Lawrence Feng

• Lawrence Feng 29 Aug 2026 4:03 UTC

  1 point

  0

  in reply to: Gonçalo Paulo’s comment on: Sublimi­nal Learn­ing Hap­pens at Every Rank, Given the Right Learn­ing Rate and Enough Data

  Could you share your hyperparams/​config/​dataset?

• Lawrence Feng 11 Jul 2026 8:17 UTC

  1 point

  0

  in reply to: dani roytburg’s comment on: Sublimi­nal Learn­ing Hap­pens at Every Rank, Given the Right Learn­ing Rate and Enough Data

  In all experiments described in the blog post, AdamW was used following prior work.

  I tried SGD and like the paper you referenced but I was unable to get it to work. I didn’t try very hard to get it to work, though.

• Lawrence Feng 16 May 2026 4:50 UTC

  4 points

  1

  on: Claude is Now Align­ment-Pretrained

  From what I understand, in “Teaching Claude Why” they explain that they are doing some sort of training on synthetic “alignment documents,” but there’s no indication that this is happening during pretraining. Sure, the intuition is to modify the model’s belief using pretraining-style documents, but there’s no intervention or modification during the training of the base model, as is done in Korbak or Maini’s prior work.