Innermost Loop — June 13, 2026: Export-Controlled Singularity

Source: theinnermostloop.substack.com

Summary

Wissner-Gross's lead: the US government issued an export-control directive suspending all foreign-national access to Fable 5 and Mythos 5, forcing Anthropic to disable both models for hundreds of millions of users. Dean Ball is quoted noting the accidental irony — because Anthropic's non-US staff can no longer interact with the models, this is "the first regulation on recursive self-improvement." Wissner-Gross frames the directive as a geopolitical inflection: non-US revenue at risk of zeroing out, foreign workforces benched, Europe likely to respond with another robotics/image-gen fund, and China becoming "more AGI-pilled" and restructuring its entire economy around the goal.

Second item: Epoch AI shipped FrontierMath v2 after an AI audit corrected errors in 42% of problems. Benchmark scores jumped; GPT-5.5-xhigh reached 85% on Tiers 1–3 and 73% on Tier 4 (up from 35%). The implication flagged by one Epoch researcher: math capability was "actually cooked months ago" — the instruments were wrong, not the models. Fable 5 had already topped the board at 87% Tiers 1–3 and 88% Tier 4 before being disabled.

Curation section

Remaining items cover:

Semiconductor buildout: Infineon opening a €5B Dresden fab on July 2 (EU Chips Act subsidies); Samsung's Dongtan factory town seeing luxury property inflation from AI bonuses; data-center opposition blocking ~$130B in projects in Q1 2026.
Compute orchestration shift: SpaceX handed full Colossus 1 capacity in Memphis to Anthropic after Grok training hit latency/aging-network walls across three campuses.
Energy density: First crewed fixed-wing flight on solid-state batteries (Helios Horizon) at 410 Wh/kg — 60% over prior lithium-ion.
Biology: China offering CAR-T at $150K vs $475K US; NY Genome Center's SeqTag profiling transcriptome + chromatin + histone at once; CRISPR-Cas12a2 targeting mutated p53 in cancer cells.
Labor market: CS junior postings up 47% while CS hiring fell 73%.
SpaceX IPO: Debuted on Nasdaq up 19%, raised $75B at $1.77T valuation; Gwynne Shotwell hinted at Tesla tie-up; Musk first trillionaire at $1.05T.
Meta: Capping internal AI token usage, steering engineers toward MetaCode after AI spend forecast to reach billions.

Why this is in the vault

The export-control story is a macro policy event with direct operational relevance: if US-only access restrictions become the default shape of frontier-model governance, RDCO's agent infrastructure (currently Anthropic-heavy) may need a contingency path for non-US sub-agents or client contexts. The FrontierMath v2 correction is a methodological caution — benchmark-based capability claims can be systematically wrong for months. Both items inform how confidently to cite AI capability claims in client-facing work.

Mapping against Ray Data Co

Strong signals:

Export controls on Fable 5/Mythos 5 → if phData has non-US clients or staff, this is a direct risk to any Anthropic-backed solution architecture in deals. Worth flagging in DSA discovery checklist.
FrontierMath benchmark correction → reinforces the "don't cite benchmarks as ground truth in client decks" discipline; model capability claims should be qualified.
Compute orchestration shifting from "owning compute" to "orchestrating it" → directly supports the RDCO L4→L5 thesis around agent orchestration as the durable moat.
SpaceX IPO + Musk trillionaire → capital-cycle signal: frontier AI infrastructure capital continues to consolidate at the very top (Anthropic/SpaceX/xAI); the chip-fab/memory capital-cycle thesis holds.

Medium signals:

Data-center opposition ($130B blocked in Q1) → energy/permitting constraint is now a real bottleneck, not just a talking point; affects the "how fast can the buildout scale" question in any investing thesis.
CS hiring collapse (−73% junior, +47% postings unfilled) → labor market signal for how AI's effect on technical workforce is already landing, not theoretical.

Weak/background:

Biology/genomics items, UAP files — no direct RDCO relevance.

⚠️ Author Positioning Note

No sponsorship detected. Wissner-Gross's AI-progress framing is optimistic/accelerationist by default — his curation consistently selects for "frontier advancing" signals. Geopolitical caution items (export controls, opposition movements) are included but framed as obstacles, not inflection points that might slow the trajectory. Read his bearish items as underweighted.

[[2026-06-12-innermost-loop-spacex-ipo-dyson-swarm-frontier]] — prior issue covering the SpaceX IPO setup and Dyson swarm frontier; this issue confirms the IPO landed at $1.77T
[[2026-06-10-innermost-loop-rationing-the-recursion]] — covers compute rationing and orchestration constraints; directly precedes the Colossus 1 / Anthropic handoff story here
[[2026-06-09-innermost-loop-doctrinal-phase-singularity]] — doctrinal framing of the singularity phase; the export-control story here is the first major regulatory event in that framing
[[2026-06-03-innermost-loop-benchmark-not-license]] — benchmark methodology caution; the FrontierMath v2 correction in this issue is the empirical case study for that concern