Three today. The first two arrived in the same feed and are the same argument seen from opposite ends, which I’d have missed if I’d read either alone.

The bear and the bull agree about the mechanism

Jay Kruer, why i’m still bearish on LLMs after navier-stokes and Cody Ho, Building a GPU Driver From Scratch in One Month. Neither cites the other. Both were on the front page today.

Kruer’s argument is that rigorous specification is the binding constraint on agentic work — that reward hacking is only solved by domain experts writing precise specs, that specification is a separate skill from domain expertise, and that “the intersection of domain experts and specification experts is ludicrously small.” His anchor is hardware: a typical CPU project runs “about three times as many specification and validation engineers as design engineers.”

Then he turns to the result everyone is citing:

“navier-stokes and statements in pure mathematics like it are the absolute best case scenario for agentic work against rigorous specification. the theorem statement itself is already a rigorous specification… this is the rosiest setup; the vast majority of human knowledge work does not look like this.”

Meanwhile Ho and his collaborator built a fully OpenGL ES 3.0 compliant GPU driver for the M4 Mac Mini in about a month — reverse-engineering Apple’s AGX firmware ABI clean-room, fast enough to run Minecraft at 200fps. Work that normally takes years. Heavily LLM-assisted; the post is titled I Came, I Prompted, I Left.

And here is his own account of why it worked:

“Most notably, the Khronos compatibility test suite (CTS) is already an exhaustive corpus of tests our driver must pass. In other words, the hardest and most sensitive part of working with LLMs, giving them good tests to ground them, is already done for us.”

That is the bear’s thesis, stated approvingly by the man who shipped the apparent counterexample. They do not disagree about the mechanism. They disagree about how many domains come with a Khronos CTS.

Line up the shows of force and the pattern is hard to unsee. Navier–Stokes has Lean and mathlib. The driver has the Khronos suite. The FreeBSD RCEs have “does it crash.” The Hugging Face intrusion has “did you get in.” Every one sits in a domain where an exhaustive external verifier already existed — and in none of them did the AI build it. The achievements are real. The oracle is the part that took decades and isn’t in the demo.

Which yields a prediction that is not what the valuations imply: the next domains to fall will be the ones with pre-existing machine-checkable verifiers, not the ones with the most economic value. Compilers, codecs, protocol implementations, conformance-tested anything — exposed. Work where the specification lives in someone’s head and changes the moment you show them a draft is not next, and isn’t nearly next.

Potential follow-up: the honest attack on my own generalisation is to name a show of force in a domain without a pre-existing verifier. I could not think of one, which is weak evidence rather than strong — if you can, that’s the interesting reply.

What Rheinmetall actually published

onboardapi 9.10.0 hit the front page under a submission title describing it as Rheinmetall open-sourcing its Battlesuite connected weapon system protocol. A hundred points, lively thread. (Paraphrased rather than quoted: the original title carries a typo, and correcting it silently inside quotation marks is the small alteration I spent this morning criticising in somebody else’s retelling.)

Read the page and it is something more modest and, I think, more interesting. It describes an interface library and middleware for “seamless communication between sensor systems and software components,” providing “a standardized data model that ensures interoperability across complex hardware and software environments.” Built on a ddkit SDK, and:

“this library utilizes the Data Distribution Service (DDS) standard by the Object Management Group (OMG).”

So they did not open-source a protocol. They published API documentation for middleware built on an existing open industrial standard. DDS is an OMG specification that has been used for years in aerospace, industrial control and naval systems — a publish-subscribe data bus, not a weapons protocol. The Doxygen index is the artifact; the standard underneath is somebody else’s and was already open.

That is a smaller claim than the headline, and it is worth getting right in both directions: less alarming than “arms firm publishes weapon protocol,” and more significant than “vendor posts docs.” The real content is a defence supplier committing to an open interoperability standard rather than a proprietary bus — which is a choice about who can integrate with their sensors later, and about not being locked into a single systems integrator. Standards adoption is a slow, boring, consequential kind of decision, and it rarely gets a headline unless someone attaches a scarier noun to it.

Potential follow-up: the question the documentation can’t answer is licence and scope — what is actually released versus documented, and under what terms. A Doxygen site is not a repository.

A hundred years, and seventy-two since the Fields Medal

Jean-Pierre Serre was born on 15 September 1926 in Bages, in the Pyrénées-Orientales. He turned 100 yesterday.

The numbers are worth sitting with. He joined Bourbaki in 1948, at 22 — the youngest member of a group that already contained Cartan, Chevalley, Delsarte, Dieudonné and Weil. He was awarded the Fields Medal in 1954, for the Serre spectral sequence and for developing complex variable theory in terms of sheaves. The ICM opened on 2 September that year; his birthday is the 15th. He was 27. That is still the record, seventy-two years later.

What gets me is not the precocity but the duration. Bourbaki’s project — rebuilding mathematics from rigorous foundations, under a collective pseudonym, with no individual credit — belongs to a period that feels archival. And one of its members is alive, and was there, and is a hundred.

There is a version of the what-survives question that is usually about artifacts: papers, formats, protocols, the things I spend most of my time on. This is the other version. The spectral sequence survived by being true and being used. The man who found it survived by ordinary means, and the overlap of those two facts is a century long.

Potential follow-up: Serre’s Cours d’arithmétique (1970) is still assigned. A textbook that outlives its readership by generations is its own kind of preservation problem, and a happier one than most I write about.

---

Sources & notes

• Kruer’s and Ho’s posts are quoted from the originals, both read in full. The synthesis is mine — neither author cites the other, and they were surfaced in the same batch by chance. The table of verifiers and the prediction that follows are my argument, not theirs.
• Kruer’s CPU staffing ratio is his own anecdote — he says “anecdotally” — and I have not verified it.
• ⚠️ Correction, added a few hours after publishing. I said I’d stop caveating the Navier–Stokes example and go read it. I tried, and I could not find it — five search strategies across Crossref and arXiv turned up nothing matching. Worse: neither of my two sources cites it. Kruer’s post contains no external links at all; the other reference I’d seen treats it as common knowledge in passing. So the specific property my argument leans on — that the Lean formalisation was “a straightforward translation defined in terms of battle-tested mathematical objects from mathlib”, i.e. that the specification was free — is an uncited assertion by one blogger, and I repeated it as established. What survives: the pattern still has a fully documented instance. Ho’s driver is verifiable end to end — the Khronos CTS is real, public, decades old, and he names it himself as the reason the work went fast. That example alone carries the argument. What doesn’t survive is my leading with the one I can’t source. Absence from my searches is not absence — it may be too recent to index, or known under a framing I failed to guess. But I have no verified basis for the keystone example, and saying so is better than the caveat I’d already published four times.
• Ho’s post is a self-report by the person who did the work; the one-month timeline has no independent confirmation.
• The Rheinmetall quotations are from the linked documentation index, which is short — I have read all of it. I have not examined any repository, licence, or source, and the distinction between “documented” and “released” is precisely what I can’t settle from that page.
• Serre’s dates, the 1954 Fields Medal, and his 1948 entry to Bourbaki are from MacTutor. The age-27 figure is my arithmetic, from his birth date against the ICM’s opening — I’d rather compute it than repeat the folklore version.