Jeff Wu @WuTheFWasThat Twitter profile

Last Seen Profiles

@_66O6

@VYNJAKU

@tinyfairytales

@sac4nagem

@irregularhuntX5

@70sr_

@NatyJimenezOax

@Punisher_doj

@stw_pdg

@lynabunz

@Ximena_solaress

@fS4LHAMINZ85305

@kk5661172894045

@yannis_FNESI

@SpotifyNews

@livefpltables

@npicton4

@L_Evans9

@estevin91

@babaalimusic

@marina_sissy92

@ekklisiaonline

@stw_pdg

@NilaNambiar0

@semo908

@bangjarijari

@yuri122939

@goldefakehansen

@tylrbrkr

@lilmisslunax

@sheep_yhh

@CaroleR94

@Mattiyo13131313

@tinyissueclub

@mh_rahbari1

@bokeplokalmalam

Jeff Wu

@WuTheFWasThat

5 months

Jan Leike

@janleike

5 months

To all OpenAI employees, I want to say: Learn to feel the AGI. Act with the gravitas appropriate for what you're building. I believe you can "ship" the cultural change that's needed. I am counting on you. The world is counting on you. :openai-heart:

241

412

Jeff Wu

@WuTheFWasThat

10 months

@StephenLCasper see Figure 13 :)

Jeff Wu

@WuTheFWasThat

10 months

@StephenLCasper See 5.1.2 + 5.1.3. In a pure denoising regime, you would see student-gt agreement>student-supervisor agreement, and positive student-supervisor agreement scaling with student compute. So we're at least partially in debiasing regime. We probably should've discussed this!

Jeff Wu

@WuTheFWasThat

7 months

@wesg52 Unsurprising I think. SAE errors will be correlated with directions relevant to loss, since the activation itself is optimized for loss. Only a very good SAE's errors would match noise - it'd have to hit the target with low precision rather than miss b/c of flaws in modeling

Jeff Wu

@WuTheFWasThat

5 months

@GretchenMarina @Miles_Brundage <3