The Mystery of Loss Functions: It's Deep Voting

iamtrask · x · 2026-08-22

The author predicts that solving AI control will reveal that the peaks and valleys in loss functions are not the magic of linear algebra, but rather algebra holding a mirror to millions of people in the training data. It is highly non-linear statistics, or "deep voting".

Related event: Solving AI control will demystify it back to statistics(3 posts)→

Original post →

More from AGI Musings

AGI Musings channel →