Andrej Karpathy's AGI prediction: about a decade out

What Karpathy actually predicts

Karpathy’s position is distinctive because it is bearish on timelines while bullish on the technology. He has repeatedly described the coming period as the “decade of agents” — pushing back on the idea that any single year is the year everything changes. In his telling, the path to AGI is real but grindy: models still make basic mistakes, lack robust memory and continual learning, and need years of unglamorous engineering before they can be trusted with real work end-to-end.

Why his view carries weight

Karpathy is not an outside skeptic. He was a founding member of OpenAI, led AI for Tesla’s Autopilot program, and remains one of the most-read educators in deep learning. When someone who has built frontier systems at two of the most aggressive AI organizations says “closer to ten years than two,” it is an insider’s read on the gap between benchmark performance and dependable autonomy — the same gap this scorecard tracks in its AGI-by-2027 verdict.

Karpathy vs. the field

On this spread, Karpathy sits between the frontier-lab optimists and the academic survey median: later than Hassabis (~50% by 2030) and Metaculus (50% by 2033) in spirit, but well before the survey’s 2040. Notably, his “about a decade” lands around the same window as Aschenbrenner’s superintelligence forecast — what one calls the start line, the other treats as the finish.

What the evidence says so far

Mid-2026 evidence gives both camps something. For the aggressive view: capability benchmarks are strong (~83% on GDPval-style knowledge work, ~80% on SWE-Bench Pro), and compute keeps scaling on trend. For Karpathy’s view: no system has autonomously conducted AI research end-to-end, and drop-in reliability still lags benchmarks — exactly the deficit he says takes a decade. The AGI-by-2027 prediction is still Open; if it resolves unfulfilled in January 2028, Karpathy’s caution will look prescient. If autonomous AI research is demonstrated before then, it will be the clearest possible refutation of the decade view.

Frequently asked questions

Karpathy puts AGI roughly a decade away. He describes the coming period as the “decade of agents” — arguing that today’s models still have real cognitive deficits and that reliable autonomous agents take years of engineering, not months.

Both extrapolate from the same technology, but Karpathy weights the gap between benchmark performance and dependable real-world autonomy far more heavily. Aschenbrenner’s 2027 assumes trendlines compound quickly into autonomous AI research; Karpathy expects a long grind of reliability engineering first.

No — he is bearish on short timelines, not on the technology. A founding member of OpenAI and former Tesla AI director, he expects transformative AI agents; he just thinks the transition takes about a decade rather than a couple of years.

Still unresolved. As of mid-2026, capability trends favor the aggressive camp (~83% GDPval, ~80% SWE-Bench Pro), but the autonomy bar that would settle AGI-by-2027 remains undemonstrated — which is the core of Karpathy’s argument. The 2027 claim resolves by January 2028.