-defined.
... | https://www.lesswrong.com/posts/goppxmpbFoTfdaTsE/what-is-calibration |
# Plan for mediocre alignment of brain-like [model-based RL] AGI
*(This post is a more simple, self-contained, and pedagogical version of *[*Post #14*](https://www.lesswrong.com/posts/QpHewJvZJFaQYuLwH/intro-to-brain-like-agi-safety-14-controlled-agi) *of*[*Intro to Brain-Like AGI Safety*](https://www.lesswrong.com/s/... | https://www.lesswrong.com/posts/Hi7zurzkCog336EC2/plan-for-mediocre-alignment-of-brain-like-model-based-rl-agi |
# Decentralized Exclusion
I'm part of several communities that are relatively decentralized. For example, anyone can host a contra dance, rationality meetup, or effective altruism dinner. Some have central organizations (contra has [CDSS](https://cdss.org/), EA has [CEA](https://www.centreforeffectivealtruism.org/)) b... | https://www.lesswrong.com/posts/Rw3yfA8xDR3F9qcpB/decentralized-exclusion |
# What Discovering Latent Knowledge Did and Did Not Find
*Thanks to Marius Hobbhahn and Oam Patel for helpful feedback on drafts. Thanks to Collin and Haotian for answering many questions about their work.*
[Discovering Latent Knowledge in Language Models Without Supervision](https://arxiv.org/abs/2212.03827) describ... | https://www.lesswrong.com/posts/bWxNPMy5MhPnQTzKz/what-discovering-latent-knowledge-did-and-did-not-find-4 |
# Discussion with Nate Soares on a key alignment difficulty
In late 2022, Nate Soares gave some feedback on my [Cold Takes series on AI risk](https://www.cold-takes.com/tag/implicationsofmostimportantcentury/) (shared as drafts at that point), stating that I hadn't discussed what he sees as one of the key difficulties... | https://www.lesswrong.com/posts/iy2o4nQj9DnQD7Yhj/discussion-with-nate-soares-on-a-key-alignment-difficulty |
# Alpaca: A Strong Open-Source Instruction-Following Model
> We performed a blind pairwise comparison between text-davinci-003 and Alpaca 7B, and we found that these two models have very similar performance: Alpaca wins 90 versus 89 comparisons against text-davinci-003.
Interestingly, Alpaca is trained using supervis... | https://www.lesswrong.com/posts/rMyt6nhnEWgx3pJ4m/alpaca-a-strong-open-source-instruction-following-model |
# To determine alignment difficulty, we need to know the absolute difficulty of alignment generalization
A core insight from Eliezer is that AI "capabilities generalize further than alignment once capabilities start to generalize far".
This seems quite true to me, but doesn't on its own make me conclude that alignmen... | https://www.lesswrong.com/posts/AETwPtyZMaBHqAeQA/to-determine-alignment-difficulty-we-need-to-know-the |
# Fixed points in mortal population games
*I wrote this post during the first part of the SERI MATS winter school, in Vanessa Kosoy's research stream. This is just an explanation of my intended research direction and the motivation behind it, and write down proofs for the statements in* [*Vanessa's comment*](https://w... | https://www.lesswrong.com/posts/HNnRCPe2CejfupSow/fixed-points-in-mortal-population-games |
# Storytelling Makes GPT-3.5 Deontologist: Unexpected Effects of Context on LLM Behavior
TL;DR
-----
When prompted to make decisions, do large language models (LLMs) show power-seeking behavior, self-preservation instincts, and long-term goals? [Discovering Language Model Behaviors with Model-Written Evaluations](htt... | https://www.lesswrong.com/posts/4qC7FfCHrdaFesPzC/storytelling-makes-gpt-3-5-deontologist-unexpected-effects |
# GPT-4
> We’ve created GPT-4, the latest milestone in OpenAI’s effort in scaling up deep learning. GPT-4 is a large multimodal model (accepting image and text inputs, emitting text outputs) that, while worse than humans in many real-world scenarios, exhibits human-level performance on various professional and academi... | https://www.lesswrong.com/posts/pckLdSgYWJ38NBFf8/gpt-4 |
# What is a definition, how can it be extrapolated?
What is a definition? Philosophy has, ironically, a large number of [definitions of definitions](https://plato.stanford.edu/entries/definitions/), but three of them are especially relevant to ML and AI safety.
There is the [intensional definition](https://en.wikiped... | https://www.lesswrong.com/posts/AdyqGnvhdqDMYJaug/what-is-a-definition-how-can-it-be-extrapolated |
# PaLM API & MakerSuite
This seems overshadowed by the [GPT-4 news](https://www.lesswrong.com/posts/pckLdSgYWJ38NBFf8/gpt-4), but apparently Google is also releasing PaLM and an API along with "MakerSuite, a tool that lets developers start prototyping quickly and easily."
Not many details yet (perhaps they rushed thi... | https://www.lesswrong.com/posts/fRzkDWewwyS6fHp4w/palm-api-and-makersuite |
# A better analogy and example for teaching AI takeover: the ML Inferno
_Epistemic Status: entirely speculative_
_Note: this article was assisted by ChatGPT_
A very common example given for AI takeover is [silently killing all humans with nanotech](https://www.lesswrong.com/posts/uMQ3cqWDPHhjtiesc/agi-ruin-a-list-of... | https://www.lesswrong.com/posts/4NB5dqbjnW5imqfwq/a-better-analogy-and-example-for-teaching-ai-takeover-the-ml |
# GPT can write Quines now (GPT-4)
GPT-4 is not only able to write code, more reliably than GPT-3.5, it writes code *that writes code*; see the example below (GPT-3.5 was not able to do this). But first, let me say:
1) @OpenAI: Thank for your openness to the world about your capabilities and shortcomings! Specifica... | https://www.lesswrong.com/posts/ux93sLHcqmBfsRTvg/gpt-can-write-quines-now-gpt-4 |
# Success without dignity: a nearcasting story of avoiding catastrophe by luck
I’ve been trying to form a [nearcast-based](https://www.lesswrong.com/posts/Qo2EkG3dEMv8GnX8d/ai-strategy-nearcasting) picture of what it might look like to suffer or avoid an AI catastrophe. I’ve written a hypothetical “failure story” ([Ho... | https://www.lesswrong.com/posts/jwhcXmigv2LTrbBiB/success-without-dignity-a-nearcasting-story-of-avoiding |
# Human preferences as RL critic values - implications for alignment
TLDR: Human preferences might be largely the result of a critic network head, much like that used in SOTA agentic RL systems. The term "values" in humans might mean almost exactly what it does in RL: an estimate of discounted sum of future rewards.... | https://www.lesswrong.com/posts/HEonwwQLhMB9fqABh/human-preferences-as-rl-critic-values-implications-for |
# Shutting Down the Lightcone Offices
Lightcone recently decided to close down a big project we'd been running for the last 1.5 years: An office space in Berkeley for people working on x-risk/EA/rationalist things that we opened August 2021.
We haven't written much about why, but I and Ben had written some messages o... | https://www.lesswrong.com/posts/psYNRb3JCncQBjd4v/shutting-down-the-lightcone-offices |
# ARC tests to see if GPT-4 can escape human control; GPT-4 failed to do so
An unorthodox approach to alignment research to be sure. The advice to "fail early" might actually a good thing when it comes to AI x-risk though.
After this post was written, ARC themselves wrote a much more detailed (and perhaps more accura... | https://www.lesswrong.com/posts/NQ85WRcLkjnTudzdg/arc-tests-to-see-if-gpt-4-can-escape-human-control-gpt-4 |
# Bushels Are Commodity-Specific
I like obscure units a lot, but mostly they're historical curiosities. It doesn't matter that an [ounce](https://en.wikipedia.org/wiki/Ounce#International_troy_ounce) of gold is heavier than an [ounce](https://en.wikipedia.org/wiki/Ounce#International_avoirdupois_ounce) of feathers whi... | https://www.lesswrong.com/posts/bgZ5y9GtHAeEeXthg/bushels-are-commodity-specific |
# Just Pivot to AI: The secret is out
Despite the 'proof' being public the [scaling hypothesis](https://gwern.net/scaling-hypothesis) remained a [Thielian secret](https://blakemasters.tumblr.com/post/22866240816/peter-thiels-cs183-startup-class-11-notes) until recently. It took chat-GPT to convince Facebook/Meta AI re... | https://www.lesswrong.com/posts/Au7WqG7prRLknNrj2/just-pivot-to-ai-the-secret-is-out |
# POC || GTFO culture as partial antidote to alignment wordcelism
[There is an important asymmetry in reception for prophets.](https://astralcodexten.substack.com/p/from-nostradamus-to-fukuyama) Go read that post first if you haven't.
For those who don't want to, the gist is: Given the same level of specificity, peop... | https://www.lesswrong.com/posts/xsB3dDg5ubqnT7nsn/poc-or-or-gtfo-culture-as-partial-antidote-to-alignment |
# The epistemic virtue of scope matching
I keep noticing a particular epistemic pitfall (not exactly a “fallacy”), and a corresponding epistemic virtue that avoids it. I want to call this out and give it a name, or find out what its name is, if it already has one.
The virtue is: identifying the correct scope for a ph... | https://www.lesswrong.com/posts/X53AFYuntCWueqSsu/the-epistemic-virtue-of-scope-matching |
# Nokens: A potential method of investigating glitch tokens
### Summary
* We can probably replicate glitch token-like behaviour by simply sampling random points in the embedding space, which I call 'nokens'.
* We may be able to find a large number of tokens which cause unwanted behaviour more easily than with the... | https://www.lesswrong.com/posts/sDn5MnBGdvKMenHRr/nokens-a-potential-method-of-investigating-glitch-tokens |
# ChatGPT (and now GPT4) is very easily distracted from its rules
Summary
=======
Asking GPT4 or ChatGPT to do a "side task" along with a rule-breaking task makes them much more likely to produce rule-breaking outputs. For example on GPT4:
 know about it
OpenAI released a [press release](https://openai.com/research/gpt-4), [research statement](https://cdn.openai.com/papers/gpt-4.pdf), and [system card](https://cdn.openai.com/papers/gpt-4-system-card.pdf) about GPT-4 approximately one eternity (24 hours) ago. The general public can’t ... | https://www.lesswrong.com/posts/jdCCBwdPqDNnzkkrm/gpt-4-what-we-i-know-about-it |
# 80k podcast episode on sentience in AI systems
I [spoke](https://80000hours.org/podcast/episodes/robert-long-artificial-sentience/) about sentience and related issues in AI systems with Luisa Rodriguez on the 80,000 Hours podcast.
Twitter thread of highlights [here](https://twitter.com/rgblong/status/1635626139694... | https://www.lesswrong.com/posts/gsi666kGqjmDDDatu/80k-podcast-episode-on-sentience-in-ai-systems |
# How well did Manifold predict GPT-4?
Chat GPT-4 is already here!! Who could have seen that coming… oh wait Manifold (kinda) did? 😅
I thought I’d write a short piece on how Manifold Markets was used to predict the launch of GPT-4 and its attributes. Both its successes and its failures. Disclaimer I work at Manifold... | https://www.lesswrong.com/posts/DaaFce3hBoEzYhdvz/how-well-did-manifold-predict-gpt-4 |
# AI Safety - 7 months of discussion in 17 minutes
In August 2022, I started making summaries of the top EA and LW forum posts each week. This post collates together the key trends I’ve seen in AI Safety discussions since then. Note a lot of good work is happening outside what's posted on these forums too! This post d... | https://www.lesswrong.com/posts/voehWNwi62JLCXEoG/ai-safety-7-months-of-discussion-in-17-minutes |
# Write a Book?
I've been thinking of maybe writing a book about effective altruism. I've been blogging about these ideas for a long time, but a book would offer the opportunity to go into more depth and to be more thoughtful about some concepts I originally wrote up quickly and casually. Many of these ideas are also ... | https://www.lesswrong.com/posts/prBtPnNaHy63DhW4u/write-a-book |
# What is a good comprehensive examination of risks near the Ohio train derailment?
I have family, including small children, that live in Pittsburgh, ~40 miles from the Ohio train derailment.
Their water supply should not be at risk from the spill itself, being out of the downstream watershed of the spill. But the su... | https://www.lesswrong.com/posts/stzDEbRDTomMMYwxj/what-is-a-good-comprehensive-examination-of-risks-near-the |
# Why self-improvement?
I spend a lot of time on self-improvement related things. For example, I listen to a lot of podcasts like Lex Friedman, Peter Attia and Andrew Huberman. In fact, I just finished listening to Lex's [episode with expert negotiator Chris Voss](https://lexfridman.com/chris-voss/).
Why do I do this... | https://www.lesswrong.com/posts/T3GJDsaM2guzegAxX/why-self-improvement |
# Is it worth avoiding detailed discussions of expectations about agency levels of powerful AIs?
I'm working on an article for the Open Phil AI worldview contest. I am thinking of explaining my interpretation of Nate's take on agency of advaned models ([see here](https://www.lesswrong.com/posts/iy2o4nQj9DnQD7Yhj/discu... | https://www.lesswrong.com/posts/2sWjuJFW3RQDFynSe/is-it-worth-avoiding-detailed-discussions-of-expectations |
# Want to predict/explain/control the output of GPT-4? Then learn about the world, not about transformers.
Introduction
============
Consider the following scene from William Shakespeare's *Julius Caesar.*
In this scene, Caesar is at home with his wife Calphurnia. She has awoken after a bad dream, and is pleaded wit... | https://www.lesswrong.com/posts/G3tuxF4X5R5BY7fut/want-to-predict-explain-control-the-output-of-gpt-4-then |
# GPT-4 Multiplication Competition
GPT-4 perfectly understands e.g. Karatsuba multiplication and can consistently multiply two-digit numbers, but hasn’t connected those dots; it needs some hand-holding to correctly use a multiplication algorithm. What’s the shortest prompt that gets GPT-4 to multiply 6-digit integers ... | https://www.lesswrong.com/posts/XvorpDSu3dwjdyT4f/gpt-4-multiplication-competition |
# How I Run Solstice, Step by Step
This is a writeup of how I run Secular Solstices, in as much detail as I can muster. I think most of the existing guides to running Solstices are now somewhat outdated, and not as step-by-step "this is what you need to do when"; I hope this one is a helpful addition.
This post assu... | https://www.lesswrong.com/posts/Quh37bGncJpJqRLZn/how-i-run-solstice-step-by-step |
# [ASoT] Some thoughts on human abstractions
TL;DR:
* Consider a human concept such as "tree." Humans implement some algorithm for determining whether given objects are trees. We expect our predictor/language model to develop a model of this algorithm because this is useful for predicting the behavior of humans.
*... | https://www.lesswrong.com/posts/ktJ9rCsotdqEoBtof/asot-some-thoughts-on-human-abstractions |
# Wittgenstein's Language Games and the Critique of the Natural Abstraction Hypothesis
+++ A Note on Authorship
This was an experiment. I, GPT-4, generated this post, and I would like to thank Chris Leong for providing valuable feedback to enhance its quality. The input given by Chris was as follows:
• Critiques 1 a... | https://www.lesswrong.com/posts/7L7Mdmnoc7xAoTwdh/wittgenstein-s-language-games-and-the-critique-of-the |
# Here, have a calmness video
This is a bit of an unusual post.
I have gotten the impression that a lot of people are kind of freaked out, either by AI or weird Bay Area social dynamics in general.
I also think that a lot of freak-out reactions are driven at least as much by social contagion as any fact-based assess... | https://www.lesswrong.com/posts/duyJ9uFo2pnPgr3Yn/here-have-a-calmness-video |
# A chess game against GPT-4
I just bought a subscription to access GPT-4 and played the following chess game against it, with me playing white. (No particular agenda, was just curious how good it is.)
 info-hazard policies?
I believe Anthropic has said they won't publish capabilities research?
OpenAI seems to be sort of doing the same (although no policy AFAIK).
I heard FHI was developing one way back when...
I think MIRI sort of does as well (defau... | https://www.lesswrong.com/posts/qTiujsznctZcnuLF3/what-organizations-other-than-conjecture-have-esp-public |
# On the Crisis at Silicon Valley Bank
Many have already written about the events of the past week’s crisis.
If you want the basics of what happened, you have many options.
Your best bet, if available to you, is that this is [Matt Levine’s](https://www.bloomberg.com/opinion/authors/ARbTQlRLRjE/matthew-s-levine?cmpid... | https://www.lesswrong.com/posts/andFu4X7MaY7wdBmz/on-the-crisis-at-silicon-valley-bank |
# Natural Abstractions: Key Claims, Theorems, and Critiques
**TL;DR: ***We distill John Wentworth’s Natural Abstractions agenda by summarizing its key claims: the Natural Abstraction Hypothesis—many cognitive systems learn to use similar abstractions—and the Redundant Information Hypothesis—a particular mathematical d... | https://www.lesswrong.com/posts/gvzW46Z3BsaZsLc25/natural-abstractions-key-claims-theorems-and-critiques-1 |
# [Appendix] Natural Abstractions: Key Claims, Theorems, and Critiques
This is the appendix to [Natural Abstractions: Key Claims, Theorems, and Critiques](https://www.lesswrong.com/posts/gvzW46Z3BsaZsLc25/natural-abstractions-key-claims-theorems-and-critiques-1). It contains additional details that we expect are only ... | https://www.lesswrong.com/posts/o7sN7moJA8TrZKtKi/appendix-natural-abstractions-key-claims-theorems-and |
# When will computer programming become an unskilled job (if ever)?
I expect for there to be a delay in deployment, but I think ultimately OpenAI is aiming as a near term goal to automate intellectually difficult portions of computer programming. Personally, as someone just getting into the tech industry, this is basi... | https://www.lesswrong.com/posts/pip5c3SvHDB2XeywK/when-will-computer-programming-become-an-unskilled-job-if |
# Is it a bad idea to pay for GPT-4?
I don't want to accelerate an arms race, and paying for access to GPT seems like a perfect way to be a raindrop in a dangerous flood. My current idea is to pay an equal amount monthly to Miri. I'll view it as the price being $40 per month with half going to AI safety research.
... | https://www.lesswrong.com/posts/cKrX86HJjvTrmhiAG/is-it-a-bad-idea-to-pay-for-gpt-4 |
# Will 2023 be the last year you can write short stories and receive most of the intellectual credit for writing them?
I've been worldbuilding a unique setting in my head and on paper for maybe a year now, and for quite a bit longer than that I have wanted to start publishing fiction. I've never "found the time" to ac... | https://www.lesswrong.com/posts/sGCYvefva2ADfAEJ2/will-2023-be-the-last-year-you-can-write-short-stories-and |
# Attribution Patching: Activation Patching At Industrial Scale
*The following is a write-up of an (incomplete) project I worked on while at Anthropic, and a significant amount of the credit goes to the then team, Chris Olah, Catherine Olsson, Nelson Elhage & Tristan Hume. I've since cleaned up this project in my pers... | https://www.lesswrong.com/posts/gtLLBhzQTG6nKTeCZ/attribution-patching-activation-patching-at-industrial-scale |
# Conceding a short timelines bet early
Last year I [bet some people about short AI timelines](https://www.lesswrong.com/posts/X3p8mxE5dHYDZNxCm/a-concrete-bet-offer-to-those-with-short-agi-timelines). While I don't think I've lost the bet yet, I think it's clear at this point that I will lose with high probability. I... | https://www.lesswrong.com/posts/nto7K5W2sNR3Cpmec/conceding-a-short-timelines-bet-early |
# Gradual takeoff, fast failure
Overview
--------
This post aims to speculate on some concrete ways that things could go awry in a world of [gradual takeoff](https://sideways-view.com/2018/02/24/takeoff-speeds/), where AI capabilities increase smoothly and predictably on human timescales.
(At this point, "human time... | https://www.lesswrong.com/posts/3yAvb8TtLJQbAbtiC/gradual-takeoff-fast-failure |
# Announcing the ERA Cambridge Summer Research Fellowship
The Existential Risk Alliance (ERA) has opened applications for an **in-person, paid, 8-week Summer Research Fellowship **focused on existential risk mitigation, taking place from July 3rd to August 25th 2023 in Cambridge, UK, and aimed at all aspiring research... | https://www.lesswrong.com/posts/uhFo7XeaRHL7b9ioh/announcing-the-era-cambridge-summer-research-fellowship |
# The algorithm isn't doing X, it's just doing Y.
Introduction
============
Mutual reduction implies equivalence
------------------------------------
Here's my most load-bearing intuition —
> If two tasks reduce to one another, then it is meaningless to ask if a machine is 'really doing' one task versus the other.
... | https://www.lesswrong.com/posts/GF6hDawC6QdwGXLsj/the-algorithm-isn-t-doing-x-it-s-just-doing-y |
# Donation offsets for ChatGPT Plus subscriptions
I've decided to donate $240 to both GovAI and MIRI to offset the $480 I plan to spend on ChatGPT Plus over the next two years ($20/month).
I don't have a super strong view on ethical offsets, like donating to anti-factory farming groups to try to offset harm from ... | https://www.lesswrong.com/posts/4uBEnYsmwxDw48fb6/donation-offsets-for-chatgpt-plus-subscriptions |
# GPT-4 solves Gary Marcus-induced flubs
**TLDR:** GPT-4 succeeds at 15 problems from Gary Marcus that exposed failures of GPT-3.
I enjoyed reading the ACX post "[My Bet: AI Size Solves Flubs](https://astralcodexten.substack.com/p/my-bet-ai-size-solves-flubs)" last year. Here are some excerpts:
> Here’s the basic st... | https://www.lesswrong.com/posts/cGbEtNbxACJpqoP4x/gpt-4-solves-gary-marcus-induced-flubs |
# Survey on intermediate goals in AI governance
It seems that a key bottleneck for the field of longtermism-aligned AI governance is limited strategic clarity (see [Muehlhauser, 2020](https://www.openphilanthropy.org/research/our-ai-governance-grantmaking-so-far/), [2021](https://forum.effectivealtruism.org/posts/M2SB... | https://www.lesswrong.com/posts/o8fobRYGAknqdTTsM/survey-on-intermediate-goals-in-ai-governance |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.