Artificial Intelligence
Models, agents, and AI–human collaboration — general-purpose capability scaling into every domain.
State of the world · updated June 2026
Right now: the frontier has converged — top models from a handful of labs cluster at similar capability, and the old benchmarks are saturating, so the contest has moved to reliability, cost, and what models can do as agents rather than to leaderboard points. Open-weight models trail that frontier by months, not years, and increasingly run on a single machine.
Watch: whether scaling keeps paying for itself or visibly plateaus; agents holding coherence across long, open-ended tasks; the next open release closing the gap to the closed frontier; and inference cost falling far enough that always-on AI becomes the default rather than a splurge.
Start here · the primer
Artificial intelligence stopped being a set of narrow tools the moment one trained model could be pointed at almost any task. The mechanism underneath is deceptively plain — predict the next token over enough data, and general capability emerges as a by-product of scale — and everything in this megatrend is a consequence of it: how far raw scaling goes before it bends (Frontier & AGI), what changes when a model is wired to take actions instead of just answer (Agents), and who is allowed to run the result (Open models). The question the field actually turns on is where the next increment of capability comes from — more compute, better feedback, or letting models touch the world — and what each one unlocks.