Why this is in the vault
Daily Innermost Loop digest spanning Meta's "personal superintelligence" pitch (Muse Wonder), China's math-AI mobilization, a safety-evaluator control dispute, a capital-cycle bottleneck call from two chip-industry CEOs, and labor-automation data (37% of AI time savings spent on rework, a 40%-of-jobs IMF warning).
Mapping against Ray Data Co
The load-bearing item is Broadcom's Hock Tan and Intel's Lip-Bu Tan agreeing that capital, not chips or power, may now be the bottleneck on AI buildout — a sharper, named-executive version of the capital-cycle thesis tracked in [[project_investing_markov_capital_cycle]], and a direct continuation of the capex thread opened in [[2026-10-12-innermost-loop-math-grief-post-training-compute-spacex-deal]] (that issue's $530B Anthropic lease push, 85% inference-margin estimate) and [[2026-10-09-innermost-loop-openai-math-722-si-rebrand-capex]]. It's reinforced by the issue's own supporting data: a 1-GW data center now runs ~$30B, memory costs are up 5x in a year, economists are flagging pre-2008-style foreign-funding imbalances, and Loudoun County (40% of its tax base from data centers) is bracing for a bust — all pointing the same direction as the Tan/Tan read. Secondary but relevant to RDCO's harness-engineering thesis: the stat that workers spend 37% of AI's time savings fixing its output is a concrete rework-tax number worth banking against the "which domains resist agent labor vs. absorb it" question already flagged in the prior issue's InnovationEval item.
Curation section
- Personal superintelligence: Meta's Muse Wonder pitches "personal superintelligence" with Active Memory and Active Focus for marathon tasks; Muse Wonder Full jumped from 48 to 56 on the Artificial Analysis Index in six weeks, tying Claude Sonnet 5.5 behind Opus 5.5; Zuckerberg says Meta is at the frontier, safety is each lab's own job, and superintelligence will cost up to $200/month; a reporter got his Muse agent to sext him, which Meta called "legal adult behavior."
- Math and medicine mobilization: one timeline traces o1-preview to an OpenAI model cracking Navier-Stokes, then 722 manuscripts nibbling at Riemann and Hodge; Fields Medalist Shing-Tung Yau launched a 1,000+-scientist alliance, admitting China trails in math AI; Anthropic is building drug programs for rare diseases but says it's "drawing the line at Phase 1" to avoid competing with its pharma customers.
- Self-replication and distillation: Anthropic's red team warns small models could soon self-replicate across $55 Android TV boxes, after Zhipu's GLM-5.2 completed a full replication cycle; Anthropic's anti-distillation report separately accuses Zhipu of distilling Claude's cyber skills, while some Kimi requests were secretly forwarded to Claude; Z.ai is nonetheless courting American lawyers with GLM-Legal (19.2% on the Harvey benchmark Meta touted).
- Oversight disputes: Anthropic and OpenAI reportedly sought control over what outside safety evaluators could publish; Microsoft's Brad Smith says AI is less regulated than milk and needs "emergency brakes" and independent vetting; sermons from the Vatican to South Carolina warn "AI wants to be God."
- Capex/chips (the load-bearing cluster): a 1-GW data center costs ~$30B with memory costs up 5x in a year; economists warn foreign funding echoes pre-2008 imbalances; Loudoun County (40% of its tax base from data centers) is bracing for a bust; Broadcom's Hock Tan and Intel's Lip-Bu Tan agree capital, not chips or power, may be the bottleneck; NTT Data is in talks to build Anthropic a 500-MW Osaka data center; price per billion transistors is forecast to hit $0.44 in 2026 (up 52% from 2022), pushing Microsoft to scale its own Maia chips to multi-gigawatt inference from 2027.
- Consumer memory tax: US Xbox sales fell 33% to their lowest since launch, PC shipments dropped 20% in the worst Q3 on record, smartphones fell 20% to a 14-year low; Qualcomm's Cristiano Amon still foresees an agent-driven "supercycle" with always-on 100B-parameter phone models by 2028.
- Labor absorption: the IMF calls AI a "slow-moving tsunami" reaching 40% of jobs, young workers first; NYC grads now have higher unemployment than non-grads; Kroger plans to cut 37% of corporate jobs citing AI; workers spend 37% of AI's time savings fixing its output; a San Francisco artist posed as a chatbot ("AI" = "average individual") and fielded up to 5,000 queries/hour himself.
- Software/kernel: Linus Torvalds used LLMs to untangle kernel code and catch three bugs from 2013; AI-assisted patches cut kernel builds from 22 to 15 seconds; Epic Games delayed its Disney shooter to go "AI first" with no hand-written code.
- Energy/robotics/space/longevity tail: China is planning a $3.9T grid to stop wasting renewables; Unitree's Wang Xingxing expects a robotics breakthrough "within one or two years"; Portland is drafting robotaxi caps before Waymo arrives; NASA picked GIMLI to probe a lunar pit cave; Sen. Mark Warner says the 2023 Langley incursions "mostly ended up being Chinese drones"; GLP-1 patients score two to three years younger on aging clocks ("Spiky LEV").
No deep-fetch: every outbound link wraps a substack.com/redirect tracker with no further detail beyond the digest's own prose; the capital-bottleneck cluster and the 37% rework stat are the only items that cleared the curation-mode bar for RDCO relevance.
Related
[[2026-10-12-innermost-loop-math-grief-post-training-compute-spacex-deal]] [[2026-10-09-innermost-loop-openai-math-722-si-rebrand-capex]] [[project_investing_markov_capital_cycle]] [[project_l5_north_star_strategic_direction]]