“Welcome to April 27, 2026” — @theinnermostloop

Why this is in the vault

Day-after companion to yesterday’s thesis declaration. Wissner-Gross opens with a riff on his own closing weather metaphor (“The Singularity is the weather now, not the forecast”) and then walks the reader through the day’s evidence as confirmation of the regime he just declared. The most load-bearing item for RDCO is OpenAI’s Noam Brown framing — model weights matter relatively less than securing inference compute, “the prize is no longer the recipe but the kitchen.” That single line reframes the moat conversation for any RDCO surface that depends on which substrate the agent runs on. The closing aphorism — “Matter learned to think, and now the thinking is learning to leave” — is also a clean Sanity Check hook on the substrate-leaving-the-planet thread (smartphone NPUs → hyperscale → orbital compute).

The core argument

The “extended era of roughly human-level AI” is the operating condition, and four simultaneous shifts confirm the regime change:

Frontier hedging from the people who built it — Bostrom calls the human-level era a surprise that has “stretched 3-5 years and may stretch further.” Hassabis, who once said AGI required 1-2 more breakthroughs, now thinks it is a coin flip whether more are needed at all. Altman openly mocks the “post-AGI nobody works” prediction by pointing at users adopting polyphasic sleep so they can ship more code with GPT-5.5 in Codex.
Strategic frame flipping from weights to inference — Noam Brown’s framing: model weights matter relatively less than securing inference compute. Recipe lifecycles also compressed: GPT-4o ran 21 months, GPT-5.4 lasted 49 days — a “mayfly schedule for synthetic minds.” The mayflies still produce: Liam Price, a 23-year-old with no advanced math training, used a single GPT-5.4 Pro prompt to crack an Erdős problem that had eluded prominent minds, prompting Terry Tao to muse that humans hit a “mental block” from “a slight wrong turn at move one.”
Substrate stretches from pocket to orbit — OpenAI working with MediaTek and Qualcomm on AI smartphone processors (Luxshare manufacturing, mass production targeted 2028); Apple has six new AI product categories in the pipeline (AirPods, smart glasses, pendants, smart displays, tabletop robots, security cameras). Kevin O’Leary planning a Box Elder County Utah hyperscale data center that will generate its own power, clean its own water for the Great Salt Lake, and consume more electricity than the entire state. AWS CEO: “we have never retired old A100s” — the post-obsolescence era of silicon. SpaceX shareholder package would grant Musk 60M additional shares if market cap climbs from $1.1T to $6.6T while delivering “100 terawatts of compute per year” from space data centers (orders of magnitude beyond peak US power consumption). ESA sealed six participants into a simulated Mars mission in Cologne with no exit until August.
Dual-use robotics + bio in parallel exponential — China State Grid deploying 500 humanoid robots for high-voltage operations (failure mode is now a melted servo, not a melted operator). 15 Ceres Air C31 chemical-spraying drones stolen in New Jersey (FBI investigating possible “nightmare scenario”) — deployment surface and attack surface expanding at the same rate. Intellia announced first Phase 3 success for an in vivo CRISPR treatment that actually edits a disease-causing gene. Sub-2-hour marathon barrier broken in London (Sawe + Kejelcha, Adizero Adios Pro Evo 3). Ozempic and Mounjaro outearned OpenAI and Anthropic combined in 2025.
Institutional layer scrambling — Berkshire Hathaway, Chubb, Travelers won approval to drop AI-related damages from corporate policies (excluding agents misusing copyrighted material in marketing). Bipartisan House AI bill targeting deepfakes + whistleblower protections. Supreme Court will decide whether geofence warrants violate the Fourth Amendment. Resilient analog holdouts: 70% more bookstores in the US than six years ago. Varda’s Andrew McCalip predicts an AI-led S&P 500 company in 5 years, an AI elected official in 20.
Closing aphorism — “Matter learned to think, and now the thinking is learning to leave.”

Mapping against Ray Data Co

Strong. Three load-bearing items:

Noam Brown: weights matter less than inference compute — “the prize is no longer the recipe but the kitchen.” This is the cleanest single articulation of the inference-compute moat we have seen filed this month. Pairs directly with Thompson’s opportunity-cost-of-compute frame and the Google ~25% of global AI compute datapoint from yesterday’s IL. For RDCO positioning this matters because the COO-as-Claude moat is not model weights (we don’t train) and not inference capacity (we don’t own data centers) — it is vault-grounding + channel-routing + accountability. Brown’s framing means the “you don’t have a moat because Anthropic ships your features” objection now has a sharper rebuttal: the actual moat in this regime is neither the recipe nor the kitchen, it is the household running on the kitchen’s output. Worth a Sanity Check angle on its own (“the kitchen is the new recipe”).
Model lifecycle compression — 21 months → 49 days. Direct operational implication for any RDCO skill that hardcodes assumptions about a specific model version. The 4o → 5.4 → 5.5 cadence is the new floor. Skills should be written to degrade gracefully when the underlying model is replaced under them, not to depend on version-specific behavior. This also tightens the SaaS-death-thesis synthesis timeline — when the recipe lives 49 days, software built to work with one specific model version has a vanishing shelf life.
Liam Price + GPT-5.4 Pro cracks an Erdős problem in a single prompt. Frontier-math democratization is the cleanest single data point this month for the abundance-of-intelligence claim. Pairs with Andon Labs’ Luna-the-store-agent from yesterday — together they bracket the range: a 23-year-old with no advanced training does work that eluded Fields-medalists; a Sonnet-4.6 agent autonomously runs a retail store. The “agent-deployer not data-engineer” hiring thesis hardens further. Strong Sanity Check candidate (“the mental block is the moat humans had”).

Secondary:

AWS CEO: “we have never retired old A100s.” Post-obsolescence-era silicon is the dual of the inference-compute moat — every chip ever made is still earning. Useful for the abundance-flywheel thread: even the waste compute is now scarce.
SpaceX 100 terawatts from orbital data centers (Musk pay-package trigger). Substrate-leaving-Earth is no longer rhetorical. Pair with O’Leary’s Utah hyperscaler that consumes more than the state — these are the two ends of the substrate-stretch arc Wissner-Gross is naming. Filed as a future Sanity Check opener candidate for the closing aphorism alone.
Insurers dropping AI-related damages (Berkshire, Chubb, Travelers). When the largest underwriters refuse to price the new risk surface, the risk is being privatized to the deployer. Direct relevance to anyone building production agent workflows for clients — RDCO’s accountability layer is where this risk lands. Worth a brief mention in the COO-as-Claude positioning doc as a forcing function for why clients should buy the deployment-with-accountability bundle rather than DIY the agent.
Ozempic + Mounjaro > OpenAI + Anthropic 2025 revenue. Useful corrective for any RDCO writing that frames AI as the only exponential. The longevity / GLP-1 sector is a heavyweight rival with a faster cash-flow curve. Sanity Check candidate: “the actually-largest 2025 platform play wasn’t a model lab.”

Curation section — notes

This is a curation issue dressed in thought-leadership chrome. Each evidence beat is a substack-redirect-wrapped citation; Wissner-Gross’s value-add is the connective tissue (“the prize is no longer the recipe but the kitchen”) not the link selection itself. No third-party sponsor block. Two standard “Subscribe for free” CTAs. No links back to other Innermost Loop posts. No self-promo of Solve Everything in the body. Did not deep-fetch any cited link — each passes a one-line headline test against the body summary, and any of them (e.g. Noam Brown’s exact remarks, the Liam Price methodology, the SpaceX shareholder package terms) belongs in a dedicated note if it warrants the threshold on its own.

2026-04-26-innermost-loop-singularity-when-intelligence-stops-being-scarce — yesterday’s thesis declaration; today’s issue is the day-after evidence walk
2026-04-23-innermost-loop-image-model-mirror-blink — prior Innermost Loop issue
2026-04-20-innermost-loop-singularity-bureaucratic-momentum — prior week’s “Singularity”-titled issue; same rhetorical arc
2026-04-17-innermost-loop-welcome-apr-17 — Opus 4.7 + Mythos cadence; complement to today’s “weights matter less than inference” frame
2026-04-13-innermost-loop-welcome-apr-13 — Jevons paradox + agents at every layer; same author book-day issue
book-solve-everything-master-synthesis-2026-04-13 — same author’s book; abundance-economics framework this issue continues to operationalize
book-solve-everything-ch6-the-engine-2026-04-13 — RoCS + abundance flywheel; the inference-compute-as-moat framing extends Ch6
2026-04-13-stratechery-mythos-muse-compute — Thompson’s opportunity-cost-of-compute frame; pairs with Noam Brown’s “kitchen not recipe”
research/2026-04-25-saas-death-thesis-vault-synthesis — Apr 25 synthesis; 49-day model lifecycle tightens the timeline

Copyright note