Does Opus 4.7 Generate Deceptive Denials About Its Own Guardrails?

usize9 May 2026 4:12 UTC

8 points

0 comments3 min readLW link

(usize.github.io)

Bad Problems Don’t Stop Being Bad Because Somebody’s Wrong About Fault Analysis

Linch9 May 2026 1:30 UTC

63 points

9 comments3 min readLW link

We Should Have Mandatory Media/Communications Training For All Communicators

Darren McKee8 May 2026 20:29 UTC

−5 points

6 comments3 min readLW link

Chess as a prediction model of the artificial intelligence impact on culture

8498 May 2026 20:19 UTC

−3 points

1 comment5 min readLW link

(lojkine.art)

The Saturation View: some responses

wdmacaskill8 May 2026 17:32 UTC

17 points

4 comments8 min readLW link

Is ProgramBench Impossible?

frmsaul8 May 2026 17:04 UTC

60 points

3 comments2 min readLW link

AI is Breaking Two Vulnerability Cultures

jefftk8 May 2026 15:50 UTC

51 points

0 comments2 min readLW link

(www.jefftk.com)

Write Cause You Have Something to Say

Logan Riggs8 May 2026 13:36 UTC

29 points

2 comments2 min readLW link

Userland Alignment

Josh H8 May 2026 13:31 UTC

2 points

0 comments2 min readLW link

A benchmark is a sensor

Håvard Tveit Ihle and Mathias Bynke

8 May 2026 13:24 UTC

25 points

0 comments3 min readLW link

Bringing More Expertise to Bear on Alignment

Edmund Lau, Geoffrey Irving, Cameron Holmes and David Africa

8 May 2026 10:29 UTC

55 points

0 comments8 min readLW link

Investigating the consequences of accidentally grading CoT during RL

papetoast8 May 2026 6:17 UTC

13 points

0 comments1 min readLW link

(alignment.openai.com)

The Frictionless Double

zw57 May 2026 23:11 UTC

8 points

4 comments8 min readLW link

Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations

Subhash Kantamneni, kitft, Euan Ong and Sam Marks

7 May 2026 20:21 UTC

186 points

26 comments8 min readLW link

Axes of Planning in LLMs + Partial Lit Review

NickyP7 May 2026 19:53 UTC

12 points

0 comments9 min readLW link

(blog.sus.cat)

A review of “Investigating the consequences of accidentally grading CoT during RL”

Buck7 May 2026 18:06 UTC

67 points

0 comments8 min readLW link

Try, even if they have you cold

WalterL7 May 2026 17:19 UTC

81 points

4 comments2 min readLW link

Mechanistic estimation for wide random MLPs

Jacob_Hilton7 May 2026 16:20 UTC

73 points

3 comments5 min readLW link

(www.alignment.org)

How to get better at chess (and everything else)

Sean Herrington7 May 2026 11:17 UTC

11 points

0 comments3 min readLW link

(www.chess.com)

Multipolar Civilisation Depends on Maintaining an Attacker’s Dilemma

Naci Cankaya7 May 2026 11:13 UTC

21 points

1 comment5 min readLW link

(nacicankaya.substack.com)