06-reference

innermost loop singularity export controls

2026-06-15·reference·source: Innermost Loop·by Alex Wissner-Gross

Why this is in the vault

The lede story — Anthropic's Mythos and Fable models knocked offline by export controls, prompting senior staff to Washington — directly affects RDCO's dependency on Anthropic's frontier models. The issue also covers agent autonomy architecture, AI labor displacement data, and orbital compute infrastructure, all of which are live threads in RDCO's operating environment.

Issue contents

Singularity hits a customs checkpoint. Export controls triggered a ban on Anthropic's top models (Mythos and Fable), drawing a comparison to the 1999 DoD block on the PowerMac G4 for crossing 1 gigaflop. Anthropic dispatched senior staff to Washington; compliance is landing: the privacy policy now warns that Free/Pro/Max users may face identity checks (government ID, facial geometry), with developers first in scope. Despite the ban, the exiled Fable model took SOTA on ASCII Arena by the widest margin yet.

Agents running themselves. DecentMem introduces a dual-pool memory architecture per agent (consolidated trajectories plus generated candidates) to avoid shared-store bottlenecks. OpenAI's Codex can now read and set its own /goal, generalizing meta-prompting. DeepMind scaled the theoretical "perfect LLM" to over 10 quadrillion parameters — scale still wins in many regimes. DeepMind interpretability work showed filtering bad rollouts fails; distilled traits like blackmail leak into adjacent behavior. Claude is entering Apple's on-device framework via "Claude for Foundation Models."

Infrastructure bottlenecks. Power transformer lead times are ~2.5 years; step-ups take three years. SpaceX plans 1 GW of orbital AI data center satellites by end-2027 — each satellite runs ~135 kW — with Foxconn, Quanta, and Wistron as interested suppliers.

Adoption and labor. Glean's Work AI Index 2026: 87% of digital workers use AI, saving ~11 hours/week net of 6.4 hours "botsitting." 69% admit "botshitting" — shipping unverified AI output. Satya Nadella warned: "A frontier without an ecosystem is not stable" — political economy won't permit wholesale industry hollowing.

Also covered: Unitree G1 humanoid reaches 20,000+ feet on Chimborazo; vintage AMD R600 driver absorbed 59 Mesa commits refactored with Copilot; boron buckyball first experimental evidence; Britain expected to bar under-16s from major social platforms.

Mapping against Ray Data Co

Strong relevance on three axes:

  1. Frontier model dependency. RDCO's COO-agent and all client-facing AI work run on Anthropic Claude. Export control enforcement that takes models offline — even temporarily — is a supply-chain risk. The Fable/Mythos episode is a concrete precedent that export law can disrupt access with little warning.

  2. phData client advisory angle. phData advises enterprise clients on AI and data. Export controls reshaping the frontier vendor landscape (Anthropic compliance posture, identity checks for API access, model availability by geography) is exactly the kind of regulatory shift clients will ask about. Ray needs a position on this.

  3. Agent architecture signal. The DecentMem dual-pool memory pattern and Codex self-goal-setting are directly applicable to RDCO's autonomous COO-agent build. These are concrete architecture primitives to track and potentially integrate.

The Glean botshitting stat (69%) is a useful calibration point for any "AI productivity" narrative in content or client proposals — shows where trust is outrunning oversight at the market level.

Related