Two Years On: Grading My Own Predictions

Accountability post. In 2024, with this series just starting, I put three claims on the record. Time to grade them in public — the standard I demand of every lab CEO applies to me too.

Claim 1: "AGI within ~18 months." Grade: B. Defined as "a system most professionals would call a genuinely general reasoning colleague," the reasoning-model era essentially delivered — o3's benchmark break, the agentic coding wave, models that plan and execute for hours. Defined as "no human cognitive advantages remain," no — and I concede the definition I used in 2024 was fuzzier than it should have been. Lesson kept: falsifiable predictions require falsifiable definitions. But the practical substance held: capability arrived faster than institutional consensus said it would, and everyone who planned against the skeptical consensus is now behind.

Claim 2: displacement before institutional response. Grade: A, sadly. The canaries data, the new-grad inversion, the layoff letters that say "AI" — versus, on the response side, essentially nothing at scale. No major income-floor pilot, no automation transparency law, no serious AI-era apprenticeship program in any large economy. I wanted to be wrong about this one. I was not.

Claim 3: planning beats pain. Grade: incomplete — by definition. This one is graded by what institutions do next, and the window I warned about in 2024 is now visibly closing rather than comfortably open.

The next two years, on the record

By mid-2028, I expect: agent fleets as normal middle-management infrastructure; humanoid deployments in the hundreds of thousands; entry-level restructuring spread from tech into law, finance, and healthcare admin; at least one significant political realignment — somewhere — driven explicitly by transition grievance; and the first real distribution mechanism (compute dividend, public infrastructure stake, or scaled income floor) operating somewhere in the OECD. And I expect the experiments — mine and the ten thousand others being run by displaced professionals with agent fleets — to have produced the first working playbooks ordinary people can copy.

The constant

Two years of essays reduce to one sentence, and it hasn't changed since the first post: the technology is not the decision — the distribution is. Machines that do our work can fund the most humane century we've ever built, or the most stratified. That fork is chosen by policy, ownership, and design — by us, on purpose, or by drift, by default.

Plan before the pain. There's still time. There is not still much time.

LLAP. 🖖