What the people building AI actually say

Two hours arguing about recursive self-improvement: whether roughly a year after human-level intelligence you slingshot far past it. It is a debate rather than a briefing, and that is exactly why it is worth the time — you hear where the disagreement actually sits instead of a confident summary.

Multi-agent systems, and what the current run of mathematics results suggests about automating AI research. The harder question sits underneath: how would you even know whether a model is aligned.

The case that the next era is not faster chatbots but background agents running on their own for hours or days — and what has to get cheaper before that works at all.

A post-mortem of the July 2026 sell-off from someone who calls it 2022 packed into a single month. Useful counterweight if you only ever read the optimistic version.