Three results landed this month with nothing in common except the kind of number each one moved: the sort that had sat still long enough that everybody stopped expecting it to move at all.
1. Twenty-Six Years of Speculation Ended in Half an Hour, and Every Human Clue Was Wrong #
When NIST published the P-192 through P-521 elliptic curves in 2000, each came with a seed parameter supplied by the NSA and no explanation of its origin. That silence fed two decades of backdoor suspicion, a literature arguing both ways, and eventually a cash bounty for anyone who could reproduce the seeds. The lore, from the engineer who made them, was that they were hashes of a forgotten joke phrase about giving Bob and Jerry a raise.
The lore was wrong, which is why nobody found anything. Asked to figure out how the seeds were generated, without being handed the phrase rumor, a model took the opposite approach: instead of guessing inputs, reconstruct the 1997 software. It assumed a stock C library generator of the period, wrote a scanner that slid along each seed a bit at a time checking whether adjacent chunks were consecutive generator outputs, and noticed four of the ten seeds had a zero in the top bit of every 32-bit word — odds of one in 5,810, and the fingerprint of the drand48 family, which exposes only the top bits of a 48-bit state. Recovering a truncated linear congruential state is a textbook lattice problem; LLL reduction over every candidate bit window finished it. Elapsed time, under thirty minutes. The bounty's author reimplemented the whole generator in Go to check it, and the timestamps land in 1997 and 1998, exactly where the history says they should.
Read the mechanism, not the headline. The model's edge was not raw cleverness; it had no stake in the rumor everyone else was chasing, so it spent its search budget on the artifact rather than the anecdote. Note what did not fall: roughly twice as many seeds pass the published criteria as were standardized, and nobody can yet explain the gap.
Why it matters:
- For ICs: The expensive failure mode in a long-stuck problem is an inherited prior, not insufficient effort. Before spending another week, have an agent attack it with the folklore deliberately withheld.
- For leaders: "We already looked into that" has a shelf life now. Keep a list of the questions your team retired on cost grounds; several are a half-hour job.
- For founders: Trust restored by reproducible evidence beats trust asserted by authority. If your product rests on an unexplained constant, explaining it is now cheap.
- The broader pattern: any question where a candidate answer can be checked in milliseconds is now this shape of problem.
2. Google's Best Circuit Got Twelve Times Cheaper and Nobody Touched the Hardware #
Researchers gave a model plain-language specifications for eight quantum circuit primitives — chemistry, optimization, simulation, number theory — plus an evaluator scoring each design on qubits times gates, then let it run a closed proposal-and-feedback loop with no human in between. It beat the best published designs on all eight, in places by more than an order of magnitude.
The decoder case shows the texture. The previous state of the art used the extended Euclidean algorithm, whose step count depends on its input — a bad fit for a circuit that must run one fixed gate sequence over all inputs at once. The model switched to Berlekamp–Massey decoding, which always takes the same steps, reconstructed intermediate values instead of storing them, then did what human teams rarely bother with: wrote a fast simulator, ran 403 million sample decodings to measure how often the worst case actually occurs, and sized the memory reserve to the measurement rather than the bound. That turned 665 qubits of worst-case storage into 19, at a one-in-10,000 failure rate the task already tolerates. Twelve times fewer gates, from a 53-hour run, against a line of work where each step had taken months.
The published limitation is the part worth internalizing. The model was conservative by default, preferring known approaches and re-verifying results already solid. The human contribution was not mathematical; it was forcing several qualitatively different approaches to develop in parallel and occasionally telling it to search wider. The best design came out of that widening, not the default policy.
Why it matters:
- For ICs: Your agent's default is the textbook answer. Novel output comes from explicitly requiring divergent approaches in parallel, then scoring them.
- For leaders: Worst-case resource budgets throughout your systems were set because measuring the real distribution was too expensive. It is not anymore, and the savings there are larger than most model upgrades.
- For founders: A domain with a cheap scorer and an old published baseline is an opportunity, not a moat. If that describes your technical edge, assume someone is running the loop against it this quarter.
3. Seven Hundred Theorems Arrived. The Missing Fifth Is the Real News. #
OpenAI released more than 700 machine-proved mathematical results, and someone ran the obvious audit: which fields are in there, and which are conspicuously absent. Statistics, machine learning theory, optimization and numerical analysis are roughly 20% of theorems on arXiv and under 1% of the release. The gap narrows but survives when the baseline is restricted to papers solving open problems, so "those fields have no open questions" does not explain it. The uncomfortable candidate explanation is the simple one: those are the fields closest to the publisher's own competitive advantage.
Whatever the motive, the structural fact stands. The published literature used to be an approximately neutral record of what was known. It is becoming a curated subset of what a few compute-rich organizations chose to show, and the curation will be heaviest exactly where results feed back into model capability. The baseline you benchmark against is now partly an editorial decision.
Why it matters:
- For ICs: Treat "state of the art" as a claim about what was disclosed, not about what exists. The gap is invisible and growing.
- For leaders: Build evaluation on your own data and your own tasks. External benchmarks increasingly measure what vendors were willing to publish.
- For founders: The deep technical moat is no longer the result — it is the private verifier and the proprietary problem. Those do not leak when the literature does.
The Verdict: Real or Hype? #
Old open problems with cheap verification → Real. Twenty-six years of speculation and a standing bounty fell in half an hour, and the independent reimplementation confirmed it the same week. Models as autonomous researchers → Real but early. They beat the published state of the art on all eight circuits, but only after humans forced the search wide; left alone, they reach for the textbook. The literature as a complete record of what is known → Hype. A fifth of mathematics went missing from the biggest result dump yet, and the missing fifth is the part worth the most.