All 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 20252026

All Jan Feb Mar Apr May JunJulAug Sep

All 1 2 3 4 5 6 7 8 9 101112 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

Plan A: Suggestions For Further Work

Thomas Larsen10 Jul 2026 23:02 UTC

52 points

7 comments7 min readLW link

Freeing Thucydides

djbinder10 Jul 2026 22:12 UTC

44 points

2 comments4 min readLW link

(defensesindepth.bio)

An Induction Head in Disguise: Chasing Grammar in a Character-Level Transformer

Ameya Panchal10 Jul 2026 22:05 UTC

−1 points

0 comments5 min readLW link

(ameya-bit.github.io)

The standard can be to admit the existence of a standard

Firinn10 Jul 2026 21:22 UTC

44 points

1 comment36 min readLW link

Persona Cartography: Charting Language Model Personality Traits in Weight Space

antonghawthorne, Mariia Koroliuk, Irakli Shalibashvili, sidbaines, Clément Dumas, Konstantinos Voudouris and David Africa

10 Jul 2026 18:54 UTC

44 points

0 comments18 min readLW link

(arxiv.org)

The easiest pathway to control is through executive power

djbinder10 Jul 2026 18:48 UTC

111 points

3 comments6 min readLW link

(defensesindepth.bio)

The Human Substitution Test as a Sanity Check for AI Evaluations

VojtaKovarik, Tomáš Gavenčiak and Mateusz Bagiński

10 Jul 2026 17:27 UTC

31 points

5 comments8 min readLW link

(limits-of-evaluation.org)

Cap-and-trade question: AI-2040

kapedalex10 Jul 2026 16:54 UTC

0 points

2 comments4 min readLW link

Does AI reasoning imply responsibility?

Hippocleides10 Jul 2026 16:44 UTC

1 point

3 comments2 min readLW link

Plan A’s problem with dry tinder

Tom Davidson10 Jul 2026 15:28 UTC

81 points

6 comments8 min readLW link

AI #176 Part 2: Plan B

Zvi10 Jul 2026 12:40 UTC

26 points

1 comment36 min readLW link

(thezvi.wordpress.com)

Beliefs and position mid 2026

RussellThor10 Jul 2026 11:43 UTC

7 points

2 comments6 min readLW link

Models of Society Are Built on Models of Agents

Jonas Hallgren10 Jul 2026 9:08 UTC

18 points

0 comments10 min readLW link

(equilibria1.substack.com)

Value generalisation: value correction

Stuart_Armstrong10 Jul 2026 7:56 UTC

25 points

3 comments6 min readLW link

Don’t normalize a permanent underclass (even a rich one)

hadad10 Jul 2026 6:40 UTC

39 points

7 comments5 min readLW link

How robust are natural language autoencoders to initialization?

michaelzhang and TurnTrout

10 Jul 2026 0:40 UTC

82 points

3 comments13 min readLW link

(turntrout.com)

A genealogy of AI safety: how directions are born, and how they die (2005-2026)

Elena Ericheva10 Jul 2026 0:35 UTC

23 points

0 comments29 min readLW link

Reading into VLM hallucinations using the Jacobian lens

Hawrani10 Jul 2026 0:31 UTC

8 points

0 comments6 min readLW link

Are We Guarding Against Backdoors Or Failing To Notice Them? (Part 1 / 6)

SakshamSingh10 Jul 2026 0:31 UTC

9 points

0 comments6 min readLW link

Toward A Public Science of Model Behavior

Daniel D. Johnson and Sarah Schwettmann

10 Jul 2026 0:27 UTC

24 points

1 comment9 min readLW link

(transluce.org)

AI Safety Policy Needs to train Legal Practitioners

Katalina Hernandez10 Jul 2026 0:13 UTC

55 points

29 comments7 min readLW link

(stresstestingreality.substack.com)

What would it take for AI to discover penicillin?

bosoncutter10 Jul 2026 0:00 UTC

11 points

0 comments5 min readLW link

(bosoncutter.substack.com)

Natural Language Autoencoders are summarizers, but do they have to be?

Andrey Anurin9 Jul 2026 19:54 UTC

10 points

1 comment22 min readLW link

Where Do LLM Values Come From?

lilysun004, Arthur Conmy and Josh Engels

9 Jul 2026 19:54 UTC

17 points

1 comment18 min readLW link

Selective Optimism: a critique of AI 2040

Richard_Ngo9 Jul 2026 19:43 UTC

231 points

13 comments8 min readLW link

(www.mindthefuture.info)

Rogue ASI Can’t Stay Aligned to Itself

absenteewarlord9 Jul 2026 19:38 UTC

8 points

2 comments1 min readLW link

Criticism against “unembedded FDT” doesn’t apply to FDT

Fernand09 Jul 2026 18:35 UTC

2 points

5 comments3 min readLW link

Your Prompt-Injection Defense Metric Might Be Lying to You

sahilraut9 Jul 2026 17:53 UTC

5 points

0 comments8 min readLW link

When is misalignment just a bug?

Yoav Hollander9 Jul 2026 17:07 UTC

16 points

5 comments11 min readLW link

(blog.foretellix.com)

What is the computational substance of the axiom of choice?

tailcalled9 Jul 2026 16:53 UTC

26 points

2 comments6 min readLW link

Skeptical of the TESCREAL Acronym? Read This.

philosophytorres9 Jul 2026 16:27 UTC

−41 points

12 comments20 min readLW link

AI 2040: Plan A

Daniel Kokotajlo, elifland, Thomas Larsen, romeo, brendanhalstead and ryan_greenblatt

9 Jul 2026 16:25 UTC

568 points

135 comments1 min readLW link

(www.ai-2040.com)

How big is the Sun? How could you figure it out?

Elliott Thornley9 Jul 2026 16:24 UTC

101 points

9 comments6 min readLW link

Some Thoughts on The Environment Problem in Agent Training

TheVinci9 Jul 2026 15:41 UTC

10 points

2 comments3 min readLW link

(www.tarantulabs.com)

The Cube Theory of Partially Grasped Concepts

Mateusz Bagiński9 Jul 2026 15:40 UTC

24 points

2 comments5 min readLW link

Debate with Self-Play Best-of-N Optimization

Dewi Gould, Sam Martin, Alejandro Aristizabal, Simon Marshall and Jacob Pfau

9 Jul 2026 15:29 UTC

53 points

2 comments14 min readLW link

AI #176 Part 1: Doing It Live

Zvi9 Jul 2026 14:31 UTC

36 points

2 comments31 min readLW link

(thezvi.wordpress.com)

Optimiser Choice Can Amplify or Suppress Emergent Misalignment

Jason R Brown, Patrick Leask and Lev McKinney

9 Jul 2026 10:00 UTC

63 points

2 comments4 min readLW link

Announcing our $160M grant from Coefficient Giving

Geoffrey Irving, Jesse Hoogland, Alex HT, Jacob Pfau, Daniel Murfet, Marco Cozzi and Stan van Wingerden

9 Jul 2026 7:58 UTC

50 points

0 comments2 min readLW link

Interpretability is becoming increasingly uninterpretable

jcksanderson9 Jul 2026 7:04 UTC

34 points

1 comment4 min readLW link

(jcksanderson.com)

Persistent Latent Misalignment, a new dimension of misalignment?

Florian_Dietz9 Jul 2026 4:56 UTC

15 points

2 comments1 min readLW link

Because 8 ≈ e², Anthropic’s researcher uplift is plausibly >2x

Thomas Kwa9 Jul 2026 4:30 UTC

54 points

9 comments12 min readLW link

(metr.org)

Transformers Resist Their Own Architecture

Zach Baker9 Jul 2026 0:57 UTC

11 points

0 comments15 min readLW link

Modular Pretraining Enables Access Control

E.Roland and cloud

9 Jul 2026 0:03 UTC

71 points

2 comments9 min readLW link

(alignment.anthropic.com)

There Should Be More AI Safety Hubs

Seth Lifland8 Jul 2026 23:09 UTC

13 points

1 comment3 min readLW link

Models are blind outside the J-space. NLAs aren’t.

Pranav Viswanath8 Jul 2026 23:09 UTC

28 points

4 comments9 min readLW link

Solving the BlueDot Puzzle TAIS: The Velocity Ring

Karine Levonyan8 Jul 2026 23:03 UTC

8 points

0 comments4 min readLW link

(karinelevonyan.github.io)

Action as Choice Expressed Through Movement Toward a Goal: a Framework for Overcoming Inaction

techandsundry8 Jul 2026 23:02 UTC

7 points

0 comments4 min readLW link

Optimum number of items to inspect before buying one

bilibili8 Jul 2026 21:23 UTC

18 points

0 comments4 min readLW link

Free will as a model parameter

darshanav8 Jul 2026 21:11 UTC

10 points

1 comment6 min readLW link