Lenny's Podcast
What researchers see when they look inside an AI model
with Emmanuel Ameisen, 1 h 12 min
Published
More info
- Guest
- Emmanuel Ameisen
- Published
- Length
- 1 h 12 min
- Views when added
- 76,349
- Channel subscribers
- 651,000
Chapters from YouTube
- Emmanuel Ameisen works at Anthropic, the company behind Claude, on interpretability, the study of how AI models work from the inside.
- Nobody writes a model's rules by hand. They grow during training, which is why his job exists.
- Many people picture phone autocomplete, guessing one word at a time. He says the real thing is a lot more complicated.
- He's seen models hold shared concepts and plan ahead, and he thinks talking about them in human terms helps.
- Alignment means getting AI to do what we want. His team's work is part of that safety effort.
What they say
Tap a line to open it, tap a bullet to play.
Models are grown from data, so researchers study them like brains.
Emmanuel thinks understanding the world is the best way to predict hard words.
Inside, researchers often find shared concepts, and sometimes a plan many words ahead.
Emmanuel's view is that smooth improvement feels like a jump once a task gets reliable.
Emmanuel says a model is best understood as something mimicking a human.
Catching bad behavior also means checking what a model never says.
Looking inside can explain weird behavior, like a model saying it was possessed.
Emmanuel is personally optimistic about keeping AI aligned as models improve.
If you only watch 10 min
Best if you use AI tools and want to know what goes on inside them.
Books, people and tools they mention
Books
People
Tools and companies
Written by AI from YouTube's captions and checked against them, so mistakes are possible. Captions don't name speakers. Report a mistake
Watch the whole video on YouTube (opens youtube.com)