Stephen Wolfram, creator of Mathematica and Wolfram|Alpha, has spent 40 years on the computational side of AI, which makes him one of the clearest voices on what today’s models actually can and can’t do. This conversation with Jon Hernandez is one of the most interesting AI conversations I’ve heard in a long time.

What stays with me: his heuristic that AI shines when getting it right 90% of the time is enough, and fails exactly where any single mistake is a disaster. Scaling won’t change that, since some problems are “computationally irreducible” and simply have to be run step by step. Less hype, more perspective.