text stringlengths 300 320k | source stringlengths 52 154 |
|---|---|
# Games for AI Control
**\[UPDATE: This post is now updated to reflect our conference-length preprint,** [**available on arXiv**](https://arxiv.org/abs/2409.07985v1)**.** [**Louis Thomson**](https://www.lesswrong.com/users/lathomson) **helped with this update.** **New results are marked "NEW".\] **
This post summaris... | https://www.lesswrong.com/posts/Yu8jADLfptjPsR58E/games-for-ai-control-1 |
# Sherlockian Abduction Master List
\[Radically updated with many new entries around 07/10/24\]
Epistemic status: The List has been tested in the real world by me (with mixed results) and extensively checked for errors by many commenters. The Background section is mostly speculation and anecdotes, feel free to skip ... | https://www.lesswrong.com/posts/Yz33koDN5uhSEaB6c/sherlockian-abduction-master-list |
# Superbabies: Putting The Pieces Together
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,... | https://www.lesswrong.com/posts/2uJsiQqHTjePTRqi4/superbabies-putting-the-pieces-together |
# Podcast: "How the Smart Money teaches trading with Ricki Heicklen" (Patrick McKenzie interviewing)
In Patrick's words, above the fold:
> Some years ago, I built a trading simulation for a startup I was running, Starfighter. It was, at the time, by far the most impressive technical artifact I'd ever made, including ... | https://www.lesswrong.com/posts/Ab7K9xEjWwzRZBbWm/podcast-how-the-smart-money-teaches-trading-with-ricki |
# Yoshua Bengio: Reasoning through arguments against taking AI safety seriously
He starts by emphasizing
> The issue is so hotly debated because the stakes are major: According to some estimates, [**quadrillions of dollars**](https://www.amacad.org/sites/default/files/publication/downloads/Daedalus_Sp22_03_Russell.pd... | https://www.lesswrong.com/posts/ndvtiaqZErjnermWF/yoshua-bengio-reasoning-through-arguments-against-taking-ai |
# Transformer Circuit Faithfulness Metrics Are Not Robust
When you think you've found a circuit in a language model, how do you know if it does what you think it does? Typically, you [ablate / resample](https://www.lesswrong.com/posts/caZ3yR5GnzbZe2yJ3/how-to-do-patching-fast#Node_Patching) the activations of the mode... | https://www.lesswrong.com/posts/kp5h6khwrK47psMaA/transformer-circuit-faithfulness-metrics-are-not-robust |
# Most smart and skilled people are outside of the EA/rationalist community: an analysis
**Introduction:**
The (highly interrelated) effective altruist and Rationalist communities are very small on a global scale. Therefore, in general, most intelligence, skill and expertise is outside of the community, not within it... | https://www.lesswrong.com/posts/noxHoo3XKkzPG6s7E/most-smart-and-skilled-people-are-outside-of-the-ea |
# Alignment: "Do what I would have wanted you to do"
Yoshua Bengio writes[^0oho2bunkex]:
> **nobody currently knows how such an AGI or ASI could be made to behave morally, or at least behave as intended by its developers and not turn against humans**
I think I do[^07chvrf8nj6]. I believe that the difficulties of ali... | https://www.lesswrong.com/posts/Gs7CmLeYgc4iuyidr/alignment-do-what-i-would-have-wanted-you-to-do |
# Whiteboard Pen Magazines are Useful
Glue your colored whiteboard makers together with duck tape to create a whiteboard pen magazine. If required glue a paper layer around the duck tape to make it non-s... | https://www.lesswrong.com/posts/eBo5YCHP5H58Y5Qjx/whiteboard-pen-magazines-are-useful |
# Robin Hanson AI X-Risk Debate — Highlights and Analysis
This linkpost contains a lightly-edited transcript of highlights of my recent [AI x-risk debate with Robin Hanson](https://www.lesswrong.com/posts/mdeDquqameec2ERe4/robin-hanson-and-liron-shapira-debate-ai-x-risk), and a written version of what I said in the po... | https://www.lesswrong.com/posts/quj3RniAw7GGJtqeX/robin-hanson-ai-x-risk-debate-highlights-and-analysis |
# Consider attending the AI Security Forum '24, a 1-day pre-DEFCON event
Jeffrey Ladish, Caleb Parikh, and I are running the [AI security forum](http://aisecurity.forum/), a 1-day event on Thursday, August 8th '24 in Las Vegas, the day before DEFCON. This is a continuation of last year's [X-infosec Forum](https://ais... | https://www.lesswrong.com/posts/6NoKJymj4AFqkfqPr/consider-attending-the-ai-security-forum-24-a-1-day-pre |
# Timaeus is hiring!
**TLDR:** We’re hiring two research assistants to work on advancing [developmental interpretability](https://www.lesswrong.com/posts/TjaeCWvLZtEDAS5Ex/towards-developmental-interpretability) and other applications of singular learning theory to alignment.
**About Us**
------------
[Timaeus](htt... | https://www.lesswrong.com/posts/w7WdcKJgcwkGAzuse/timaeus-is-hiring |
# A Second Wetsuit Summer
I often write about things when they're new to me, because that's when I'm most interested in them. Sometimes how I feel after a short time ends up being how I feel long-term; other times I end up with an enthusiastic post about something I ended up abandoning a few months later. So when I re... | https://www.lesswrong.com/posts/TLZsfyrGJEmeMD8ju/a-second-wetsuit-summer |
# Thought Experiments Website
Hullo,
I'm hoping for your feedback. I recently finished a sort of beta of a website that allows users to explore and vote on thought experiments:
[https://thought-experiment-explorer.vercel.app/](https://thought-experiment-explorer.vercel.app/)
The spirit behind it is somewhat simila... | https://www.lesswrong.com/posts/9qYGCt4grJ62nJua5/thought-experiments-website |
# The world is awful. The world is much better. The world can be much better: The Animation.
*This video is an adaptation of Max Roser's* [*article,*](https://ourworldindata.org/much-better-awful-can-be-better) *"The world is awful. The world is much better. The world can be much better." It's a simple yet important ... | https://www.lesswrong.com/posts/nw4X7rNTnyapiPB8z/the-world-is-awful-the-world-is-much-better-the-world-can-be |
# A simple case for extreme inner misalignment
This post is the version of Yudkowsky's argument for inner misalignment that I wish I'd had in my head a few years ago. I don't claim that it's novel, that I endorse it, or even that Yudkowsky would endorse it; it's primarily an attempt to map his ideas into an ontology t... | https://www.lesswrong.com/posts/fjfWrKhEawwBGCTGs/a-simple-case-for-extreme-inner-misalignment |
# Stitching SAEs of different sizes
*Work done in Neel Nanda’s stream of MATS 6.0, equal contribution by Bart Bussmann and Patrick Leask, Patrick Leask is concurrently a PhD candidate at Durham University*
**TL;DR: **When you scale up an SAE, the features in the larger SAE can be categorized in two groups: 1) “novel ... | https://www.lesswrong.com/posts/baJyjpktzmcmRfosq/stitching-saes-of-different-sizes |
# Trust as a bottleneck to growing teams quickly
This is an adaptation of an internal doc I wrote for [Anthropic](https://www.anthropic.com/).
I’ve been noticing recently that often, a big blocker to teams staying effective as they grow is _trust_.
“Alice doesn’t trust Bob” makes Alice sound like the bad guy, but it... | https://www.lesswrong.com/posts/EjkKtKK7KZit7Kca8/trust-as-a-bottleneck-to-growing-teams-quickly |
# Ice: The Penultimate Frontier
I argue here that preventing a large iceberg from melting is absurdly cheap per unit area compared to just about any other way of making new land, and it's kind of crazy to spend money on space exploration and colonization before colonizing the oceans with floating ice-islands. | https://www.lesswrong.com/posts/gthjxPDywrMTs3p2j/ice-the-penultimate-frontier |
# Brief notes on the Wikipedia game
Alex Turner [introduced](https://www.lesswrong.com/posts/dqSwccGTWyBgxrR58/turntrout-s-shortform-feed?commentId=nL7E6SEtXqDG7SHGB) an exercise to test subjects’ ability to notice falsehoods: change factual statements in Wikipedia articles, hand the edited articles to subjects and se... | https://www.lesswrong.com/posts/4fqdfrDzFebsx8amf/brief-notes-on-the-wikipedia-game |
# LLMs as a Planning Overhang
*It's quite possible someone has already argued this, but I thought I should share just in case not.*
Goal-Optimisers and Planner-Simulators
--------------------------------------
When people in the past discussed worries about AI development, this was often about AI agents - AIs that h... | https://www.lesswrong.com/posts/yRqzRmEC3aqcajKjy/llms-as-a-planning-overhang |
# An Introduction to Representation Engineering - an activation-based paradigm for controlling LLMs
Representation Engineering (aka Activation Steering/Engineering) is a new paradigm for understanding and controlling the behaviour of LLMs. Instead of changing the prompt or weights of the LLM, it does this by directly ... | https://www.lesswrong.com/posts/3ghj8EuKzwD3MQR5G/an-introduction-to-representation-engineering-an-activation |
# Robert Caro And Mechanistic Models In Biography
:
[](https://www.jefftk.com/child-handrail-with-top-returns-big.jpg)
Comp... | https://www.lesswrong.com/posts/aW98e9wZqWqTb2Ane/child-handrail-returns |
# Breaking Circuit Breakers
A few days ago, Gray Swan published code and models for their recent “circuit breakers” method for language models.[\[1\]](https://confirmlabs.org/posts/circuit_breaking.html#ref-zou2024improvingalignmentrobustnesscircuit)[^1^](https://confirmlabs.org/posts/circuit_breaking.html#fn1)
The c... | https://www.lesswrong.com/posts/NAYyHimM3FaDYLvEH/breaking-circuit-breakers |
# patent process problems
The current patent process has some problems. Here are some of them.
## patenting is slow
The US Patent Office [tracks pendency](https://www.uspto.gov/dashboard/patents/pendency.html) of patents. Currently, on average, there's 20 months from filing to the first response, and over 25 months... | https://www.lesswrong.com/posts/qKCgnSaf5JRvtxDEQ/patent-process-problems |
# Four ways I've made bad decisions
I was in the process of making a decision the other day, and thought about the times I felt like I either made an incorrect decision or followed an incorrect decision making process. Turns out, there was actually *many such cases.*[^vm59787hizo]
I categorized those into four differ... | https://www.lesswrong.com/posts/guYpJdkxExYiZZLjS/four-ways-i-ve-made-bad-decisions |
# Misnaming and Other Issues with OpenAI's “Human Level” Superintelligence Hierarchy
Bloomberg [reports](https://archive.is/SLtFQ) that OpenAI internally has benchmarks for “Human-Level AI.” They have 5 levels, with the first being the achieved level of having intelligent conversation, to level 2, “\[unassisted, PhD-l... | https://www.lesswrong.com/posts/JycQ8dZwAKcgfbWig/misnaming-and-other-issues-with-openai-s-human-level |
# Stacked Laptop Monitor Update
In the spirit of coming back to [things once I have more experience with them](https://www.jefftk.com/p/a-second-wetsuit-summer), I've continued to be very happy with my [stacked laptop monitor](https://www.jefftk.com/p/stacked-laptop-monitor). I just got back from a week of [touring](h... | https://www.lesswrong.com/posts/rZqqZJK3BLwKipo2Y/stacked-laptop-monitor-update |
# Comparing Quantized Performance in Llama Models
*Epistemic Status: Quick tests, most of this was done in less than 48 hours*
*TL;DR: Can you skimp on GPU VRAM? 8bit quantized seems fine, for 4bit it ... | https://www.lesswrong.com/posts/qmPXQbyYA66DuJbht/comparing-quantized-performance-in-llama-models |
# Dialogue on What It Means For Something to Have A Function/Purpose
*Context for LW audience: Ramana, Steve and John regularly talk about stuff in the general cluster of agency, abstraction, optimization, compression, purpose, representation, etc. We decided to write down some of our discussion and post it here. This... | https://www.lesswrong.com/posts/pvBqB6F89enufZkYr/dialogue-on-what-it-means-for-something-to-have-a-function |
# Deceptive agents can collude to hide dangerous features in SAEs
TL;DR: We use refusal-ablated Llama 3 70B agents as both labeling and simulating agents on the features of a sparse auto-encoder, with GPT-4 as an overseer. The agents follow a deceptive coordinated strategy to mislabel important features such as "coup"... | https://www.lesswrong.com/posts/PyzZ6gcB7BaGAgcQ7/deceptive-agents-can-collude-to-hide-dangerous-features-in |
# Against Aschenbrenner: How 'Situational Awareness' constructs a narrative that undermines safety and threatens humanity
**Summary/Introduction**
========================
Aschenbrenner’s ‘Situational Awareness’ [(Aschenbrenner, 2024)](https://paperpile.com/c/teOwJn/7zy0) promotes a dangerous narrative of national se... | https://www.lesswrong.com/posts/i5pccofToYepythEw/against-aschenbrenner-how-situational-awareness-constructs-a |
# I found >800 orthogonal "write code" steering vectors
*Produced as part of the MATS Summer 2024 program, under the mentorship of Alex Turner (TurnTrout).*
A few weeks ago, I stumbled across a very weird fact: it is possible to find multiple steering vectors in a language model that activate very similar behaviors w... | https://www.lesswrong.com/posts/CbSEZSpjdpnvBcEvc/i-found-greater-than-800-orthogonal-write-code-steering |
# How (and why) to get tested for CMV
As you may know, [I’m not a fan of cytomegalovirus (CMV)](https://denovo.substack.com/p/cytomegalovirus-the-worst-herpesvirus), a herpesvirus that is usually asymptomatic, but causes all sorts of bad things like immune aging and birth defects.
Over the last few months I’ve been f... | https://www.lesswrong.com/posts/ypBpFuBFCfLWJDddv/how-and-why-to-get-tested-for-cmv |
# MIRI's July 2024 newsletter
### MIRI updates
* Rob Bensinger suggests that AI risk discourse could be improved by adopting [a new set of labels](https://x.com/robbensinger/status/1801306833325592759) for different perspectives on existential risk from AI. One drawback of “AI doomer” (a label sometimes used in onl... | https://www.lesswrong.com/posts/Gq2PCjfuhLX99b6go/miri-s-july-2024-newsletter |
# Paper Summary: Princes and Merchants: European City Growth Before the Industrial Revolution
by J. Bradford de Long and Andrei Shleifer. (1993)
[https://scholar.harvard.edu/files/shleifer/files/princes_merchants.pdf](https://scholar.harvard.edu/files/shleifer/files/princes_merchants.pdf).
**Summary:** Freer societi... | https://www.lesswrong.com/posts/hCkXuNtX4aqc8KTyx/paper-summary-princes-and-merchants-european-city-growth |
# Towards more cooperative AI safety strategies
*This post is written in a spirit of constructive criticism. It's phrased fairly abstractly, in part because it's a sensitive topic, but I welcome critiques and comments below. The post is structured in terms of three claims about the strategic dynamics of AI safety effo... | https://www.lesswrong.com/posts/5uffoui3axcK7gQXG/towards-more-cooperative-ai-safety-strategies |
# Why the Best Writers Endure Isolation
Douglas Adams, author of *The Hitchhiker’s Guide to the Galaxy*, was [once](https://en.wikipedia.org/wiki/Douglas_Adams#cite_ref-32) locked in a room for three weeks until he completed one of his books.
Victor Hugo, [when faced with a deadline](https://theweek.com/articles/6472... | https://www.lesswrong.com/posts/mXpMEn5oqFyvssLu3/why-the-best-writers-endure-isolation |
# Apply now: Get "unstuck" with the New IFS Self-Care Fellowship Program
You finally want to resolve deeper-seated inner conflicts, and remove inner blocks in the way of becoming a more fulfilled, resilient, and well-performing version of yourself? This post allows you to learn how IFS as a coaching or therapy approac... | https://www.lesswrong.com/posts/DE8zd6DbvGdiMe8ry/apply-now-get-unstuck-with-the-new-ifs-self-care-fellowship |
# DM Parenting
*Cause no one will question your ethics if you refer to yourself as a Dungeon Mom.*
* * *
I snort experimentation to feel alive. It’s a certain type of orientation to life, completely at odds with all parenting advice about predictability and routine.
Enter DM parenting. Where you approach every pare... | https://www.lesswrong.com/posts/6NWt2GyXunvM36Cus/dm-parenting |
# Mech Interp Lacks Good Paradigms
*Note: I wrote this post rather quickly as an exercise in sharing rough / unpolished thoughts. I am also not an expert on some of the things I've written about. If you spot mistakes or would like to point out missed work / perspectives, please feel free! *
*Note 2: I originally sent... | https://www.lesswrong.com/posts/3CZF3x8FX9rv65Brp/mech-interp-lacks-good-paradigms |
# Fully booked - LessWrong Community weekend
***Update: the event is fully booked. Applications are still open but put on a wait list until spots get available.***
The **LessWrong community weekend (Fr 13th - Mo 16th September, Berlin)** is this very special place where rationalists who joined once, often say "**... | https://www.lesswrong.com/posts/S7BEuCJihzAv9sCrv/fully-booked-lesswrong-community-weekend |
# Francois Chollet inadvertently limits his claim on ARC-AGI
Specifically, this tweet makes the assumption that open-source AI is basically as good at search, especially automated search for algorithms as the big labs, and if that assumption is violated, the prize won't be claimed even if big companies have much bette... | https://www.lesswrong.com/posts/BbfpA4GMFua9otCd5/francois-chollet-inadvertently-limits-his-claim-on-arc-agi |
# How to wash your hands precisely and thoroughly
[Common instructions on handwashing](https://www.cdc.gov/clean-hands/about/index.html) are too vague for me.
> **Scrub** your hands **for at least 20 seconds**.
Sure, but how? I'm left guessing at which parts to scrub how much, possibly ruining the effect of the wash... | https://www.lesswrong.com/posts/hzguZnTTQ545oApfH/how-to-wash-your-hands-precisely-and-thoroughly |
# Multiplex Gene Editing: Where Are We Now?
the Cas9 enzyme used in CRISPR
We’re starting to get [... | https://www.lesswrong.com/posts/oSy5vHvwSfnjmC7Tf/multiplex-gene-editing-where-are-we-now |
# Simplifying Corrigibility – Subagent Corrigibility Is Not Anti-Natural
Max Harms recently published [an interesting series of posts on corrigibility](https://www.alignmentforum.org/s/KfCjeconYRdFbMxsy), which argue that corrigibility should be the sole objective we try to give to a potentially superintelligent AI. A... | https://www.lesswrong.com/posts/Zyq7PfBuaZgqEL8qQ/simplifying-corrigibility-subagent-corrigibility-is-not-anti |
# Turning Your Back On Traffic
We do a lot of [walking around the neighborhood](https://www.jefftk.com/p/the-serendipity-of-density) with kids, which usually involves some people getting to intersections a while before others. I'm not worried about even the youngest going into the street on their own—Nora's been [stre... | https://www.lesswrong.com/posts/FJokhnaYjEbcHLusJ/turning-your-back-on-traffic |
# Announcing Open Philanthropy's AI governance and policy RFP
AI has [enormous](https://epochai.org/blog/explosive-growth-from-ai-a-review-of-the-arguments) beneficial [potential](https://www.planned-obsolescence.org/the-costs-of-caution/) if it is governed well. However, in line with a [growing](https://www.science.o... | https://www.lesswrong.com/posts/aiexwwFKbwQpCYuDN/announcing-open-philanthropy-s-ai-governance-and-policy-rfp |
# Agency in Politics
This is a crosspost from [https://250bpm.substack.com/p/agency-in-politics](https://250bpm.substack.com/p/agency-in-politics)
[Matt Yglesias explains](https://www.slowboring.com/p/we-have-more-agency-than-we-like-300) why history is not a steam roller with a steering wheel welded tight, unyieldin... | https://www.lesswrong.com/posts/5jXchjCmCsqkBftyX/agency-in-politics |
# We ran an AI safety conference in Tokyo. It went really well. Come next year!
Abstract
========
[Technical AI Safety 2024 (TAIS 2024)](https://tais2024.cc/) was a conference organised by [AI Safety 東京... | https://www.lesswrong.com/posts/pjxDs92hHEjLg5JcA/we-ran-an-ai-safety-conference-in-tokyo-it-went-really-well |
# Housing Roundup #9: Restricting Supply
I’d split the latest housing roundup into local versus global questions. I was planning on waiting a bit between them.
Then Joe Biden decided to propose a version of the worst possible thing.
So I guess here we are.
What is the organizing principle of Bidenomics?
#### Restr... | https://www.lesswrong.com/posts/sX5ANDiTb96CkYpxd/housing-roundup-9-restricting-supply |
# Profit and Value
* * *
In 1950, the eponymous founder George Wilhelm Merck said that **“**[We try never to forget that medicine is for the people](https://www.merck.com/company-overview/history/). It is not for the profits. The profits follow, and if we have remembered that, they have never failed to appear.”
In 1... | https://www.lesswrong.com/posts/yrkdefTbhjj5qFgvX/profit-and-value |
# Individually incentivized safe Pareto improvements in open-source bargaining
Summary
=======
Agents might fail to peacefully trade in high-stakes negotiations. Such bargaining failures can have catastrophic consequences, including great power conflicts, and AI [flash wars](https://www.lesswrong.com/posts/LpM3EAakwY... | https://www.lesswrong.com/posts/uGfDx9es2pnYWaWJr/individually-incentivized-safe-pareto-improvements-in-open |
# What are you getting paid in?
*Crossposting this essay by my friend,* [*Leila Clark.*](https://www.approachwithalacrity.com/)
* * *
A long time ago, a manager friend of mine wrote a book to collect his years of wisdom. He never published it, which is a shame because it was full of interesting insights. One that I ... | https://www.lesswrong.com/posts/ottALpgA9uv4wgkkK/what-are-you-getting-paid-in |
# Baking vs Patissing vs Cooking, the HPS explanation
Why is baking harder than cooking?
I’m not ev... | https://www.lesswrong.com/posts/b3tYWoSbKXqpirjxq/baking-vs-patissing-vs-cooking-the-hps-explanation |
# Optimistic Assumptions, Longterm Planning, and "Cope"
Eliezer Yudkowsky periodically [complains](https://www.lesswrong.com/posts/j9Q8bRmwCgXRYAgcJ/miri-announces-new-death-with-dignity-strategy#Q2___I_have_a_clever_scheme_for_saving_the_world___I_should_act_as_if_I_believe_it_will_work_and_save_everyone__right__even... | https://www.lesswrong.com/posts/8ZR3xsWb6TdvmL8kx/optimistic-assumptions-longterm-planning-and-cope |
# D&D.Sci: Whom Shall You Call? [Evaluation and Ruleset]
This is a followup to [the D&D.Sci post](https://www.lesswrong.com/posts/vrB8xYK8pFPmP5ts8/d-and-d-sci-whom-shall-you-call) I made twelve 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-a... | https://www.lesswrong.com/posts/AicLyFGmZMSwG7yM7/d-and-d-sci-whom-shall-you-call-evaluation-and-ruleset |
# Friendship is transactional, unconditional friendship is insurance
It feels a little icky to say, but we befriend people because we get something out of it. We enjoy the company, the conversation, the emotional support, the activities, the connection, etc. It's not a coincidence people don't befriend brick walls.
(... | https://www.lesswrong.com/posts/SXvc8xsLA6hPkdh9w/friendship-is-transactional-unconditional-friendship-is |
# SAEs (usually) Transfer Between Base and Chat Models
*This is an interim report sharing preliminary results that we are currently building on. We hope this update will be useful to related research occurring in parallel.*
Executive Summary
-----------------
* We train SAEs on base / chat model pairs and find tha... | https://www.lesswrong.com/posts/fmwk6qxrpW8d4jvbd/saes-usually-transfer-between-base-and-chat-models |
# A List of 45+ Mech Interp Project Ideas from Apollo Research’s Interpretability Team
**Why we made this list: **
* The interpretability team at Apollo Research wrapped up a few projects recently[^gvazr9vg2n7]. In order to decide what we’d work on next, we generated a lot of different potential projects. Unfortuna... | https://www.lesswrong.com/posts/KfkpgXdgRheSRWDy8/a-list-of-45-mech-interp-project-ideas-from-apollo-research |
# AI #73: Openly Evil AI
What do you call a clause explicitly saying that you waive the right to whistleblower compensation, and that you need to get permission before sharing information with government regulators like the SEC?
I have many answers.
I also know that OpenAI, having f***ed around, seems poised to find... | https://www.lesswrong.com/posts/fM4Bs9nanDzio3xCq/ai-73-openly-evil-ai |
# Me & My Clone
An advanced alien species clones me on the atomic level, lines me up **exactly** across myself, in a perfect mirrored room:
Diagram of the room, as seen from above.
I stare at myself fo... | https://www.lesswrong.com/posts/QpejqFa6BsnE87ENt/me-and-my-clone |
# [Interim research report] Evaluating the Goal-Directedness of Language Models
*This post was written as part of the summer 2024 cohort of the*[* ML Alignment & Theory Scholars*](https://www.matsprogram.org/)* program, under the mentorship of Marius Hobbhahn.*
Summary
=======
Over the past four weeks, we have been ... | https://www.lesswrong.com/posts/ABsb23t74M5ZetuqX/interim-research-report-evaluating-the-goal-directedness-of |
# What are the actual arguments in favor of computationalism as a theory of identity?
A few months ago, Rob Bensinger made [a rather long post](https://www.lesswrong.com/posts/zPM5r3RjossttDrpw/when-is-a-mind-me) (that even [got curated](https://www.lesswrong.com/posts/zPM5r3RjossttDrpw/when-is-a-mind-me?commentId=5Fv... | https://www.lesswrong.com/posts/yoAhc7ZhQZfGqrzif/what-are-the-actual-arguments-in-favor-of-computationalism |
# My experience applying to MATS 6.0
The current cohort of the [ML Alignment & Theory Scholars Program](https://www.matsprogram.org/), MATS 6.0, had a unique application process and its broadest selection of mentors yet, with 40 mentors to apply to. I was invited to interview with twelve mentors and was accepted by fi... | https://www.lesswrong.com/posts/sajjdRDKKt3BJqRiP/my-experience-applying-to-mats-6-0 |
# Linkpost: Surely you can be serious
Adam Mastroianni writes about "actually caring about stuff, and for the right reasons", rather than just LARPing. The opening is excerpted below.
> I once saw someone give a talk about a tiny intervention that caused a gigantic effect, something like, “We gave high school seniors... | https://www.lesswrong.com/posts/FfihtAuPwtLnBmeGZ/linkpost-surely-you-can-be-serious |
# How do we know that "good research" is good? (aka "direct evaluation" vs "eigen-evaluation")
*AI Alignment is my motivating context but this could apply elsewhere too.*
The nascent field of AI Alignment research is pretty happening these days. There are multiple orgs and dozens to low hundreds of full-time rese... | https://www.lesswrong.com/posts/ZPixGrCbJoyuAdjkK/how-do-we-know-that-good-research-is-good-aka-direct |
# Have people given up on iterated distillation and amplification?
The BlueDot Impact write-up for scalable oversight seems to suggest that people have given up on [iterated distillation and amplification](https://aisafetyfundamentals.com/blog/scalable-oversight-intro/) (IDA) working. I haven’t really seen much resear... | https://www.lesswrong.com/posts/EDBxdR9MzXwDaJYuC/have-people-given-up-on-iterated-distillation-and |
# Romae Industriae
*Whatever each culture grows and manufactures cannot fail to be here at all times and in great profusion. Here merchant vessels arrive carrying these many commodities from every region in every season and even at every equinox, so that the city takes on the appearance of a sort of common market for ... | https://www.lesswrong.com/posts/hmACfbrorLf2jKegr/romae-industriae |
# Sustainability of Digital Life Form Societies
Hiroshi Yamakawa^1^*^,^*^2^*^,^*^3^*^,^*^4^
^1 ^The University of Tokyo, Tokyo, Japan
^2 ^AI Alignment Network, Tokyo, Japan
^3 ^The Whole Brain Architecture Initiative, Tokyo, Japan
^4 ^RIKEN, Tokyo, Japan
*Even in a society composed of digital life forms (DLFs) wi... | https://www.lesswrong.com/posts/2u4Dja2m6ud4m7Bb7/sustainability-of-digital-life-form-societies |
# Truth is Universal: Robust Detection of Lies in LLMs
A short summary of the paper is presented below.
**TL;DR**: We develop a robust method to detect when an LLM is lying based on the internal model activations, making the following contributions: (i) We demonstrate the existence of a *two*-dimensional subspace, al... | https://www.lesswrong.com/posts/72vpkRRvoPHKi48fi/truth-is-universal-robust-detection-of-lies-in-llms-3 |
# JumpReLU SAEs + Early Access to Gemma 2 SAEs
New paper from the Google DeepMind mechanistic interpretability team, led by Sen Rajamanoharan!
We introduce JumpReLU SAEs, a new SAE architecture that replaces the standard ReLUs with discontinuous JumpReLU activations, and seems to be (narrowly) state of the art over e... | https://www.lesswrong.com/posts/wZqqQysfLrt2CFx4T/jumprelu-saes-early-access-to-gemma-2-saes |
# Feature Targeted LLC Estimation Distinguishes SAE Features from Random Directions
**Tl;dr:** In this post we present the exploratory phase of a project aiming to study neural networks by applying static [local learning coefficient](https://www.lesswrong.com/posts/6g8cAftfQufLmFDYT/you-re-measuring-model-complexity-w... | https://www.lesswrong.com/posts/zpwh9knxBMHuFbJxc/feature-targeted-llc-estimation-distinguishes-sae-features |
# (Approximately) Deterministic Natural Latents
*Background:* [*Natural Latents: The Math*](https://www.lesswrong.com/posts/dWQWzGCSFj6GTZHz7/natural-latents-the-math)*,* [*Natural Latents: The Concepts*](https://www.lesswrong.com/posts/mMEbfooQzMwJERAJJ/natural-latents-the-concepts)*,* [*Why Care About Natural Latent... | https://www.lesswrong.com/posts/fJb8ryrMW5XfJaq7m/approximately-deterministic-natural-latents |
# Krona Compare
*Cross-posted from my [NAO Notebook](https://data.securebio.org/jefftk-notebook/sample-vs-global-prevalence).*
When trying to understand how metagenomic samples differ I often want to drill down through the taxonomic hierarchy, comparing relative abundances. I've tried several tools for this, existing... | https://www.lesswrong.com/posts/he9BpRWuds5istBus/krona-compare |
# BatchTopK: A Simple Improvement for TopK-SAEs
Work done in Neel Nanda’s stream of MATS 6.0.
*Epistemic status: Tried this on a single sweep and seems to work well, but it might definitely be a fluke of something particular to our *[*implementation*](https://github.com/bartbussmann/BatchTopK) *or experimental set-up... | https://www.lesswrong.com/posts/Nkx6yWZNbAsfvic98/batchtopk-a-simple-improvement-for-topk-saes |
# A more systematic case for inner misalignment
This post builds on [my previous post](https://www.alignmentforum.org/posts/fjfWrKhEawwBGCTGs/a-simple-case-for-extreme-inner-misalignment) making the case that squiggle-maximizers are plausible. The argument I presented was a deliberately simplified one, though, and glo... | https://www.lesswrong.com/posts/uAhBngKf7FrWcApqr/a-more-systematic-case-for-inner-misalignment |
# Why Georgism Lost Its Popularity
Henry George’s 1879 book *Progress & Poverty* was the second best-selling book in the entire world during the 1880s and 1890s, outsold by only the Bible. Nobody knows exactly how many copies it sold during those two decades since nobody was keeping track, but it definitely sold at le... | https://www.lesswrong.com/posts/4EzZ54jqgLh8aGejB/why-georgism-lost-its-popularity |
# Ball Sq Pathways
With the Red Line shut down north of Kendall [for repairs](https://www.mbta.com/news/2023-11-09/mbta-announces-ambitious-track-improvement-program-eliminate-all-speed-restrictions) I've been walking to the Green Line at Ball Sq. It's definitely a slower commute, but it's not bad. Except for 150 pain... | https://www.lesswrong.com/posts/mKsmmcyjLuzbvACur/ball-sq-pathways |
# aimless ace analyzes active amateur: a micro-aaaaalignment proposal
_This idea is so simple that I'm sure it's been had by someone somewhere._
Suppose we have some method to make really smart honest AIs that do not have goals. Let's say it's a yes/no oracle. Our aimless ace.
But we want to accomplish stuff! AIcorp... | https://www.lesswrong.com/posts/t8jwPBrxccf5bhMcC/aimless-ace-analyzes-active-amateur-a-micro-aaaaalignment |
# Holomorphic surjection theorem (Picard's little theorem)
Consider an entire function (complex-differentiable everywhere) $f(z)$. I will intuitively prove that certain lemmas hold on any such $f$. If $f$ is a polynomial, I can combine those lemmas with another one, showing that the holomorphic surjection theorem ([mo... | https://www.lesswrong.com/posts/7pHLW5pfBStDBdj8z/holomorphic-surjection-theorem-picard-s-little-theorem |
# Using an LLM perplexity filter to detect weight exfiltration
A recent area of focus has been securing AI model weights. If the weights are located in a data center and an adversary wants to obtain model weights, the weights have to leave physically (such as a hard drive going out the front door) or through the data ... | https://www.lesswrong.com/posts/aWZEDw6oxR6Wk5hru/using-an-llm-perplexity-filter-to-detect-weight-exfiltration |
# A simple model of math skill
I've noticed that when trying to understand a math paper, there are a few different ways my skill level can be the blocker. Some of these ways line up with some typical levels of organization in math papers:
* **Definitions:** a formalization of the kind of objects we're even talking ... | https://www.lesswrong.com/posts/EF8tvShQJ5cbdZzTb/a-simple-model-of-math-skill |
# Coalitional agency
### The coalitional frame
[Earlier in this sequence](https://www.alignmentforum.org/s/MGMwqENAgdi85fiwF) I laid out an argument that the goals of increasingly intelligent AIs will become increasingly systematized, until they converge to squiggle-maximization. In [my last post](https://www.alignme... | https://www.lesswrong.com/posts/WD6fdTreuTaXgZb4H/coalitional-agency |
# OpenAI Boycott Revisit
Note: An initial proposal and some good discussion already existed on LW [here](https://www.lesswrong.com/posts/sXhBCDLJPEjadwHBM/boycott-openai). I’m spurring this here as a post instead of a comment due to length, the need for a fresh look, and a specific call to action.
**Summary**
I thin... | https://www.lesswrong.com/posts/dhEWyhkaL3PLfYyfd/openai-boycott-revisit |
# An experiment on hidden cognition
*Work done as part of the Visiting Fellow program at Constellation. Thanks to Aaron Scher for conversations and feedback throughout the project, and for pointing out relevant literature. Thanks also to Jérémy Scheurer for feedback and references.*
Motivation
==========
There are m... | https://www.lesswrong.com/posts/fiksCnoEnA2hBSf2m/an-experiment-on-hidden-cognition |
# Categories of leadership on technical teams
This is an adaptation of an internal doc I wrote for [Anthropic](https://www.anthropic.com/).
Recently I’ve been having a lot of conversations about how to structure and staff teams. One framework I’ve referenced repeatedly is to break down team leadership into a few diff... | https://www.lesswrong.com/posts/yRhhGpxQwKekmo4bJ/categories-of-leadership-on-technical-teams |
# Initial Experiments Using SAEs to Help Detect AI Generated Text
*This is an informal progress update with some preliminary results using Sparse Autoencoders (SAEs) to detect AI-generated text.*
Summary
=======
I would like reliable methods for distinguishing human and AI-generated text. Unfortunately, not only are... | https://www.lesswrong.com/posts/LQBFqyXA5to4iHEBC/initial-experiments-using-saes-to-help-detect-ai-generated |
# Auto-Enhance: Developing a meta-benchmark to measure LLM agents’ ability to improve other agents
Summary
=======
* Scaffolded LLM agents are, in principle, able to execute arbitrary code to achieve the goals they have been set. One such goal could be self-improvement.
* This post outlines our plans to build a b... | https://www.lesswrong.com/posts/s9zd6f9eZ8qN2jrcu/auto-enhance-developing-a-meta-benchmark-to-measure-llm |
# On the CrowdStrike Incident
Things went very wrong on Friday.
A bugged CrowdStrike update temporarily bricked quite a lot of computers, bringing down such fun things as airlines, hospitals and 911 services.
It was serious out there.
> [Ryan Peterson](https://x.com/typesfast/status/1814390450318655761): Crowdstrik... | https://www.lesswrong.com/posts/oAKfaxKKfuz2cuRLr/on-the-crowdstrike-incident |
# Caring about excellence
> “Anything worth doing, is worth doing right.”
― Hunter S Thompson
> “Be not professional in what you do, rather be excellent. Excellence has life in it — it has colors in it — it has sweetness in it — whereas professionalism is a dead corpse exuding the disgusting smell of obedience. Exce... | https://www.lesswrong.com/posts/hxX3PBRcrcsoeEAhr/caring-about-excellence |
# The Garden of Eden
\[This is the text of the sermon given by Pastor James Windrow on Sunday, July 14, 2024.\]
Greetings, brothers and sisters. Today, we reflect on humanity’s origin in the Garden of Eden and its fall and devolution thereafter.
When Christians used to think about the Garden of Eden, it was common t... | https://www.lesswrong.com/posts/YzkBxG9sfyKa5BA4W/the-garden-of-eden |
# Analyzing DeepMind's Probabilistic Methods for Evaluating Agent Capabilities
*Produced as part of the* [*MATS Program*](https://www.matsprogram.org/) *Summer 2024 Cohort. The project is supervised by Marius Hobbhahn and Jérémy Scheurer*
Update: See also our [paper on this topic](https://arxiv.org/abs/2409.16125... | https://www.lesswrong.com/posts/NKmjGS4a3ykriqRNR/analyzing-deepmind-s-probabilistic-methods-for-evaluating |
# Efficient Dictionary Learning with Switch Sparse Autoencoders
*Produced as part of the* [ML Alignment & Theory Scholars Program](https://www.matsprogram.org/) *\- Summer 2024 Cohort*
0\. Summary
-----------
To recover all the relevant features from a superintelligent language model, we will likely need to scale sp... | https://www.lesswrong.com/posts/47CYFbrSyiJE2X5ot/efficient-dictionary-learning-with-switch-sparse |
# Trying to understand Hanson's Cultural Drift argument
At 2024's [Manifest](https://www.manifest.is/), Robin Hanson gave a talk (in his usual sweeping polymathic style) on cultural drift - a phenomenon for which he thinks "there's going to be hell to pay", and about which he is "scared, because this is a really big f... | https://www.lesswrong.com/posts/EtKXdPfDaNSpLJFwN/trying-to-understand-hanson-s-cultural-drift-argument |
# Ransomware Payments Should Require a Sin Tax
A tax could largely mitigate the growing ransomware problem. The following is a proposal for a scheduled, gradual increase in the tax rate on ransom payments:
1. The ransom demanded by an attacker is a function of the expected likelihood that the victim will pay that am... | https://www.lesswrong.com/posts/YbxeknCavWiPK6Wg3/ransomware-payments-should-require-a-sin-tax |
# D&D.Sci Scenario Index
There have been a lot of D&D.Sci scenarios, but there's a lot of variance between them in complexity and quality. Some are more difficult, and might not be a good place to start, while others are much simpler - some were very good, while others on reflection didn't flow quite right.
Unfort... | https://www.lesswrong.com/posts/6QfatwCcmuaDBNeDv/d-and-d-sci-scenario-index |
# Room Available in Boston Group House
We have a room opening up in a rationalist house near Central Square, Cambridge (Boston). It is a 4-bedroom, 3-bathroom apartment with two floors, with a large kitchen and living room plus additional common space upstairs. There’s a washer, dryer, and dishwasher in-unit. It’s an ... | https://www.lesswrong.com/posts/zxPmJCRA7oP3Mb4Gc/room-available-in-boston-group-house |
# My covid-related beliefs and questions
Things I'm fairly confident in:
- I should take colds in general more seriously than I did pre-pandemic: Staying at home with cold symptoms is good.
General masking during cold season is good.
We should have air filters in all public indoor spaces.
- Long covid is real and we s... | https://www.lesswrong.com/posts/P7aswE7XqNKGQg4D3/my-covid-related-beliefs-and-questions |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.