TurnTrout

Karma: 20,288

My name is Alex Turner. I’m a research scientist at Google DeepMind on the Scalable Alignment team. My views are strictly my own; I do not represent Google. Reach me at alex[at]turntrout.com

Ban development of unpredictable powerful models?

TurnTroutJun 20, 2023, 1:43 AM

46 points

25 comments4 min readLW link

Mode collapse in RL may be fueled by the update equation

TurnTrout and MichaelEinhorn

Jun 19, 2023, 9:51 PM

53 points

10 comments8 min readLW link

Think carefully before calling RL policies “agents”

TurnTroutJun 2, 2023, 3:46 AM

134 points

38 comments4 min readLW link 1 review

Steering GPT-2-XL by adding an activation vector

TurnTrout, Monte M, David Udell, lisathiergart and Ulisse Mini

May 13, 2023, 6:42 PM

437 points

98 comments50 min readLW link 1 review

Residual stream norms grow exponentially over the forward pass

StefanHex and TurnTrout

May 7, 2023, 12:46 AM

77 points

24 comments11 min readLW link

Behavioural statistics for a maze-solving agent

peligrietzer and TurnTrout

Apr 20, 2023, 10:26 PM

46 points

11 comments10 min readLW link

[April Fools’] Definitive confirmation of shard theory

TurnTroutApr 1, 2023, 7:27 AM

170 points

8 comments2 min readLW link

Maze-solving agents: Add a top-right vector, make the agent go to the top-right

TurnTrout, peligrietzer and lisathiergart

Mar 31, 2023, 7:20 PM

101 points

17 comments11 min readLW link

Understanding and controlling a maze-solving policy network

TurnTrout, peligrietzer, Ulisse Mini, Monte M and David Udell

Mar 11, 2023, 6:59 PM

333 points

28 comments23 min readLW link

Predictions for shard theory mechanistic interpretability results

TurnTrout, Ulisse Mini and peligrietzer

Mar 1, 2023, 5:16 AM

105 points

10 comments5 min readLW link

Parametrically retargetable decision-makers tend to seek power

TurnTroutFeb 18, 2023, 6:41 PM

172 points

10 comments2 min readLW link

(arxiv.org)

Some of my disagreements with List of Lethalities

TurnTroutJan 24, 2023, 12:25 AM

70 points

7 comments10 min readLW link

Positive values seem more robust and lasting than prohibitions

TurnTroutDec 17, 2022, 9:43 PM

52 points

13 comments2 min readLW link

Inner and outer alignment decompose one hard problem into two extremely hard problems

TurnTroutDec 2, 2022, 2:43 AM

149 points

22 comments47 min readLW link 3 reviews

Alignment allows “nonrobust” decision-influences and doesn’t require robust grading

TurnTroutNov 29, 2022, 6:23 AM

62 points

41 comments15 min readLW link

Don’t align agents to evaluations of plans

TurnTroutNov 26, 2022, 9:16 PM

48 points

49 comments18 min readLW link

Don’t design agents which exploit adversarial inputs

TurnTrout and Garrett Baker

Nov 18, 2022, 1:48 AM

72 points

64 comments12 min readLW link

People care about each other even though they have imperfect motivational pointers?

TurnTroutNov 8, 2022, 6:15 PM

33 points

25 comments7 min readLW link

A shot at the diamond-alignment problem

TurnTroutOct 6, 2022, 6:29 PM

95 points

67 comments15 min readLW link

Four usages of “loss” in AI

TurnTroutOct 2, 2022, 12:52 AM

46 points

18 comments4 min readLW link