Lun

• LunJul 11, 2025, 7:03 PM

  6 points

  4 votes

  Overall karma indicates overall quality.

  2

  1 vote

  Agreement karma indicates agreement, separate from overall quality.

  in reply to: Cedar’s comment on: Lun’s Shortform

  fwiw I made my account in January, which I guess is still very new relative to average age of account here but hopefully means you can trust I didn’t make this account just to drop a link to the lumina post.

Lun’s Shortform

• LunJul 11, 2025, 4:38 AM

  26 points

  14 votes

  Overall karma indicates overall quality.

  1

  4 votes

  Agreement karma indicates agreement, separate from overall quality.

  on: Lun’s Shortform

  Someone has posted about a personal case of vision deterioration after taking lumina and a proposed mechanism of action. I learned about lumina on lesswrong a few years back, so sharing this link.

  https://​​substack.com/​​home/​​post/​​p-168042147

  For the past several months I have been slowly losing my vision, and I may be able to trace it back to taking the Lumina Probiotic. Or rather, one of its byproducts that isn’t listed in the advertising

  I don’t know enough about this to make an informed judgement on the accuracy of the proposed mechanism.

• LunJan 24, 2025, 10:13 AM

  1 point

  1 vote

  Overall karma indicates overall quality.

  0

  0 votes

  Agreement karma indicates agreement, separate from overall quality.

  on: Why mesa-op­ti­miza­tion is less likely un­der gra­di­ent de­scent than nat­u­ral selection

  Gradient descent generates updates by suggesting algorithmic improvements for single training examples, thereby exerting much less pressure for generality than evolution does.

  A recent technique Gradient Agreement Filtering filters out gradients that disagree between samples, which if I’m understanding correctly is intentionally breaking this crux and pushing for more generalization /​ less memorization of specific samples.
  
  Vague intuition that with typical LM pretraining which involves batches of far more than 1 sample + optimizers with momentum this might already not hold, non-generalizing /​ noisy /​ disagreeing updates are unimportant over the training run and the generalizing, agreeing updates stick around.