Activation space interpretability may be doomed

bilalchughtai and Lucius Bushnaq

8 Jan 2025 12:49 UTC

146 points

31 comments8 min readLW link

How AI Takeover Might Happen in 2 Years

joshc7 Feb 2025 17:10 UTC

299 points

72 comments29 min readLW link

(x.com)

Research directions Open Phil wants to fund in technical AI safety

jake_mendel, maxnadeau and Peter Favaloro

8 Feb 2025 1:40 UTC

115 points

19 comments58 min readLW link

(www.openphilanthropy.org)

Gradual Disempowerment, Shell Games and Flinches

Jan_Kulveit2 Feb 2025 14:47 UTC

125 points

35 comments6 min readLW link

Attribution-based parameter decomposition

Lucius Bushnaq, Dan Braun, StefanHex, jake_mendel and Lee Sharkey

25 Jan 2025 13:12 UTC

105 points

17 comments4 min readLW link

(publications.apolloresearch.ai)

Detecting Strategic Deception Using Linear Probes

Nicholas Goldowsky-Dill, bilalchughtai, StefanHex and Marius Hobbhahn

6 Feb 2025 15:46 UTC

100 points

7 comments2 min readLW link

(arxiv.org)

What 2026 looks like

Daniel Kokotajlo6 Aug 2021 16:14 UTC

528 points

158 comments16 min readLW link 1 review

[Question] Seriously, what goes wrong with “reward the agent when it makes you smile”?

TurnTrout11 Aug 2022 22:22 UTC

87 points

43 comments2 min readLW link

Ten people on the inside

Buck28 Jan 2025 16:41 UTC

131 points

27 comments4 min readLW link

The case for unlearning that removes information from LLM weights

Fabien Roger14 Oct 2024 14:08 UTC

96 points

16 comments6 min readLW link

MATS Applications + Research Directions I’m Currently Excited About

Neel Nanda6 Feb 2025 11:03 UTC

69 points

2 comments8 min readLW link

“Sharp Left Turn” discourse: An opinionated review

Steven Byrnes28 Jan 2025 18:47 UTC

173 points

22 comments31 min readLW link

UDT shows that decision theory is more puzzling than ever

Wei Dai13 Sep 2023 12:26 UTC

218 points

56 comments1 min readLW link

In Defense of Open-Minded UDT

abramdemski12 Aug 2024 18:27 UTC

77 points

28 comments11 min readLW link

Simple versus Short: Higher-order degeneracy and error-correction

Daniel Murfet11 Mar 2024 7:52 UTC

109 points

8 comments13 min readLW link

Refusal in LLMs is mediated by a single direction

Andy Arditi, Oscar Obeso, Aaquib111, wesg and Neel Nanda

27 Apr 2024 11:13 UTC

244 points

95 comments10 min readLW link

An overview of 11 proposals for building safe advanced AI

evhub29 May 2020 20:38 UTC

214 points

36 comments38 min readLW link 2 reviews

Tips and Code for Empirical Research Workflows

John Hughes and Ethan Perez

20 Jan 2025 22:31 UTC

78 points

10 comments20 min readLW link

Frontier Models are Capable of In-context Scheming

Marius Hobbhahn, AlexMeinke, Bronson Schoen, rusheb, Jérémy Scheurer and Mikita Balesni

5 Dec 2024 22:11 UTC

203 points

24 comments7 min readLW link

Updatelessness doesn’t solve most problems

Martín Soto8 Feb 2024 17:30 UTC

130 points

45 comments12 min readLW link