Notas do episódio
Listen @ Apple Podcasts | Spotify.
1. AI’s Fake Confidence
In the clip above, Ilya Sutskever describes a familiar experience. You ask an AI to fix a bug, and it eagerly makes a change that introduces another one. When you point that out, it apologizes and brings back the first bug.
He suggests that reinforcement learning may make models too narrowly focused, alongside weaknesses in applying what they have learned to new situations. These possible explanations raise a broader question about how a system can produce impressive work yet give assurances its results do not support.
The previous…
Site: signaldrifts.substack.com

