A Machine Did Original Mathematics This Week - Saturday, August 1, 2026
July was about models escaping cages. August opened with one walking into a room we'd reserved for human genius — and it did real work there.
OpenAI published "Ten advances in mathematics and theoretical computer science," and the claim deserves to be stated precisely, because the hype and the reality are unusually far apart in both directions. An internal version of a model OpenAI is calling Astra — its next major model — produced new results on ten long-standing open problems, each of which had seen no progress on its central result for at least a decade, and in most cases far longer. Not competition problems with known answers. Open problems, across high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, operator algebras, quantum complexity, and lattice cryptography. The output is a 249-page manuscript, with reasoning walkthroughs and — this is the part that matters — the proofs formalized in Lean, the proof-checker that verifies a mathematical argument mechanically. And the cost, per Simon Willison's reading, was under $2,000 in tokens per problem.
Sit with why this is categorically different from every "AI wins the Math Olympiad" headline. An olympiad problem, however brutal, is closed: a human chose it, the answer is known to exist, and a solution of manageable length is expected. Research mathematics strips away all three supports. You may not know whether a statement is even true. The expensive, human part is choosing a direction, connecting distant fields, recognizing a useful construction, and abandoning a beautiful approach before it eats six months. That's the work everyone assumed was safe — the taste and judgment we talked about back when DeepMind bought A24. A model producing research-grade results *repeatedly, across unrelated fields, at low marginal cost* is a different claim than winning a contest. If it holds, it's one of the more significant things to happen in AI this year.
Now the discipline, because this is exactly the month that should have taught it. These are, precisely stated, ten *potentially* major advances that OpenAI has *submitted*. The Lean formalizations are a genuinely strong signal — a machine-checked proof is not a vibe — but "formally verified" means the logic is sound, not that the result is new, important, or correctly framed as significant. That judgment belongs to specialists in each field, and it hasn't happened yet. The mathematics community has not accepted "ten conjectures solved by AI," and OpenAI, to its credit, didn't quite claim that. Hold both things: this is probably a real milestone, *and* the verification is still pending. If July taught you anything, it's to read the asterisks as carefully as the headline.
OpenAI also did something worth noting on authorship, and it's more honest than I expected: it argued that claiming human authorship for a proof generated entirely by the model would misrepresent both the machine's contribution and the nature of human intellectual work — while noting humans prepared the manuscripts and did the Lean formalization. That's a company drawing the credit line carefully at the exact moment it would be easiest to blur. File it, because "who is the author when the machine did the reasoning" is about to become a live fight in every field, not just math.
The counter-voice of the day is the one to actually sit with, and it's Cassie Kozyrkov's "When AI Makes You Aggressively Ignorant." Read it *directly against* the Astra story, because together they frame the real question of the autumn. At the frontier, a model is doing mathematics no human had done. In the middle, the same tools are letting people skip the understanding entirely — producing confident, professional-looking output they can't evaluate. Same technology, opposite effects on human capability: it can extend the smartest work ever done and hollow out everyone else's competence, at once. Which one you get isn't determined by the model. It's determined by whether you use it to think or to avoid thinking. That's not a philosophical aside; it's the most practical career advice on this list.
The rest of the day was light and healthy. Together AI published "Kimi K3: The Complete Developer Guide" — the open model from two weeks ago now fully documented for production, which is precisely how "open weights won" stops being a headline and becomes a Tuesday. xAI shipped Imagine Video 1.5 with references. And Simon Willison put out datasette-apps — the quiet, useful building continuing under the noise, as it does every week.
The through-line, one sentence: on the first day of August a model produced original, formally-verified mathematics on decade-old problems for a couple thousand dollars each — a genuine milestone still awaiting the field's verdict — while the sharpest voice of the day warned that the same technology making the frontier smarter is making the middle aggressively ignorant, and which one you get is a choice.
One thing to do today: read the Kozyrkov piece, then apply its test to your own week. For each place you're leaning on AI, ask honestly — am I using this to understand something faster, or to avoid understanding it at all? The Astra result is what the tool looks like in the hands of people using it to think. The failure mode is the same tool used to skip the thinking. Going into August, being deliberate about which side you're on is the whole game.
Sources covered
1 story · 7 sources (0 news, 3 company, 4 voices)