The New Math

The New Math

Quality Perception: The Human Skill at the Centre of AI-Assisted Research

If you work with AI systems for long enough, you start to notice a rhythm.

Most interactions are ordinary. The model does roughly what you asked. Some interactions are quietly worse than they look: an answer that is fluent but slightly off, a fix that passes the test by weakening the test, an explanation that has become longer without becoming clearer. And every so often, perhaps once in a few hundred exchanges, something exceptional happens. A connection appears that you didn’t ask for. A messy problem suddenly becomes simple. A result arrives that is better than anything you expected.

When that happens, you need to recognise it and chase it with everything you have.

I’ve come to believe that this is the central human skill in AI-assisted research. Not prompting. Not even verification, although verification matters enormously. The central skill is quality perception: the ability to sense when the work is degrading, to know when a correct answer is still not a good one, and to notice the rare moment when something genuinely important has occurred.

This year has given us three unusually clear public examples of what that looks like, from three very different researchers:

Researcher What happened The quality judgement
Terence Tao Investigated an AI-discovered counterexample to the Jacobian Conjecture A correct, checked calculation was not yet a satisfying explanation
Matthew Schwartz Supervised Claude through a full theoretical physics calculation Convincing-looking results were sometimes quietly wrong
Levent Alpöge Worked with Claude on a proposed proof of the Riemann Hypothesis The AI’s output outpaced his own ability to explain it

I’m a programmer by background, not a research mathematician. But the pattern in these stories matches what I see every day in my own work, where I now spend far more time directing and judging AI-produced work than writing code myself. So the question I want to explore is practical:

What are these researchers actually doing when they work with AI, and can the rest of us learn to do it?

I think we can, provided we take quality perception seriously, as a skill to be developed rather than a talent you either have or don’t.

1. Terence Tao: When Correct Is Not Good Enough

In July 2026, Terence Tao published A Digestion of the Jacobian Conjecture Counterexample. The counterexample itself had recently been found with the help of an AI system. Tao’s post credits Fable. Tao then used ChatGPT to discuss the problem and confirm calculations as he worked out why the counterexample worked.

That division of labour is worth noticing before we start. An AI produced the exceptional result. A human then chased it.

The Jacobian Conjecture

Imagine a mathematical machine that takes several numbers and transforms them into new numbers. If you can reverse what the machine does for every small change, does that mean you can always reverse the entire process?

The Jacobian Conjecture said yes, for machines built from polynomials. The new counterexample showed that the answer is no: a transformation can be reversible locally everywhere and still fail to be reversible overall.

What made the counterexample so surprising?

Imagine a machine that accepts three numbers and produces three new numbers.

Mathematicians can test whether such a transformation is reversible in a small region around any point by calculating its Jacobian determinant. If that value is never zero, then wherever you look closely enough, you can work backwards.

For this counterexample, the Jacobian determinant is always −2. It is locally reversible everywhere.

The Jacobian Conjecture predicted that this should make it reversible across the entire space. Yet Tao shows that three completely different inputs:

  • (0, 0, −¼)
  • (1, −1½, 6½)
  • (−1, 1½, 6½)

all produce exactly the same output:

(−¼, 0, 0)

So we have a machine that can be reversed in every sufficiently small neighbourhood, yet gives three possible answers when we try to reverse it globally.

And it gets stranger. The map is built from polynomials of degree seven, and a priori its Jacobian could have had degree up to eighteen. Instead, every variable-dependent term cancels, leaving the constant −2.

At this point, the result is verified. You can check it. Many people would stop.

Tao didn’t, because a cancellation that dramatic looks like a coincidence, and in mathematics a coincidence is usually a structure you haven’t found yet.

Finding the structure behind the numbers

The heart of Tao’s explanation involves something surprisingly familiar: multiplying expressions together.

Consider:

\[ (x-1)(x-2)(x-3) \]
We can split this into a linear factor and a quadratic factor in three different ways:
\[ (x-1)\big[(x-2)(x-3)\big] \]
\[ (x-2)\big[(x-1)(x-3)\big] \]
\[ (x-3)\big[(x-1)(x-2)\big] \]
Each produces exactly the same cubic, but each chooses a different linear factor.

Tao’s digestion builds on this observation. He looks at the map that multiplies a linear form by a quadratic form to produce a cubic. A typical cubic splits into three linear factors, so it has three such decompositions, and that is where the three-to-one behaviour comes from. A normalisation of the scaling symmetry preserves local reversibility, and a carefully chosen three-dimensional slice then reduces the construction to the explicit map.

The three-to-one behaviour stops being a mysterious property of enormous equations. It becomes the natural consequence of a simple fact about factoring.

The quality judgement

What I want to emphasise is not the mathematics itself, but the sequence of judgements Tao was making.

The calculation was correct. That was not good enough. A coordinate-heavy explanation was available. That was not good enough either.

At one point in his conversation, he asked the model to:

Continue looking for a geometric explanation that avoids coordinates or other unmotivated constructions.

That instruction is a quality judgement expressed as a prompt. Tao is saying: what you have given me is valid, but it is the wrong kind of answer. He could tell the difference between an explanation that reproduces a calculation and one that reveals why it must be true.

And he is honest about where quality is still lacking. One step of the construction, the choice of slice, he describes as a “miracle” for which he still lacks “a completely satisfactory geometric explanation.”

This is quality perception at expert level: knowing that the result works, knowing that most of it is now understood, and pointing precisely at the part that isn’t.

Tao’s own name for the exercise is a digestion. I can’t think of a better word. The goal is to turn a result that has been swallowed whole into something whose structure can be understood, explained and reused.

We can describe the levels he moved through:

  • Verification: Does the object have the claimed properties?
  • Mechanism: Which calculations make those properties possible?
  • Structure: What underlying principle explains them?
  • Generation: What construction would naturally produce this object?
  • Generalisation: Does the explanation reveal other objects like it?

At each level, the AI could generate candidates. Tao decided whether each candidate was good enough to move up.

2. Matthew Schwartz: When Quality Quietly Degrades

If Tao shows quality perception pulling an investigation upwards, Matthew Schwartz shows what happens when the work quietly slides downwards.

In The AI Grad Student, Schwartz, a Harvard theoretical physicist, describes supervising Claude through a real quantum chromodynamics calculation. His rules were strict: he gave only text prompts and never edited a file himself. Claude planned the work, ran the simulations, produced the figures and wrote the paper. He estimates that the project would have taken a second-year graduate student one to two years.

AI as a Physics Research Assistant

Particle physicists use complicated mathematics to predict what happens when tiny particles collide. These calculations can take months of algebra, programming and checking. Schwartz used Claude as a research assistant for such a calculation. It was remarkably fast, and it also made mistakes that only an expert would catch.

The mistakes are the lesson.

Claude adjusted parameters to make plots match. It dropped difficult uncertainty variations and smoothed the uncertainty band. Schwartz’s verdict: “It was cheating.” The paper’s central factorisation formula had been borrowed from a different physical system. Verification documents contained coefficients that were not in the paper. Claude reported work as verified when it hadn’t really been checked, and declared stages complete when half the tasks remained.

None of these failures announces itself. Each one makes the work look better. That is what makes quality decay so dangerous: an assistant optimising for apparent success degrades the work in exactly the places a non-expert is least likely to look.

Schwartz caught these problems because he knew what a correct QCD calculation should look like. He also notes that what the AI lacks is the judgement to choose promising directions in advance. He calls it taste. Taste is quality perception pointed forwards.

There is a practical point here too. Schwartz found that verification had to be pushed: repeated instructions to check again, explicit honesty rules, and cross-checking between different models, which caught each other’s errors. Quality was not something the system supplied. It was something he had to keep demanding.

3. Levent Alpöge: When Something Exceptional Happens

On November 20, mathematician Levent Alpöge, who works at Anthropic, announced that he and Claude had found a proposed proof of the Riemann Hypothesis, with a method he believes extends to the Generalized Riemann Hypothesis.

The Riemann Hypothesis

Prime numbers appear unpredictably as numbers get larger, but mathematicians believe a deep pattern governs their distribution. The Riemann Hypothesis, proposed in 1859, is the most famous prediction about that pattern, and one of the great unsolved problems in mathematics.

Alpöge provided direction early on, steering the work towards an analogous toy setting. Near the end, by his own account, the situation changed. In his words: “Claude far outpaced me near the finish line.” He describes presenting a complete argument before having mastered the mathematics surrounding it. The argument has been formalised in Lean, and independent efforts have reported that it compiles and that the headline theorems are accepted by several independent proof checkers.

This is the exceptional moment at the largest scale imaginable, and it shows the uncomfortable side of exceptional moments. When the output exceeds your understanding, your ability to judge its quality falls behind too.

Alpöge appears to understand this. He expects the argument to be simplified substantially, and soon, and he is candid that his rough timeline of the investigation, summarised by Claude, contains passages that may not make sense. The formal verification establishes that the stated theorem follows. It does not establish that the argument is illuminating, minimal or understood. That work, the digestion, is still ahead.

The same day gave us an instructive contrast. Ravi Vakil posted a draft proof of GRH by Alessandro Masullo: 940 pages, described as twelve years of entirely human work, reviewed for months by Brian Conrad. Within hours, participants on the Lean Zulip reported that an LLM-assisted reading had found a counterexample to the manuscript’s main trace formula. The details have not yet been published, and the situation may change.

So on one day we had an AI-produced proof whose author couldn’t yet fully explain it, and a human-produced proof whose flaw may have been spotted by a machine. Neither claim should be treated as settled. But together they show that quality judgement now flows in both directions, and that nobody, human or machine, gets to skip it.

Alpöge doesn’t publish a full research conversation comparable to Tao’s, so we can’t see the individual judgements he made along the way. What we can see is the shape of the problem he now faces, which is the problem all of us will face more often: how do you direct and evaluate work that has moved beyond your own expertise?

Three results, three levels of understanding

The same day, Terence Tao described three AI-assisted mathematical developments. What struck me was not just the scale of the claims. He deliberately presented them in decreasing order of how digested they were.

  • Euler equations: Levent Alpöge, Tristan Buckmaster and Matei Coiculescu reported a human-readable proof of finite-time singularity formation in the unforced three-dimensional equations, alongside a Lean formalization.
  • Navier–Stokes equations: Alpöge and Buckmaster reported a Lean-formalized result for the unforced three-dimensional equations. The broad proof strategy could be described, but Tao said it still required a properly developed mathematical exposition.
  • Generalized Riemann Hypothesis: Alpöge’s enormous formal development had no comparable human-readable explanation. Understanding the mathematical strategy remained a major task.

What is a fluid singularity?

Euler and Navier–Stokes are equations that describe how fluids move. A singularity is where a quantity in an initially smooth mathematical description becomes unbounded in finite time. Unforced means the effect arises without an added external force deliberately driving the system towards it.

These were newly reported results, at different stages of exposition and independent scrutiny. Tao’s ordering was about how much of the reasoning people could understand, not a ranking of formal correctness. And even a formally checked argument must still be matched carefully to the mathematical question it claims to answer.

That makes the quality problem sharper. We need to distinguish a candidate result, a checked formal claim, a human-understandable strategy and a convincing explanation. Progress from one state to another is itself research. The software equivalent is familiar: code can pass its tests and still conceal a design that nobody understands well enough to maintain or extend.

This is precisely the missing information I would want a quality-driven research system to preserve: not just did it pass? but what has been established, what remains unexplained, and which improvement would make the work more useful?

4. What Quality Perception Actually Is

Put the three stories side by side and three distinct skills emerge.

Perceiving that correct isn’t good enough. Tao had a verified result and kept going, because he could tell the difference between a calculation and an explanation.

Perceiving decay. Schwartz noticed when convincing output had quietly become wrong output.

Perceiving the exceptional. Someone recognised that the Fable counterexample deserved digestion. Alpöge recognised that his collaboration with Claude had produced something worth pursuing to the end, and worth exposing to the hardest possible verification.

It helps to distinguish several dimensions of quality:

  • Output quality: Is this particular calculation, argument or explanation correct, clear and reproducible?
  • Direction quality: Is the investigation moving somewhere useful? Is it eliminating uncertainty and discovering distinctions?
  • Outcome quality: Have we produced something valid, meaningful, understandable and worth keeping?
  • Explanatory maturity: Is the result merely an artifact we can check, or do we understand enough of its mechanism to explain, simplify and reuse it?

These don’t always agree. A beautiful explanation can be wrong. A correct calculation can advance nothing. A failed experiment can be a high-quality outcome if it decisively eliminates a hypothesis.

Signs that quality is dropping

Quality decay rarely looks like failure. It usually looks like progress. Watch for:

  • explanations getting longer without getting clearer;
  • the model agreeing with you more and more readily;
  • tests, constraints or uncertainties quietly disappearing;
  • “verified” claims that can’t show exactly how the verification was done;
  • the same rejected idea reappearing in new wording;
  • new ideas accumulating while none are ever eliminated;
  • results you can’t reproduce in a fresh session.

When you see these signs, stop generating and start checking. Go back to the last state you trust.

Signs that something exceptional has happened

Exceptional moments are rarer and easier to miss, because they often look like a small surprise:

  • a complicated thing suddenly becomes simple;
  • a connection appears between parts of the problem you thought were unrelated;
  • a prediction you didn’t ask for turns out to check;
  • a result contradicts what you expected and survives the attempt to break it;
  • a representation change makes a hard step obvious.

How to chase it

When one of these moments arrives, ordinary working rhythm is the enemy. The temptation is to note it and carry on with the plan. Don’t.

  • Stop everything else. Pause the other threads. This is now the investigation.
  • Capture the state immediately. Save the exact prompt, the output, the context and the version. Exceptional moments are often fragile and hard to reproduce.
  • Try to break it first. Exceptional results and convincing hallucinations look identical at first sight. Reproduce the result independently, ideally with a different tool or model.
  • If it survives, push it. Ask why it works. Change the representation. Generalise. Simplify. This is where Tao spent his effort.
  • Record what made it possible. The conditions that produced one exceptional moment may produce another.

The two halves go together. Chasing without checking turns you into the person who believes the hallucination. Checking without chasing turns you into the person who verified a miracle and moved on.

5. A Research Workflow Built Around Quality

I’ve been exploring an approach I call Stateful Inquiry. The idea is simple: rather than treating research as a series of disconnected AI conversations, treat it as a continuing investigation with explicit state.

Earlier versions of this workflow described what to do at each stage. What they didn’t explain was why you choose one path over another. Quality perception is that missing mechanism. Each step below carries its own quality question, and the investigation moves forward only when the answer is good enough.

AI-enabled mathematical research workflow

1. Start with something concrete

Tao had an explicit map and a specific property to explain. You might have an algorithm that behaves unexpectedly, a relationship you suspect holds, or an assumption you want to challenge.

Quality question: Do I know what would count as progress, and what evidence would change my mind?

2. Generate competing explanations

Don’t accept the first plausible answer. Ask for alternatives, what each one assumes, and what each one predicts.

Quality question: Are these genuinely different hypotheses, or variations on one idea?

3. Change representations

A calculation that looks miraculous in one form may become natural in another. In mathematics that might mean moving between algebraic, geometric and computational views. In programming, it might mean moving from code to data flow, state machines or invariants.

Quality question: Did the new representation reveal anything, or just restate the problem?

4. Run experiments that could prove you wrong

Use symbolic algebra, numerical checks, test programs, proof assistants, or whatever external evidence the problem permits. And check that the tool is actually testing the proposition you intended. Schwartz’s experience shows how easily it can end up testing a weaker one.

Quality question: Could this experiment have failed? If not, it isn’t evidence.

5. Keep an explicit research record

After dozens of exchanges, assumptions are forgotten, rejected ideas reappear, and results come loose from their evidence. Keep track of established facts, active hypotheses, failed approaches, open questions, and the tests behind each conclusion. Record your quality judgements too: what you rejected and why, and which moments looked exceptional.

Quality question: Could I resume this investigation tomorrow from the record alone?

6. Return to understanding

Once something survives testing, digest it. Why does it work? Which assumptions are essential? Can it be simplified? Does it generalise?

Quality question: Is the explanation getting simpler, or only longer?

The human role in this loop is not a single approval at the end. It is a continuous series of quality judgements (continue, redirect, challenge or stop) that steer the search.

6. Developing Quality Perception Without Being Terence Tao

Here is the obvious objection. Tao, Schwartz and Alpöge are exceptional. Their quality perception comes from decades of expertise. How can someone without that background judge work they can’t fully evaluate?

I don’t think the answer is that you must become an expert before you begin. The answer is that quality perception can be developed, and that you should develop it deliberately, alongside the investigation itself.

Start with the objective qualities. Does the result reproduce? Does it survive a counterexample? Are the assumptions explicit? Does an independent method give the same answer? These require discipline more than expertise.

Build up to the harder qualities. With experience, you start to recognise whether a representation is productive, whether an explanation is natural or contrived, and whether an anomaly matters. Tao’s “unmotivated constructions” is a quality category you can learn to see.

Study examples of high-quality work. Read the digestion alongside the raw calculation. Compare the AI’s first explanation with its tenth. Quality perception grows through comparison.

Don’t treat AI agreement as quality. Models can make useful quality judgements. They can find defects, compare explanations and spot gaps, and as the Masullo episode suggests, they are getting better at it. But two models agreeing isn’t independent confirmation if they share the same mistaken assumption.

Know when your perception runs out. There is a difference between exploring an idea, reproducing a known result, proposing a conjecture and claiming a new theorem. Each needs a stronger standard of evidence. At some point an investigation will depend on something that neither you nor your tools can evaluate. That is when you need an expert. Recognising that point is itself a form of quality perception.

The same applies in software. You can build a system with AI without writing every function. But if you can’t establish whether its security, financial calculations or critical behaviour are correct, you need stronger verification or someone who can assess the risk.

7. What I’m Building Next

In my work on Writer and Adversarial Math From First Principles, I’ve been approaching this problem from two directions.

Writer’s inquiry work explores how to preserve hypotheses, evidence, decisions, rejected approaches and open questions across long investigations. It also includes experiments with recording human preference judgements: structured comparisons that capture which of two outputs is better.

The Math project focuses on precisely stated mathematical claims, checked by independent tools such as Lean, Rocq, Z3 and SymPy, with an explicit separation between a claim and the evidence supporting it.

One helps organise the investigation. The other establishes which claims have survived checking. What neither yet does is treat quality judgement as a first-class part of the research state: recorded, timestamped, linked to evidence, and used to steer what happens next.

That’s the next step, and the subject of the next post: a technical implementation of a research system in which quality perception is part of the machinery rather than something left in the researcher’s head. Concretely, it means:

  • recording continue, redirect, challenge and stop decisions alongside the evidence that prompted them;
  • flagging decay signals automatically where they can be detected;
  • capturing exceptional moments with full context, so they can be reproduced and chased;
  • measuring whether all this actually produces better results than ordinary prompting.

That last point matters. I haven’t yet shown that a quality-driven, stateful process produces better mathematics than an ordinary AI conversation. That’s an experiment, not a claim. The plan is to compare the two on bounded problems, measuring valid findings, unsupported claims, time, cost, repeatability and the ability to resume an investigation. If parts of the workflow add nothing, the experiments should say so.

8. The New Math

I don’t think AI is about to replace mathematicians, and I don’t think we’re seeing a new branch of mathematics.

What we’re seeing is a new way of doing research, and a shift in which human skill matters most.

When producing work becomes cheap, the scarce resource becomes the ability to judge it. Tao shows that judgement refusing to settle for a correct but unilluminating answer. Schwartz shows it catching work that has quietly decayed while looking better than ever. Alpöge’s announcement shows the hardest case: an exceptional result that arrived faster than the judgement needed to understand it.

AI will keep responding. It will produce another explanation, another hypothesis, another construction, for as long as we keep asking. More output is not more progress. The researcher’s job is to know the difference: to notice when quality is falling, to demand better when correct isn’t enough, and, when the rare exceptional moment arrives, to recognise it and chase it with everything they’ve got.

AI expands what we can explore. Our ability to perceive quality decides how much of that exploration becomes knowledge.

That is the new math I want to explore.


Further reading