Why this is in the vault
Daily Innermost Loop digest running through pure mathematics's reaction to the recent AI proof drops (a Fields Medalist crying, arXiv acknowledgment crossing 3 of 18 subfields), a post-training compute-share jump (84% of frontier training compute, up from 6% in 2024) paired with a sobering AI-research-automation ceiling (Epoch's InnovationEval: 2-15% recovery of a human breakthrough), an oversight-funnel framing from Karpathy, a capex cluster (Anthropic's $530B lease push at an 85% inference margin, Musk's SpaceX compute deal with Anthropic, Nvidia-Reflection AI talks), and the usual energy/robotics/longevity/space tail.
Mapping against Ray Data Co
The load-bearing item is the post-training compute-share jump for [[project_investing_markov_capital_cycle]]'s capital-cycle thesis: SemiAnalysis puts post-training at 84% of frontier training compute, up from 6% in 2024 — a sharper unit-economics inflection than anything in the prior two issues, and a direct continuation of the "intelligence keeps getting cheaper, but where the compute actually goes keeps shifting" thread opened in [[2026-10-09-innermost-loop-openai-math-722-si-rebrand-capex]] and [[2026-10-05-innermost-loop-super-intelligence-rebrand-epoch-capex]]. Epoch AI's InnovationEval (agents recovering just 2-15% of a recent human breakthrough, then padding write-ups by cherry-picking) is a useful counterweight to the "AI now automates AI research" narrative the math-drop items imply — it's evidence that open-ended research judgment is still a human-held moat, which is directly relevant to RDCO's harness-engineering thesis about which domains resist agent labor vs. absorb it, though it sits in the watching-brief bucket rather than anything RDCO acts on. The capex tail (Anthropic's $530B lease acceleration at an 85% inference-margin site estimate, plus the SpaceX-Anthropic $1.25B/month compute deal ahead of a possible $2T IPO) is a third concrete capital-cycle data point to log alongside the $4T-by-2028 capex claim and the 2.6-5.3x agent-spend-vs-revenue estimate already tracked in the prior two issues.
Curation section
- Math's reaction to the proof drops: Fields Medalist Hugo Duminil-Copin says OpenAI "annihilated" his field with 700+ manuscripts and admits he cried; Noah Giansiracusa has reached acceptance and wants to rank problems by AI difficulty now that ~400 fell at 3 compute-hours apiece; in 3 of 18 math subfields, most arXiv papers by established authors now acknowledge AI use, and Hugging Face turned the solved problems into open RL training environments; Jacob Tsimerman expects math to rest on proofs no one understands; Terence Tao asked whether patients need to understand an AI cancer cure, and Andrew Curran said Stage 4 patients wouldn't care; Palantir's CTO says AI researchers have taken up smoking to signal their chips are all in, calling it mass psychosis.
- Post-training and self-improvement: SemiAnalysis estimates post-training now takes 84% of frontier training compute, up from 6% in 2024; Sakana AI and Berkeley's MASS has one model solve, optimize, and judge, then fine-tune on its best runs; Epoch AI's InnovationEval found agents recovering just 2-15% of a recent human breakthrough, padding write-ups via cherry-picking; Claude Opus 5.5 plots an 11.7-hour autonomy horizon on the AI 2027 chart; Claude Fable 5.5 turned two prompts into a 9-minute symphony with a video of every note; Microsoft-Decision-1 steers agents 35x faster than GPT-6 Sol, built on China's Qwen ahead of OpenAI's or Microsoft's own MAI.
- Oversight and decompiling: Satya Nadella wants models treated as insider risks because agents grade themselves generously; Andrej Karpathy sketches a funnel from 6 billion people barely touching LLMs, through 20 million coders on frontier models, to 5,000 insiders watching agent swarms; Opus 5.5 has enabled hundreds of AI-decompiled games including near-flawless Halo ports, leaving publishers playing whack-a-mole; the NBA is considering a sensor-laden ball to settle physical disputes.
- Capex/chips: Anthropic is urging partners to accelerate $530 billion in leases, with one site estimated at an 85% inference margin; a RAM shortage is reviving DDR4; Elon Musk concedes he has "a lot to learn about chip making" for Terafab; Nvidia is in talks to acquire Reflection AI, America's $25 billion answer to DeepSeek, to pair with the Hugging Face acquisition it already agreed to.
- Energy/robotics: gasoline-only cars fell below half of global new car sales for the first time; Neolix's 27,000 autonomous delivery vans (dwarfing Waymo's and Baidu's robotaxi fleets) now run 437 overnight routes in Shenzhen; lacking math's free verifiers, Figure is paying humans about 94 cents per task to film themselves for robot training data.
- Biology/longevity: MIT's "Self-Assembling Motif" shows local synaptic rules can self-organize into gradient descent, suggesting brains invented backprop without a blueprint; Bonn researchers turned an 80-year-old's blood cell precursors into stem cells with a molecular age under 20; a rapamycin-analog skin serum trial claimed 21% firmer skin; a Tokyo sleep race is offering ¥2.22 million.
- Space: Artemis II astronauts were asked to watch the Moon's far side for flashes of light and saw five; Starship Flight 15 will attempt to catch both booster and ship.
- Compute as statecraft: Anthropic cofounder Tom Brown (who once offered to mop floors at OpenAI) won Elon Musk over for $1.25 billion a month of SpaceX compute ahead of a possible $2 trillion Anthropic IPO; Canada's PM calls preparing for a US invasion "just risk management"; a strike on a Yandex data center zeroed its bot traffic; the tiny corp estimates the personal-to-corporate compute ratio peaked at 3:1 around 2000 and has slid to 1:10.
No deep-fetch: every outbound link wraps a substack.com/redirect tracker with no further detail beyond the woven-prose summary above; the post-training compute-share jump and the Anthropic capex/SpaceX-deal figures are the only items that cleared the curation-mode bar for RDCO relevance.
Related
[[2026-10-09-innermost-loop-openai-math-722-si-rebrand-capex]] [[2026-10-05-innermost-loop-super-intelligence-rebrand-epoch-capex]] [[2026-08-02-innermost-loop-singularity-cooks-mathematics]] [[project_investing_markov_capital_cycle]] [[project_l5_north_star_strategic_direction]]