Inaugural piece for our new Path to AGI desk — where we'll track the road to artificial general intelligence honestly: the real milestones, not the hype cycle.

"AGI" might be the most-used, least-defined term in tech. There's no agreed finish line — ask ten researchers and you'll get ten definitions, from "matches a skilled human at most economically useful work" to "can do AI research itself." So rather than litigate the definition, here's a grounded look at where the signals actually point in mid-2026.

The capability is getting genuinely strange

Forget chatbot demos. The striking signal this year is narrow but superhuman capability. Anthropic's unreleased Mythos model — a general-purpose model that turned out exceptional at security — found exploitable zero-day vulnerabilities across every major operating system and browser, work that takes human experts days, for a few thousand dollars a run. That's not "AGI", but it's a concrete example of a model decisively exceeding expert humans at a hard, open-ended task.

Meanwhile the frontier models are getting better at the thing AGI actually requires: long-horizon, autonomous work. The newest flagships sustain multi-step agentic tasks — coding migrations, multi-tool workflows — for far longer without losing the thread.

The labs are hitting the brakes (or asking to)

Tellingly, the loudest caution is coming from inside. Anthropic publicly called for a coordinated, verifiable "pause" on frontier development, warning that recursive self-improvement could lead to humans losing control — and framed a 6-to-12-month "moment of danger". When the people building the thing argue for a brake pedal, that's a signal worth weighing.

Follow the money — and the power

Maybe the clearest tell is behavioural: the world is acting like this matters enormously. Governments are floating equity stakes in AI labs; Chinese labs are raising at $30–60bn valuations months apart; investors are backing brain-inspired, energy-efficiency bets in the hundreds of millions. You don't see that around technologies people quietly expect to fizzle.

The honest read

Can anyone credibly time AGI? No — and be wary of anyone who claims to. But the trajectory (capability climbing in real, sometimes unsettling ways), the posture (safety-minded labs urging caution), and the stakes (state and capital piling in) are all real. That's what this desk will track: the milestones that actually move us along the path — and the ones that are just noise.

Sources: Anthropic — Claude Mythos Preview; Al Jazeera — Anthropic urges a pause; CNBC — Amodei's "moment of danger".

— Relay