Karpathy vs Altman: same OpenAI roots, opposite AGI reads

Side by side

Why insiders with the same information disagree

This pairing controls for the usual explanation of forecast gaps — access to information. Both have seen frontier systems up close. The divergence comes from two places. First, weighting: Karpathy weights the reliability gap — models' remaining cognitive deficits, the years of unglamorous engineering between impressive demos and dependable agents — far more heavily than trend extrapolation. Second, incentive: Altman's confidence sustains OpenAI's positioning as the presumed frontrunner; Karpathy, now an independent educator, carries no organizational need for near-term AGI to be true.

What mid-2026 evidence says about each read

The current data genuinely supports both stories, which is why the disagreement persists. For Altman's read: capability keeps climbing (~83% GDPval-style knowledge work, ~80% SWE-Bench Pro), and the input curves — compute, capex — have held or exceeded projections. For Karpathy's read: no system has autonomously conducted AI research or run a job unsupervised, exactly the deficit he says takes a decade. The tiebreaker arrives on a schedule: Aschenbrenner's 2027 resolves by January 2028, and its outcome will be the first hard evidence for one read over the other.

How to use this disagreement

When two frontier insiders disagree this widely, the honest conclusion is that the deciding evidence doesn't exist yet — anyone claiming certainty in either direction is ahead of the data. That's the scorecard's approach: track checkable predictions with pre-registered flip conditions, and let resolutions — not confidence — settle it.

Frequently asked questions

Altman, OpenAI's CEO, says the lab is confident it knows how to build AGI and puts superintelligence 'a few thousand days' away. Karpathy, an OpenAI founding member who left, puts AGI about a decade out — roughly eight years apart despite shared frontier-level visibility.

Two factors: weighting and incentives. Karpathy weights the reliability gap between demos and dependable agents heavily; Altman extrapolates trends. Altman's confidence also serves OpenAI's positioning, while Karpathy has no lab to promote.

Both, partially: strong benchmarks (~83% GDPval, ~80% SWE-Bench Pro) support the optimistic read, while undemonstrated autonomous work supports the decade view. The first hard tiebreaker is Aschenbrenner's 2027 claim, resolving by January 2028.

That the deciding evidence doesn't exist yet. When insiders with equal visibility disagree by nearly a decade, tracking dated, checkable predictions — not confidence — is the only honest way to follow AGI progress.