BRIEF №1 · AUGUST 2026

The 5-Year Disagreement

Insiders say 2027–28. Markets price 2033. Someone with real money is very wrong — and this week's evidence board leans one way.

BY SARVESH PATEL · AUGUST 2026

The spread

Ask the people building frontier AI when machines match top human experts across the board, and — per ex-OpenAI researcher Daniel Kokotajlo's July interview — the private answers cluster around 2027–28. Ask the broadest public forecasting market, Metaculus, and the community puts 25% odds on AGI by 2029 and a 50% mark out at 2033. Five years of disagreement about the most economically consequential event in history is not a philosophical curiosity. It is a mispricing, and every allocator holding AI-exposed assets — which is now every allocator — is implicitly on one side of it.

What the board says this week

Three series carry the argument. First, METR's task-horizon measure — the length of expert work a frontier agent completes autonomously — sits at 16–20 hours and its doubling time has compressedfrom seven months to roughly four. Metrics don't accelerate into a wall. Second, SWE-bench Verified, the standard coding benchmark, is saturated at 96%: coding automation is no longer capability-limited, only deployment-limited. Third, the money confirms the software: one frontier lab's revenue ran from roughly $1B to $60B in a year, and announced compute capex now runs past $100B. Capex is a hard-to-fake belief.

Against that: headline unemployment (US 4.2%, UK 5%) shows nothing, and the loudest public narrative remains "bubble." Note that the fast-timeline scenarios themselves predict exactly this quiet — labs automate their own research first, invisibly, before anything shows up in labor statistics. A lagging indicator failing to move is not evidence against a leading indicator that is moving.

The referee's note

We are not endorsing the insiders. Insiders have incentives — fear sells safety budgets, hype sells capacity. That is precisely why this desk scores rather than opines: as of this issue, every claim on the board — Kokotajlo's 50%-by-2029-30, the private 2027-28 whispers, Metaculus's 2033 — enters the public ledger with resolution criteria attached. In three years the arithmetic, not the rhetoric, will say who deserved your attention.

Allocation-relevant thresholds to watch

Not advice — arithmetic. (1) If the 4-month horizon doubling holds, week-long autonomous work arrives mid-2027; that is the threshold at which agentic labor stops being a demo and becomes a line item, with obvious implications for application-software margins. (2) Watch contracted gigawatts, not model announcements — power is the physical-world confirmation of the software story, and cancellations would be the first honest stall signal. (3) The tell on labor is entry-level hiring in exposed categories, not the headline rate — by the time the headline moves, the fast scenario says the repricing is over.