06-reference

innermost loop singularity cooks mathematics

2026-08-02·reference·source: Innermost Loop·by Alex Wissner-Gross
ai-safetyharness-engineeringai-capexai-mathformal-verification

"Welcome to August 2, 2026" — @theinnermostloop

Why this is in the vault

A sustained single-argument issue (not the usual paragraph-cluster digest) tracing OpenAI's ten-solved-open-problems math result through its social/epistemic fallout, plus a harness/model-coupling argument from Anthropic and a fresh chip-capex data point — worth filing on its own rather than folding into the running capex-roundup series.

The core argument

Alex frames the week as mathematics being visibly "cooked" by frontier models: after OpenAI's models settled ten decade-plus-old open problems (sphere packing toward the Cohn–Elkies threshold, a disproved 150-year-old-adjacent conjecture among them), number theorist Daniel Litt conceded — four years early — his bet that AI couldn't produce Annals-quality number theory under $100k/paper. Manifold now prices an AI-solved Millennium Prize Problem at 31% by 2027, 52% by 2028. The proofs read strangely: Fields medalist-adjacent commentary flags results once called "science fiction" months ago now landing as boilerplate-buried writeups, with critics (Henry Yuen) complaining models bury the technical crux "as if this were the obvious thing to do."

The social reaction splits into grief and dismissal. Several mathematicians describe a "dark night of mathematics" — discovery as how humans touched the ineffable, math functioning as a quasi-religious order now being "slaughtered" — while skeptics wave it off as a "less-than-a-cat-level intelligence" plagiarizing solutions (the LeCun-penguin joke). Fernando Borretti's "Mathematics Without Mathematicians" catalogues and refutes the copes directly. Even the trophies are unstable: a DeepMind researcher notes a Fields-worthy result could turn AI-trivial before the medal ceremony.

Underneath the math story, two structural threads: Anthropic's Jess Yan argues peak performance is "impossible" without tying harness and model together — read by one VC as labs signaling they'll compete with their own customers on the harness layer, not just the model. And the compute substrate keeps compounding: Epoch AI estimates 20 million AI chips doubling every nine months toward 200 million H100-equivalents by 2028, data-center power quadrupling by 2030, $1T invested by 2029 — with used EVs already appreciating 7% this year as a downstream energy-price signal.

A smaller but concrete aside: Apple capped vulnerability-bounty report submissions after AI-generated reports (mixing real flaws with slop) overwhelmed its human review capacity, stranding at least one startup's six-figure legitimate exploit chain even as AI-assisted patches shipped 5x the normal fix volume.

Mapping against Ray Data Co

The Apple vulnerability-report story is the most concrete connection: a system that used to gate on human review got flooded by AI-generated submission volume until the humans couldn't keep the gate closed, and a real six-figure legitimate finding got lost in the slop. That's the exact failure mode the fresh-eyes critic stack (verify-vault-write, verify-strategic-output, verify-dispatch, station-critic) exists to prevent for RDCO's own output volume — evidence that "run more of it through a human/model gate" breaks down at volume unless the gate itself is load-tested for throughput, not just correctness. Worth a note in the next critic-stack review: has anyone stress-tested gate throughput, or only gate accuracy?

Jess Yan's harness/model-coupling argument is a second, more direct hit: it's the commercial-incentive mirror of the harness-engineering thesis already tracked in [[2026-05-11-innermostloop-harness-eats-the-model]] — if labs believe peak performance requires owning both layers, that's a structural argument for why RDCO's own harness investment (skills, sub-agent brigades, the critic gates) isn't a stopgap until models get good enough to not need it, but a permanent surface competing directly with what labs will eventually try to vertically integrate.

The chip-capex figures ($1T invested by 2029, 200M H100-equivalents by 2028, power quadrupling by 2030) are a same-week data point for the hyperscaler-capex anchor feeding /investing-edgar-watch and the chip-fab/memory capital-cycle thesis — directionally consistent with, not a revision of, the standing Phase 2 placement.

Related