The Drop of Oil

0. A Word That Must Not Be Used

Get this out of the way immediately.

Almost every news piece about this event begins with: "Artificial intelligence has once again shocked the mathematical community." That is a template opener. It carries zero information; it substitutes emotion for fact. Under the constraint rules we are using, such collective exclamations of "shock", "stupefaction", or "vertigo" are forbidden as empty rhetoric.

An opening should be a concrete action or a concrete date.

A Usable Opening

In late August 2026, a preprint on prime gaps appeared on arXiv. The author field read "GPT‑5.6 Sol". Three weeks later, the mathematician Boris Alexeev announced that he had formalised the core proof in the Lean theorem prover. As of this writing, Lean compiles without error.

That establishes a fact: a mathematical proof generated by an artificial intelligence has passed a machine‑checked logical verification. The content of that proof is—on a classic number‑theory problem with a history of nearly a century, initiated by Erdős and broken open in 2014 by Tao and others—an advance on the "best known lower bound" previously held jointly by five leading mathematicians.

No "shock". No "stupefaction". Just dates, names, and verifiable subsequent actions.

1. One Thing

First, the most basic question: what is this actually about?

Prime gaps are the distances between consecutive primes. Between 2 and 3 it is 1, between 3 and 5 it is 2, between 89 and 97 it is 8. As primes grow larger, they become sparser on average, with the average gap roughly a multiple of the number of digits. But the question is: how large can a gap be in extreme cases?

The problem has a recorded mathematical history dating back at least to 1931, when Westzynthius proved that gaps can be arbitrarily large relative to the average. Not "infinite" in absolute size, but the ratio can grow without bound. Erdős and Rankin pushed the lower bound further in the 1930s. Erdős even offered a prize for a particular directional breakthrough.

Then a long stagnation.

In 2014, James Maynard and, independently, Ford–Green–Konyagin–Tao solved Erdős's open question. That was a genuine breakthrough—they proved that a previously fixed constant could actually go to infinity. The papers appeared in the Annals of Mathematics.

In 2018, the same five authors collaborated and pushed the quantitative lower bound to a specific scale.

Eight years later, GPT‑5.6 Sol pushes that scale one notch further.

This narrative is linear. Every node has a year and a name. There is no fuzzy "the mathematical community widely believes". That is the skeleton.

2. One Modification

Now the unavoidable question: what exactly did GPT do?

This requires extreme care. Once we start explaining mathematics, we easily slide into two traps: either scare the reader with formulas, or use empty adjectives like "ingenious", "elegant", or "profound".

The constraint rule is: no sentence that could be dropped into another story.

So write it this way.

The FGKMT proof—that 2018 five‑author work—is a machine. It works in two layers.

Layer 1: a sieve that runs a random selection over all small primes, removing a large batch of integers and leaving behind some "survivors".

Layer 2: for the hardest survivors—those that behave almost like primes—an extremely complex covering strategy, like spreading a net, captures them one by one.

GPT‑5.6 Sol did not rebuild the machine. It took one component of the first‑layer sieve and replaced it with a new one.

The old component was "uniform random choice"—for each small prime, every residue class had the same probability of being chosen. The new component is "biased choice"—some classes are chosen with higher probability.

That is the only change.

But that change produces a cascade: under the new sieve, the survival probability of those "prime‑like composites" can be estimated so precisely that one can prove almost all of them are eliminated in the first round. Thus the second layer has much less work—it only needs to handle genuine primes, and for those, GPT directly reuses the FGKMT method, unchanged.

So the technical essence is: within an existing powerful framework, one input parameter was retuned; the effect of that retuning, after careful estimates, pushes the final output forward by one step.

This is not new physics, nor is it a new paradigm. It is not even "a tool humans never thought of". Humans thought of all the tools. Humans just did not think of tuning that parameter that way.

This fact matters. But it requires no embellishment. Embellishment would turn it into a different story.

3. A Verification History

Three weeks after the preprint appeared, Boris Alexeev translated the proof into the Lean language. It compiled.

How to write this without over‑ or under‑selling?

Lean verifies logical integrity. If every step is translated faithfully and compilation succeeds, there is no explicit inferential break. This is an extremely strong baseline guarantee in mathematical proof—countless human errors hide behind the word "clearly", and Lean does not allow "clearly".

But Lean cannot verify three other aspects.

First, whether the Lean version faithfully represents the original preprint. That requires human checking.

Second, whether the proof is "deep". Depth is not a logical property; it is an aesthetic judgment of mathematical structure. The computer does not participate in that judgment.

Third, whether an AI‑generated proof is readable. Several mathematicians have remarked that the proof is extremely hard to read—not because it is wrong, but because its exposition does not follow human mathematical conventions. It was not written for human readers, even though it is about human mathematics.

This creates a peculiar tension: logical verification passed, but human "understanding" is not fully achieved—not understanding the content, but understanding why certain paths were chosen, why a particular estimate was made at a particular node. Traditional mathematical proofs are not only correct; they are also comprehensible. This paper relied on external human translation to partially restore comprehensibility in that dimension.

This detail matters, because it touches the essence of "AI doing mathematics". It is a productivity issue, but it is also an epistemological one.

4. A Distance

Now we must address the misjudgment that easily arises: how far is this breakthrough from the "real answer"?

The "real answer" in the prime‑gap field has a widely cited reference point—the Cramér conjecture. It gives a scale far larger than any currently known lower bound. If we picture the ultimate limit suggested by Cramér as the top of a high tower, then both the 2018 FGKMT result and the 2026 GPT‑5.6 Sol result are still on some platform above the base.

This is a judgment of scale, not a belittlement.

"Lower bound" means "at least this large". "Upper bound" or "true scale" means "at most that large". Pushing a lower bound to a larger lower bound, and approaching the true scale from below, are two qualitatively different problems. The former can often be achieved by optimisation within an existing framework; the latter usually requires entirely new mathematical constructions.

So the sentence can be written this way: GPT‑5.6 Sol's advance is a substantive step along an established path, but it does not touch the central mystery—the ultimate scale of prime gaps remains an unknown abyss.

This sentence contains no forbidden word. It has a judgment ("substantive step"), a qualification ("along an established path"), and a clear boundary of the unknown ("remains an unknown abyss"). The reader can calibrate expectations accordingly.

5. An Evaluation

Finally, answer the most contentious question: does this count as the first time AI has "beaten" top human mathematicians on a classic mainstream problem?

The word "beaten" should first be banned. It belongs to competitive sports, not to knowledge production.

A more accurate formulation: this is the first time that AI has made a substantive improvement on the "best known lower bound" on a problem with a century‑long history, advanced by multiple generations of leading mathematicians.

Every qualification in that sentence has meaning. "Century‑long history" gives the temporal depth. "Multiple generations" gives the classic status. "Best known lower bound" gives the precise nature of the improvement. "Substantive" distinguishes it from mere parameter tweaking.

But one must add the complementary judgment: the mathematical framework of this work entirely inherits the 2018 FGKMT methods; the core novelty is concentrated in a redesigned input parameter.

Placing both judgments together yields a complete and measured evaluation.

As for the more distant judgment—whether this means AI will replace mathematicians—all current evidence points elsewhere: at this stage, AI behaves more like an extremely sharp parameter explorer, able to find overlooked corners within frameworks already built by humans. It has not proposed new frameworks, invented new concepts, or offered a deeper explanation of why the existing framework works.

It is like someone finding, in an already structurally complete symphony, a harmonic position never noticed before, and striking a note that is unexpected but perfectly valid. That note gives the whole piece a new layer. But the composer of the symphony remains human.

6. An Ending

The constraint rules impose a hard requirement: no uplifting conclusion.

No uplifting conclusion means no sentence like "this marks yet another widening of the frontier of human intelligence" or "we stand at the threshold of a new era". These sentences carry no information; they correspond to no concrete action.

An honest ending should return to some specific, verifiable detail.

For example, this:

The preprint from late August 2026 is still under peer review. Lean verification passed, but mathematicians are still reading the paper. Some of them say they need more time to decide whether there is any hidden problem—not because they found an error, but because the exposition is unfamiliar.

That is the current state. An unfamiliar proof passed a machine check, but has not yet fully passed human reading.

If one day it passes, it will become an important footnote in mathematical history—not because it overturned anything, but because it demonstrated one thing: where humans thought they had squeezed a framework dry, a machine could still find a drop of oil.

That is what machines are good at.

But explaining why that drop matters—that remains human work.

And every judgment in this essay, if replaced with another subject—say, "GPT improved another record in another problem"—would fail. Because every judgment here attaches to a specific year, a specific name, a specific action. That is the test of flesh‑and‑blood uniqueness.