[jostylr]

I'd be curious about their analysis of the Nvidia self-driving car project which uses world models(?) to train them in far more extensive scenarios, though simulated, than possible in real world case. That keynote was after this article of course.

But I did check their dismissive claim about the 90% coding at anthropic by watching the link they provided. The Anthropic guy said that 90% was achieved at various teams within Anthropic and also hedged about the exact nature of it; it is a messy metric to be precise about. I thought the other was not generous in interpreting it which makes me skeptical of the edge of the rest of the article.