Lenny's Podcast Video posted an update
How researchers check whether AI is actually honest | Emmanuel Ameisen (Anthropic)
Why it mattersThe description says Emmanuel Ameisen's team examined intermediate activations and saw the number 36 light up before the model read the math. It presents this as an alignment challenge: a model may represent the right answer yet be optimized to give a wrong one.
Discuss: In the AI interpretability example, what might the early activation of 36 reveal about the gap between a model's internal answer and what it says?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.