text stringlengths 300 320k | source stringlengths 52 154 |
|---|---|
# D&D.Sci December 2022 Evaluation and Ruleset
This is a followup to [the D&D.Sci post](https://www.lesswrong.com/posts/v3KJBXye83KYuX6er/d-and-d-sci-december-2022-the-boojumologist) I made ten days ago; if you haven’t already read it, you should do so now before spoiling yourself.
[Here](https://h-b-p.github.io/d-an... | https://www.lesswrong.com/posts/8h5yhnSD9biPxdJMi/d-and-d-sci-december-2022-evaluation-and-ruleset |
# Concept extrapolation for hypothesis generation
*[Posted initially](https://www.aligned-ai.com/post/concept-extrapolation-for-hypothesis-generation) on the [Aligned AI website](https://www.aligned-ai.com/). Authored by Patrick Leask, Stuart Armstrong, and Rebecca Gorman.*
There’s an [apocryphal story](https://www.g... | https://www.lesswrong.com/posts/jjiDARiv4hybZXeXL/concept-extrapolation-for-hypothesis-generation |
# 12 career-related questions that may (or may not) be helpful for people interested in alignment research
*Epistemic status: Some people tell me that these kinds of questions are helpful sometimes*.
At EAGxBerkeley, people interested in alignment research often asked me for career advice. I am not a technical align... | https://www.lesswrong.com/posts/Q37Ay82dfb3wnKjTr/12-career-related-questions-that-may-or-may-not-be-helpful |
# An exploration of GPT-2's embedding weights
*I wrote this doc in December 2021, while working at Redwood Research. It summarizes a handful of observations about GPT-2-small's weights -- mostly the embedding matrix, but also the LayerNorm gain parameters -- that I found while doing some open-ended investigation of th... | https://www.lesswrong.com/posts/BMghmAxYxeSdAteDc/an-exploration-of-gpt-2-s-embedding-weights |
# Revisiting algorithmic progress
How much progress in ML depends on algorithmic progress, scaling compute, or scaling relevant datasets is relatively poorly understood. [In our paper](https://arxiv.org/abs/2212.05153), we make progress on this question by investigating algorithmic progress in image classification on ... | https://www.lesswrong.com/posts/Cvr723R7DmKcLiAzQ/revisiting-algorithmic-progress |
# Alignment with argument-networks and assessment-predictions
|  Here is a simplistic diag... | https://www.lesswrong.com/posts/gvm8WvLmazj2NdTPk/alignment-with-argument-networks-and-assessment-predictions |
# Take 10: Fine-tuning with RLHF is aesthetically unsatisfying.
*As a writing exercise, I'm writing an AI Alignment Hot Take Advent Calendar - one new hot take, written every day for 25 days. Or until I run out of hot takes. This take owes a lot to the Simulators discussion group.*
Fine-tuning a large sequence model ... | https://www.lesswrong.com/posts/QujNmRy3uFyrkfqb7/take-10-fine-tuning-with-rlhf-is-aesthetically-unsatisfying |
# AI alignment is distinct from its near-term applications
I work on AI alignment, [by which I mean the technical problem of building AI systems that are trying to do what their designer wants them to do](https://ai-alignment.com/clarifying-ai-alignment-cec47cd69dd6).
There are many different reasons that someone cou... | https://www.lesswrong.com/posts/Hw26MrLuhGWH7kBLm/ai-alignment-is-distinct-from-its-near-term-applications |
# Last day to nominate things for the Review. Also, 2019 books still exist.
Two quick reminders:
One: You have about one more day to nominate posts! Posts need at least two positive upvotes to make it to phase 2, so check out the [voting dashboard](https://www.lesswrong.com/reviewVoting/2021) to upvote things that s... | https://www.lesswrong.com/posts/dp3HHfpozpKvyWuHp/last-day-to-nominate-things-for-the-review-also-2019-books |
# Bay 2022 Solstice
Join us on Saturday, December 17th to celebrate humanity and reflect on our place in the universe. We'll sing songs, hear speeches, and connect with each other.
TICKETS:
Tickets are pay-what-you-want, and are available on [EventBrite](https://www.eventbrite.com/e/bay-area-winter-solstice-2022-ti... | https://www.lesswrong.com/events/zty9JsFjhLeHoqAbR/bay-2022-solstice |
# Okay, I feel it now
I’ve been coming to LessWrong for a while. I’ve read most of the arguments for how and why things might go wrong.
I’ve been keeping across most developments. I’ve been following alignment efforts. I’ve done some thinking about the challenges involved.
But now I *feel* it.
Spending time observi... | https://www.lesswrong.com/posts/zyudvnRruMQMxM3NE/okay-i-feel-it-now |
# Existential AI Safety is NOT separate from near-term applications
Stephen Casper, [scasper@mit.edu](mailto:scasper@mit.edu). Thanks to Alex Lintz and Daniel Dewey for feedback.
This is a reply but not an objection to a recent post from Paul Christiano titled [AI alignment is distinct from its near term application... | https://www.lesswrong.com/posts/yKzyCw5EjabyZRkbJ/existential-ai-safety-is-not-separate-from-near-term |
# Is the ChatGPT-simulated Linux virtual machine real?
Context: [Building A Virtual Machine inside ChatGPT](https://www.engraved.blog/building-a-virtual-machine-inside/)
It very much triggers my "roll to disbelieve" reflex, but people seem to be talking about it as if it's real and not *that* surprising. I'd try it ... | https://www.lesswrong.com/posts/u3SueTC44tgKFMMNs/is-the-chatgpt-simulated-linux-virtual-machine-real |
# [Interim research report] Taking features out of superposition with sparse autoencoders
*We're thankful for helpful comments from Trenton Bricken, Eric Winsor, Noa Nabeshima, and Sid Black. *
*This post is part of the work done at *[*Conjecture*](https://www.conjecture.dev/research)*.*
**TL;DR**: Recent result... | https://www.lesswrong.com/posts/z6QQJbtpkEAX3Aojj/interim-research-report-taking-features-out-of-superposition |
# It Takes Two Paracetamol?
*Note: Depending on where you live paracetamol is also known as acetaminophen, APAP and Tylenol. *
I have had a version of the following conversation a surprising number of times:
> "I have a headache."
>
> "Aah, that sucks. Did you take some paracetamol?"
>
> "Yes, but it still hurts."... | https://www.lesswrong.com/posts/6npSCaALdFaoeFx59/it-takes-two-paracetamol |
# Applications open for AGI Safety Fundamentals: Alignment Course
The [AGI Safety Fundamentals (AGISF)](https://www.agisafetyfundamentals.com/?utm_source=EA+Forum&utm_medium=launch+post&utm_campaign=AGISF): Alignment Course is designed to introduce the key ideas in AGI safety and alignment, and provide a space and sup... | https://www.lesswrong.com/posts/mXKjuquNC8ivpiKWz/applications-open-for-agi-safety-fundamentals-alignment-1 |
# Best introductory overviews of AGI safety?
I'm interested what people think are the **best overviews of AI risk for various types of people**. Below I've listed as many good overviews as I could find (excluding some drafts), splitting based on "good for a popular audience" and "good for AI researchers." I'd also lik... | https://www.lesswrong.com/posts/T98kdFL5bxBWSiE3N/best-introductory-overviews-of-agi-safety |
# How can one literally buy time (from x-risk) with money?
[This post](https://www.lesswrong.com/posts/bkpZHXMJx3dG5waA7/ways-to-buy-time) talks about types of projects that can buy us more time to work on AI alignment. But like, I am just definitely not going to pivot to working on any of those projects right now. Wh... | https://www.lesswrong.com/posts/wWgYe5tqXSn9JCy2f/how-can-one-literally-buy-time-from-x-risk-with-money |
# Trying to disambiguate different questions about whether RLHF is “good”
(A few of the words in this post were written by Ryan Greenblatt and Ajeya Cotra. Thanks to Oliver Habryka and Max Nadeau for particularly helpful comments.)
Sometimes people want to talk about whether RLHF is “a promising alignment strategy”, ... | https://www.lesswrong.com/posts/NG6FrXgmqPd5Wn3mh/trying-to-disambiguate-different-questions-about-whether |
# Proof as mere strong evidence
Proofs at first exude a strong aroma of certainty.
On closer inspection though, more subtle notes emerge. Despite the dreams of formalists, the overwhelming majority of proofs in use and circulation are not minute low-level logical derivations, but rather intuitive high-level explanati... | https://www.lesswrong.com/posts/NqEmPZDSBgEpF8oHG/proof-as-mere-strong-evidence |
# Beyond a better world
*As I take man’s last step from the surface, back home for some time to come — but we believe not too long into the future — I’d like to just say what I believe history will record: that America’s challenge of today has forged man’s destiny of tomorrow. And, as we leave the Moon at Taurus–Littr... | https://www.lesswrong.com/posts/ysDGYjmCXWmejhvja/beyond-a-better-world |
# COVID China Personal Advice (No mRNA vax, possible hospital overload, bug-chasing edition)
So, uh, things in China are getting a bit interesting.
There's a general consensus that the public COVID numbers no longer reflect reality. Baidu started hiding their case tracking display. Their last updates had weird incons... | https://www.lesswrong.com/posts/exznkk3Ha4matbAah/covid-china-personal-advice-no-mrna-vax-possible-hospital |
# Discovering Latent Knowledge in Language Models Without Supervision
> Existing techniques for training language models can be misaligned with the truth: if we train models with imitation learning, they may reproduce errors that humans make; if we train them to generate text that humans rate highly, they may output e... | https://www.lesswrong.com/posts/kCEjcu53EEiqBH4gN/discovering-latent-knowledge-in-language-models-without |
# Key Mostly Outward-Facing Facts From the Story of VaccinateCA
I had the privilege of getting an advance draft of [Patrick McKenzie’s very long story founding and running VaccinateCA](https://www.worksinprogress.co/issue/the-story-of-vaccinateca/), an organization dedicated to providing Americans information on where... | https://www.lesswrong.com/posts/tTXfS3H3gNeBqCJk5/key-mostly-outward-facing-facts-from-the-story-of |
# Extracting and Evaluating Causal Direction in LLMs' Activations
*This post was written by Fabien at SaferAI*[^w9a0hn8g8b]*. Simeon has prompted Fabien in relevant directions and has provided valuable feedback.*
*Thanks to Jean-Stanislas Denain, Alexandre Variengien, Charbel-Raphael Segerie, and Nicole Nohemi for pr... | https://www.lesswrong.com/posts/mtZEjDortt8E8Apjs/extracting-and-evaluating-causal-direction-in-llms |
# My AGI safety research—2022 review, ’23 plans
*“Our greatest fear should not be of failure, but of succeeding at something that doesn't really matter.” –DL Moody (allegedly)*
**The short version:** In this post I’m briefly summarizing how I spent my work-time in 2022, and what I’m planning for 2023.
* The first... | https://www.lesswrong.com/posts/qusBXzCpxijTudvBB/my-agi-safety-research-2022-review-23-plans |
# Chair Hanging Peg
Our house is now back up to full occupancy, which meant some things that had been hanging out in unused rooms needed better homes. Four of those things were [folding chairs](https://www.ikea.com/us/en/p/terje-folding-chair-black-00222440/), and now they are up and out of the way:
[ whether I wanted to have children eventually, I would've said yes, without hesitation. In recent years I've been telling myself: I don't know how likely these AI doom predictions are, but I'm going to fo... | https://www.lesswrong.com/posts/wx793JXieh97AXtg6/is-the-ai-timeline-too-short-to-have-children |
# «Boundaries», Part 3b: Alignment problems in terms of boundaries
AI alignment is a notoriously murky problem area, which I think can be elucidated by rethinking its foundations in terms of boundaries between systems, including soft boundaries and directional boundaries. As anticipated previously in this [«Boundari... | https://www.lesswrong.com/posts/SajYfrsoTHxiXPNtf/boundaries-part-3b-alignment-problems-in-terms-of-boundaries |
# AI Neorealism: a threat model & success criterion for existential safety
Threat Model
============
There are *many* ways for AI systems to cause a catastrophe from which Earth-originating life could never recover. All of the following seem plausible to me:
* **Misuse**: An AI system could help a human or group o... | https://www.lesswrong.com/posts/Wi4RAJCbh3qD9fynj/ai-neorealism-a-threat-model-and-success-criterion-for |
# Basic building blocks of dependent type theory
This post is the first in a sequence of posts meant to be an introduction to pragmatic type theory. “Type theory” here refers to *Martin-Löf dependent type theory* (MLTT), as opposed to theories like the simply-typed lambda calculus; the main difference being that in ML... | https://www.lesswrong.com/posts/ccbsYSpTcTqXwukH8/basic-building-blocks-of-dependent-type-theory |
# The next decades might be wild
*I’d like to thank Simon Grimm and Tamay Besiroglu for feedback and discussions.*
**Update (early April 2023):** I now think the timelines in this post are too long and expect the world to get crazy faster than described here. For example, I expect many of the things suggested for 203... | https://www.lesswrong.com/posts/qRtD4WqKRYEtT5pi3/the-next-decades-might-be-wild |
# Covid 12/15/22: China’s Wave Begins
Cases increased this week, as expected. Deaths _also_ increased, which I did not expect so soon, indicating a bigger and faster winter wave. That was not the big event in the West this week with respect to Covid.
The big event this week in the West’s Covid experience was a memory... | https://www.lesswrong.com/posts/hCHPsgLMuNESpTeQt/covid-12-15-22-china-s-wave-begins |
# High-level hopes for AI alignment
_Click lower right to download or find on Apple Podcasts, Spotify, Stitcher, etc._
In previous pieces, I argued that there's a real and large risk of AI systems' ... | https://www.lesswrong.com/posts/7BWmLhFtqzqEPs8d5/high-level-hopes-for-ai-alignment |
# How "Discovering Latent Knowledge in Language Models Without Supervision" Fits Into a Broader Alignment Scheme
Introduction
------------
A few collaborators and I recently released a new paper: [Discovering Latent Knowledge in Language Models Without Supervision](https://arxiv.org/abs/2212.03827). For a quick summa... | https://www.lesswrong.com/posts/L4anhrxjv8j2yRKKp/how-discovering-latent-knowledge-in-language-models-without |
# How is ARC planning to use ELK?
Let's say we arrive at a worst-case solution for ELK, how are we planning to use it? My initial guess was that ELK is meant to help make IDA viable so that we may be able to use it for some automated alignment-type approach. However, this might not be it. Can someone clarify this? Tha... | https://www.lesswrong.com/posts/tz4ZGANPxADmmt8hS/how-is-arc-planning-to-use-elk |
# Consider working more hours and taking more stimulants
Epistemic status: amphetamines
> Sooner or later, the great men turn out to be all alike. They never stop working. They never lose a minute. It is very depressing.
>
> — V.S. Pritchett
EA produces much talk about an obligation to donate some or even most of y... | https://www.lesswrong.com/posts/36jdYNJRR9DyMRWDN/consider-working-more-hours-and-taking-more-stimulants |
# Masking to Avoid Missing Things
About a year ago I [wrote](https://www.jefftk.com/p/personal-response-to-omicron) that I'd now:
> follow official guidance and mask regulations, cheerfully go along with precautions others need, and test+isolate when sick, but I'm not going to go above and beyond to attempt to reduce... | https://www.lesswrong.com/posts/SjyDXu36pBFqh2mxM/masking-to-avoid-missing-things |
# Will Machines Ever Rule the World? MLAISU W50
Watch this week's episode on [YouTube](https://youtu.be/saiqSZkDBSk) or listen to the audio version [here](https://share.transistor.fm/s/44069318).
Hopes and fears of the current AI safety paradigm, GPU performance predictions and popular literature on why machines will... | https://www.lesswrong.com/posts/wAekjDysTiTivfKBe/will-machines-ever-rule-the-world-mlaisu-w50 |
# Paper: Transformers learn in-context by gradient descent
The paper argues that auto-regressive transformers implement in-context learning via gradient-based optimization on in-context data.
The authors... | https://www.lesswrong.com/posts/firtXAWGdvzXYAh9B/paper-transformers-learn-in-context-by-gradient-descent |
# Introducing Shrubgrazer
Since starting using Mastodon a month ago, a few different things haven't been a good fit for how I like to use social media:
* I'm relatively free with following people, but there are some people whose posts I don't want to miss. I'd like these to go [at the top of my feed](https://www.je... | https://www.lesswrong.com/posts/Doj5zwgdTouuTGNiA/introducing-shrubgrazer |
# How important are accurate AI timelines for the optimal spending schedule on AI risk interventions?
Summary
-------
I present an extension to my [*optimal timing of spending on AGI safety*](https://forum.effectivealtruism.org/posts/Ne8ZS6iJJp7EpzztP/the-optimal-timing-of-spending-on-agi-safety-work-why-we#Explainin... | https://www.lesswrong.com/posts/rq9BnzvnXjn8srZrv/how-important-are-accurate-ai-timelines-for-the-optimal |
# College Selection Advice for Technical Alignment
Hi all, I'm a highschool senior trying to make some college-related decisions, and I'd like to ask for some advice.
My current situation is:
* I want to work on technical alignment. For exogenous reasons, not going to college (e.g., taking a year off, just being a... | https://www.lesswrong.com/posts/4fSn7Jrbdwmp3GhLg/college-selection-advice-for-technical-alignment |
# Proper scoring rules don’t guarantee predicting fixed points
*Johannes Treutlein and Rubi Hudson worked on this post while participating in* [*SERI MATS*](https://www.serimats.org/)*, under Evan Hubinger's and Leo Gao's mentorship respectively. We are grateful to Marius Hobbahn, Erik Jenner, and Adam Jermyn for usef... | https://www.lesswrong.com/posts/Aufg88v7mQ2RuEXkS/proper-scoring-rules-don-t-guarantee-predicting-fixed-points |
# Can we efficiently explain model behaviors?
ARC’s [current plan](https://ai-alignment.com/mechanistic-anomaly-detection-and-elk-fb84f4c6d0dc) for solving ELK (and maybe also deceptive alignment) involves three major challenges:
1. Formalizing probabilistic heuristic argument as an operationalization of “explanatio... | https://www.lesswrong.com/posts/dQvxMZkfgqGitWdkb/can-we-efficiently-explain-model-behaviors |
# Beyond the moment of invention
Derek Thompson has a feature in *The Atlantic* this week on “[Why the Age of American Progress Ended](https://www.theatlantic.com/magazine/archive/2023/01/science-technology-vaccine-invention-history/672227/)”—thoughtful and worth reading. Some comments and reactions.
One of the ideas... | https://www.lesswrong.com/posts/XmENzEFmgbmHHFmKy/beyond-the-moment-of-invention |
# Vaguely interested in Effective Altruism? Please Take the Official 2022 EA Survey
**The 2022 Effective Altruism Survey is now live at the following link:** [**https://rethinkpriorities.qualtrics.com/jfe/form/SV_1NfgYhwzvlNGUom?source=lw**](https://rethinkpriorities.qualtrics.com/jfe/form/SV_1NfgYhwzvlNGUom?source=lw... | https://www.lesswrong.com/posts/7sMk5wDdwCpZBzJ3D/vaguely-interested-in-effective-altruism-please-take-the |
# Paper: Constitutional AI: Harmlessness from AI Feedback (Anthropic)
The authors propose a method for training a harmless AI assistant that can supervise other AIs, using only a list of rules (a "constitution") as human oversight. The method involves two phases: first, the AI improves itself by generating and revisin... | https://www.lesswrong.com/posts/aLhLGns2BSun3EzXB/paper-constitutional-ai-harmlessness-from-ai-feedback |
# Using Information Theory to tackle AI Alignment: A Practical Approach
\[Edited\]: I have removed the experiments after realizing that I had implement the algorithm incorrectly. I have replaced the experiment with different research papers that use similar objectives.
**Introduction**
================
I have been w... | https://www.lesswrong.com/posts/mdWByqAjT6p3rg9o8/using-information-theory-to-tackle-ai-alignment-a-practical |
# "Starry Night" Solstice Cookies
There are no traditional aspiring-rationalist foods that I know of. Here's a potential candidate for one. I started with a friend's family recipe and adjusted it to be v... | https://www.lesswrong.com/posts/eubXuy634qqBvsGKz/starry-night-solstice-cookies |
# [Link] Escape the Echo Chamber (2018)
Related to: [Epistemic Learned Helplessness](https://slatestarcodex.com/2019/06/03/repost-epistemic-learned-helplessness/), [Bounded Distrust](https://astralcodexten.substack.com/p/bounded-distrust)
This seems like a particularly insightful essay that expresses very well an ide... | https://www.lesswrong.com/posts/DFgaDCpKWxhuLWWbX/link-escape-the-echo-chamber-2018 |
# Predictive Processing, Heterosexuality and Delusions of Grandeur
**Predictive processing** is the theory that biological neurons minimize free energy. In this context, free energy isn't physical energy like the energy in your laptop battery. Instead, free energy is an informatic concept. It is useful to think about ... | https://www.lesswrong.com/posts/hL5pDhEjw8PkjzvJJ/predictive-processing-heterosexuality-and-delusions-of |
# Ordinary human life
I woke up early and read a book. At 7 I turned on the radio. A bit later, I got up, fed and walked the dog, playing a game on my smartphone. I skinned a mango for breakfast and started work. A friend came round for lunch and we talked about a paper we’re working on. Tonight would normally be choi... | https://www.lesswrong.com/posts/ozKY6pX8Fzdeq6gv5/ordinary-human-life |
# How to Convince my Son that Drugs are Bad
Hello.
My son (16m, henceforth referred to as John) has monologued about this site a few times over the past couple of months, so I figured, based on my brief impression of the community, you might be able to help me with an issue. Given the topical nature here, I am not su... | https://www.lesswrong.com/posts/ubPAo3zGeJNqtZDqT/how-to-convince-my-son-that-drugs-are-bad |
# Looking for an alignment tutor
Hey, this is [me](https://oatml.cs.ox.ac.uk//members/jan_brauner/). I’d like to understand AI X-risk better. Is anyone interested in being my “alignment tutor”, for maybe 1 h per week, or 1 h every two weeks? I’m happy to pay.
Fields I want to understand better:
* Anything related ... | https://www.lesswrong.com/posts/3us74zNBGgFJeAXTo/looking-for-an-alignment-tutor |
# Why write more: improve your epistemics, self-care, & 28 other reasons
Writing is like exercise. It’s just good on so many levels.
* Writing can be high impact
* Writing can be inspiring
* Writing can help you learn and grow
* Writing can help your career
* Writing can be meaningful and deeply satisfying... | https://www.lesswrong.com/posts/4LJHPRNMESePFY8fm/why-write-more-improve-your-epistemics-self-care-and-28 |
# Positive values seem more robust and lasting than prohibitions
*Originally* [*a shortform comment*](https://www.lesswrong.com/posts/dqSwccGTWyBgxrR58/turntrout-s-shortform-feed?commentId=cMs3hAyfsdLr7YRv3)*.*
Imagine we train an AI on realistic situations where it can kill people, and penalize it when it does so. S... | https://www.lesswrong.com/posts/cHnQ4bBFr3cX6rBxh/positive-values-seem-more-robust-and-lasting-than |
# Bad at Arithmetic, Promising at Math
$n$-Cohesive Rings
------------------
***Definition**: Let *$n$* be a positive integer. We define an*$n$***-cohesive ring** to be a commutative ring *$S$* such that, for every prime *$$$p$* dividing the characteristic of *$S$*, *$p^n$* divides the order of the multiplicative gro... | https://www.lesswrong.com/posts/qy5dF7bQcFjSKaW58/bad-at-arithmetic-promising-at-math |
# Take 11: "Aligning language models" should be weirder.
*As a writing exercise, I'm writing an AI Alignment Hot Take Advent Calendar - one new hot take, written ~~every day~~ some days for 25 days. I have now procrastinated enough that I probably have enough hot takes.*
People often talk about aligning language mode... | https://www.lesswrong.com/posts/tYbusKv4Yci3GaBiM/take-11-aligning-language-models-should-be-weirder |
# Boston Solstice 2022 Retrospective
Last night was our first full in-person Boston Secular Solstice since [2019](https://www.jefftk.com/p/boston-solstice-2019-retrospective), and I was back in the "music director and host" role while Taymon did most of the organizing. Overall I feel like it went pretty well! People h... | https://www.lesswrong.com/posts/Ds6SZALnsqQp3rWdN/boston-solstice-2022-retrospective |
# Why Are Women Hot?
Cross-posted, as always, from [Putanumonit](https://putanumonit.com/2022/12/13/why-are-women-hot/).
* * *
Why are women hot?
*They just are* is not an unreasonable answer, but it’s not like women are hot entirely by *coincidence*. Women put a lot of effort into their appearance. They care a gre... | https://www.lesswrong.com/posts/3iYiXDvTa2BnrxNL3/why-are-women-hot |
# Event [Berkeley]: Alignment Collaborator Speed-Meeting
If you’re working in alignment research, or would like to be, and you would like help finding researchers with whom to collaborate, this event is for you.
Inspired by the structure of a speed-dating event, participants will spend the first two hours in short on... | https://www.lesswrong.com/posts/hQZcoSBeAHLSzxhYi/event-berkeley-alignment-collaborator-speed-meeting |
# The Risk of Orbital Debris and One (Cheap) Way to Mitigate It
I worked as a researcher on orbital debris during university and now work full time on this and adjacent problems. My full write-up on the approach for using cheap radar data to assess orbital debris is [here on my blog.](https://locationtbd.home.blog/202... | https://www.lesswrong.com/posts/JpvxqmjFDLXg3hkBn/the-risk-of-orbital-debris-and-one-cheap-way-to-mitigate-it |
# The True Spirit of Solstice?
A year ago at the Winter Solstice afterparty, I had a 5 hour discussion with various folk about what things are or aren't appropriate for Solstice. Some example debates:
* "Is Solstice primarily a rationality holiday? An EA holiday? The broader secular community?"
* "How essential ... | https://www.lesswrong.com/posts/K5kL7jttaGDEnnNEK/the-true-spirit-of-solstice |
# An Effective Grab Bag
Preface
=======
I've been exploring low cost ways to increase resilience in the AI safety community as part of my work on [Alignment Continuity](https://alignmentcontinuity.org/); grab bags are one such intervention.
Since the Russia-Ukraine war kicked off, I've had a few EAs ask me about th... | https://www.lesswrong.com/posts/dZrqheHkQqRMoxbYk/an-effective-grab-bag |
# CEA Disambiguation
In 2011 effective altruists in Oxford had two main organizations: [Giving What We Can](https://www.givingwhatwecan.org/) and [80,000 Hours](https://80000hours.org/). They wanted to incorporate, and created the [Centre for Effective Altruism](https://www.centreforeffectivealtruism.org/) as an umbre... | https://www.lesswrong.com/posts/2KupwzZEjeoTke8rN/cea-disambiguation |
# Next Level Seinfeld
Thanks to [Jordan Schneider](https://twitter.com/jordanschnyc) gifting a last minute ticket for an amazing seat, serendipity led me to go to Friday night’s performance by Jerry Seinfeld at the Beacon Theater.
If I had to describe the show in one word, I’d say it was _unsurprising._ Jerry was Jer... | https://www.lesswrong.com/posts/fh5zACqJBbQ8R4tk7/next-level-seinfeld |
# Conditions for Superrationality-motivated Cooperation in a one-shot Prisoner's Dilemma
Summary
=======
It has been argued that, if two very similar agents follow decision theories allowing for [superrationality](https://www.lesswrong.com/tag/superrationality) (e.g., [EDT](https://www.lesswrong.com/tag/evidential-de... | https://www.lesswrong.com/posts/HLXiJgqxuMpwamdar/conditions-for-superrationality-motivated-cooperation-in-a |
# Does ChatGPT’s performance warrant working on a tutor for children? [It’s time to take it to the lab.]
Of course we already have sophisticated robot toys and companion robots for, e.g. older people, but I have no direct experience with any of these. Tutoring is quite different.
Here’s a dialog I had yesterday with ... | https://www.lesswrong.com/posts/Xg6gvec97KmFcxJ33/does-chatgpt-s-performance-warrant-working-on-a-tutor-for |
# Results from a survey on tool use and workflows in alignment research
In March 22nd, 2022, we released a [survey](https://forms.office.com/pages/responsepage.aspx?id=DQSIkWdsW0yxEjajBLZtrQAAAAAAAAAAAAN__kY0foxUQklQUktaRlRaR0hKQk1ZOE1GRjRVVllUUi4u) with an accompanying [post](https://www.alignmentforum.org/posts/ebYi... | https://www.lesswrong.com/posts/a2io2mcxTWS4mxodF/results-from-a-survey-on-tool-use-and-workflows-in-alignment |
# Computational signatures of psychopathy
I think that among the worst imaginable AI development outcomes would be if humans were to build artificially intelligent systems that exhibited psychopathic behavior. Though this sort of concern might appear to some as anthropomorphizing or otherwise far-fetched, I hope to de... | https://www.lesswrong.com/posts/NGA3F85ZxECabFfNo/computational-signatures-of-psychopathy |
# Towards Hodge-podge Alignment
*ABSTRACT: What could AI alignment look like if we had 6000+ full-time researchers and software developers? I've spent a few days thinking about that question, and this report is my first-draft attempt to answer it. This report sketches a highly-scalable research agenda: studying the ho... | https://www.lesswrong.com/posts/YnGRBADQwpYRbuCbz/towards-hodge-podge-alignment-1 |
# AGI Timelines in Governance: Different Strategies for Different Timeframes
Summarization Table
===================
| Timelines | **Pre-2030** | **Post-2030** |
| --- | --- | --- |
| **Expectations** | AGI will be built by an organization that’s already trying to build it (85%) | Some governments will be in the race... | https://www.lesswrong.com/posts/6CjnFcsRHJesR9MEA/agi-timelines-in-governance-different-strategies-for |
# Shard Theory in Nine Theses: a Distillation and Critical Appraisal
**TL;DR:** [*Shard theory*](https://www.alignmentforum.org/tag/shard-theory)* is a new research program started by Quintin Pope and Alex Turner. Existing introductions tend to be relatively long winded and aimed at an introductory audience. Here, I o... | https://www.lesswrong.com/posts/8ccTZ9ZxpJrvnxt4F/shard-theory-in-nine-theses-a-distillation-and-critical |
# The "Minimal Latents" Approach to Natural Abstractions
Background: The Language-Learning Argument
------------------------------------------
How many functions are there from a 1 megabyte image to a yes/no answer to the question “does this image contain an apple?”.
Well, there are 8M bits in the image, so $2^{8000... | https://www.lesswrong.com/posts/N2JcFZ3LCCsnK2Fep/the-minimal-latents-approach-to-natural-abstractions |
# More notes from raising a late-talking kid
My previous post on the topic was [here](https://www.lesswrong.com/posts/WqjXPeALsLH6Dxfys/late-talking-kids-and-einstein-syndrome). These are just some quick notes and observations, in no particular order.
Professionals pathologizing normal kid behaviors
=================... | https://www.lesswrong.com/posts/txuzuDmM4re32inRa/more-notes-from-raising-a-late-talking-kid |
# Take 12: RLHF's use is evidence that orgs will jam RL at real-world problems.
*As a writing exercise, I'm writing an AI Alignment Hot Take Advent Calendar - one new hot take, written ~~every day~~ some days for 25 days. I have now procrastinated enough that I probably have enough hot takes.*
*I felt like writing th... | https://www.lesswrong.com/posts/QQMzxSJDgWkAhupi5/take-12-rlhf-s-use-is-evidence-that-orgs-will-jam-rl-at-real |
# Notice when you stop reading right before you understand
Here is a bad habit I noticed in myself.
* I encounter something I want to learn about. For example, let's pick transformers.
* I scroll through the tweet/paper/article and get some sort of general gist. *"Okay, there's positional encoding, and an encod... | https://www.lesswrong.com/posts/CvxC8NpKPh3YWDhSn/notice-when-you-stop-reading-right-before-you-understand |
# [Fiction] Unspoken Stone
*Author's note: I got the seed of an idea for a story. As an exercise, I wrote it down. I have no expectation of returning to this; I have* [*a larger writing project*](https://www.lesswrong.com/s/HMs2yT9D6LjYR5jQT) *I care about more. But, it was fun to write, so maybe you'll find it fun to... | https://www.lesswrong.com/posts/JMjnTG4qYsqZ9FRhi/fiction-unspoken-stone |
# EA & LW Forums Weekly Summary (12th Dec - 18th Dec 22')
*Supported by Rethink Priorities*
This is part of a weekly series summarizing the top posts on the EA and LW forums - you can see the full collection [here.](https://forum.effectivealtruism.org/s/W4fhpuN26naxGCBbN) The first post includes some details on p... | https://www.lesswrong.com/posts/vkQPvedvk3NXYjCFD/ea-and-lw-forums-weekly-summary-12th-dec-18th-dec-22 |
# Under-Appreciated Ways to Use Flashcards - Part I
By Tereza Ruzickova
Those who know me well sometimes call me a “flashcard freak,” - and they are not wrong.
I first discovered flashcards at the end of high school, as I was getting frustrated with my own forgetting curve. I learned so many interesting facts as a s... | https://www.lesswrong.com/posts/BwZiNv8YuionaGAh5/under-appreciated-ways-to-use-flashcards-part-i |
# An Open Agency Architecture for Safe Transformative AI
*Edited to add (2024-03): This early draft is largely outdated by my ARIA programme thesis,* [*Safeguarded AI*](https://www.aria.org.uk/wp-content/uploads/2024/01/ARIA-Safeguarded-AI-Programme-Thesis-V1.pdf)*. I, davidad, am no longer using "OAA" as a proper nou... | https://www.lesswrong.com/posts/pKSmEkSQJsCSTK6nH/an-open-agency-architecture-for-safe-transformative-ai |
# (Extremely) Naive Gradient Hacking Doesn't Work
Tldr
====
A super simple form of gradient hacking might be to perform badly on any change to the value of the model's mesa-objective, in order to preserve the mesa-objective. I use some simple properties of SGD to explain why this method of gradient hacking cannot wor... | https://www.lesswrong.com/posts/Nnb5AqcunBwAZ4zac/extremely-naive-gradient-hacking-doesn-t-work |
# Marvel Snap: Phase 2
\[NOTE: This is a pure gaming post, if you’re not into that please do skip it.\]
Previously: [Marvel Snap Review: Phase 1](https://thezvi.substack.com/p/marvel-snap-review-phase-1)
At the end of my Phase 1 review, I removed Marvel Snap from my phone.
I am _very_ happy with that decision.
I d... | https://www.lesswrong.com/posts/LyRXu98ALnGhypSat/marvel-snap-phase-2 |
# Note on algorithms with multiple trained components
**Example 1:** consider a [GAN](https://en.wikipedia.org/wiki/Generative_adversarial_network). There’s a generator and a discriminator. As an intuitive mnemonic, we can say
* The “purpose” of the generator is to trick the discriminator,
* The “purpose” of the ... | https://www.lesswrong.com/posts/hRuE66SXobGhrbwxR/note-on-algorithms-with-multiple-trained-components |
# Properties of current AIs and some predictions of the evolution of AI from the perspective of scale-free theories of agency and regulative development
In this essay, I try to convey a set of related ideas that emerged in my mind upon reading the following papers:
* "[Dream of Being: Solving AI Alignment Problems ... | https://www.lesswrong.com/posts/oSPhmfnMGgGrpe7ib/properties-of-current-ais-and-some-predictions-of-the |
# Discovering Language Model Behaviors with Model-Written Evaluations
“[Discovering Language Model Behaviors with Model-Written Evaluations](https://www.anthropic.com/model-written-evals.pdf)” is a new Anthropic paper by Ethan Perez et al. that I (Evan Hubinger) also collaborated on. I think the results in this paper ... | https://www.lesswrong.com/posts/yRAo2KEGWenKYZG9K/discovering-language-model-behaviors-with-model-written |
# Podcast: Tamera Lanham on AI risk, threat models, alignment proposals, externalized reasoning oversight, and working at Anthropic
***TLDR:** I interviewed Tamera Lanham. You can listen to our conversation* [*here*](https://open.spotify.com/episode/4Twhw4Dw9hwYB6T27fDFwI?si=ZGpqtDWDR9-NHHtE7nkDCQ&nd=1) *or read some ... | https://www.lesswrong.com/posts/m92aaNdkiKFwmLmBo/podcast-tamera-lanham-on-ai-risk-threat-models-alignment |
# Progress links and tweets, 2022-12-20
I’ve been traveling, so this one is covering the last few weeks.
**The Progress Forum**
----------------------
* [Patrick McKenzie AMA](https://progressforum.org/posts/ManARnErMRFEvzffo/patrick-mckenzie-ama). See also his [retrospective on VaccinateCA](https://www.worksinpro... | https://www.lesswrong.com/posts/JCinZQhsHYiq4axDe/progress-links-and-tweets-2022-12-20 |
# [DISC] Are Values Robust?
Epistemic Status
----------------
[Discussion question](https://www.lesswrong.com/posts/zhhYwM7gk8LZsDzxj/dragongod-s-shortform?commentId=vYqEZxYZwEnsiSKJ6).
Related Posts
-------------
See also:
* [Complexity of Value](https://www.lesswrong.com/tag/complexity-of-value)
* [Value is ... | https://www.lesswrong.com/posts/YoFLKyTJ7o4ApcKXR/disc-are-values-robust |
# CIRL Corrigibility is Fragile
$$**Tl;dr:** An optimal CIRL agent is corrigible under a certain problem formulation and given certain assumptions. By examining three key assumptions—that the game is played only once, that the human is rational, and that human feedback is free—we demonstrate how minor changes can lead... | https://www.lesswrong.com/posts/PGK3AJtNG4rPHuZxy/cirl-corrigibility-is-fragile |
# Logging Shell History in Zsh
By default most shells don't log your history in a detailed durable way, which gives up one of the big advantages of working on the command line. Good history lets you look back at things you did months or years ago, in a searchable and skimmable fashion, so you can answer questions like... | https://www.lesswrong.com/posts/6WGyKKQnfaGa3YtMn/logging-shell-history-in-zsh |
# Local Memes Against Geometric Rationality
[Earlier](https://www.lesswrong.com/posts/DMxe4XKXnjyMEAAGw/the-geometric-expectation) in the [sequence](https://www.lesswrong.com/s/4hmf7rdfuXDJkxhfg), I presented the claim that humans are evolved to be naturally inclined towards geometric rationality over arithmetic ratio... | https://www.lesswrong.com/posts/EJ2teayyxzGA4MmxG/local-memes-against-geometric-rationality |
# Sazen
Purpose of post: describe and (hopefully) popularize a concept I've found highly useful.
* * *
Last year, my partner Logan Strohl wrote a [sequence](https://www.lesswrong.com/s/evLkoqsbi79AnM5sz) to introduce the "naturalism" concept they've been developing and teaching for the past decade or so.
That seque... | https://www.lesswrong.com/posts/k9dsbn8LZ6tTesDS3/sazen |
# Google Search loses to ChatGPT fair and square
I mentioned before that Google search is less than satisfactory even on uncomplicated queries: [https://www.lesswrong.com/posts/naAs59xiGfr7fPjej/google-search-as-a-washed-up-service-dog-i-halp](https://www.lesswrong.com/posts/naAs59xiGfr7fPjej/google-search-as-a-washed... | https://www.lesswrong.com/posts/xdswLKp4cW5EwgMvd/google-search-loses-to-chatgpt-fair-and-square |
# A Comprehensive Mechanistic Interpretability Explainer & Glossary
*This is a linkpost for a [very long doc](https://neelnanda.io/glossary) defining, explaining, and giving intuitions and conceptual frameworks for all the concepts I think you should know about when engaging with mechanistic interpretability. If you f... | https://www.lesswrong.com/posts/vnocLyeWXcAxtdDnP/a-comprehensive-mechanistic-interpretability-explainer-and |
# Price's equation for neural networks
Price's equation is a fundamental equation in genetics, which can be used to predict how traits will change due to evolution. It can be phrased in many ways, but for the current post I will use the following simplified continuous-time variant:
$\frac{d}{dt} x = gcov(x, f) = (\na... | https://www.lesswrong.com/posts/f5g5rH549rN7bGEAP/price-s-equation-for-neural-networks |
# You become the UI you use
Epistemic status: exploration
*
People instinctively copy other people they interact with. But the communication tools we use determine what we see and don't see (what we can copy and what we cannot copy) and often also who we interact with (whom we copy).
Written communication is a filt... | https://www.lesswrong.com/posts/8XmiWyd3A5qaCCGTm/you-become-the-ui-you-use |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.