Thomas Kwa

Karma: 6,249

Member of technical staff at METR.

Previously: Vivek Hebbar’s team at MIRI → Adrià Garriga-Alonso on various empirical alignment projects → METR.

I have signed no contracts or agreements whose existence I cannot mention.

Tsinghua paper: Does RL Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?

Thomas KwaMay 5, 2025, 6:56 PM

68 points

21 comments2 min readLW link

(arxiv.org)

Should CA, TX, OK, and LA merge into a giant swing state, just for elections?

Thomas KwaNov 6, 2024, 11:01 PM

115 points

35 comments1 min readLW link

The murderous shortcut: a toy model of instrumental convergence

Thomas KwaOct 2, 2024, 6:48 AM

37 points

0 comments2 min readLW link

Goodhart in RL with KL: Appendix

Thomas KwaMay 18, 2024, 12:40 AM

12 points

0 comments6 min readLW link

Catastrophic Goodhart in RL with KL penalty

Thomas Kwa and Adrià Garriga-alonso

May 15, 2024, 12:58 AM

62 points

10 comments7 min readLW link

[Question] Is a random box of gas predictable after 20 seconds?

Thomas Kwa and habryka

Jan 24, 2024, 11:00 PM

37 points

35 comments1 min readLW link

[Question] Will quantum randomness affect the 2028 election?

Thomas Kwa and habryka

Jan 24, 2024, 10:54 PM

66 points

52 comments1 min readLW link

Thomas Kwa’s research journal

Thomas Kwa and Adrià Garriga-alonso

Nov 23, 2023, 5:11 AM

79 points

1 comment6 min readLW link

Thomas Kwa’s MIRI research experience

Thomas Kwa, peterbarnett, Vivek Hebbar, Jeremy Gillen, Bird Concept and Raemon

Oct 2, 2023, 4:42 PM

174 points

53 comments1 min readLW link

Catastrophic Regressional Goodhart: Appendix

Thomas Kwa and Drake Thomas

May 15, 2023, 12:10 AM

25 points

1 comment9 min readLW link

When is Goodhart catastrophic?

Drake Thomas and Thomas Kwa

May 9, 2023, 3:59 AM

180 points

29 comments8 min readLW link 1 review

Challenge: construct a Gradient Hacker

Thomas Larsen and Thomas Kwa

Mar 9, 2023, 2:38 AM

39 points

10 comments1 min readLW link

Failure modes in a shard theory alignment plan

Thomas KwaSep 27, 2022, 10:34 PM

26 points

2 comments7 min readLW link

Utility functions and probabilities are entangled

Thomas KwaJul 26, 2022, 5:36 AM

15 points

5 comments1 min readLW link

Deriving Conditional Expected Utility from Pareto-Efficient Decisions

Thomas KwaMay 5, 2022, 3:21 AM

24 points

1 comment6 min readLW link

Most problems don’t differ dramatically in tractability (under certain assumptions)

Thomas KwaMay 4, 2022, 12:05 AM

8 points

0 comments3 min readLW link

The case for turning glowfic into Sequences

Thomas KwaApr 27, 2022, 6:58 AM

87 points

29 comments5 min readLW link

[Question] (When) do high-dimensional spaces have linear paths down to local minima?

Thomas KwaApr 22, 2022, 3:35 PM

12 points

7 comments1 min readLW link

How dath ilan coordinates around solving alignment

Thomas KwaApr 13, 2022, 4:22 AM

65 points

46 comments5 min readLW link

5 Tips for Good Hearting

Thomas KwaApr 1, 2022, 7:47 PM

25 points

10 comments1 min readLW link