Was Aschenbrenner right about AGI?

The scorecard at a glance

What he got right

The trendlines he bet on are mostly intact. Capital expenditure on AI infrastructure has exceeded even his aggressive projections. Effective compute scaling has held near his ~0.5 orders-of-magnitude-per-year pace (Epoch AI tracking). And frontier models now score around 83% on knowledge-work benchmarks like GDPval, with agentic coding near 80% on SWE-Bench Pro — broadly the "outpaces college graduates" bar he described.

What he got wrong

His clearest miss is the prediction that open-source models would fade and proprietary algorithms would form a durable US moat. As of mid-2026 the opposite is happening: open-weight models like DeepSeek V4 and Qwen 3.7 Max sit roughly 3–6 months behind the frontier, at a fraction of the cost, with genuine architectural innovation. This matters because a third of the essay's policy program ("Lock Down the Labs") rests on that assumption.

What's still undecided

The two biggest claims remain open. AGI by 2027 has 18 months left to resolve — agentic coding is strong, but no system has autonomously conducted AI research end-to-end. A formal US government AGI project hasn't materialized, though national-security involvement is growing.

Frequently asked questions

Partly. Two years in (mid-2026): 3 of his major predictions are on track, 1 is clearly wrong (open source fading), 2 are open, and 2 are too early to grade. His headline AGI-by-2027 prediction resolves by January 2028.

His prediction that open-source AI would fade and proprietary labs would hold a durable moat. As of mid-2026, open-weight models (DeepSeek V4, Qwen 3.7 Max) trail the frontier by only ~3–6 months at far lower cost.

By January 1, 2028. As of mid-2026 it is graded Open — agentic capability is strong but autonomous AI research is undemonstrated.