RSS

lewis smith

Karma: 732

Nega­tive Re­sults for SAEs On Down­stream Tasks and Depri­ori­tis­ing SAE Re­search (GDM Mech In­terp Team Progress Up­date #2)

lewis smith, Senthooran Rajamanoharan, Arthur Conmy, CallumMcDougall, Tom Lieberum, János Kramár, Rohin Shah and Neel Nanda
Mar 26, 2025, 7:07 PM
113 points
15 comments29 min readLW link
(deepmindsafetyresearch.medium.com)

A Prob­lem to Solve Be­fore Build­ing a De­cep­tion Detector

Eleni Angelou and lewis smith
Feb 7, 2025, 7:35 PM
71 points
12 comments14 min readLW link

lewis smith’s Shortform

lewis smithAug 30, 2024, 9:51 AM
12 points
7 commentsLW link

The ‘strong’ fea­ture hy­poth­e­sis could be wrong

lewis smithAug 2, 2024, 2:33 PM
231 points
19 comments17 min readLW link

Im­prov­ing Dic­tionary Learn­ing with Gated Sparse Autoencoders

Senthooran Rajamanoharan, Arthur Conmy, lewis smith, Tom Lieberum, Vikrant Varma, János Kramár, Rohin Shah and Neel Nanda
Apr 25, 2024, 6:43 PM
63 points
38 comments1 min readLW link
(arxiv.org)

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

Neel Nanda, Arthur Conmy, lewis smith, Senthooran Rajamanoharan, Tom Lieberum, János Kramár and Vikrant Varma
19 Apr 2024 19:06 UTC
79 points
10 comments8 min readLW link

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

Neel Nanda, Arthur Conmy, lewis smith, Senthooran Rajamanoharan, Tom Lieberum, János Kramár and Vikrant Varma
19 Apr 2024 19:06 UTC
72 points
0 comments3 min readLW link

Dropout can cre­ate a priv­ileged ba­sis in the ReLU out­put model.

lewis smith28 Apr 2023 1:59 UTC
24 points
3 comments5 min readLW link