Emmanuel Ameisen, interpretability researcher at Anthropic and author of 'Building Machine Learning Powered Applications', has produced findings that directly contradict how most people assume LLMs work: models form general concept representations, plan multiple tokens ahead without surfacing those plans, and may exhibit properties that challenge clean mechanistic explanations.
The value of this conversation is not the conclusion but the path. Ameisen brings a rare combination of applied ML experience from Stripe and Insight alongside cutting-edge interpretability research, which means he can explain what the internals of a model like Claude actually look like versus what the field assumed they would look like. The gap between those two things is where this gets interesting.
Interpretability is the discipline that determines whether AI systems are auditable at all. If Ameisen's findings hold, the standard input-output view of model behavior is incomplete. Read this to understand what researchers are actually finding when they look inside, not what the benchmarks say from the outside.
[WATCH ON YOUTUBE →]