text stringlengths 300 320k | source stringlengths 52 154 |
|---|---|
# How security and cryptography can aid AI safety [sequence]
Overview
========
In this review sequence, I lay out why I think security, cryptography, and auxiliary approaches are still neglected in AI safety, specific insights those communities can offer for AI safety, and some potential next steps.
I am writing... | https://www.lesswrong.com/posts/FRQNC7rJJ4jBjonii/how-security-and-cryptography-can-aid-ai-safety-sequence |
# Introducing AlignmentSearch: An AI Alignment-Informed Conversional Agent
Authors: Henri Lemoine, Thomas Lemoine, and Fraser Lee
We are excited to introduce [AlignmentSearch](https://alignmentsearch.up.railway.app/), an attempt to create a conversational agent that can answer questions about AI alignment. We built t... | https://www.lesswrong.com/posts/bGn9ZjeuJCg7HkKBj/introducing-alignmentsearch-an-ai-alignment-informed |
# AI infosec: first strikes, zero-day markets, hardware supply chains, adoption barriers
**\[This part 1 of a** [**5 part sequence on security and cryptography areas relevant for AI safety**](https://www.lesswrong.com/posts/FRQNC7rJJ4jBjonii/four-security-and-cryptography-areas-relevant-for-ai-safety)**, published and... | https://www.lesswrong.com/posts/kvk2ZorXui4YB4zvc/ai-infosec-first-strikes-zero-day-markets-hardware-supply |
# Quaker Practice for the Aspiring Rationalist
I wrote this up for a meetup I ran recently, where we held a Quaker-style Meeting for Worship session. I recommend doing this in a group setting, but you can also do it by yourself if the ideas appeal to you.
Why Quaker meeting
------------------
I’ve had depressive epi... | https://www.lesswrong.com/posts/ybDjpuu99Mwp5ddEv/quaker-practice-for-the-aspiring-rationalist-1 |
# 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*](https://www.lesswrong.com/post... | https://www.lesswrong.com/posts/rH9sXupnoR8wSmRe9/ai-safety-via-luck-2 |
# Repairing the Effort Asymmetry
Sometimes, you will say a thing!
You: "Proposition A."
A seems to you to be obviously true. It accords with your experience, doesn't violate any of the rules and laws of reality, makes sense with the rest of your model of the world. You'd like to move on *from* A to the actual more i... | https://www.lesswrong.com/posts/LrCt2T5sDn6KcSJgM/repairing-the-effort-asymmetry |
# A policy guaranteed to increase AI timelines
The number of years until the creation of powerful AI is a major input to our thinking about risk from AI and which approaches are most promising for mitigating that risk. While there are downsides to transformative AI arriving many years from now, rather than few years f... | https://www.lesswrong.com/posts/fDJyKk323ob22WhEC/a-policy-guaranteed-to-increase-ai-timelines |
# Shutting down AI is not enough. We need to destroy all technology.
A [TIME article published recently](https://time.com/6266923/ai-eliezer-yudkowsky-open-letter-not-enough/) calls for an “indefinite and worldwide” moratorium on new large AI training runs.
This moratorium would be better than no moratorium. I have r... | https://www.lesswrong.com/posts/aZd9s5mXkdrGJNJi6/shutting-down-ai-is-not-enough-we-need-to-destroy-all |
# A Confession about the LessWrong Team
The team and I had agreed that no one had to know, that in fact doing so would be bad for the world via proving unequivocally to companies the economic value of AI and thereby spurring more investment (I think this is possible on the margin), but also one must remember to [speak... | https://www.lesswrong.com/posts/ZSRAoaW4ErMdxdeEm/a-confession-about-the-lesswrong-team |
# A report about LessWrong karma volatility from a different universe
*In a far away universe, a news report is written about LessWrong. The following passages have been lifted over and written into this post...*
**Early one morning** [**all voting on LessWrong was halted**](https://www.coindesk.com/business/2022/11/... | https://www.lesswrong.com/posts/THtqYL8adZSwZCtQ9/a-report-about-lesswrong-karma-volatility-from-a-different |
# Policy discussions follow strong contextualizing norms
*tl;dr: because policy discussions follow contextualizing norms, value judgements are often interpreted as endorsements of (uncooperative) actions; people used to decoupling norms should carefully account for this in order to communicate effectively.*
One diffi... | https://www.lesswrong.com/posts/cbQih72wbKkrSX7yx/policy-discussions-follow-strong-contextualizing-norms |
# Analysis of GPT-4 competence in assessing complex legal language: Example of Bill C-11 of the Canadian Parliament. - Part 1
If your at all familiar with Canadian politics over the past few years you may have heard of this mysterious '[C-11](https://www.parl.ca/DocumentViewer/en/44-1/bill/C-11/third-reading)' (a.k.a.... | https://www.lesswrong.com/posts/4hz6Rswmu6qLEHAei/analysis-of-gpt-4-competence-in-assessing-complex-legal |
# Some lesser-known megaproject ideas
Below is a list of megaproject ideas that didn’t quite make the cut enough to be added to the [database](https://docs.google.com/spreadsheets/d/1VLEiWx0_sFwjhCMao3vT_QxdWCufuq71XqaL2joVj8Y/edit#gid=1763987441). I compiled this list last year and sent it last November to a prominen... | https://www.lesswrong.com/posts/c5fQsxq4j2W4uKfwv/some-lesser-known-megaproject-ideas |
# When should a neural network-based approach for plant control systems be preferred over a traditional control method?
As a trained control engineer, I understand that when the plant model is well understood (or even with many uncertainty in the plant dynamic model), designing a control system using traditional metho... | https://www.lesswrong.com/posts/Ric9wkpotFAe5i6XK/when-should-a-neural-network-based-approach-for-plant |
# Transparency for Generalizing Alignment from Toy Models
*Status: Some rough thoughts and intuitions.*
*🪧 indicates [signposting](https://www.qub.ac.uk/graduate-school/Filestore/Filetoupload,597684,en.pdf)*
**TL;DR**
If we make our optimization procedures transparent, we might be able to analyze them in toy enviro... | https://www.lesswrong.com/posts/9iHwqnH4ZeqkGDbrb/transparency-for-generalizing-alignment-from-toy-models-1 |
# Ultimate ends may be easily hidable behind convergent subgoals
*[Metadata: crossposted from [https://tsvibt.blogspot.com/2022/12/ultimate-ends-may-be-easily-hidable.html](https://tsvibt.blogspot.com/2022/12/ultimate-ends-may-be-easily-hidable.html). First completed December 18, 2022.]*
Thought and action in pursuit... | https://www.lesswrong.com/posts/8qCKZj24FJotm3EKd/ultimate-ends-may-be-easily-hidable-behind-convergent |
# Exposure to Lizardman is Lethal
This is a short, experimentally-off-the-cuff post about one way in which social groups and institutions are effectively destroyed, in practice. It's fairly straightforward, but I don't try to connect all the dots (or even name all of them).
* * *
In a recent FB post, I made the foll... | https://www.lesswrong.com/posts/gbdqMaADqxMDwtgaf/exposure-to-lizardman-is-lethal |
# Orthogonality is Expensive
[@beren](https://www.lesswrong.com/users/beren-1?mention=user) discusses the assumption that intelligent systems would be well factored into a world model, objectives/values and a planning system.
He highlights that this factorisation doesn't describe intelligent agents created by ML syst... | https://www.lesswrong.com/posts/emjZiBFZYftzTHcao/orthogonality-is-expensive-1 |
# Slack Group: Rationalist Startup Founders
*Update: I decided to discontinue this group. Not enough people with concrete enough plans to start a startup joined, and there was a general lack of activity.*
I have a friend who is a top poker player. In asking him about his experiences and how he got so good, one thing ... | https://www.lesswrong.com/posts/Xb3mmHhcz3DpufhzQ/slack-group-rationalist-startup-founders |
# The Friendly Drunk Fool Alignment Strategy
Let's set some context!
### A Philosophic Frame (half ironic, or maybe level two sarcasm)
Systematicity isn't all that. I mean to say: formal systems only go so far.
Physicists gave up on having a single physics, so they just use two of them that contradict in general b... | https://www.lesswrong.com/posts/EqzbyW5EReJ7TNXhK/the-friendly-drunk-fool-alignment-strategy |
# Planecrash Podcast
I used AWS' text-to-speech service to create a podcast out of [Eliezer's Planecrash Glowfix](https://www.lesswrong.com/posts/SA9hDewwsYgnuscae/projectlawful-com-eliezer-s-latest-story-past-1m-words).
It is available here: [https://anchor.fm/s/a48073e4/podcast/rss](https://anchor.fm/s/a48073e4/pod... | https://www.lesswrong.com/posts/MX8SAwa2rDJFSiiyW/planecrash-podcast |
# Orthogonality is expensive
*Crossposted from* [*my personal blog*](https://www.beren.io/2023-03-19-Orthogonality-is-expensive/)
A common assumption about AGI is the orthogonality thesis, which argues that goals/utility functions and the core intelligence of an AGI system are orthogonal or can be cleanly factored ap... | https://www.lesswrong.com/posts/WXLJASckbjJcoaEmx/orthogonality-is-expensive |
# AI-kills-everyone scenarios require robotic infrastructure, but not necessarily nanotech
Tl;DR: Nanotech seems the only option for AI to kill everybody; but we can imagine a prosaic-AI-killing-everyone using three steps: taking over a country, building robotic infrastructure, releasing biological weapons.
*Argument... | https://www.lesswrong.com/posts/QBTdEyL3tDaJY3LNa/ai-kills-everyone-scenarios-require-robotic-infrastructure |
# Exploratory Analysis of RLHF Transformers with TransformerLens
**TL;DR:** I demonstrate how to use RLHF models trained with the TRLX library with TransformerLens, and how to take an exploratory look at how RLHF changes model internals. I also use activation patching to see what RLHF activations are sufficient to rec... | https://www.lesswrong.com/posts/Ky3WnDwQbLAucGrXf/exploratory-analysis-of-rlhf-transformers-with |
# Exploring non-anthropocentric aspects of AI existential safety
**Epistemic status:** These are some initial thoughts.
**Goal:** Anthropocentric approaches dominate the field of AI existential safety.
In particular, AI existential safety is usually formulated in terms of _alignment to human goals and values_.
It m... | https://www.lesswrong.com/posts/WJuASYDnhZ8hs5CnD/exploring-non-anthropocentric-aspects-of-ai-existential |
# Monthly Roundup #5: April 2023
As AI eats the world and my time and news feeds, it is always good to know other things also happen. Here are some of them, that did not seem to fit anywhere else and did not seem worth standing on their own even with my new lower thresholds for that.
#### Bad News
Not only do you on... | https://www.lesswrong.com/posts/JeK4ztN4nNBgjXpNS/monthly-roundup-5-april-2023 |
# Towards empathy in RL agents and beyond: Insights from cognitive science for AI Alignment
This is a talk I gave at the recent AI Safety Europe Retreat (AISER) on my research on obtaining insights from the cognitive science of empathy and applying them to RL agents and LLMs.
**Talk link:** [https://clipchamp.com/wa... | https://www.lesswrong.com/posts/bNpqBNvfgCWixB2MT/towards-empathy-in-rl-agents-and-beyond-insights-from-1 |
# If interpretability research goes well, it may get dangerous
I've [historically](https://twitter.com/robbensinger/status/1602835145488113664) been pretty publicly [supportive](https://www.lesswrong.com/posts/3pinFH3jerMzAvmza/on-how-various-plans-miss-the-hard-bits-of-the-alignment) of interpretability research. I'm... | https://www.lesswrong.com/posts/BinkknLBYxskMXuME/if-interpretability-research-goes-well-it-may-get-dangerous |
# Communicating effectively under Knightian norms
*tl;dr: rationalists concerned about AI risk often make claims that others consider not just unjustified, but* unjustifiable *using their current methodology, because of high-level disagreements about epistemology. If you actually want to productively discuss AI risk, ... | https://www.lesswrong.com/posts/tG9BLyBEiLeRJZvX6/communicating-effectively-under-knightian-norms |
# Apply to the Cavendish Labs Fellowship (by 4/15)
[Cavendish Labs](https://www.lesswrong.com/posts/JiGwL5nDMTwehP62N/announcing-cavendish-labs) is a new research organization in Vermont focused on technical work on existential risks. We'd like to invite you to [apply to our fellowships](https://cavendishlabs.org/fell... | https://www.lesswrong.com/posts/qgxQuuFpdQzAcPf9x/apply-to-the-cavendish-labs-fellowship-by-4-15 |
# Complex Systems are Hard to Control
The deployment of powerful deep learning systems such as ChatGPT raises the question of how to make these systems safe and consistently aligned with human intent. Since building these systems is an engineering challenge, it is tempting to think of the safety of these systems prima... | https://www.lesswrong.com/posts/iwCRYnGYMvxgzrCMf/complex-systems-are-hard-to-control |
# Steering systems
## Preface: who this post is for
This post is mainly targeted at people who are already familiar with the concepts and terminology in alignment, AI x-risk, and particularly the case for high probabilities of doom, but disagree with the arguments or have bounced off of them for one reason or another... | https://www.lesswrong.com/posts/LsqvMKnFRBQh4L3Rs/steering-systems |
# AI Summer Harvest
**Trending metaphor**
I've noticed that AI safety advocates have recently been appealing to a new metaphor: the AI summer harvest.
(1) Here's the FLI open letter *Pause Giant AI Experiments:*
> Humanity can enjoy a flourishing future with AI. Having succeeded in creating powerful AI systems, we ... | https://www.lesswrong.com/posts/P98i7kAN2uWuy7mhD/ai-summer-harvest |
# Consider applying to a 2-week alignment project with former GitHub CEO
Signal-boosting this tweet from Nat Friedman, former github CEO, who says:
> I'm seeking a researcher/journalist with an interest in AI/rationalism/xrisk for a 1-2 week project. To apply please email [job@nat.org](mailto:job@nat.org) with evide... | https://www.lesswrong.com/posts/mR44SBdDYpxzQX672/consider-applying-to-a-2-week-alignment-project-with-former |
# Best Ways to Try to Get Funding for Alignment Research?
Hey Everyone! I recently left my FAANG job to split my time between doing Alignment Research (70%) and investigating start-up ideas (30%).
If I decide to fully commit to Alignment Research, what is the best way to go about applying for and/or getting funding? ... | https://www.lesswrong.com/posts/y9f79C3wLDDQjmqZa/best-ways-to-try-to-get-funding-for-alignment-research |
# Invocations: The Other Capabilities Overhang?
*Abstract: An LLM’s invocation is the non-model code around it that determines when and how the model is called. I illustrate that LLMs are already used under widely varying invocations, and that a model’s capabilities depend in part on its invocation. I discuss several ... | https://www.lesswrong.com/posts/SLGu37Wug7HmdGpbD/invocations-the-other-capabilities-overhang |
# Running many AI variants to find correct goal generalization
*Problem: An AI may incorrectly generalize a goal we present it.*
*Proposal: Run many slightly different instances of AI and find the “Schelling point” around correct generalization.*
The idea is that if we run many similar AIs or copies of the same AI i... | https://www.lesswrong.com/posts/HcgzWjnGaLkRfE26f/running-many-ai-variants-to-find-correct-goal-generalization |
# Given the Restrict Act, Don’t Ban TikTok
While others have already written similar posts, given my previous position it seems necessary for me to do so as well – the Restrict Act is a no good very bad bill, and having seen the bill I realize that I was wrong to support banning TikTok.
For a while I have been in fav... | https://www.lesswrong.com/posts/bxebfujNqpmpnj5kz/given-the-restrict-act-don-t-ban-tiktok |
# Steelman / Ideological Turing Test of Yann LeCun's AI X-Risk argument?
Yann LeCun has been saying a lot of things on social media recently about this topic, only some of which I've read. He's also written and talked about it several times in the past. Most of what I've seen from him recently seems to not be addressi... | https://www.lesswrong.com/posts/4t8SZ4R4AAkb5etHT/steelman-ideological-turing-test-of-yann-lecun-s-ai-x-risk |
# Play My Futarchy/Prediction Market Mafia Game
In the social deduction game [Mafia](https://en.wikipedia.org/wiki/Mafia_(party_game)) (broadly the same game as Werewolf, Town of Salem, etc) the minority mafiosi—who know each other's identities—play against the majority villagers in alternating day phases, where the p... | https://www.lesswrong.com/posts/D9aGhRSbiJHYRvxHA/play-my-futarchy-prediction-market-mafia-game |
# Giant (In)scrutable Matrices: (Maybe) the Best of All Possible Worlds
It has become common on LW to refer to "giant inscrutable matrices" as a problem with modern deep-learning systems.
To clarify: deep learning models are trained by creating giant blocks of random numbers -- blocks with dimensions like 4096 x 512 ... | https://www.lesswrong.com/posts/GkC6YTu4DWp2zwf9k/giant-in-scrutable-matrices-maybe-the-best-of-all-possible |
# Penalize Model Complexity Via Self-Distillation
When you self-distill a model (e.g. train a new model using predictions from your old model), the resulting model represents a less complex function. After many rounds of self-distillation, you essentially end up with a constant function. [This](https://proceedings.neu... | https://www.lesswrong.com/posts/fzGbKHbSytXH5SKTN/penalize-model-complexity-via-self-distillation |
# Excessive AI growth-rate yields little socio-economic benefit.
Basic argument
==============
In short —
1. There exists an **adaptive limit** representing the maximum rate at the economy can adapt to new technology.
2. Exceeding this limit yields no socio-economic benefit, in the general case.
3. Exceeding this... | https://www.lesswrong.com/posts/WbdLYgbpxfrSXCBS6/excessive-ai-growth-rate-yields-little-socio-economic |
# LW Team is adjusting moderation policy
Lots of new users have been joining LessWrong recently, who seem more filtered for "interest in discussing AI" than for being bought into any particular standards for rationalist discourse. I think there's been a shift in this direction over the past few years, but it's gotten ... | https://www.lesswrong.com/posts/kyDsgQGHoLkXz6vKL/lw-team-is-adjusting-moderation-policy |
# New survey: 46% of Americans are concerned about extinction from AI; 69% support a six-month pause in AI development
YouGov America released a survey of 20,810 American adults. Highlights below. Note that I didn't run any statistical tests, so any claims of group differences are just "eyeballed."
* 46% say that t... | https://www.lesswrong.com/posts/p8suSXSwEoKTfGbp9/new-survey-46-of-americans-are-concerned-about-extinction |
# "Corrigibility at some small length" by dath ilan
Although this appears to be some sort of fictional role-playing server, I quite like the explanation of corrigibility given. It is clearly a direct reference to the AI alignment problem. They even included a [tl;dr, which I included at the bottom of the post](https:/... | https://www.lesswrong.com/posts/5sRK4rXH2EeSQJCau/corrigibility-at-some-small-length-by-dath-ilan |
# Has anyone thought about how to proceed now that AI notkilleveryoneism is becoming more relevant/is approaching the Overton window?
I figure that this would be a pretty important question since getting governments on board, and getting people to coordinate in general, could buy us a lot of time, but I haven't seen a... | https://www.lesswrong.com/posts/doH7ZThCujnCQsSDF/has-anyone-thought-about-how-to-proceed-now-that-ai |
# AGI deployment as an act of aggression
I started a draft of this post some days ago, but then a lot of things happened and so I'm rewriting it from scratch. Most importantly, [a TIME editorial in which Eliezer Yudkowski talks of bombing AI training data centres happened](https://www.lesswrong.com/posts/Aq5X9tapacnk2... | https://www.lesswrong.com/posts/otArJmyzWgfCxNZMt/agi-deployment-as-an-act-of-aggression |
# ICA Simulacra
###
Pre-preface
This post was written about a year ago, but I decided not to publish it with concerns of burning timelines. Since then, after the hype caused by ChatGPT, similar systems have been proposed, implemented and even tested; so I think that “the secret”, even if there was one, is out, and... | https://www.lesswrong.com/posts/cLKR7utoKxSJns6T8/ica-simulacra |
# AI safety and the security mindset: user interface design, red-teams, formal verification
**\[This is part 2 of a** [**5 part sequence on how security and cryptography can help AI safety**](/posts/FRQNC7rJJ4jBjonii/four-security-and-cryptography-areas-relevant-for-ai-safety)**, published and linked here a few days a... | https://www.lesswrong.com/posts/RKCfBMoMXr6JsgHAy/ai-safety-and-the-security-mindset-user-interface-design-red |
# Universality and Hidden Information in Concept Bottleneck Models
Summary
-------
I use the CUB dataset to finetune models to classify images of birds, via a layer trained to predict relevant concepts (e.g. "has\_wing\_pattern::spotted"). These models, when trained end-to-end, naturally put additional relevant infor... | https://www.lesswrong.com/posts/uTPetRBbkP4p6FhGf/universality-and-hidden-information-in-concept-bottleneck |
# Progress links and tweets, 2023-04-05
**The Progress Forum**
----------------------
* [Wizards and prophets of AI](https://progressforum.org/posts/baPwkBDa9f9b8ja2z/wizards-and-prophets-of-ai-draft-for-comment). I posted this for comment, then decided to rewrite it, then ended up [posting the core argument of the... | https://www.lesswrong.com/posts/esabThofG2gby7CeP/progress-links-and-tweets-2023-04-05 |
# Eliezer Yudkowsky’s Letter in Time Magazine
[FLI put out an open letter](https://thezvi.substack.com/p/on-the-fli-ai-risk-open-letter), calling for a 6 month pause in training models more powerful than GPT-4, followed by additional precautionary steps.
[Then Eliezer Yudkowsky put out a post in Time](https://time.co... | https://www.lesswrong.com/posts/eo8odvou4efc9syrv/eliezer-yudkowsky-s-letter-in-time-magazine |
# "On Living in an Atomic Age", by C.S. Lewis (1948)
*My note: This essay was written for people worried about nuclear weapons during the Cold War. Although I do not agree with the supernaturalism towards the end, I find the first three paragraphs in particular to be incredibly prescient and helpful for many people wo... | https://www.lesswrong.com/posts/fMwvWu3Bvf2dw8kHT/on-living-in-an-atomic-age-by-c-s-lewis-1948 |
# Why Are Maximum Entropy Distributions So Ubiquitous?
If we measure the distribution of particle velocities in a thin gas, we’ll find that they’re roughly normally distributed. Specifically, the probability density of velocity $v$ will be proportional to $e^{-\frac{1}{2} mv^2 /(k_B T)}$ \- or, written differently, $e... | https://www.lesswrong.com/posts/p6Qkq8qFdMejy2JFL/why-are-maximum-entropy-distributions-so-ubiquitous |
# Auto-GPT: Open-sourced disaster?
Sharing this here doesn't seem like an infohazard at this point. This is all over my YouTube feed anyway.
Description from the authors:
> Auto-GPT is an experimental open-source application showcasing the capabilities of the GPT-4 language model. This program, driven by GPT-4, auto... | https://www.lesswrong.com/posts/s9JWqgnv7xT2mxmE7/auto-gpt-open-sourced-disaster |
# Someone already tried "Chaos-GPT"
With the new Auto-GPT release, I was surprised to see how quickly some people turned towards using it for nefarious goals, apparently just for fun. Is this the first piece of clear cut, empirical evidence for the argument against open sourcing these soon-to-be very powerful general ... | https://www.lesswrong.com/posts/MAgQ8owro57eYFhNu/someone-already-tried-chaos-gpt |
# Yoshua Bengio: "Slowing down development of AI systems passing the Turing test"
Full text of the blog post below:
> I recently signed an [**open letter**](https://futureoflife.org/open-letter/pause-giant-ai-experiments/) asking to slow down the development of giant AI systems more powerful than GPT-4 –those that c... | https://www.lesswrong.com/posts/tq8uMdSDj8iRnGmTE/yoshua-bengio-slowing-down-development-of-ai-systems-passing |
# No Summer Harvest: Why AI Development Won't Pause
Cleo Nardo recently highlighted Yudkowsky and others discussing an "[AI Summer Harvest](https://www.lesswrong.com/posts/P98i7kAN2uWuy7mhD/ai-summer-harvest)" (or simply "Summer Harvest"). This potential near-future era would involve a pause in research on state-of-th... | https://www.lesswrong.com/posts/hYYdpgag6Hbs7jLre/no-summer-harvest-why-ai-development-won-t-pause |
# What's the deal with Effective Accelerationism (e/acc)?
I've been hearing murmurs about a recently formed philosophy called "Effective Accelerationism", described as:[^2y02pjfiae5]
> ...an ideology that draws from Nick Land's theories of accelerationism to advocate for the belief that artificial intelligence and LL... | https://www.lesswrong.com/posts/2ss6gomAJdqjwdSCy/what-s-the-deal-with-effective-accelerationism-e-acc |
# Dual-Useness is a Ratio
A lot of AI-risk-concerned people are struggling with how to relate to dual-use research, and relatedly, to doing alignment research inside of AI orgs. There's a pretty simple concept that seems, to me, to be key to thinking about this coherently: the dual-useness ratio. Most prosaic alignmen... | https://www.lesswrong.com/posts/vquDf3fbGyzH7Ryav/dual-useness-is-a-ratio |
# Revisiting the Horizon Length Hypothesis
**Summary**: As part of my work at [Epoch](https://epochai.org/), I investigated the horizon length hypothesis - the idea that the horizon length of a task is predictive of the training compute needed to learn that task. My current (weak) conclusion is that the horizon length... | https://www.lesswrong.com/posts/BGtjG6PzzmPngCgW9/revisiting-the-horizon-length-hypothesis |
# Is "Recursive Self-Improvement" Relevant in the Deep Learning Paradigm?
Disclaimer
----------
I wrote this three months ago, and abandoned it[^l22qk49d8pa]. I currently do not plan to return to it anytime soon, but [I have nonetheless found myself wanting](https://twitter.com/CineraVerinia/status/164369672356534681... | https://www.lesswrong.com/posts/oyK6fYYnBi5Nx5pfE/is-recursive-self-improvement-relevant-in-the-deep-learning |
# The Computational Anatomy of Human Values
This is crossposted from my [personal blog](https://www.beren.io/2023-04-03-Computational-anatomy-of-human-values/).
**Epistemic Status**: *Much of this draws from my studies in neuroscience and ML. Many of the ideas in this post are heavily inspired by the work of *[*Steve... | https://www.lesswrong.com/posts/pZHpq6dBQzCZjjMgM/the-computational-anatomy-of-human-values |
# Stupid Questions - April 2023
Welcome to the "Stupid Questions" thread! Feel free to ask any questions, regardless of whether they seem obvious, tangential, silly, or what-have-you. Don't be shy - everyone has gaps in their knowledge, and the goal here is to help reduce them.
Please remember to be respectful when s... | https://www.lesswrong.com/posts/25yi9PeqzDXPn2qpq/stupid-questions-april-2023 |
# AI #6: Agents of Change
If you didn’t have any future shock over the past two months, either you weren’t paying attention to AI developments or I am very curious how you managed that.
I would not exactly call this week’s pace of events _slow._ It was still distinctly _slower_ than that which we have seen in the pre... | https://www.lesswrong.com/posts/DXByibGtvFixSsyfs/ai-6-agents-of-change |
# Do we get better or worse at adapting to change?
Verner Vinge, in [a classic 1993 essay](https://frc.ri.cmu.edu/~hpm/book98/com.ch1/vinge.singularity.html), described “the Singularity” as an era where progress becomes “an exponential runaway beyond any hope of control.”
The idea that technological change might acce... | https://www.lesswrong.com/posts/vZaxYBZFJS5xkhD8v/do-we-get-better-or-worse-at-adapting-to-change |
# Eliezer on The Lunar Society podcast
Another podcast appearance. Great to see Eliezer on a real media blitz tour.
> For 4 hours, I tried to come up reasons for why AI might not kill us all, and Eliezer Yudkowsky explained why I was wrong.
> We also discuss his call to halt AI, why LLMs make alignment harder, what ... | https://www.lesswrong.com/posts/vTRi3mNH8Abzxq3xv/eliezer-on-the-lunar-society-podcast |
# Pause AI Development?
I like the basic idea of a pause in training increasingly powerful AIs.
Yet I\'m quite dissatisfied with any specific plan that I can think of.
AI research is proceeding at a reckless pace. There\'s massive
disagreement among intelligent people as to how dangerous this is.
I participate... | https://www.lesswrong.com/posts/hTF9h3ufbSuMuZfwj/pause-ai-development |
# I asked my senator to slow AI
All my favorite twitter accounts agree that AI capabilities are advancing too fast.
One way to slow down AI progress is to ask politicians for help. Politicians are... | https://www.lesswrong.com/posts/tcBEh9ZtEWvoHNKZ4/i-asked-my-senator-to-slow-ai |
# AISafety.world is a map of the AIS ecosystem
 It turns out you can ask GPT4 to compress a message in a way it can understand. The end result is usually not understandable for a human. It's interesting, since it allows... | https://www.lesswrong.com/posts/bNCDexejSZpkuu3yz/you-can-use-gpt-4-to-create-prompt-injections-against-gpt-4 |
# Misgeneralization as a misnomer
Here's two different ways an AI can turn out [unfriendly](https://www.lesswrong.com/posts/BSee6LXg4adtrndwy/what-does-it-mean-for-an-agi-to-be-safe):
1. You somehow build an AI that cares about "making people happy". In training, it tells people jokes and buys people flowers and off... | https://www.lesswrong.com/posts/dkjwSLfvKwpaQSuWo/misgeneralization-as-a-misnomer |
# Anthropic is further accelerating the Arms Race?
Anthropic is raising even more funds and the pitch deck seems scary. A choice quote from the article:
> These models could begin to automate large portions of the economy,” the pitch deck reads. “We believe that companies that train the best 2025/26 models will be to... | https://www.lesswrong.com/posts/BbMJo5vQEc7PoRKMn/anthropic-is-further-accelerating-the-arms-race |
# If Alignment is Hard, then so is Self-Improvement
Let’s accept that aligning very intelligent artificial agents is hard. In that case, if we build an intelligent agent with some goal (which probably won’t be the goal we intended, because we’re accepting alignment is hard) and it decides that the best way to achieve ... | https://www.lesswrong.com/posts/4KHPxsJgGfxwCNSCC/if-alignment-is-hard-then-so-is-self-improvement |
# Catching the Eye of Sauron
*(edit 7/24/2023: Certain sections of this post I no longer endorse, but the central dilemma of the Eye remains) *
The decision to reach out to the broad public isn't - or shouldn't - be one that comes lightly. However, once you are actively vying for the Eye of Sauron - writing in TI... | https://www.lesswrong.com/posts/CqvwtGpJZqYW9qM2d/catching-the-eye-of-sauron |
# Polio Lab Leak Caught with Wastewater Sampling
I hadn't seen this discussed here:
> The near complete eradication of wildtype polioviruses (WPV) means that strict containment by facilities for essential work with infectious WPV is required. In the Netherlands, we have implemented environmental surveillance around a... | https://www.lesswrong.com/posts/nhb3CpnnpnLxWBn5x/polio-lab-leak-caught-with-wastewater-sampling |
# Goal alignment without alignment on epistemology, ethics, and science is futile
Originally posted as a comment to the post "[Misgeneralization as a misnomer](https://www.alignmentforum.org/posts/dkjwSLfvKwpaQSuWo/misgeneralization-as-a-misnomer)" by Nate Soares.
* * *
I start to suspect that the concepts of "goals... | https://www.lesswrong.com/posts/fqfAmAGFLKpsnjfJB/goal-alignment-without-alignment-on-epistemology-ethics-and |
# Beren's "Deconfusing Direct vs Amortised Optimisation"
Preamble
--------
I heavily recommend [@beren](https://www.alignmentforum.org/users/beren-1?mention=user)'s "[Deconfusing Direct vs Amortised Optimisation](https://www.lesswrong.com/posts/S54HKhxQyttNLATKu/deconfusing-direct-vs-amortised-optimization)". It's a ... | https://www.lesswrong.com/posts/5YDczJcLZ6RmN5SSz/beren-s-deconfusing-direct-vs-amortised-optimisation-2 |
# Stampy's AI Safety Info - New Distillations #1 [March 2023]
This is our first check-in from the [AI Safety Info distillation fellowship](https://www.lesswrong.com/posts/arkQaWauCvkTvgcRH/ai-safety-info-distillation-fellowship). We are working on distilling all manner of content to help make it easier for people to e... | https://www.lesswrong.com/posts/FYRYhkdAQoQibasNB/stampy-s-ai-safety-info-new-distillations-1-march-2023 |
# Pre-registering a study
This post is my [pre-registration](https://www.acf.hhs.gov/opre/blog/2022/08/pre-registering-studies-what-it-how-do-you-do-it-and-why#:~:text=Pre%2Dregistration%20is%20the%20practice,submitting%20it%20to%20a%20registry.) of a study I will be running to continue the exploratory work I started ... | https://www.lesswrong.com/posts/hfrktmuEMfDZJYsHq/pre-registering-a-study |
# Reliability, Security, and AI risk: Notes from infosec textbook chapter 1
I recently read the first chapter of [Building Secure & Reliable Systems](https://static.googleusercontent.com/media/sre.google/en//static/pdf/building_secure_and_reliable_systems.pdf) (the book from the [EA infosec book club](https://forum.ef... | https://www.lesswrong.com/posts/4qyqvKj3N87EQXo4Y/reliability-security-and-ai-risk-notes-from-infosec-textbook |
# Superintelligence Is Not Omniscience
### The Power of Intelligence
It is often implicitly assumed that the power of a superintelligence will be practically unbounded. There seems like there could be “ample headroom” above humans, i.e. that a superintelligence will be able to vastly outperform us across virtually al... | https://www.lesswrong.com/posts/qpgkttrxkvGrH9BRr/superintelligence-is-not-omniscience |
# Environments for Measuring Deception, Resource Acquisition, and Ethical Violations
Models are rapidly being deployed in the real world. How do we evaluate models, especially ones as complex as GPT-4, to ensure that they behave safely in pursuit of their objectives? Can we design models that robustly avoid any harms ... | https://www.lesswrong.com/posts/smDeWfgeYDg9eGq5G/environments-for-measuring-deception-resource-acquisition |
# Pausing AI Developments Isn't Enough. We Need to Shut it All Down
(*Published in *[*TIME*](https://time.com/6266923/ai-eliezer-yudkowsky-open-letter-not-enough/) *on March 29.*)
An [open letter](https://futureoflife.org/open-letter/pause-giant-ai-experiments/) published today calls for “all AI labs to immediately p... | https://www.lesswrong.com/posts/oM9pEezyCb4dCsuKq/pausing-ai-developments-isn-t-enough-we-need-to-shut-it-all-1 |
# Consider The Hand Axe
A long time ago, some primitive apes got addicted to rocks.
The earliest [stone tools](https://en.wikipedia.org/wiki/Stone_tool) were crude bastards, made by smashing large river pebbles together and calling it a day.
[ have shown impressive... | https://www.lesswrong.com/posts/wLjSr66XAfAQeFp6Y/upcoming-changes-in-large-language-models |
# Bringing Agency Into AGI Extinction Is Superfluous
Epistemic status: Maybe I am missing something here, normally I'd publish this on my own blog, but I'd actually love to get some feedback before I signal boost it.
I’ve seen a lot of debates around AGI risk hedge on the question of whether or not AGI will be “agent... | https://www.lesswrong.com/posts/hwoCkQkL4ixGqScWp/bringing-agency-into-agi-extinction-is-superfluous |
# All AGI Safety questions welcome (especially basic ones) [April 2023]
**tl;dr: Ask questions about AGI Safety as comments on this post, including ones you might otherwise worry seem dumb!**
Asking beginner-level questions can be intimidating, but everyone starts out not knowing anything. If we want more people in t... | https://www.lesswrong.com/posts/wqeStKQ3PGzZaeoje/all-agi-safety-questions-welcome-especially-basic-ones-april-1 |
# Guidelines for productive discussions
I enjoyed reading Duncan_Sabien's post, [Basics of Rationalist Discourse](https://www.lesswrong.com/posts/XPv4sYrKnPzeJASuk/basics-of-rationalist-discourse-1), as well as the response posts and comments. The guidelines probably wouldn't have made sense to my beginner-self though... | https://www.lesswrong.com/posts/wmvCcQD4naApTvrFY/guidelines-for-productive-discussions |
# SERI MATS - Summer 2023 Cohort
Applications have opened for the Summer 2023 Cohort of the [SERI ML Alignment Theory Scholars Program](https://www.serimats.org/)! Our mentors include [Alex Turner](https://www.lesswrong.com/users/turntrout), [Dan Hendrycks,](https://www.lesswrong.com/users/dan-h) [Daniel Kokotajlo,](h... | https://www.lesswrong.com/posts/aEQBkDPZi6L2LMpnC/seri-mats-summer-2023-cohort |
# The benevolence of the butcher
A few days ago I published [this post](https://www.lesswrong.com/posts/otArJmyzWgfCxNZMt/agi-deployment-as-an-act-of-aggression) on the risks of powerful transformative AGI (by which I meant AGI that takes off fast and pretty much rules the world in no time), even if aligned. Among the... | https://www.lesswrong.com/posts/2ujT9renJwdrcBqcE/the-benevolence-of-the-butcher |
# Relative Abstracted Agency
Note: This post was pasted without much editing or work put into formatting. I may come back and make it more presentable at a later date, but the concepts should still hold.
Relative abstracted agency is a framework for considering the extent to which a modeler models a target as an a... | https://www.lesswrong.com/posts/JqWQxTyWxig8Ltd2p/relative-abstracted-agency |
# Can we evaluate the "tool versus agent" AGI prediction?
In 2012, Holden Karnofsky[^MISSING-ID] [critiqued MIRI](https://www.lesswrong.com/posts/6SGqkCgHuNr7d4yJm/thoughts-on-the-singularity-institute-si) (then SI) by saying "SI appears to neglect the potentially important distinction between 'tool' and 'agent' AI." ... | https://www.lesswrong.com/posts/3YqBGnRLtHDGHLgSb/can-we-evaluate-the-tool-versus-agent-agi-prediction |
# The surprising parameter efficiency of vision models
*Crossposted from* [*my personal blog*](https://www.beren.io/2023-04-04-Surprising-parameter-efficiency-vision-models/)*.*
**Epistemic status:** *This is a short post meant to highlight something I do not yet understand and therefore a potential issue with my mod... | https://www.lesswrong.com/posts/qiXs9bEuT2iTYTF2f/the-surprising-parameter-efficiency-of-vision-models |
# GPTs are Predictors, not Imitators
(Related text posted to [Twitter](https://twitter.com/ESYudkowsky/status/1644760694771048449); this version is edited and has a more advanced final section.)
Imagine yourself in a box, trying to predict the next word - assign as much probability mass to the next token as possible ... | https://www.lesswrong.com/posts/nH4c3Q9t9F3nJ7y8W/gpts-are-predictors-not-imitators |
# ELCK might require nontrivial scalable alignment progress, and seems tractable enough to try
Written quickly rather than not at all, I was thinking about this a couple of days ago, and decided to commit to writing something by today rather than adding this idea to my list of a million things to write up.
This post ... | https://www.lesswrong.com/posts/ugjPKpLenCS8iZmvD/elck-might-require-nontrivial-scalable-alignment-progress |
# Feature Request: Right Click to Copy LaTeX
Something I do frequently is run into a block of LaTeX that I'm not sure I fully grasp.
At such a point, what I'd like to do is just copy the equation and paste it to GPT-4 and have it explain it to me step by step, then iterate on the feedback until I'm sure I understood ... | https://www.lesswrong.com/posts/xx79KgauajF4zpYZi/feature-request-right-click-to-copy-latex |
# A decade of lurking, a month of posting
This post is a look back on my first month or so as an active contributor on LessWrong, after lurking for over a decade. My experience so far has been overwhelmingly positive, and one purpose of this post is to encourage other lurkers to do the same.
The reason I decided to s... | https://www.lesswrong.com/posts/Rqok2cFnjYrLiFdst/a-decade-of-lurking-a-month-of-posting |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.