"Welcome to May 26, 2026" — Alex Wissner-Gross

Why this is in the vault

Innermost Loop is a daily AI/frontier-tech digest written as one flowing prose stream that threads ~25 third-party links across six "frontiers" (compute/AI, semiconductors-quantum, robotics, energy-space, biology, institutions). It is curation with a light editorial voice, not an original argument. Filed because two items cross the RDCO relevance threshold on agent architecture and one corroborates a thesis already in the vault (orbital compute). The lead item (Microsoft SkillOpt) is already deeply mapped in a standalone note from 2026-05-26, so this digest's value is the second item ("Language Models Need Sleep") plus the cross-reference, not a re-litigation of SkillOpt.

Issue contents

Six thematic clusters, connective-tissue framing only (no thesis):

• AI / recursion: "Language Models Need Sleep" (CMU/UMD) — periodic offline consolidation of recent context into persistent fast weights in SSM blocks before clearing KV cache; BenchBench (can a model author a benchmark stronger models can't clear; GPT-5.2 leads); Microsoft SkillOpt (skill doc as trainable state of a frozen agent).
• Semiconductors / quantum: imec fabricates a quantum-dot qubit via High-NA EUV (~6nm gate gaps); IBM + Dept. of Commerce back "Anderon," a pure-play quantum foundry ($1B CHIPS + $1B IBM match).
• Robotics: Hyundai/Boston Dynamics humanoids playing soccer; Waymo robotaxis as an independence tool for blind passengers.
• Energy / space: Aalo Atomics 10MW microreactor hardware complete; Ferrari "Luce" EV ($640k, Jony Ive interior); JAXA Mach-5 ramjet ground trial; hydrogen studied for wormhole signatures (ER=EPR).
• Space-as-compute: Musk to present an AI-satellite design "within weeks"; SpaceX selling AI compute as a service post Anthropic partnership, orbital data centers on roadmap; orbital solar scaling 10MW → 100MW → 1,000MW per Starlink generation.
• Biology: Eli Lilly VERVE-102 one-dose permanent LDL lowering; Westlake handheld cancer-lab device (~10,000x ELISA sensitivity); Lilly weight-loss revenue ≈ ChatGPT + Claude combined; a 12-fold rise in fabricated references in biomedical literature over three years.
• Institutions: federal "anti-technology extremist" framing post CEO-attacks; China restricting overseas travel for senior AI staff; frontier-lab "patron saint" jokes re: Vatican/Anthropic encyclical; AI supercharging pro se litigation; Tether's Georgia stablecoin; ~70,000 MA rideshare drivers win union certification.

Mapping against Ray Data Co

Medium overall — one strong item already covered, one genuinely new and on-thesis, the rest off-thesis frontier color.

• "Language Models Need Sleep" — strong, and NEW to the vault. Offline consolidation of recent context into persistent fast weights before clearing cache is a mechanism-level answer to the context-rot problem RDCO manages by convention (CLAUDE.md rule 4: route long artifacts through subagents; the Thariq 1M-context guidance). The paper's framing — push extra compute to "sleep" so wake-time latency stays flat — is the research analog of RDCO's "process big artifacts out-of-band, keep parent context lean" discipline. Worth a future watch: if consolidation-into-weights matures, it changes the calculus on what must be subagent-routed.
• Microsoft SkillOpt — strong, but already mapped. Cross-references the standalone note [[2026-05-26-skillopt-self-evolving-agent-skills]] (validation-gate discipline, bounded edits, compact skills, the /improve critique). No new analysis needed here; the digest merely surfaces it again.
• SpaceX orbital compute / Anthropic partnership — medium, corroborates existing thesis. Same-day Stratechery piece [[2026-05-27-stratechery-spacex-ipo-data-centers-in-space]] covers this with depth; the digest is a thinner confirmation, not independent signal.
• 12-fold rise in fabricated biomedical references — medium, indirect. Reinforces RDCO's source-fidelity and verification posture (verify-* critics, source_fidelity frontmatter): as AI-generated citations pollute the literature, independent verification of any cited claim becomes load-bearing, not optional.
• Everything else (qubits, Ferrari EV, ramjets, wormholes, soccer robots, stablecoins, rideshare union) — skip. Frontier-tech interesting, off the RDCO agent-deployer / data-infra thesis. No forced mapping.

Curation section — notes

All links are third-party (arXiv papers, company announcements, news outlets). No sponsor placements, no paid items detected. Self-promo is limited to the standard Substack "Subscribe for free" boilerplate (free newsletter, no paywall). Editorial voice is the only first-party content — thematic riffs ("Quantum has officially exited the lab and entered the fab"; "exaflops to go before we sleep") that frame but do not argue.

Deep-checked third-party items (cap 2):

• "Language Models Need Sleep" — arXiv 2605.26099 (CMU/UMD). Verified: offline recurrent passes update fast weights in SSM blocks via a learned local rule; tested on cellular automata, multi-hop graph retrieval, and a math-reasoning task where vanilla transformers and SSM-attention hybrids fail.
• SkillOpt — arXiv 2605.23904, microsoft/SkillOpt. Verified: text-space optimizer, validation-gated edits, best-or-tied in 52/52 settings, +19.1 pts inside Claude Code. (Already detailed in the standalone note.)

Related

• [[2026-05-26-skillopt-self-evolving-agent-skills]] — the digest's lead item, already deeply mapped
• [[2026-04-15-thariq-claude-code-session-management-1m-context]] — context-rot management; "Need Sleep" is the research-side mechanism
• [[2026-02-23-every-chatgpt-memory-context-rot]] — prior vault framing of context degradation
• [[2026-05-27-stratechery-spacex-ipo-data-centers-in-space]] — same-day deep coverage of the orbital-compute item
• [[2026-05-07-alphasignal-stanford-deep-learning-throttling-multiagent]] — adjacent daily-digest source covering SpaceX/Anthropic + multiagent