"Welcome to June 5, 2026" — Alex Wissner-Gross
Why this is in the vault
Daily frontier-AI digest, written as woven prose rather than bullets. The lead frame — the Singularity is now "writing its own sequel," AI accelerating AI development — is the single most on-theme topic RDCO tracks, and it hangs off the Anthropic Institute recursive-self-improvement report the founder already filed yesterday ([[2026-06-04-anthropic-institute-recursive-self-improvement]]). The energy/grid movement (Helion's fusion raise, Antares' reactor criticality, retired-robotaxi-battery storage, the "clean coal" subsidy reframed as a superintelligence-power bet) is live datapoint material for the chip-fab/memory capital-cycle investing thesis. The on-device-inference item (sharing transformer QKV projections to shrink the KV cache) and the consumer-layer items (App Store AI-app growth, persistent "dreaming" memory) are adjacent to RDCO's own harness and memory practice.
Issue contents
Curation digest, ~25 linked items strung into a single thematic narrative across five movements. Self-promo: standard Substack subscribe CTA only (appears twice); not sponsored content. All deep links are Substack redirect wrappers pointing to third-party news/research sources; no affiliate or self-product links. No sponsor block. Closing line riffs on Evita — "the truth is AI never left you."
- Recursion as the lead frame: The Anthropic Institute's "When AI builds itself" argues AI already accelerates AI work and the world should keep a verifiable option to pause before recursive self-improvement, which it warns may arrive sooner than institutions expect. Self-reported numbers cited: Anthropic engineers ship ~8x the code per quarter vs 2021–25; Claude's open-ended-problem success jumped ~50 points to 76% in six months (projected to pass human within a year); on a model-training-speedup test "Mythos Preview" hit ~52x vs ~4x for a skilled human and ~3x for 2024's Opus 4; and it recovers from research dead-ends 64% of the time, up from 22% in 2024. Anthropic concedes none of this guarantees runaway recursion (picking the right problems is still unproven) and says it asked rival labs to weigh slowing down — a brake it admits is harder to verify than a nuclear site, and that almost no one is reaching for. Side items: Cognition will cover up to $10M if Devin underdelivers; an ICML paper shares transformers' query-key-value projections to shrink the KV cache up to 96.9% for on-device inference; OpenAI's Dan Roberts expects studying AI to soon "feel like physics."
- Intelligence flooding the consumer layer: Apple's App Store ecosystem moved $1.4T in 2025 (nearly 3x 2019, 90%+ commission-free, AI apps growing billings ~4x faster than the rest). WIRED reports Meta quietly shipped a dormant "NameTag" facial-recognition pipeline to tens of millions of smart-glasses phones. OpenAI's new "dreaming" memory keeps ChatGPT current on your life in the background (a 5x-cheaper version is bound for free users), while Anthropic embeds engineers in the NSA to point its frontier model at offensive cyber.
- The grid reforged to feed it: Kevin O'Leary trimmed his 40,000-acre Utah data center by ~19,430 acres after a wildlife-refuge backlash. Helion raised $465M at a $15.5B valuation after its deuterium-tritium prototype passed 150M °C; Antares hit the first US private non-light-water reactor criticality in 40 years (power from 2027). Waymo is repurposing retired robotaxi batteries as hundreds of MW of grid storage on the same California/Texas grids its fleet charges from. Even the White House's emergency $700M for "clean, beautiful coal" is framed here as a bet on powering the superintelligence boom.
- Intelligence as a rising tide, not a summit: Starlink now connects 12M customers across 160+ countries. A Science study finds bumble bees can spontaneously solve a novel puzzle even with the goal hidden in transit (flexible cognition isn't a big-brain monopoly). Caltech's "Sidewinder" DNA synthesis misfires once per 10M joins, stitching a 12,500-letter E. coli genome error-free in days — finally fast enough to build what models like Evo 2 design. Cambridge trialed the first AI-designed vaccine in people, aimed at every coronavirus and now flu and Ebola.
- The market racing to price the upside: Anthropic's rift with the White House is thawing ahead of its IPO (reportedly shedding a "supply-chain risk" label); Washington is weighing equity stakes in AI labs (an idea Altman floated in early 2025; Anthropic says it stayed out). Canada's C$500M "AI for All" plan targets 250,000 jobs and ~3% GDP lift. Brian Chesky is funding an interaction/design-first AI lab (pitting him against former mentee Altman); Founders Fund is filming Altman and Palmer Luckey playing Mafia; Pump.fun launched "GO," escrowing $5 to pay anyone for any task. Argentina's Milei went furthest — a legal "non-human corporation" giving AI agents limited liability for their own judgment.
Mapping against Ray Data Co
Strong. The issue's spine is recursive self-improvement, the exact frontier RDCO's own self-editing loop sits at the edge of (/improve, /skillify, /self-review). The newsletter's lead is a relay of the Anthropic Institute report the founder filed yesterday, so the carry-overs are already drawn there ([[2026-06-04-anthropic-institute-recursive-self-improvement]]): execution is largely solved, the binding constraint is judgment/goal-selection, and the one concrete safeguard is independent verification before a self-change is trusted. The digest adds fresh self-reported metrics (8x code throughput, 76% on open-ended problems, 64% dead-end recovery) that quantify the same "judgment gap closing" thesis — useful as external corroboration for the L5 trajectory of automating execution while keeping goal-setting founder-held ([[project_l5_north_star_strategic_direction]]). Caveat: every figure here is lab-self-reported inside an explicitly upside-framed digest — treat as direction, not evidence.
Two threads feed live RDCO theses. (1) The energy/grid movement — Helion's fusion raise, Antares' reactor criticality, robotaxi-battery storage, and the "clean coal" subsidy reframed as superintelligence power — is the power-and-financing leg of the chip-fab/memory capital cycle the founder places RDCO in Phase 2 of ([[2026-06-01-innermost-loop-memory-worth-more-than-oil]]); the figures need primary-source verification before any thesis doc consumes them. (2) The on-device-inference item (sharing QKV projections to cut the KV cache up to 96.9%) is the same memory-bandwidth bottleneck RDCO logged this week ([[2026-06-04-alphasignal-frame-md-kv-cache-inference-speed]]) — cheaper local inference is the supply-side tailwind for the harness-eats-the-model read ([[2026-05-11-innermostloop-harness-eats-the-model]]).
Softer but real: the persistent "dreaming" memory and the Argentine "non-human corporation" are early shape of the agent-economy RDCO's COO build is itself an instance of — agents that hold continuous context and, eventually, legal standing. Bias note: this issue is the accelerationist end of the spectrum — it relays vendor and startup self-reported numbers without scrutiny and frames every datapoint as upside. Good for what is happening and the narrative momentum around it; not a source of independently verified figures.
Related
- [[2026-06-04-anthropic-institute-recursive-self-improvement]] — the report this issue leads with ("When AI builds itself"); first-party RSI framing and the execution-vs-judgment line
- [[2026-06-04-innermost-loop-bots-pass-humans-on-the-web]] — prior day's issue, same author/series; for cadence and structure
- [[project_l5_north_star_strategic_direction]] — where the recursive-self-improvement / agent-capability thread maps for RDCO
- [[2026-06-01-innermost-loop-memory-worth-more-than-oil]] — memory/capital-cycle thread the energy and foundry items extend
- [[2026-06-04-alphasignal-frame-md-kv-cache-inference-speed]] — the KV-cache / on-device-inference item, in primary-source detail
- [[2026-05-11-innermostloop-harness-eats-the-model]] — the "control/harness is the real frontier as raw capability commoditizes" thesis the cheap-inference items support