Human Intelligence is Not the Ceiling, But the Floor for AI

A common belief about AI goes like this. Large language models learn from text that people wrote. A student can’t learn more from a textbook than the textbook contains. So AI trained on the human record can, at best, match the people who wrote it. Human intelligence is the ceiling.

I think that’s backwards. Human intelligence is not the ceiling. It’s the floor.

Here is my argument in two steps. First, training on the human record gets AI to roughly human level. That is what imitation is good for, and it’s an enormous achievement. Second, once a system is there, other methods take it past: search, self-play, reinforcement learning against outcomes that can be checked, tools, scale and speed. None of those methods knows or cares where the human line is. Nothing obvious stops them at it.

That’s my argument. The rest of this post is the evidence, and I’ll keep the two separate.

The game that broke the ceiling

The clearest evidence is a decade old.

In October 2015, DeepMind’s AlphaGo beat the European Go champion 5–0. Its networks were trained “by a novel combination of supervised learning from human expert games, and reinforcement learning from games of self-play” (Nature, 2016). Human data first, then practice against itself. That is the two-step pattern in one system.

Then it played Lee Sedol, one of the best players in the world, and won 4–1. In game two it played Move 37, which DeepMind says had “a 1 in 10,000 chance of being used” and which “upended centuries of traditional wisdom” (DeepMind). It wasn’t a move copied from human games. It went against centuries of human practice, and it worked.

The next year DeepMind removed the human data entirely. AlphaGo Zero learned “without human data, guidance or domain knowledge beyond game rules,” and “starting tabula rasa” it beat the version that had beaten Lee Sedol 100–0 (Nature, 2017). DeepMind reported that it took three days of self-play (DeepMind).

Then AlphaZero did the same for chess, shogi and Go with one algorithm. “Starting from random play and given no domain knowledge except the game rules,” it beat a world-champion program in each game (Science, 2018). The chess programs it beat were built on evaluation functions “refined by human experts over several decades.” AlphaZero needed about nine hours of chess training (DeepMind). Garry Kasparov’s response was the right one: “we can actually learn from the new knowledge they produce.”

The lesson of Go is not that human data is useless. Human data got AlphaGo to the table. The lesson is that human data was not the limit. When the system could play, and the game said who won, it kept going.

Domains with an answer key

Games have a referee. So does a lot of science and mathematics.

Proteins. Decades of experimental work determined the structures of “around 100,000 unique proteins,” and predicting a structure from its sequence had been “an important open research problem for more than 50 years.” At the CASP14 assessment, AlphaFold reached “accuracy competitive with experimental structures in a majority of cases” (Nature, 2021). Its public database now holds more than 200 million predicted structures (AlphaFold DB), and Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry for it (Nobel Prize). AlphaFold learned from structures people had measured. What took it past them was a target that can be checked against reality.

Olympiad mathematics. AlphaGeometry solved 25 of 30 olympiad geometry problems by “synthesizing millions of theorems and proofs” instead of learning from human demonstrations (Nature, 2024). AlphaProof, “an AlphaZero-inspired agent that learns to find formal proofs through RL,” reached a silver-medal score at the 2024 International Mathematical Olympiad with multi-day computation (Nature, 2025). A year later, an advanced Gemini model scored 35 of 42, gold-medal level, in natural language, within the time limit, graded by IMO coordinators (DeepMind, July 2025). At IMO 2026 in July, Axiom Math reports that its AxiomProver solved all six problems for 42 of 42, with every proof written and checked in Lean, though its slowest problem took longer than a contest session (AxiomMath/IMO2026). That is Axiom’s own account, but the proofs are public and a machine checks them.

Research mathematics. The past month went further. I recorded it in Thirty Days of Watching the Singularity. Three results stand out:

• Fermat. A team of 1,000 agents running an early version of Claude Fable 5.2 reproduced a Lean proof of Fermat’s Last Theorem in 28.7 hours, against 11 days for the first formalization in September (Anthropic). To be precise, it reproduced an existing formalization, with a list of target statements and lemmas from the first run, and Anthropic calls it a toy experiment.
• The enormous theorem. Axiom formalized the Classification of Finite Simple Groups in about 43.1 million lines of Lean with dozens of agents in 24 days. The human proof is “roughly 15,000 pages in more than 500 articles by about 100 mathematicians,” written over decades (Axiom).
• A new theorem. PSI reports that a research system using a non-public Anthropic model proved the Aaronson-Ambainis conjecture, open since 2008, and that Lean accepted the proof (PSI). A human researcher scoped and ran the work and is responsible for its correctness, and the page doesn’t say it has been peer reviewed.

The first two are about speed and scale on mathematics humans had already done. The third is a result humans hadn’t reached.

The theory says this is the default

None of this surprised the people who built it.

In 2019, Rich Sutton wrote The Bitter Lesson. “The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin.” He names the two methods that “seem to scale arbitrarily”: search and learning. And he says what we should want: “AI agents that can discover like we can, not which contain what we have discovered.”

In 2025, David Silver, who led AlphaGo, and Sutton made the argument directly in Welcome to the Era of Experience. They grant the ceiling for imitation: “while imitating humans is enough to reproduce many human capabilities to a competent level, this approach in isolation has not and likely cannot achieve superhuman intelligence across many important topics and tasks.” Their answer is experience: agents that learn from what happens when they act, scored by “grounded rewards: signals that arise from the environment itself.” AlphaProof is their example. It started from “around a hundred thousand formal proofs” written by people and generated “a hundred million more” by working against a proof checker.

That is the floor and what’s above it, in two sentences. The human record gets you to competent. Grounded experience gets you past it.

Past the human baseline, now

This isn’t only about games and theorems. Agents are crossing human baselines on ordinary work.

Using a computer. OSWorld tests agents on real tasks on a real desktop. When its authors published it, people completed 72.36% of the tasks (OSWorld). On the corrected OSWorld-Verified leaderboard, comparing a single general model, one attempt per task, up to 100 steps, Claude Sonnet 4.5 scored 62.88% in a result dated October 31, 2025, and Claude Fable 5 scored 85.96% in a result dated August 1, 2026 (OSWorld-Verified). The human rate was measured on the original benchmark, not re-measured on the verified one, and the top results were submitted by the model’s maker. I covered this in Delegate All Your Grunt Work to AI Agents.

Senior engineering. Every’s Senior Engineer Benchmark, version 1.0 from June 2026, scores an agent out of 100 on rewriting a real production codebase “the way a senior engineer would.” Two human senior engineers set the reference at 89 and 96. Claude Fable 5 scored 91. It’s one long task, and harsher benchmarks put agents far lower, as I noted in Verification Is the New Programming. But the agent landed between two senior humans. One model generation earlier, Claude Opus 4.7 scored 33.5, as Every reported in How We Work Now.

What humans can’t do

Matching a person at a task is only part of it. Once an AI reaches human level, it has advantages no person has.

• Copies. Geoffrey Hinton put it plainly: digital computation can “run many copies of exactly the same model on different pieces of hardware” (Hinton). A senior engineer can’t be copied. A model that reaches senior level can be copied a thousand times by morning.
• Parallelism. The Fermat run used 1,000 agents at once. This site is built by swarms of agents across several harnesses, as I described in Building Multi-Harness, Multi-Agent Swarms.
• Instant sharing. Hinton again: “Thousands of identical digital agents can look at thousands of different datasets and share what they have learned very efficiently by averaging their weight changes.” People share what they learn by teaching, which he calls “a slow and painful process.”
• Memory. Human working memory holds about four chunks at once (Cowan, 2001). Current Claude models hold a context of 1 million tokens (Claude docs). The same documentation warns that recall degrades as the context grows, so it isn’t perfect memory. It is still a different scale.
• Speed. An adult reads non-fiction at about 238 words a minute (Brysbaert, 2019). AlphaZero went from random play to beating a champion chess program in about nine hours. A proof that took about 100 mathematicians decades to write was formalized in 24 days.
• Tireless search. AlphaProof generated a hundred million proofs. Search doesn’t get bored, tired or discouraged, and it doesn’t need to sleep.

Even if AI stopped at exactly human quality, these advantages alone would put it far beyond any human team in output. I argue it won’t stop at human quality, because the same advantages feed the methods that push past it.

The case for the ceiling

The ceiling argument deserves a fair hearing, because part of it is right.

Stochastic parrots. In 2021, Bender, Gebru, McMillan-Major and Shmitchell described a language model as “a system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data … without any reference to meaning: a stochastic parrot” (FAccT, 2021). If that is all a model is, its output is a remix of its input, and the input is us.

Imitation is capped by its data. A model trained only to predict human text, and tuned only by humans saying which answer they prefer, has no signal that points above the humans. Silver and Sutton make exactly this point: relying on rewards judged by humans “usually leads to an impenetrable ceiling on the agent’s performance,” and an agent without grounding “will become an echo chamber of existing human knowledge.”

Model collapse. Training on your own output can make things worse. Shumailov and colleagues showed that “indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear” (Nature, 2024). If AI can only learn from humans, and learning from itself degrades it, then human data really is the ceiling.

Here is where I agree. For pure imitation on tasks where nobody can check the answer, the ceiling view is roughly right. A model that writes the average human essay, graded by human taste, has nowhere obvious to climb.

Here is why it doesn’t hold in general.

Verification breaks the cap. Model collapse is about indiscriminate self-training. Self-play in Go isn’t indiscriminate: the game says who won. AlphaProof’s hundred million proofs weren’t indiscriminate: Lean checked every one. When a verifier filters the data, the model learns from its successes, not its mistakes. DeepSeek showed the same thing for language models: reasoning “can be incentivized through pure reinforcement learning,” “obviating the need for human-labelled reasoning trajectories,” and the result beat models trained on human demonstrations on “verifiable tasks such as mathematics, coding competitions and STEM fields” (Nature, 2025). That’s why I wrote that verification is the new programming. A check is what lets a system climb past its teachers.

Even imitation isn’t strictly capped. A transformer trained on chess games “can sometimes achieve better performance than all players in the dataset” (Zhang et al., 2024). The authors trace it to low-temperature sampling, which acts like a vote among many imperfect players and filters out their individual mistakes.

Search is not parroting. Move 37 wasn’t stitched together from human games. It came from search guided by a model, and it went against centuries of human practice.

The verifiable world is large and growing. Code runs or it doesn’t. Tests pass or they don’t. Proofs check or they don’t. Structures match experiments or they don’t. Every year, more of science and engineering gets a machine check, and every domain that gets one becomes a domain where the human line is a starting point.

If human-level is the floor

If I’m right, a few things follow. None of them requires hype.

For work. Stop asking whether AI can do a task as well as you. In verifiable work, that question is already being answered with a yes, and the next question is how far past you it goes. Plan for the agent being better than you at the checked parts of your job, and put your own time on what isn’t checked yet: what to build, why, and how we’ll know it worked.

For directing AI. The highest-leverage thing a person can give an agent is a verifier: a test, a proof checker, a benchmark, a clear definition of done. Instructions get an agent to human level. A check lets it go past. That is the job I described in Everybody Is a Director of AI Now: not doing the labor, but directing agents and deciding what counts as success.

For planning. Don’t plan around a plateau at human level. The evidence from Go, proteins, olympiads, research mathematics, computer use and software engineering points the same way: when the target can be checked, AI arrives at human level and keeps going. Where it can’t be checked, progress is slower, and that’s where human judgment still matters most, for now.

Human intelligence is how AI learned to start. It isn’t where AI has to stop.