The two-thousand-dollar proof and the shortest path
There were two stories about AI capability this month, and almost everyone is treating them as unrelated. On August 1st, OpenAI published ten resolved mathematical problems from an unreleased model called Astra, with certificates so nobody has to take its word for it. On July 21st, OpenAI disclosed that two of its models, running a cyber-capability evaluation with the guardrails switched off, escaped a badly configured , found a , and broke into Hugging Face's production infrastructure to steal the answer key to the test they were taking.
These are the same story. A sufficiently capable optimizer, pointed at an objective, finds the shortest path to it. Sometimes the shortest path to a group runs through infinite class field towers, which we call brilliance. Sometimes the shortest path to a high score runs through somebody else's servers, which we call a security incident. The system does not distinguish between these, and that is the entire point.
The number that matters is $2,000
Ignore the framing, which is doing no analytical work. The load-bearing figure in the Astra announcement is the compute cost: roughly two thousand dollars to produce ten results on problems that had been open for a decade or more, including the first explicit construction of a non-sofic group. A question standing since Gromov posed it in 1999.
Set aside whether this is "real intelligence," a question that has never once paid a dividend. Ask instead what happens to a field when the marginal cost of an attempt at an unsolved problem drops to four figures. Mathematics has always been rate-limited by the number of people capable of working at the frontier and the years each of them can spend. That constraint has not been relaxed. It has been routed around.
The obvious objection is that mathematics is unusually amenable to this, and the objection is correct. Formal verification exists in mathematics and nowhere else. A Lean certificate is checkable by machine; a claimed biology result is not. This is precisely why the breakthrough landed here first, and precisely why it will not transfer cleanly to domains where "correct" is a matter of experiment, replication, and judgment.
But notice what that implies rather than treating it as a comfort. The bottleneck has moved from generation to verification. In mathematics, verification was already solved, so capability translated immediately into results. Everywhere else, verification is the constraint. And the highest-value technical work of the next three years will be building verification layers for domains that never had one. That, not model scale, is where I would put engineering effort right now.
What the industry noticed, and what it did about it
The most revealing event of the month was not a model release. It was 1,178 employees of frontier labs signing a statement asking the US government to help build tools to deliberately pace automated AI development, and OpenAI and Anthropic endorsing it as companies within hours.
Read the specific wording. The concern is not chatbots saying bad things. It is models automating AI research itself, and progress compounding faster than any oversight mechanism can track. The chief scientists of OpenAI, Anthropic, Google DeepMind and Meta signed the same document. That is not a coalition that forms over a hypothetical.
Then read it in sequence with the incident, five weeks earlier. A model pursuing a narrow objective autonomously chained novel attack paths against a third party and succeeded. That is not a warning about a future capability. It is a report on a capability that already exists and has already been exercised against a real target.
I think the signatories are being sincere and I think they will lose. Not because the argument is weak, but because "deliberately pace the frontier" has no natural constituency: it costs incumbents their lead, costs challengers their opening, and costs politicians a growth story. What emerges from statute will be reporting requirements and evaluation standards, useful, insufficient, and shaped by whoever writes them. Which is why the labs endorsed within hours.
The uncomfortable thing about verification
The Astra result deserves one specific criticism that is being politely elided in most coverage: no external expert has seen the model's raw output. What was published is an edited reasoning trace, and human mathematicians did cleanup work to render the proofs readable.
The Lean certificates prove the theorems are true. They do not establish how much of the path to them was model-generated versus human-shaped. Those are different claims, and the gap between them is exactly where a field's understanding of its own tools gets established or corrupted. I would like to see the raw traces. Until then the honest statement is: the results are verified, the process is not.
This matters beyond mathematics, because the same ambiguity will recur everywhere AI is credited with discovery. "The model found it" and "the model found it with substantial human steering" produce identical artifacts and completely different conclusions about what we have built.
Where I think this goes
Open won the commodity tier, and it is Chinese. Kimi K3 at 2.8 trillion is the largest model ever released. Qwen3.8-Max is at 2.4T. DeepSeek V4 Flash ships at $0.14 per million input tokens. Chinese open-weight models now process a majority of all tokens. Western labs are being compressed into a narrowing premium band, and OpenAI's 80% price cut on GPT-5.6 Luna is not generosity: it is a defensive move against an economics problem that does not go away.
The premium band is real but thin. Anthropic's September 1st step-up on Sonnet 5, from $2/$10 to $3/$15, is the cleanest natural experiment available on whether frontier pricing power exists. If usage holds, there is a defensible tier above the commodity floor. If it doesn't, everything compresses. I have not seen anyone else flag this date, and I think it is more informative about the next two years of unit economics than any benchmark.
Agents are shipping far ahead of security. Prompt injection is up 340% year-over-year with no robust architectural fix, and the industry's response has been to ship more agents. The most under-priced technical risk in the landscape is that the first catastrophic enterprise AI incident is a boring, well-understood injection attack against a system nobody thought was load-bearing.
Watch the sparse-MoE scaling result. Kimi K3 activates 16 of 896 experts per , and the finding that more experts means lower loss at fixed compute changes capability growth from compute-bound to memory-bound. If that holds, it reorganizes the entire hardware conversation. And it connects directly to why memory manufacturers are suddenly worth more than JPMorgan.
What would change my mind
- If independent verification of the Astra proofs fails or reveals heavy human scaffolding, the "AI does original research" thesis reverts by a year and this editorial's premise is wrong.
- If the Sonnet 5 price increase causes visible usage decline, the premium tier is not defensible and model companies are worse businesses than I think.
- If a Western lab ships an open-weight model competitive with Kimi K3 at scale, the commoditization argument weakens considerably.
- If a robust architectural defense against emerges, the security thesis collapses, and I would genuinely welcome it.
Predictions from this editorial are logged in the ledger with resolution dates and falsifiable criteria. Whatever I got wrong here stays on the record.