text stringlengths 300 320k | source stringlengths 52 154 |
|---|---|
# Somerville Porchfest Thoughts
This Saturday was Porchfest in Somerville, an annual festival where musicians around the city play on their porches and people wander around listening. As in the past few years Cecilia and I ( [Kingfisher](https://www.kingfisherband.com/)) played for contra dancing:
[ an improved LLM called GPT-4o.
From their website
> Today, GPT-4o is much better than any existing model at understanding and discussing the images you share. For example, you can now take a picture of a menu... | https://www.lesswrong.com/posts/zKjhwt6pPCBMEJ9KJ/gpt-4o-is-out |
# OpenAI releases GPT-4o, natively interfacing with text, voice and vision
* Until now ChatGPT dealt with audio through a pipeline of 3 models: audio transcription, then GPT-4, then text-to-speech. GPT-4o is apparently trained on text, voice and vision so that everything is done natively. You can now interrupt it mi... | https://www.lesswrong.com/posts/KtCgWunZqgbTnP94a/openai-releases-gpt-4o-natively-interfacing-with-text-voice |
# Environmentalism in the United States Is Unusually Partisan
*This is the first in a sequence of four posts taken from my recent report:* [*Why Did Environmentalism Become Partisan?*](http://aiimpacts.org/wp-content/uploads/2023/04/Why-Did-Environmentalism-Become-Partisan-1.pdf)
Introduction
------------
In the Uni... | https://www.lesswrong.com/posts/5nfTXn4LrxnTmBWsb/environmentalism-in-the-united-states-is-unusually-partisan |
# D&D.Sci Long War: Defender of Data-mocracy Evaluation & Ruleset
This is a follow-up to [last week's D&D.Sci scenario](https://www.lesswrong.com/posts/hveHKFcs3LZh2rJXh/d-and-d-sci-long-war-defender-of-data-mocracy): if you intend to play that, and haven't done so yet, you should do so now before spoiling yourself.
... | https://www.lesswrong.com/posts/JAmvLDQGr9wL8rEqQ/d-and-d-sci-long-war-defender-of-data-mocracy-evaluation-and |
# How to do conceptual research: Case study interview with Caspar Oesterheld
[Caspar Oesterheld](https://www.andrew.cmu.edu/user/coesterh/) came up with two of the most important concepts in my field of work: [Evidential Cooperation in Large Worlds](https://longtermrisk.org/multiverse-wide-cooperation-via-correlated-d... | https://www.lesswrong.com/posts/QEfy9Dqin7nEJ9Fbs/how-to-do-conceptual-research-case-study-interview-with |
# How much AI inference can we do?
Suppose you have a bunch of GPUs. How many LLM forward passes can you do with them?[^n189of3zsp]
This is relevant to figuring out how profitable AI will be in the short-term, how powerful AI systems might be able to come in the near future, how large the compute overhang will be and... | https://www.lesswrong.com/posts/g7H2sSGHAeYxCHzrz/how-much-ai-inference-can-we-do |
# A Narrative History of Environmentalism's Partisanship
*This is the second in a sequence of four posts taken from my recent report:* [*Why Did Environmentalism Become Partisan?*](http://aiimpacts.org/wp-content/uploads/2023/04/Why-Did-Environmentalism-Become-Partisan-1.pdf)
*Many of the specific claims made here ar... | https://www.lesswrong.com/posts/ZuKfs5Ank2zbgCcPG/a-narrative-history-of-environmentalism-s-partisanship |
# Teaching CS During Take-Off
I stayed up too late collecting way-past-deadline papers and writing report cards. When I woke up at 6, this anxious email from one of my g11 Computer Science students was already in my Inbox.
*Student:* Hello Mr. Carle, I hope you've slept well; I haven't.
I've been seeing a lot of new... | https://www.lesswrong.com/posts/D6nTSEdCcbGQKCfc2/teaching-cs-during-take-off |
# MIRI's May 2024 Newsletter
**Update (5-15-2024):** I wrote that “it appears that not all of the leading AI labs are honoring the voluntary agreements they made at \[AI Safety Summit\],” citing a [Politico article](https://www.politico.eu/article/rishi-sunak-ai-testing-tech-ai-safety-institute/). However, after seein... | https://www.lesswrong.com/posts/vLBW5wMxvRLZwA4Wo/miri-s-may-2024-newsletter |
# my note system
I've been told that my number of blog posts is impressive, but my personal notes are much larger than my blog, over a million words and with higher information density. Since I've had a bit of practice taking notes, I thought I'd describe the system I developed. It's more complex than some integrated... | https://www.lesswrong.com/posts/dD63tGi88KvgC8cmx/my-note-system |
# Ilya Sutskever and Jan Leike resign from OpenAI [updated]
Ilya Sutskever and Jan Leike have resigned. They led OpenAI's alignment work. Superalignment will now be led by John Schulman, it seems. Jakub Pachocki replaced Sutskever as Chief Scientist.
Reasons are unclear (as usual when safety people leave OpenAI).
Th... | https://www.lesswrong.com/posts/JSWF2ZLt6YahyAauE/ilya-sutskever-and-jan-leike-resign-from-openai-updated |
# Catastrophic Goodhart in RL with KL penalty
TLDR: In the [last](https://www.lesswrong.com/posts/fuSaKr6t6Zuh6GKaQ/when-is-goodhart-catastrophic) [two](https://www.lesswrong.com/posts/GdkixRevWpEanYgou/catastrophic-regressional-goodhart-appendix) posts, we showed that optimizing for a proxy can fail to increase true ... | https://www.lesswrong.com/posts/pEZoTSCxHY3mfPbHu/catastrophic-goodhart-in-rl-with-kl-penalty |
# Was Partisanship Good for the Environmental Movement?
*This is the third in a sequence of posts taken from my recent report:* [*Why Did Environmentalism Become Partisan?*](http://aiimpacts.org/wp-content/uploads/2023/04/Why-Did-Environmentalism-Become-Partisan-1.pdf)
Summary
-------
Rising partisanship did not mak... | https://www.lesswrong.com/posts/kFjkX6ve738bAvWCu/was-partisanship-good-for-the-environmental-movement-1 |
# [Linkpost] Please don't take Lumina's anticavity probiotic
Update:
Trevor Klee (author of the linked post) has [published an update](https://trevorklee.substack.com/p/luminas-legal-threats-and-my-about) in which he (arguably) moderates his view (or at least that which he expresses publicly). Specifically, he states... | https://www.lesswrong.com/posts/jajbBbSuJmHe2ZJ92/linkpost-please-don-t-take-lumina-s-anticavity-probiotic |
# How is GPT-4o Related to GPT-4?
GPT-4o both has a [new tokenizer](https://community.openai.com/t/whats-the-new-tokenization-algorithm-for-gpt-4o/746708) and was [trained directly on audio](https://openai.com/index/hello-gpt-4o/) (whereas my understanding is that GPT-4 was trained only on text and images). Is there p... | https://www.lesswrong.com/posts/cBgWimXiggbEq5q2f/how-is-gpt-4o-related-to-gpt-4 |
# Why I'll Keep My Crummy Drawings—How Generative AI Art Won't Supplant... Art.
The other day the post below caught my attention. In its characteristic bad-grammar-ambiguous-origin-meme style, Neo’s sentiment is one I share.
 or do-what-I... | https://www.lesswrong.com/posts/7NvKrqoQgJkZJmcuD/instruction-following-agi-is-easier-and-more-likely-than |
# Why you should learn a musical instrument
I have liked music very much since I was a teenager. I spent many hours late at night in Soulseek chat rooms talking about and sharing music with my online friends. So, I tend to just have some music floating around in my head on any given day. But, I never learned to play a... | https://www.lesswrong.com/posts/EHRXKxk2YMKa7oGaw/why-you-should-learn-a-musical-instrument |
# Let's Design A School, Part 2.3
School as Education - The Curriculum (Phase 2, Specific)
In the previous post, we outlined three *phases* that students would go through, where each student matriculated through them at their own speed.
Phase 1 was literacy and numeracy.
Phase 2 was core civilizational requirements ... | https://www.lesswrong.com/posts/DS7cNqrhXtC4xvQ5A/let-s-design-a-school-part-2-3-school-as-education-the |
# Taming Infinity (Stat Mech Part 3)
The canonical example of quantum mechanics in action is the harmonic oscillator, which is something like a mass on a spring. In classical mechanics, it wobbles back and forth periodically when it is given energy, if it's at a position $x$, wobbling about $x_0$ and moving with veloc... | https://www.lesswrong.com/posts/X2XMFPg6pTeGC3YFd/taming-infinity-stat-mech-part-3 |
# Introducing Statistical Utility Mechanics: A Framework for Utility Maximizers
*This is independent alignment research. A decent grasp on stat mech is needed to understand this post fully, but I have recently written a stat mech sequence (up to part 3) which should do the trick to catch someone up to speed rapidly.*
... | https://www.lesswrong.com/posts/2MTtKzbxeXDXTJhC2/introducing-statistical-utility-mechanics-a-framework-for |
# Tackling Moloch: How YouCongress Offers a Novel Coordination Mechanism
[Moloch](https://www.lesswrong.com/posts/TxcRbCYHaeL59aY7E/meditations-on-moloch), as articulated by Scott Alexander, represents the coordination problems that lead to outcomes that leave everyone worse off. While prediction markets explore what ... | https://www.lesswrong.com/posts/4KjiZeAWc7Yv9oyCb/tackling-moloch-how-youcongress-offers-a-novel-coordination |
# Do you believe in hundred dollar bills lying on the ground? Consider humming
Introduction
============
\[Reminder: I am an internet weirdo with no medical credentials\]
A few months ago, I published some [crude estimates](https://acesounderglass.com/2024/02/04/nitric-oxide-for-covid-and-other-viral-infections/) of... | https://www.lesswrong.com/posts/NBZvpcBx4ewqkdCdT/do-you-believe-in-hundred-dollar-bills-lying-on-the-ground-1 |
# How can I make the most of Less Online/Camp/Manifest?
I spent several weeks psyching myself into buying tickets to what is essentially my first 'vacation' in my adult life (I'm 27 and I typically dislike traveling and enjoy my job more than average).
I'm optimistic it will be enjoyable but social events are not so... | https://www.lesswrong.com/posts/PQiRgcuECS5w3fKZW/how-can-i-make-the-most-of-less-online-camp-manifest |
# Against "argument from overhang risk"
Epistemic status: I wrote this in August 2023, got some feedback I didn't manage to incorporate very well, and then never published it. There's been less discussion of overhang risk recently but I don't see any reason to keep sitting on it. Still broadly endorsed, though there... | https://www.lesswrong.com/posts/TAab4g9DKZdyFd8XH/against-argument-from-overhang-risk |
# A case for fairness-enforcing irrational behavior
There's a long-standing and possibly unsolvable puzzle about how AIs should behave in game-theoretic situations with each other. The simplest example is the Ultimatum Game, where player A proposes how a dollar should be split between A and B, and B either accepts or ... | https://www.lesswrong.com/posts/w2EAEsvL9zEPZtMqr/a-case-for-fairness-enforcing-irrational-behavior |
# The Dunning-Kruger of disproving Dunning-Kruger
In an online discussion elsewhere today someone linked [this article](https://www.psychologytoday.com/us/blog/how-do-you-know/202012/dunning-kruger-isnt-real) which in turn linked the paper Gignac & Zajenkowski, *The Dunning-Kruger effect is (mostly) a statistical arte... | https://www.lesswrong.com/posts/6Tqm8Jet9mzo6buj9/the-dunning-kruger-of-disproving-dunning-kruger |
# Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems
Authors: David "[davidad](https://www.lesswrong.com/users/davidad)" Dalrymple, Joar Skalse, [Yoshua Bengio](https://www.lesswrong.com/users/yoshua-bengio), Stuart Russell, [Max Tegmark](https://www.lesswrong.com/users/max-tegmark), S... | https://www.lesswrong.com/posts/wvgwYQv9B4jioqgqg/towards-guaranteed-safe-ai-a-framework-for-ensuring-robust |
# FMT: a great opportunity for (soon-to-be) parents
**Executive summary**
=====================
**Fecal Microbiota Transplant (FMT)** is a procedure that involves transferring** the stool of healthy people to the guts of unhealthy people**. The bacteria in the healthy person’s stool helps to rebalance the unhealthy p... | https://www.lesswrong.com/posts/jt47HsikDuBAYKhGS/fmt-a-great-opportunity-for-soon-to-be-parents-1 |
# AI #64: Feel the Mundane Utility
It’s happening. The race is on.
Google and OpenAI both premiered the early versions of their fully multimodal, eventually fully integrated AI agents. Soon your phone experience will get more and more tightly integrated with AI. You will talk to your phone, or your computer, and it w... | https://www.lesswrong.com/posts/29fswYuy6KB8Edbjm/ai-64-feel-the-mundane-utility |
# GPT-4o My and Google I/O Day
At least twice the speed! At most half the price!
That’s right, it’s GPT-4o My.
[Some people’s expectations](https://twitter.com/robinhanson/status/1789471660648669578) for the OpenAI announcement this week were very high.
> Spencer Schiff: Next week will likely be remembered as one o... | https://www.lesswrong.com/posts/bqa5wmrwPL5zbfgxH/gpt-4o-my-and-google-i-o-day |
# Ninety-five theses on AI
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:st... | https://www.lesswrong.com/posts/zMHifwvZB8pwcTZbx/ninety-five-theses-on-ai |
# Advice for Activists from the History of Environmentalism
*This is the fourth in a sequence of posts taken from my recent report:* [*Why Did Environmentalism Become Partisan?*](http://aiimpacts.org/wp-content/uploads/2023/04/Why-Did-Environmentalism-Become-Partisan-1.pdf)
*This post has more of my personal opinions... | https://www.lesswrong.com/posts/CD6gWDbgKftFW37gs/advice-for-activists-from-the-history-of-environmentalism-1 |
# Feeling (instrumentally) Rational
*Contra* [*this post*](https://www.lesswrong.com/posts/SqF8cHjJv43mvJJzx/feeling-rational) *from the Sequences*
In Eliezer's sequence post, he makes the following (excellent) point:
> I can’t find any theorem of probability theory which proves that I should appear ice-cold and exp... | https://www.lesswrong.com/posts/yeezBxTLaEcDorYLr/feeling-instrumentally-rational |
# To an LLM, everything looks like a logic puzzle
I keep seeing this meme doing the rounds where people present ChatGPT with a common logic problem or riddle, only with some key component changed to make it trivial. ChatGPT has seen the original version a million times, so it gives the answer to the original, not the ... | https://www.lesswrong.com/posts/qNXXe7EDGyveC4SCp/to-an-llm-everything-looks-like-a-logic-puzzle |
# D&D.Sci (Easy Mode): On The Construction Of Impossible Structures
*This is a D&D.Sci scenario: a puzzle where players are given a dataset to analyze and an objective to pursue using information from that dataset.*
Duke Arado’s obsession with physics-defying architecture has caused him to run into a small problem. H... | https://www.lesswrong.com/posts/Syfq6MwgdZhHg9vha/d-and-d-sci-easy-mode-on-the-construction-of-impossible |
# My Hammer Time Final Exam
**Epistemic Status:** I thought about and wrote each paragraph in 10 minutes total, with slight editing afterwards.
I hope I'm not too late to the party!
I wrote this up quite a few months ago and found that I delayed indefinitely editing it before publication. I decided it's probably be... | https://www.lesswrong.com/posts/2jjRcBcFWk5rWcS4H/my-hammer-time-final-exam |
# Is There Really a Child Penalty in the Long Run?
A couple of weeks ago three European economists published [this paper](https://docs.iza.org/dp16959.pdf) studying the female income penalty after childbirth. The surprising headline result: there is no penalty.
### Setting and Methodology
The paper uses Danish data ... | https://www.lesswrong.com/posts/3FqgRqgadJ9EwyPBE/is-there-really-a-child-penalty-in-the-long-run |
# AISafety.com – Resources for AI Safety
There are many resources for those who wish to contribute to AI Safety, such as [courses](https://www.aisafety.com/courses), [communities](https://www.aisafety.com/communities), [projects](https://www.aisafety.com/projects), [jobs](https://www.aisafety.com/jobs), [events and tr... | https://www.lesswrong.com/posts/WNZGqeLMjPGFp78wX/aisafety-com-resources-for-ai-safety |
# Identifying Functionally Important Features with End-to-End Sparse Dictionary Learning
*A short summary of the paper is presented below.*
*This work was produced by Apollo Research in collaboration with Jordan Taylor (MATS + University of Queensland) .*
**TL;DR:** We propose end-to-end (e2e) sparse dictionary lear... | https://www.lesswrong.com/posts/xzJK3nENopiLmo77H/identifying-functionally-important-features-with-end-to-end |
# DeepMind: Frontier Safety Framework
DeepMind's RSP is here: [blogpost](https://deepmind.google/discover/blog/introducing-the-frontier-safety-framework/), [full document](https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/introducing-the-frontier-safety-framework/fsf-technical-report.pdf). Compare to [An... | https://www.lesswrong.com/posts/AFQt6uByLYNrNgyBb/deepmind-frontier-safety-framework |
# Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems
I want to draw attention to a [new paper](https://arxiv.org/abs/2405.06624), written by myself, David "davidad" Dalrymple, Yoshua Bengio, Stuart Russell, Max Tegmark, Sanjit Seshia, Steve Omohundro, Christian Szegedy, Ben Goldhaber, ... | https://www.lesswrong.com/posts/LkECxpbjvSifPfjnb/towards-guaranteed-safe-ai-a-framework-for-ensuring-robust-1 |
# Language Models Model Us
$$*Produced as part of the MATS Winter 2023-4 program, under the mentorship of* [*@Jessica Rumbelow*](https://www.alignmentforum.org/users/jessica-rumbelow?mention=user)
*One-sentence summary: On a dataset of human-written essays, we find that gpt-3.5-turbo can accurately infer demographic ... | https://www.lesswrong.com/posts/dLg7CyeTE4pqbbcnp/language-models-model-us |
# Goodhart in RL with KL: Appendix
This is the appendix to the [previous post](https://www.lesswrong.com/s/6rhjdbnEXoek4YiH7/p/pEZoTSCxHY3mfPbHu) on Goodhart’s Law and KL regularization, containing all of our proofs.
Theorem about distributions
---------------------------
***Theorem 1:*** Given any heavy-tailed refe... | https://www.lesswrong.com/posts/HBn95kqYq2nYKK5qT/goodhart-in-rl-with-kl-appendix |
# International Scientific Report on the Safety of Advanced AI: Key Information
I thought that the recently released [International Scientific Report on the Safety of Advanced AI](https://assets.publishing.service.gov.uk/media/66474eab4f29e1d07fadca3d/international_scientific_report_on_the_safety_of_advanced_ai_interi... | https://www.lesswrong.com/posts/md8DJ5smqjHdJs65Z/international-scientific-report-on-the-safety-of-advanced-ai |
# DeepMind's "Frontier Safety Framework" is weak and unambitious
[FSF blogpost](https://deepmind.google/discover/blog/introducing-the-frontier-safety-framework/). [Full document](https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/introducing-the-frontier-safety-framework/fsf-technical-report.pdf) (just ... | https://www.lesswrong.com/posts/y8eQjQaCamqdc842k/deepmind-s-frontier-safety-framework-is-weak-and-unambitious |
# "If we go extinct due to misaligned AI, at least nature will continue, right? ... right?"
* \[memetic status: stating directly despite it being a clear consequence of core AI risk knowledge because many people have "but nature will survive us" antibodies to other classes of doom and misapply them here.\]*
Unfortuna... | https://www.lesswrong.com/posts/6C3ndLd3nkrfy4K6j/if-we-go-extinct-due-to-misaligned-ai-at-least-nature-will |
# Scientific Notation Options
When working with numbers that span many orders of magnitude it's very helpful to use some form of [scientific notation](https://en.wikipedia.org/wiki/Scientific_notation). At its core, scientific notation expresses a number by breaking it down into a decimal ≥1 and <10 (the "significand"... | https://www.lesswrong.com/posts/Hpmc2hmakfzutXLWa/scientific-notation-options |
# To Limit Impact, Limit KL-Divergence
### TL;DR
Run a potentially-harmful model alongside a known-harmless model, such that their action-spaces (e.g. output token sets) are equivalent. Combine the output probabilities so as to limit the KL-divergence between the resulting token probabilities and the harmless model's... | https://www.lesswrong.com/posts/t8S8y3jbAGydfme3J/to-limit-impact-limit-kl-divergence |
# Fund me please - I Work so Hard that my Feet start Bleeding and I Need to Infiltrate University
*Thanks to Taylor Smith for doing some copy-editing this.*
*In this article, I tell some anecdotes and present some evidence in the form of research artifacts about how easy it is for me to work hard when I have collabor... | https://www.lesswrong.com/posts/2D74Ctr5Aj3Sb5f69/fund-me-please-i-work-so-hard-that-my-feet-start-bleeding |
# On Privilege
The forum has been very much focused on AI safety for some time now, thought I'd post something different for a change. Privilege.
Here I define Privilege as an **advantage over others that is invisible to the** ~~**be**~~**holder**. \[EDIT: thanks to JenniferRM for pointing out that "beholder" is a wr... | https://www.lesswrong.com/posts/YtDtJC7vdgyLiDCwB/on-privilege |
# Some "meta-cruxes" for AI x-risk debates
*\[Epistemic status: As I say below, I've been thinking about this topic for several years and I've worked on it as part of my PhD research. But none of this is based on any rigorous methodology, just my own impressions from reading the literature.\]*
I've been thinking abou... | https://www.lesswrong.com/posts/bXQjSaYH9NRsjPinS/some-meta-cruxes-for-ai-x-risk-debates |
# Hot take: The AI safety movement is way too sectarian and this is greatly increasing p(doom)
The movement to reduce AI x-risk is overly purist. This is leading to a lot of sects to maintain each individual sect's platonic level of purity and is actively (greatly) harming the cause.
How the Safety Sects Manifest... | https://www.lesswrong.com/posts/FK5ctN989MzADufoM/hot-take-the-ai-safety-movement-is-way-too-sectarian-and |
# The consistent guessing problem is easier than the halting problem
The halting problem is the problem of taking as input a Turing machine M, returning true if it halts, false if it doesn't halt. This is known to be uncomputable. The [consistent guessing problem](https://scottaaronson.blog/?p=710) (named by Scott Aar... | https://www.lesswrong.com/posts/8kghiWcnxpjhraDgE/the-consistent-guessing-problem-is-easier-than-the-halting |
# Anthropic: Reflections on our Responsible Scaling Policy
Last September we published our first [Responsible Scaling Policy (RSP)](https://www.anthropic.com/news/anthropics-responsible-scaling-policy) [[LW discussion](https://www.lesswrong.com/posts/6tjHf5ykvFqaNCErH/anthropic-s-responsible-scaling-policy-and-long-t... | https://www.lesswrong.com/posts/vAopGQhFPdjcA8CEh/anthropic-reflections-on-our-responsible-scaling-policy |
# Why I find Davidad's plan interesting
### Abstract
The aim of this text is to give an overview of [Davidad’s safety plan](https://www.lesswrong.com/posts/mnoc3cKY3gXMrTybs/a-list-of-core-ai-safety-problems-and-how-i-hope-to-solve), while also outlining some of the limitations and challenges. Additionally, I’ll exp... | https://www.lesswrong.com/posts/wQxjf67tnfJPG33ju/why-i-find-davidad-s-plan-interesting |
# D&D.Sci (Easy Mode): On The Construction Of Impossible Structures [Evaluation and Ruleset]
This is a followup to [the D&D.Sci post](https://www.lesswrong.com/posts/Syfq6MwgdZhHg9vha/d-and-d-sci-easy-mode-on-the-construction-of-impossible) I made last Friday; if you haven’t already read it, you should do so now befor... | https://www.lesswrong.com/posts/oGXmwzsDqKM9uP5dA/d-and-d-sci-easy-mode-on-the-construction-of-impossible-1 |
# Jaan Tallinn's 2023 Philanthropy Overview
to follow up my [philantropic pledge](https://www.lesswrong.com/posts/ZMRCpfGD4iXs5Mi5Q/jaan-tallinn-s-philanthropic-pledge) from 2020, i've updated my [philanthropy page](https://jaan.info/philanthropy/) with 2023 results.
in 2023 my donations funded [$44M worth of endpoin... | https://www.lesswrong.com/posts/bjqDQB92iBCahXTAj/jaan-tallinn-s-2023-philanthropy-overview |
# Infra-Bayesian haggling
Preface
=======
I wrote this post during my scholarship at [MATS](https://www.matsprogram.org/). My goal is to describe a research direction of the [learning theoretic agenda](https://www.lesswrong.com/posts/ZwshvqiqCvXPsZEct/the-learning-theoretic-agenda-status-2023) (LTA). Namely, a natura... | https://www.lesswrong.com/posts/gBgw7mbLNYv5qzDLG/infra-bayesian-haggling |
# OpenAI: Exodus
Previously: [OpenAI: Facts From a Weekend](https://thezvi.substack.com/p/openai-facts-from-a-weekend?utm_source=publication-search), [OpenAI: The Battle of the Board](https://thezvi.substack.com/p/openai-the-battle-of-the-board?utm_source=publication-search), [OpenAI: Leaks Confirm the Story](https://... | https://www.lesswrong.com/posts/ASzyQrpGQsj7Moijk/openai-exodus |
# NAO Updates, Spring 2024
Now that the [NAO blog](https://naobservatory.org/blog) is up, we’re taking the opportunity to post some written updates on the work our team has done over the past ~6 months. We’re hoping to make similar updates something like quarterly. Since this post covers a longer period it’s a bit lon... | https://www.lesswrong.com/posts/L7mKt4okQLWdv7mu5/nao-updates-spring-2024 |
# The Local Interaction Basis: Identifying Computationally-Relevant and Sparsely Interacting Features in Neural Networks
*This is a linkpost for our two recent papers:*
1. *An exploration of using degeneracy in the loss landscape for interpretability* [*https://arxiv.org/abs/2405.10927*](https://arxiv.org/abs/2405.1... | https://www.lesswrong.com/posts/bchjSwxBTxZBFXBXs/the-local-interaction-basis-identifying-computationally |
# Interpretability: Integrated Gradients is a decent attribution method
*A short post laying out our reasoning for using integrated gradients as attribution method. It is intended as a stand-alone post based on our LIB papers* [*\[1\]*](https://arxiv.org/abs/2405.10927) [*\[2\]*](https://arxiv.org/abs/2405.10928)*. T... | https://www.lesswrong.com/posts/Rv6ba3CMhZGZzNH7x/interpretability-integrated-gradients-is-a-decent |
# Some perspectives on the discipline of Physics
I wrote the linked post, and I’m posting a lightly edited version here for discussion. I plan to attend LessOnline, and this is my first attempt at blogging to understand and earnestly explain and is also gauging interest in the topic in case someone at LessOnline wants... | https://www.lesswrong.com/posts/6gMvyKuxZSECMyzah/some-perspectives-on-the-discipline-of-physics |
# [Linkpost] Statement from Scarlett Johansson on OpenAI's use of the "Sky" voice, that was shockingly similar to her own voice.
Scarlett Johansson makes a statement about the "Sky" voice, a voice for GPT-4o that OpenAI recently pulled after less than a week of prime time.
tl;dr: OpenAI made an offer last Septemb... | https://www.lesswrong.com/posts/qZGgLiyheoh8f7Cga/linkpost-statement-from-scarlett-johansson-on-openai-s-use |
# Harmony Intelligence is Hiring!
Hey folks! Pleased to announce we have a new open position for a Founding Research Engineer at Harmony Intelligence. You’ll be responsible for measuring and identifying dangerous AI capabilities across various domains: cybersecurity, biosecurity, persuasion and manipulation, self-exfi... | https://www.lesswrong.com/posts/PRjqTjzqwLnibxzFv/harmony-intelligence-is-hiring |
# What's Going on With OpenAI's Messaging?
*This is a quickly-written opinion piece, of what I understand about OpenAI. I first posted it to Facebook, where it had* [*some discussion*](https://www.facebook.com/ozzie.gooen/posts/pfbid02jMvzgzwCgbaK7dRZPYfgPAoBCy25CCvCUEwwxzh5DZuwCu8xuM88C95zUeoYMJp6l)*. *
**Some argum... | https://www.lesswrong.com/posts/cy99dCEiLyxDrMHBi/what-s-going-on-with-openai-s-messaging |
# The Problem With the Word ‘Alignment’
*This post was written by Peli Grietzer, inspired by internal writings by TJ (tushita jha), for AOI*[^ps56qyd7s5]*. The original post, published on Feb 5, 2024, can be found here:* [*https://ai.objectives.institute/blog/the-problem-with-alignment*](https://ai.objectives.institut... | https://www.lesswrong.com/posts/p3aL6BwpbPhqxnayL/the-problem-with-the-word-alignment-1 |
# Scorable Functions: A Format for Algorithmic Forecasting
You can think of a *scorable function* as a "*box*" that can produce large clusters of complex forecasts. Credit to Dall-E 3.
Introduction... | https://www.lesswrong.com/posts/otFDNWGN3zhNXXGrH/scorable-functions-a-format-for-algorithmic-forecasting |
# My Dating Heuristic
I don’t have to practice being afraid of a lion charging at me—my instincts tell me to run. But when I started dating, my instincts weren’t that reliable when attempting to attract a partner. They needed to be recalibrated. Author Matthew Hussey talks about retraining your (likely faulty) dating ... | https://www.lesswrong.com/posts/gmysPZ3t5Rz9nzcCr/my-dating-heuristic |
# New voluntary commitments (AI Seoul Summit)
Basically the companies commit to make [responsible scaling policies](https://metr.org/blog/2023-09-26-rsp/).
Part of me says this is amazing, the best possible commitment short of all committing to a specific RSP. It's certainly more real than almost all other possible k... | https://www.lesswrong.com/posts/qfEgzQ9jGEk9Cegvy/new-voluntary-commitments-ai-seoul-summit |
# Is deleting capabilities still a relevant research question?
I've had it suggested that a good criterion for whether interpretability is on the right track is if we can do surgical "deletions" of model capabilities, e.g. removing its ability to build bombs and such.
Obviously in one sense this is fairly trivial sin... | https://www.lesswrong.com/posts/HkcMLfaWmEc9HtSQp/is-deleting-capabilities-still-a-relevant-research-question |
# On Dwarkesh’s Podcast with OpenAI’s John Schulman
[Dwarkesh Patel recorded a Podcast with John Schulman](https://www.youtube.com/watch?v=Wo95ob_s_NI&ab_channel=DwarkeshPatel), cofounder of OpenAI and at the time their head of current model post-training. [Transcript here](https://www.dwarkeshpatel.com/p/john-schulma... | https://www.lesswrong.com/posts/rC6CXZd34geayEH4s/on-dwarkesh-s-podcast-with-openai-s-john-schulman |
# rough draft on what happens in the brain when you have an insight
*Epistemic status: It is better to be wrong than to have no model at all. I recommend the footnotes.*[^j7lj9ak3e3g]
🍵
On my current models of theoretical[^azvig7psxdh] insight-making, it looks something like this:
1. A gradual build-up and pr... | https://www.lesswrong.com/posts/jmqJvJ2SSLR4bvtCp/rough-draft-on-what-happens-in-the-brain-when-you-have-an |
# Mitigating extreme AI risks amid rapid progress [Linkpost]
In a new [*Science* paper](https://www.science.org/doi/10.1126/science.adn0117), the authors provide concise summaries of AI risks and offer recommendations for governments.
I think the piece is quite well-written. It concisely explains a lot of relevant ar... | https://www.lesswrong.com/posts/BikZyjiEgFmo7HQHm/mitigating-extreme-ai-risks-amid-rapid-progress-linkpost |
# EIS XIII: Reflections on Anthropic’s SAE Research Circa May 2024
Part 13 of 12 in the [Engineer’s Interpretability Sequence](https://www.alignmentforum.org/s/a6ne2ve5uturEEQK7).
TL;DR
=====
On May 5, 2024, [I made a set of 10 predictions](https://x.com/StephenLCasper/status/1787270794017702045) about what the next... | https://www.lesswrong.com/posts/pH6tyhEnngqWAXi9i/eis-xiii-reflections-on-anthropic-s-sae-research-circa-may |
# Anthropic announces interpretability advances. How much does this advance alignment?
Anthropic just published a pretty impressive set of results in interpretability. This raises for me, some questions and a concern: Interpretability helps, but it isn't alignment, right? It seems to me as though the vast bulk of alig... | https://www.lesswrong.com/posts/M3QqgcbXr3mgQKnBD/anthropic-announces-interpretability-advances-how-much-does |
# ARIA's Safeguarded AI grant program is accepting applications for Technical Area 1.1 until May 28th
**Note**: I am completely unaffiliated with [ARIA](https://www.lesswrong.com/posts/qweXJ6v9heSn4wvdk/davidad-s-provably-safe-ai-architecture-aria-s-programme). I figured I'd post this since applications are closing so... | https://www.lesswrong.com/posts/yfyjD9aCgNvqydv8J/aria-s-safeguarded-ai-grant-program-is-accepting |
# "Which chains-of-thought was that faster than?"
Here's some good advice from Eliezer:
TAP: "How could I have thought that faster?"
--------------------------------------------
* WHEN[^jf37bd9flbt] *you complete a chain-of-thought*
* THEN *ask yourself,* [*"**how could I have thought that faster?**"*](https://w... | https://www.lesswrong.com/posts/7oGfJG2BuvTgdCHQH/which-chains-of-thought-was-that-faster-than |
# Announcing Human-aligned AI Summer School
*The fourth Human-aligned AI Summer School will be held in Prague from 17^th^ to 20^th^ July 2024.* We will meet for four intensive days of talks, workshops, and discussions covering latest trends in AI alignment research and broader framings of AI alignment research. [***Ap... | https://www.lesswrong.com/posts/7RtAc6drC7Jtuzpqx/announcing-human-aligned-ai-summer-school |
# Cicadas, Anthropic, and the bilateral alignment problem
There have been a number of [responses](https://www.lesswrong.com/posts/pH6tyhEnngqWAXi9i/eis-xiii-reflections-on-anthropic-s-sae-research-circa-may) to today's [Anthropic interpretability research](https://transformer-circuits.pub/2024/scaling-monosemanticity/... | https://www.lesswrong.com/posts/MuyCbad9ZHW8b3rMP/cicadas-anthropic-and-the-bilateral-alignment-problem |
# Procedural Executive Function, Part 3
*The Off Road project has since been folded into Rethink Wellbeing, but I’ve continued working to better understand and treat Executive Dysfunction. You can read more about the project’s origins* [*here.*](https://forum.effectivealtruism.org/posts/8ekqxkowckzEKu3uG/off-road-supp... | https://www.lesswrong.com/posts/HAPhon49ofEwT3LCR/procedural-executive-function-part-3 |
# Should we be concerned about eating too much soy?
Parts of the internet say that, especially for men, eating too much soy is unhealthy while other parts of the internet advocate that soy is really great way for vegans to consume their protein.
Has anyone made a deep dive into the evidence base and formed an opinion... | https://www.lesswrong.com/posts/qWXviyhy6FZJuao3p/should-we-be-concerned-about-eating-too-much-soy |
# Do Not Mess With Scarlett Johansson
I repeat. Do not mess with Scarlett Johansson.
[You](https://letterboxd.com/film/black-widow-2021/) [would](https://letterboxd.com/film/lucy-2014/) [think](https://letterboxd.com/film/marriage-story-2019/) [her](https://letterboxd.com/film/her/) [movies](https://letterboxd.com/fi... | https://www.lesswrong.com/posts/N8aRDYLuakmLezeJy/do-not-mess-with-scarlett-johansson |
# Offering service as a sensayer for simulationist-adjacent beliefs.
A Sensayer is[^zlf37uy98up] a specialist in the private discussion of religion[^cfa8hzcympv] in small groups, or one on one, a bit like a councilor with theological acumen. Many of us, believe it or not, harbor private beliefs. Sometimes those belief... | https://www.lesswrong.com/posts/6CHeqNeXd6fffQKyA/offering-service-as-a-sensayer-for-simulationist-adjacent |
# A Bi-Modal Brain Model
When I am programming, writing, reading, browsing the internet, watching a movie, or playing a game, my brain is in a different mode of operation compared to when I am sitting in an empty room with nothing to do.
In the empty room, my brain will continuously generate fragments of language and... | https://www.lesswrong.com/posts/kyXLzHXGWfnLg39ev/a-bi-modal-brain-model |
# Higher-Order Forecasts
Higher-order forecasting could be a useful concept for prediction markets and forecasting systems more broadly.
The core idea is straightforward:
**Nth-order forecasts are forecasts about (N-1)th order forecasts.**
*
While the [NAO](https://naobservatory.org/) has primarily focused on wastewater, we're now spinning up a [swab sampling effort](https://naobservatory.or... | https://www.lesswrong.com/posts/KpaXzkhAng5xsPmns/quick-thoughts-on-our-first-sampling-run |
# Paper in Science: Managing extreme AI risks amid rapid progress
[https://www.science.org/doi/10.1126/science.adn0117](https://www.science.org/doi/10.1126/science.adn0117)
**Authors:**
Yoshua Bengio, Geoffrey Hinton, Andrew Yao, Dawn Song, Pieter Abbeel, Yuval Noah Harari, Ya-Qin Zhang, Lan Xue, Shai Shalev-Shwartz... | https://www.lesswrong.com/posts/82f3o2SuS3pwaZt8Y/paper-in-science-managing-extreme-ai-risks-amid-rapid |
# Big Picture AI Safety: Introduction
*tldr: I conducted 17 semi-structured interviews of AI safety experts about their big picture strategic view of the AI safety landscape: how will human-level AI play out, how things might go wrong, and what should the AI safety community be doing. While many respondents held “trad... | https://www.lesswrong.com/posts/yMTNjeEHfHcf2x7nY/big-picture-ai-safety-introduction |
# What will the first human-level AI look like, and how might things go wrong?
This is the first of 3 posts summarizing what I learned when I interviewed 17 AI safety experts about their "big picture" of the existential AI risk landscape: how will AGI play out, how things might go wrong, and what the AI safety communi... | https://www.lesswrong.com/posts/wHRMZizqfdW9RjrCY/what-will-the-first-human-level-ai-look-like-and-how-might |
# What should AI safety be trying to achieve?
This is the second of three posts summarizing what I learned when I interviewed 17 AI safety experts about their "big picture" of the existential AI risk landscape: how will artificial general intelligence (AGI) play out, how things might go wrong, and what the AI safety c... | https://www.lesswrong.com/posts/XfnnkK8XEjTqtuXGM/what-should-ai-safety-be-trying-to-achieve |
# What mistakes has the AI safety movement made?
This is the third of three posts summarizing what I learned when I interviewed 17 AI safety experts about their "big picture" of the existential AI risk landscape: how AGI will play out, how things might go wrong, and what the AI safety community should be doing. See [h... | https://www.lesswrong.com/posts/cGzQBRDrpNHoYtbKN/what-mistakes-has-the-ai-safety-movement-made |
# AI #65: I Spy With My AI
In terms of things that go in AI updates, this has been the busiest two week period so far. Every day ends with more open tabs than it started, even within AI.
As a result, some important topics are getting pushed to whenever I can give them proper attention. Triage is the watchword.
In pa... | https://www.lesswrong.com/posts/jkWvyzRzZQoaeq4mG/ai-65-i-spy-with-my-ai |
# Executive Dysfunction 101
This is an intro post for the [Procedural Executive Function](https://www.lesswrong.com/s/oTMuc9C3KKdZ6FcCB) sequence, which is now complete, to help give some general background of how to orient to executive dysfunction both philosophically and practically.
First things first; “executive ... | https://www.lesswrong.com/posts/ZgvdirAdsP8ofABeG/executive-dysfunction-101 |
# Are most people deeply confused about "love", or am I missing a human universal?
(In the following I am talking about "love" towards human beings only, not love of other things (such as music or food or God).)
A pet topic of mine is that the term *love* is so ambiguous as to be nigh-useless in rational discourse. B... | https://www.lesswrong.com/posts/GjheyeGCSACmBXC2K/are-most-people-deeply-confused-about-love-or-am-i-missing-a |
# The case for stopping AI safety research
TLDR: AI systems are failing in obvious and manageable ways for now. Fixing them will push the failure modes beyond our ability to understand and anticipate, let alone fix. The AI safety community is also doing a huge economic service to developers. Our belief that our minds ... | https://www.lesswrong.com/posts/vkzmbf4Mve4GNyJaF/the-case-for-stopping-ai-safety-research |
# Quick Thoughts on Scaling Monosemanticity
1\. How Many Features are Active at Once?
-----------------------------------------
Previously I’ve seen the rule of thumb “20-100 for most models”. Anthropic says:
> For all three SAEs, the average number of features active (i.e. with nonzero activations) on a given token... | https://www.lesswrong.com/posts/XpysWvyoqF4Cmx4yC/quick-thoughts-on-scaling-monosemanticity |
# Mentorship in AGI Safety (MAGIS) call for mentors
*Tldr: If you are working on AI Safety and are willing to help someone to start their career in AI Safety by sharing your experience at 1:1 meetings consider applying as a mentor*
In the last year, we’ve seen a surge of interest in AI safety. Many young professiona... | https://www.lesswrong.com/posts/D9Q4nXfxTdWckF2RX/mentorship-in-agi-safety-magis-call-for-mentors |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.