Four LLM loss functions → four flavors of LLM misalignment

Steven Byrnes10 Aug 2026 16:16 UTC

275 points

25 comments6 min readLW link

FAQ: Isn’t AGI coming too soon for reprogenetics to help?

TsviBT8 Aug 2026 8:54 UTC

147 points

68 comments14 min readLW link

The Long (Self-)Correction

Wei Dai24 Jul 2026 21:01 UTC

241 points

62 comments2 min readLW link

LLMs are (still) mostly powered by imitative learning, not RL

Steven Byrnes24 Jul 2026 14:26 UTC

227 points

39 comments9 min readLW link

AI 2040: Plan A

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

9 Jul 2026 16:25 UTC

551 points

130 comments1 min readLW link

(www.ai-2040.com)

Surprising facts about the slave trade

Joseph Miller26 Jun 2026 1:34 UTC

462 points

51 comments8 min readLW link

The worthlessness of vitamin D is mildly exaggerated

dynomight23 Jun 2026 19:39 UTC

190 points

28 comments20 min readLW link

(dynomight.net)

A Mechanistic Explanation of Prompt Injection (and why you should study roles)

charlescye and Jasmine C.

22 Jun 2026 14:09 UTC

398 points

60 comments16 min readLW link

The LLM shoggoth meme is weirder than you think

HedonicEscalator19 Jun 2026 23:35 UTC

259 points

48 comments7 min readLW link

(hedonicescalator.substack.com)

Guardian Angels: LLM Personalization for Productivity and Security

gwern17 Jun 2026 3:21 UTC

192 points

40 comments2 min readLW link

(gwern.net)

Estimating No-CoT Task-Completion Time Horizons of Frontier AI Models

Anders Cairns Woodruff, Francis Rhys Ward, Dewi Gould, Rauno Arike, Jason R Brown, Jo Jiao, wlanderson, ariana_azarbal, harrymayne, Patrick Leask, Twm Stone, Josh Hills, Ida Caspary, Shubhorup Biswas and Julian Stastny

10 Jun 2026 17:58 UTC

278 points

23 comments4 min readLW link

Trees are mostly made of air and a generalizable lesson for AI safety

Zephaniah Roe29 May 2026 4:08 UTC

324 points

62 comments4 min readLW link

Mnemonic portraits for 19,023 human genes

Brinedew28 May 2026 22:16 UTC

361 points

29 comments15 min readLW link

Models finding software vulnerabilities is not the primary source of cybersecurity risk

Dean Valentine14 May 2026 3:39 UTC

315 points

25 comments2 min readLW link

Empowerment, corrigibility, etc. are simple abstractions (of a messed-up ontology)

Steven Byrnes11 May 2026 17:48 UTC

193 points

74 comments16 min readLW link

Who Got Breasts First and How We Got Them

rba11 May 2026 13:11 UTC

165 points

59 comments10 min readLW link

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

Linch9 May 2026 1:30 UTC

268 points

76 comments3 min readLW link

x-risk-themed

kave6 May 2026 15:16 UTC

251 points

26 comments3 min readLW link

(kaverennedy.substack.com)

Irretrievability; or, Murphy’s Curse of Oneshotness upon ASI

Eliezer Yudkowsky4 May 2026 22:11 UTC

370 points

132 comments22 min readLW link

How Go Players Disempower Themselves to AI

Ashe Vazquez Nuñez1 May 2026 23:24 UTC

744 points

79 comments8 min readLW link