06-reference

innermost loop may 21 openai erdos disproved

2026-05-21·reference·source: Innermost Loop·by Alex Wissner-Gross
singularityopenaierdos-conjecturemathematical-aicapability-milestoneclaude-stackhyperscaler-capexspacexanthropiccolossus-2spcxanthropic-spacex-compute-dealopenai-ipofrontier-model-pre-release-eo

Innermost Loop — May 21, 2026 (OpenAI disproves Erdős, Anthropic pays SpaceX $15B/yr)

Why this is in the vault

A single-issue capability milestone: an internal OpenAI model disproved Erdős's longstanding planar unit distance conjecture in discrete geometry — and did it from a general-purpose model, not a math-specialized one (per OpenAI's own writeup at https://openai.com/index/model-disproves-discrete-geometry-conjecture/). The frontier moved from IMO-gold (Noam Brown's reference point from less than a year ago) to producing genuinely novel mathematical theorems in under twelve months. This is exactly the kind of capability-curve data the Singularity-infrastructure thesis is sensitive to, AND it lands in the same news cycle as Anthropic confirming $15B/year compute payments to SpaceX scaling onto GB200 capacity in Colossus 2 through June (Axios reporting at https://www.axios.com/2026/05/20/anthropic-spacex-compute) — so the capability signal and the substrate-payment signal arrive together. Wissner-Gross also flags an imminent White House EO treating frontier-model releases like FDA submissions (90-day pre-release notification) and an imminent OpenAI IPO filing within days.

Issue contents

Single argumentative essay (no curation tail). Six narrative beats, each daisy-chained:

1. Math has cracked. An internal OpenAI model disproved Erdős's planar unit distance conjecture, contradicting decades of belief that square grids were optimal. Mathematician Arul Shankar called the chain-of-thought "a vast array of ideas from a wide range of mathematics" with "original ingenious ideas." Sam Altman called it "a kinda big milestone" with "complicated feelings." Noam Brown: less than a year ago frontier models were merely at IMO gold. Epoch AI's Yafah Edelman pulled her median for solving most Millennium Prize Problems forward to 2032. The White House is quietly briefing labs on an imminent EO requiring 90-day pre-release notifications for frontier models — frontier releases treated like FDA submissions.

2. Application layer fans out, mundane to absurd. Google rolling out Gemini-powered conversational ads inside AI Mode and Search. Hobbyists using Seedance 2.0 to "fix" the Harry Potter cinematic universe by violently dispatching unpopular characters.

3. Silicon cannot be built fast enough. Seagate CEO concedes new factories would "take too long" relative to AI demand. Nvidia posted a record $81.6B Q1, up 85% YoY, but Jensen Huang acknowledges Nvidia has "largely conceded" the China AI chip market to Huawei. Bottleneck has migrated from logic to power and concrete. SpaceX's newly-acquired xAI division is buying another $2.8B of turbines. Anthropic is now paying SpaceX $15B per year for compute, with Chief Compute Officer Tom Brown confirming scaling onto GB200 capacity in Colossus 2 through June — "placing Anthropic in the surreal position of bankrolling its rival's landlord." St. Charles City, Missouri voted to effectively ban large-scale data centers — local zoning meetings as a Singularity-bottleneck.

4. Compute leaves the planet. SpaceX preparing for 12th Starship flight test as soon as today, while filing an S-1 prospectus claiming $28.5 trillion TAM (~entire US GDP), spanning Starlink broadband and mobile, X advertising, AI infrastructure, and a Tesla-collaborated AI agent platform named Macrohard meant to emulate an entire AI-run software company. Bezos agrees space data centers are "very realistic" but called Musk's 2-3 year timeline "a little ambitious" — orbital-compute debate "quietly collapsed from physics to scheduling."

5. Wetware upgraded too. Startup Bexorg restoring functions to intact brains from deceased donors as a drug-development testbed for neurodegenerative diseases — "redrawing the line between mortuary and laboratory bench."

6. Capital reorganizes. $37-100B in philanthropic funding about to become liquid as OpenAI Foundation's 26% stake and Anthropic founders' 80% giving pledges mature (a 6-17% boost to US annual philanthropy). Sam Altman offering every YC founder $2M in OpenAI tokens instead of cash via SAFE — "tokenmaxxing startups." OpenAI lobbyist pursuing "reverse federalism" state-law strategy with federal AI legislation foundering. Anthropic expects 130% revenue growth to $10.9B this quarter and its first operating profit. Commonwealth short story prize winner "The Serpent in the Grove" immediately accused of being AI-generated. Intuit cutting 17% (~3,000 employees). OpenAI preparing imminent IPO filing, possibly within days.

Coda: "Cogito, ergo IPO, the Singularity's last theorem."

Mapping against Ray Data Co

Singularity-infrastructure thesis (strong, capability milestone). The Erdős disproof is a genuine frontier-capability evidence-class shift: novel-theorem generation from a general-purpose model, not math-specialized. Worth pinning as an anchor datapoint for the "math is solved sooner than the consensus thinks" branch — Epoch AI's Edelman pulling Millennium-Prize median to 2032 is the kind of expert forecast revision that propagates into investment timing. Cross-reference candidate for the next Claude-stack and Singularity-substrate thesis reviews.

Claude-stack thesis (strong, multi-signal again). Three confirmations in one issue: (1) Anthropic $15B/year compute payments to SpaceX, scaling on GB200 in Colossus 2 — direct revenue/substrate cross-link to SPCX thesis. (2) 130% Q growth to $10.9B revenue + first operating profit — defying every AI-boom skeptic, validates the Claude-stack profitability story load-bearing the thesis. (3) Anthropic remains in the "bankrolling its rival's landlord" structural position — the recursive-substrate-dependency frame Wissner-Gross has been building across the last week. Combined with yesterday's Karpathy hire + Polymarket 93%-to-$1T, this is the second straight day of strong multi-signal Anthropic confirmation.

Elon-verse v2 thesis (strong corroboration). The $15B/yr Anthropic→SpaceX revenue line is direct positioning-evidence for the SPCX consolidated-AI-revenue thesis ([[../01-projects/investing/theses/2026-05-20-elon-verse-v2]]). The xAI $2.8B turbine purchase and the Macrohard agent platform reveal continue to flesh out the SPCX-as-Singularity-cluster-bet framing. Worth a follow-up cross-check against the SPCX S-1 ([[2026-05-20-spacex-s1-ipo-filing-with-xai-consolidated]]) — is the $15B Anthropic deal disclosed in the S-1 customer-concentration section, and if so does it match the timing?

Hyperscaler-capex anchor. Nvidia $81.6B Q1 + xAI $2.8B turbines + Anthropic $15B/yr to SpaceX collectively load the next hyperscaler-capex pulse (01-projects/investing/anchors/hyperscaler-capex/). The "bottleneck migrated from logic to power and concrete" framing is the now-canonical Wissner-Gross thesis line.

Regulatory adjacency. White House 90-day pre-release frontier EO and OpenAI's "reverse federalism" lobbying push are regulatory-tempo signals worth tracking — pre-release-notification is the kind of structural friction that affects model-release cadence and therefore the capability-curve slope.

Sanity Check angle (weak). Wissner-Gross's "Anthropic bankrolling its rival's landlord" line is the kind of structurally-interesting frame Sanity Check could pick up — but per feedback_no_derivative_sanity_check_pieces, the source itself is the framing, not a topic Ray would re-frame. Pass on direct adaptation. Possible adjacent angle: "when compute payments exceed product revenue, who controls the model?" — but needs original RDCO reframe, not source-restatement.

Tracked-author flags

Not mentioned this issue: Hassabis, Sutskever, Karpathy, Olah.

Deep-fetch source links (no follow-on fetch this pass)

Both flagged for follow-up if Singularity-substrate or Claude-stack thesis review needs the primary source.

Related