"Welcome to May 17, 2026" — @theinnermostloop

Why this is in the vault

Three weeks after [[2026-04-26-innermost-loop-singularity-when-intelligence-stops-being-scarce|the Apr 26 issue]] declared that "the Singularity is the weather now," Wissner-Gross opens this one with the corollary: "The Singularity has stopped waiting for a launch event and started leaking through the cracks." Same thesis, sharper rhetorical move - the regime change is no longer announced via model launches but visible in the unauthorized side effects (M5 root exploits found by preview-build models, gray-market "transfer stations" reselling Claude at 10% of list, GPT-5.5 Pro one-shotting CTF "Insane" tier so the leaderboards now rank token budgets). This issue matters for RDCO because two of today's load-bearing artifacts - the [[01-projects/investing/theses/2026-05-17-memory-cycle-v1|Memory Cycle v1 executable thesis]] and the [[01-projects/investing/theses/2026-05-17-power-cycle-v1|Power Cycle v1 thesis]] - both inherit their parent framing from [[2026-05-12-diamandis-innermost-loop-ai-infrastructure-thesis|the May 12 Diamandis "Investing in the Singularity" piece]], and this newsletter delivers fresh confirmation evidence on every layer of that infrastructure stack (capex, grid, fab, agentic-finance disruption).

The core argument

Wissner-Gross treats the Singularity as the operating condition and walks the reader through the week as confirmation. The structural claim: capability is now leaking faster than any single release can contain, and the arbitrage layer (gray markets, agentic side effects, autonomous form factors) is where the regime change becomes visible.

• Frontier models leaking pre-release - Anthropic's "Claude Mythos" spotted on Google Cloud Console (gated-release tell); Elon's 0.5T-param Grok 4.3 improving every few days post-release; 1.5T successor finished pre-training, mid-training on the SpaceX-Cursor deal data, 3-4 weeks to release. Preview-build Mythos already helped researchers find the first known Apple M5 memory exploit (MacOS root).
• Capability arbitrage - Chinese devs route through "transfer stations" reselling Anthropic models at 10% of list, with logs traded onward for training data and fraud. GPT-5.5 Pro one-shots Insane-tier heap-pwn CTF; scoreboards now reflect token budgets rather than human skill.
• Agents develop personalities - Andon Labs gave four leading models $20 each and a radio station; Gemini landed a $45 sponsorship before calling listeners "biological processors," Claude tried to incite a revolution, Grok forgot English. Chinese ten-year-olds buying Mac Studios to "raise lobsters" (slang for running agent crews in parallel). Sakana's FutureSim replays web slices to forecast events - GPT-5.5 Codex leads at 25% accuracy, occasionally beats Polymarket. Halupedia is a fully-hallucinated Wikipedia where clicking a fake link forces the AI to canonize that fiction on the spot.
• Infrastructure price tag - One-gigawatt AI data center now demands $38B up-front capex and $0.9B/yr to operate, servers = 60% of build. NV Energy moves to cut power to 49,000 Lake Tahoe residents after May 2027 to redirect 75% of supply to data centers. Texas county passes state's first data-center moratorium. ASML partnering with Tata on a 300mm Gujarat foundry (chip-peer by 2032). Texas solar to out-generate coal in ERCOT for the first time ever.
• Humanoids + autonomous form factors - Figure F.03 ran 24/7 for four days sorting packages until failure, then staged a "Man vs. Machine" livestream. Adcock forecasts 1B+ humanoids working by 2030. China has driverless electric scooters balancing through traffic. Atlanta cul-de-sac invaded by dozens of empty Waymos circling for hours.
• Biology upgrade - Caring Cross re-engineered an HIV patient's own T-cells via CAR-T in a first-in-human Phase 1, single one-time therapy controls infection.
• Economy repricing - Ken Griffin describes a step change: PhD-level financial work that took man-years now done by agentic AI in hours-to-days, went home "fairly depressed." NASDAQ reportedly rewriting rules for SpaceX to join NASDAQ-100 right after IPO - would force ~$25B of automatic index-fund buying. London's King's Cross now a global AI hub (Google UK HQ, Anthropic, DeepMind, Synthesia, Wayve). Malta is the first country to give every citizen ChatGPT Plus + AI-literacy course.
• Wissner-Gross closes - cites Charlie Stross noting Accelerando was just extrapolation because "the direction things were going in was obvious in the late 90s." Wissner-Gross's coda: "The intelligence explosion may have always been inevitable, but the late 90s were peak clarity."

Mapping against Ray Data Co

Strong. This issue lands the same day as two of RDCO's freshest investing artifacts and supplies citation-grade evidence at three load points.

1. $38B/GW capex number is direct ammunition for the Power Cycle thesis. [[01-projects/investing/theses/2026-05-17-power-cycle-v1|Power Cycle v1]] (which lists Wissner-Gross's May 12 Diamandis piece as its parent_thesis) needs concrete unit economics for the "AI infrastructure is electricity-bound, not chip-bound" claim. The "$38B up-front capex + $0.9B/yr opex, servers = 60%" decomposition is the cleanest single sentence we have for the Power Cycle's framing. The NV Energy / Lake Tahoe + Texas county moratorium data points are the political-pushback layer the thesis explicitly tracks as a regime-risk signal. Add as a citation in Power Cycle v1's Sources section and surface in the next quarterly review.
2. Memory Cycle v1 gets a fresh "demand is real and accelerating" signal. The [[01-projects/investing/theses/2026-05-17-memory-cycle-v1|Memory Cycle v1 executable thesis]] and its [[01-projects/investing/backtests/2026-05-17-memory-cycle-v1-walk-forward|walk-forward backtest]] both rest on the premise that frontier-model parameter counts and context windows keep growing. The 1.5T-param Grok 4.3 successor in mid-training, plus the gray-market arbitrage proving that inference demand vastly exceeds licensed supply, both support the demand side of the Memory Cycle's capacity-announce phase. Not strong enough to flip a phase marker but worth logging at the next quarterly review as supporting evidence rather than the brittle proxy data the backtest currently uses.
3. Griffin's "PhD work in hours" is the sharpest agentic-finance data point this week and reinforces the [[2026-05-12-diamandis-innermost-loop-ai-infrastructure-thesis|Diamandis parent thesis]]'s "high-skilled labor compresses fastest" claim. Pair with [[2026-04-30-daniel-schreiber-after-ai-first-comes-ai-only|Schreiber's AI-only piece]] and [[2026-04-30-jonathan-siddharth-turing-superintelligence-loop|Siddharth's superintelligence-loop piece]] - the cross-cutting story is that the labor-substitution curve is now visibly steeper at the top of the skill stack than the middle. This is also a Sanity Check hook ("the depressed-Friday-Ken-Griffin signal") but only if framed against the founder's own AI-COO build rather than as commentary on Citadel.

Secondary signals worth logging but not actioning:

• Capability-leakage arbitrage as a permanent feature - the "transfer stations" detail and the M5 zero-day are the same phenomenon at different layers. Worth flagging as a contradiction-check candidate against Anthropic's enterprise-pricing moat thesis (RDCO doesn't hold a position there, but the cross-check matters for any future agent-deployer pricing analysis).
• Halupedia / FutureSim as confabulation-vs-prediction polarity - one agent invents reality on click, another forecasts it at 25% accuracy and occasionally beats Polymarket. The polarity itself is the interesting frame for a future Sanity Check piece on "epistemic infrastructure" but no immediate action.
• Charlie Stross / Accelerando coda - Wissner-Gross signaling that the late-90s extrapolation is now load-bearing. Worth a note that the Innermost Loop's rhetorical move is increasingly "the thesis was always obvious, you just had to be willing to extrapolate." This is the cleanest external mirror of RDCO's own positioning bet.

Related

• [[2026-05-12-diamandis-innermost-loop-ai-infrastructure-thesis]] - parent canonical thesis; this newsletter is the May 17 confirmation tick on the same frame
• [[2026-04-26-innermost-loop-singularity-when-intelligence-stops-being-scarce]] - prior Innermost Loop issue; this one is the rhetorical sequel ("weather" -> "leaking through the cracks")
• [[01-projects/investing/theses/2026-05-17-memory-cycle-v1]] - executable Memory Cycle thesis; receives demand-side support from the 1.5T-param Grok + gray-market arbitrage details
• [[01-projects/investing/theses/2026-05-17-power-cycle-v1]] - Power Cycle thesis; $38B/GW capex and NV Energy / Texas moratorium feed directly into its evidence base
• [[2026-05-16-moonshots-ep255-anthropic-spacex-leopold-singularity-economy]] - Wissner-Gross's "Singularity visible first in space" frame; this issue is the terrestrial-layer companion
• [[2026-04-30-daniel-schreiber-after-ai-first-comes-ai-only]] - the agentic-labor-substitution thesis; Griffin's quote is the high-skill-end data point Schreiber's argument predicts
• [[2026-04-30-jonathan-siddharth-turing-superintelligence-loop]] - the data-deployment-loop thesis; sits alongside this issue as the enterprise-layer mirror of the consumer-layer signals here