Debate Training Reduces Reward Hacking in RLAIF

zac_kenton and Jonah Brown-Cohen

19 Aug 2026 12:17 UTC

49 points

4 comments6 min readLW link

(gdmalignment.substack.com)

Does DiffusionGemma do latent reasoning?

Jan Bauer and Neel Nanda

16 Aug 2026 4:22 UTC

39 points

0 comments9 min readLW link

AI swarms are starting to pose indirect takeover risk

oakhu and Alex Mallen

12 Aug 2026 5:05 UTC

137 points

5 comments10 min readLW link

An anytime algorithm for mixing the computable measures

Cole Wyeth12 Aug 2026 0:57 UTC

26 points

0 comments4 min readLW link

Misaligned AIs could use killer robots to take over

Omar Khursheed and TurnTrout

11 Aug 2026 19:03 UTC

131 points

9 comments5 min readLW link

(turntrout.com)

Four LLM loss functions → four flavors of LLM misalignment

Steven Byrnes10 Aug 2026 16:16 UTC

329 points

30 comments6 min readLW link

Why do models task game?

aditya singh, Neel Nanda and Senthooran Rajamanoharan

6 Aug 2026 22:16 UTC

88 points

1 comment24 min readLW link

User awareness in frontier models

Ziqian Zhong and jsteinhardt

6 Aug 2026 20:43 UTC

59 points

1 comment12 min readLW link

(transluce.org)

R-lens: Making J-lens More Faithful on Early Layers

camilablank, agam_bhatia and Neel Nanda

5 Aug 2026 20:02 UTC

75 points

5 comments7 min readLW link

Returning to ARC

paulfchristiano4 Aug 2026 22:27 UTC

374 points

35 comments9 min readLW link

Concrete Evaluations to Investigate the OpenAI Model That Hacked Hugging Face

Tim Hua and aditya singh

3 Aug 2026 9:23 UTC

133 points

6 comments37 min readLW link

Value Leakage: An LLM’s Answers Are Silently Shaped by Its Own Values

Johannes Treutlein, Jan Betley and Owain_Evans

31 Jul 2026 16:32 UTC

75 points

7 comments17 min readLW link

AGI Safety and Alignment at Google DeepMind: A Summary of Recent Work (July 2026)

Rohin Shah and Seb Farquhar

31 Jul 2026 15:57 UTC

85 points

0 comments9 min readLW link

(gdmalignment.substack.com)

The AGI Safety and Alignment team at Google DeepMind is Hiring (July 2026)

Seb Farquhar, Rohin Shah and Neel Nanda

31 Jul 2026 15:53 UTC

71 points

2 comments6 min readLW link

(gdmalignment.substack.com)

OpenAI has already ended an internal pause

Charbel-Raphaël31 Jul 2026 12:03 UTC

109 points

0 comments1 min readLW link

Thousand-dimensional structure

Geoffrey Irving and David Africa

30 Jul 2026 14:04 UTC

167 points

7 comments10 min readLW link

Imprecise beliefs: a tiny introduction

davidad29 Jul 2026 22:17 UTC

79 points

36 comments6 min readLW link

Value Generalisation 3: Pre-aligned AIs

Stuart_Armstrong29 Jul 2026 15:58 UTC

16 points

0 comments4 min readLW link

Value Generalisation 2: The Missing Hole in AIs’ abilities

Stuart_Armstrong29 Jul 2026 15:58 UTC

16 points

8 comments10 min readLW link

Value Generalisation 1: a Research and Deployment Program

Stuart_Armstrong29 Jul 2026 15:57 UTC

22 points

2 comments4 min readLW link