06-reference

innermost loop anthropic ipo rsi gdpx moratorium

2026-11-02·reference·source: The Innermost Loop·by Alex Wissner-Gross
ai-capexanthropic-iporecursive-self-improvementagent-evalsdata-centersvibe-manufacturing

Why this is in the vault

Daily curation digest: Anthropic's IPO filing shows internal Claude usage (priced at retail) has out-valued its own cash payroll since April while writing >90% of its new code, and two independent real-job agent evals (SemiAnalysis's GDPx, OpenAI's LifeVal) both show agents still floundering outside narrow domains (12% on GDPx, zero on finance/legal) even as one domain-specific result — Claude Opus 5.5 producing a permit-quality house blueprint set in 12 hours — lands as a concrete data point for the founder's "vibe manufacturing" direction.

The core argument

Capability and compute-politics are both accelerating past their current container. On the capability side, Anthropic's self-described recursive-self-improvement loop (Claude writing its own replacement code, internal usage value exceeding payroll) is now a disclosed financial fact, not a claim — paired with $895.4B in committed compute through 2045. But two separate "put agents in real jobs" evals (GDPx, LifeVal) both show the frontier is domain-dependent and uneven: near-zero on finance/legal, 80%+ on caregiving/paperwork, and a single standout (Opus 5.5 architecture blueprints) that reads as an existence proof rather than a general capability. Meanwhile the politics of the buildout caught up to the capex: data-center moratoriums are now a live electoral issue in 21 of 36 governor races, pushing OpenAI toward a "AI compute campuses" rebrand and self-throttling grid pledges, and compute itself is being treated as a strategic state asset (Putin's 10% capacity reservation order).

Curation section

Mapping against Ray Data Co

The architecture item is the most concrete connection. Claude Opus 5.5 producing a permit-quality house blueprint set in 12 hours for a few hundred dollars is a second, independent data point for the founder's post-10-05 "vibe manufacturing" / programmable-engineering direction ([[project_ambition_bigger_problem_network]]), after the ScanRuler/CNC Kitchen MCP-driven scan-to-CAD loop already filed at [[2026-10-31-cnc-kitchen-free-tool-3d-scans-to-cad-elegoo-m1-ultra]]. Both show the AI-design layer clearing a specific, narrow, high-stakes domain (permit-grade CAD output) rather than general design competence — directly consistent with the gap the [[06-reference/research/2026-10-15-vibe-manufacturing-trend-map]] already mapped, and worth a line in [[01-projects/positioning/2026-10-18-prompt-to-product-founder-note]].

The GDPx/LifeVal split (12% on real-job agent tasks, zero on finance/legal, vs. 80%+ on caregiving/paperwork) is a clean, sourced data point for the founder's "can Ray do X" targeting filter ([[feedback_targeting_system_prioritization_filter]]) and for phData's "domain = agents in production" credibility positioning ([[project_credibility_for_phdata_sales]]): current frontier agents are competence-uneven by domain, not uniformly capable, which is the exact nuance that positioning needs to hold instead of flattening into "agents can do X now."

The data-center-moratorium electoral item (21 of 36 governor races) is a fresh, concrete instance of the political-constraint thread this newsletter has tracked since August and that RDCO's L5 north-star framing treats as the binding variable on bet timing ([[project_l5_north_star_strategic_direction]]), continuing directly from yesterday's capital-tightening issue.

Related