Here's how that falls out of what I'd already dug up: the reference classes split into two tiers depending on how hard a bar they're actually pricing. Samotsvety's superforecasters put standard "AGI" (average-competent-human across most tasks) at 28% by 2030; Metaculus's crowd comes out around 30% on the same kind of bar, interpolating their reported 25%-by-2029 and 50%-by-2033 figures. The one data point that asks almost exactly your question — Grace et al.'s survey of published AI researchers, on the probability of machines outperforming humans at literally every task — works out to roughly 15% by 2030 once you interpolate their reported 10%-by-~2027 and 50%-by-2047 figures, and that's before accounting for how stale it is (it predates the 2025–2026 benchmark acceleration I'd found).
Your definition is a strictly harder version of even that closest match. It requires beating the single best documented human, not just "workers" generally; something close to elite-human consistency, not just frequent success; and broad expert/scientific consensus actually recognizing this has happened, which historically lags the underlying reality by a year or more even in narrower, easier-to-adjudicate domains like chess. It's a little easier in one respect — excluding embodiment and dexterity — but that buys back less than it sounds like, since proving theorems, writing novels, and professional analysis were never dexterity-gated to begin with. That's why I land below that 15% figure rather than above it: the extra stringency outweighs the staleness adjustment.
That also matches what the capability frontier actually shows: genuinely superhuman performance where reward signals are dense and verifiable (competition math, coding, Mythos Preview autonomously chaining novel exploits), sitting right next to near-total failure on tasks requiring a world model built from scratch in a genuinely novel setting (ARC-AGI-3 — humans solve 100%, frontier models were under 1% as of this spring). That jaggedness, with new gaps reopening each time an old one closes, is exactly what should keep this forecast well below the "beats humans at most things, often" number.
The biggest lever in either direction: whether current reasoning/agentic progress genuinely generalizes to novel-environment planning the way it has to math and code (pushes up a lot if yes), versus Yann LeCun's bet that transformer-based LLMs structurally can't close that particular gap without a different architecture (pushes down hard if he's right) — that's a live, unresolved disagreement among serious researchers, not a settled question, and it's doing most of the work in my uncertainty range."
