Refusal in LLMs is mediated by a single direction

Andy Arditi, Oscar Balcells Obeso, Aaquib111, wesg and Neel Nanda

27 Apr 2024 11:13 UTC

19 points

0 comments9 min readLW link

Superposition is not “just” neuron polysemanticity

LawrenceC26 Apr 2024 23:22 UTC

25 points

0 comments13 min readLW link

An Introduction to AI Sandbagging

Teun van der Weij, Felix Hofstätter and Francis Rhys Ward

26 Apr 2024 13:40 UTC

28 points

0 comments8 min readLW link

AXRP Episode 29 - Science of Deep Learning with Vikrant Varma

DanielFilan25 Apr 2024 19:10 UTC

18 points

1 comment63 min readLW link

Improving Dictionary Learning with Gated Sparse Autoencoders

Neel Nanda, Senthooran Rajamanoharan, Arthur Conmy, lsgos, Tom Lieberum, Vikrant Varma, János Kramár and Rohin Shah

25 Apr 2024 18:43 UTC

60 points

23 comments1 min readLW link

(arxiv.org)

Simple probes can catch sleeper agents

Monte M, Carson Denison, Zac Hatfield-Dodds, David Duvenaud, Sam Bowman, Ethan Perez and evhub

23 Apr 2024 21:10 UTC

118 points

14 comments1 min readLW link

(www.anthropic.com)

Dequantifying first-order theories

jessicata23 Apr 2024 19:04 UTC

39 points

8 comments8 min readLW link

(unstableontology.com)

ProLU: A Nonlinearity for Sparse Autoencoders

Glen Taggart23 Apr 2024 14:09 UTC

29 points

2 comments8 min readLW link

Time complexity for deterministic string machines

alcatal21 Apr 2024 22:35 UTC

14 points

0 comments21 min readLW link

Inducing Unprompted Misalignment in LLMs

Sam Svenningsen, evhub and Henry Sleight

19 Apr 2024 20:00 UTC

35 points

6 comments16 min readLW link

[Full Post] Progress Update #1 from the GDM Mech Interp Team

Neel Nanda, Arthur Conmy, lsgos, Senthooran Rajamanoharan, Tom Lieberum, János Kramár and Vikrant Varma

19 Apr 2024 19:06 UTC

70 points

8 comments8 min readLW link

[Summary] Progress Update #1 from the GDM Mech Interp Team

Neel Nanda, Arthur Conmy, lsgos, Senthooran Rajamanoharan, Tom Lieberum, János Kramár and Vikrant Varma

19 Apr 2024 19:06 UTC

68 points

0 comments3 min readLW link

Discriminating Behaviorally Identical Classifiers: a model problem for applying interpretability to scalable oversight

Sam Marks18 Apr 2024 16:17 UTC

99 points

5 comments12 min readLW link

LLM Evaluators Recognize and Favor Their Own Generations

Arjun Panickssery, Sam Bowman and Shi Feng

17 Apr 2024 21:09 UTC

43 points

1 comment3 min readLW link

(tiny.cc)

Transformers Represent Belief State Geometry in their Residual Stream

Adam Shai16 Apr 2024 21:16 UTC

304 points

63 comments12 min readLW link

Speedrun ruiner research idea

lukehmiles13 Apr 2024 23:42 UTC

4 points

11 comments2 min readLW link

AXRP Episode 27 - AI Control with Buck Shlegeris and Ryan Greenblatt

DanielFilan11 Apr 2024 21:30 UTC

67 points

10 comments107 min readLW link

The theory of Proximal Policy Optimisation implementations

salman.mohammadi11 Apr 2024 13:00 UTC

3 points

1 comment6 min readLW link

(salmanmohammadi.github.io)

How I select alignment research projects

Ethan Perez, Henry Sleight and Mikita Balesni

10 Apr 2024 4:33 UTC

34 points

4 comments24 min readLW link

PIBBSS is hiring in a variety of roles (alignment research and incubation program)

Nora_Ammann, Lucas Teixeira and DusanDNesic

9 Apr 2024 8:12 UTC

47 points

0 comments3 min readLW link