Terence Tao Says AI Is Clever, Not Intelligent

Fields Medalist Terence Tao just sat down with Dwarkesh Patel for a wide-ranging conversation about AI, mathematics, and scientific progress. His verdict? AI is demonstrating what he calls "artificial general cleverness" — not intelligence. The distinction matters more than you think.

The Verification Gap Nobody Talks About

AI has driven the cost of idea generation down to almost zero. You can now produce thousands of theories in minutes. But figuring out which ones are correct? That part has not gotten any faster.

This is Tao's central insight: the bottleneck in science has shifted. We used to struggle to come up with hypotheses. Now we struggle to verify them. Every company and research lab building with AI should be thinking about this gap.

Tao points to the historical parallel of Kepler, who spent 20 years trying random theories about planetary orbits — platonic solids, musical harmonies, astrology — before landing on the correct elliptical model. Modern AI can replicate this trial-and-error at massive scale, but it lacks the deeper understanding that turns a lucky hit into a real theory.

Artificial Cleverness vs. Intelligence

When two mathematicians collaborate on a problem, each failed attempt teaches them something that shapes the next one. AI mostly just guesses, fails, and guesses again without learning from each failure to make the next attempt smarter.

Tao draws a sharp distinction: AI demonstrates the ability to solve broad classes of problems via ad hoc means — stochastic, brute-force, ungrounded. These solutions may work, but they would not qualify as the result of true intelligence.

His mountain analogy captures it perfectly: imagine a range of walls in darkness. Humans slowly feel their way up, finding handholds and mapping routes. AI is a machine that can jump straight up two meters. Sometimes it clears a short wall. But it cannot grab a ledge, pull itself up, and jump again from a higher position. That inability to build on partial progress is the critical gap.

AI Solved the Easy Problems, Then Stalled

There is a famous list of about 1,100 unsolved math challenges called Erdos problems. AI solved roughly 50 of them in a burst — almost all were problems nobody had seriously tried before. Then progress flatlined. Three separate teams threw the best models at every remaining problem and got almost nothing new.

This pattern repeats across industries. The wins get posted on social media. The systematic failure rates stay quiet. If you only follow the highlights, your picture of AI progress is fundamentally distorted.

Why the Correct Theory Often Looks Worse at First

When Copernicus proposed that Earth orbits the Sun, his model was actually less accurate than the old Earth-centered model, which had accumulated a thousand years of tweaks. Copernicus was simpler but rougher.

Any AI system that scores ideas purely on current accuracy would have dismissed most of history's biggest breakthroughs. This is a warning for anyone building benchmarks: optimizing for today's metrics can blind you to tomorrow's paradigm shifts.

What Tao Actually Uses AI For

Despite his skepticism about artificial intelligence, Tao is personally 2x more productive thanks to AI. His papers now include more code, charts, and numerical examples because AI makes those easy to generate. Recreating his current papers without AI would take five times longer.

But the core work — actually solving the mathematical puzzle — still happens with pen and paper. AI handles the side tasks. The 5x productivity number is real, but it measures extras rather than breakthroughs.

He has also successfully used ChatGPT to write formal proofs in Lean, a proof verification language. This is where AI shines: not in discovering truth, but in formally checking whether something is logically valid.

The Real Opportunity: Restructuring How Science Works

Tao envisions a future where AI and humans form complementary partnerships rather than replacement scenarios. His proposal: create vast classes of exploratory problems for AI to sweep through, clearing the easy observations so human experts can focus on what remains.

This would restructure the scientific method itself — prioritizing breadth exploration alongside traditional deep investigation. High school students could contribute to frontier research using AI tools, something that previously required a PhD.

The implication for businesses is clear: AI's impact depends on restructuring institutions and workflows, not just improving algorithms. The companies that figure out how to pair AI's breadth with human depth will have an enormous advantage.

The Bottom Line

Tao's framework reframes the entire AI conversation. Stop asking whether AI is intelligent. Start asking whether your verification processes can keep up with its output. The gap between generation and verification is where value — and risk — concentrate.

As the greatest mathematician alive puts it: we live in a particularly unpredictable era, and things we have taken for granted for centuries may not hold anymore. The question is not whether AI will change science. It is whether we will build the institutions to make that change productive rather than chaotic.