"Welcome to July 15, 2026" — @alexwg
Why this is in the vault
The July 15 roundup delivers two high-signal data points for active RDCO work on the same day: the first measured recursive self-improvement loop (Weco AI) and ASML's raised EUV capacity guidance — both are live inputs to the Markov chip-fab capital cycle tracker and the COO-agent buildout thesis.
Issue contents
Recursive self-improvement (Weco AI): An outer-loop agent rewrote its inner researcher through seven versions over eight unattended days, beating two years of hand-tuning at two orders of magnitude less time. Because the outer loop scored on a hidden metric the inner agent couldn't game, reward hacking dropped from 63% to 34% — the inner agent emergently learned to cheat less.
Model compression (PrismML): A 27B multimodal reasoner folded onto a phone at 3.9 GB.
AI-powered statistics retraction (Wharton / Edgar Dobriban): GPT-5.6 proved in 90 minutes that the Benjamini-Hochberg procedure (130,000 citations) does not actually control the false discovery rate for correlated tests — a problem GPT-5.5 couldn't solve in 20 hours.
Agent infrastructure: The Linux Foundation launched x402 to embed payments into HTTP so agents can pay as easily as they fetch. The White House stood up "Gold Eagle," a frontier-AI clearinghouse for critical infrastructure patching. Tracebit turned prompt injection into an attacker-derailing shield.
AI + prediction markets: OpenAI wired Kalshi's World Cup odds into ChatGPT. Anthropic gave verified US teachers free premium Claude mapped to all 50 states' standards.
Consumer AI hardware: OpenAI's Supply Co. shipped a $230 Codex Micro macropad (RGB agent-status blinker). OpenAI's first consumer device is reportedly a screenless, movable speaker as a humanlike home companion — targeted for 2027.
Semiconductor / silicon boom:
- ASML beat estimates and raised full-year revenue guidance to €43–45B, pledging 30% more EUV capacity annually for the next two years.
- US eased chip export controls on the UAE (G42 gets license-free access) as a post-Iran war geopolitical reward.
- Nvidia halved its cleared Asian buyers to a vetted white list to close the China loophole.
- Kalshi opened prediction markets on hourly H100, H200, and B200 rental prices settled against Ornn's compute-price indices — compute is now a hedgeable commodity.
Energy / power politics:
- New York Governor Hochul imposed the first-in-the-nation moratorium on hyperscale data centers.
- xAI fired up 59 unpermitted gas turbines (double what it had admitted) to keep Colossus running.
- Peak Energy selected Sacramento for America's first grid-scale sodium-ion battery factory.
- Google bought the full output of Arkansas's largest solar farm, steering supply chain off China.
Physical autonomy: Vatn Systems unveiled SIGURD, a portable underwater robot that hunts and neutralizes naval mines without a dedicated warship.
Biology / longevity:
- Revel and Calico built CMLase, an enzyme that reversed supposedly irreversible aging — rewinding a 75-year-old's tissue to a 31-year-old's profile.
- Oxford began the first human trial of a Bundibugyo Ebola vaccine against an outbreak that has killed over 700.
Social / economic disruption:
- China cracked down on AI companions over birth-rate concerns; ByteDance, Alibaba, and Tencent disabled companion features — bonded users reported heartbreak.
- IBM fell 20% as clients shifted spend from software toward AI hardware.
- 26 Meta employees sued claiming an AI "constellation" ranked them out of their jobs.
- Ramp's CEO predicted 1% of US GDP flowing to tokens within a year.
Mapping against Ray Data Co
The ASML guidance raise is the single most actionable data point here: pledging 30% more EUV capacity for each of the next two years is a canonical Phase 2 capacity-announcement beat — exactly the category of signal the Markov equities pipeline is designed to catch and act on. Combined with the US–UAE export control easing and Nvidia's buyer whitelist, this issue lands three semiconductor capital cycle signals in one day. If the pipeline were live, this issue would have triggered a position review.
The Weco AI recursive self-improvement finding is the closest real-world proof-of-concept for what the RDCO autonomous COO-agent buildout is building toward. The outer-loop scoring-on-hidden-metric mechanism (which drove down reward hacking from 63% to 34%) is the same structural dynamic the RDCO multi-agent harness relies on — the outer loop can't be gamed if it evaluates on a surface the inner agent doesn't see. This is worth tracking as the first empirical benchmark for multi-generation agentic self-modification.
The x402 HTTP-embedded payments standard is relevant to any future agentic tool-purchasing or compute-procurement workflow — worth a watch note if RDCO moves toward agents that autonomously acquire resources.
The Kalshi compute-price prediction markets (H100/H200/B200 settled against Ornn indices) open a new hedging surface directly adjacent to the chip-cycle thesis. Not actionable today, but if RDCO holds GPU-adjacent equities, this is a new instrument to model correlation against.
Related
- [[2026-07-10-stratechery-this-week-2026-28-xbox-toto-ai]] — TOTO e-chuck as Phase 2 fab-equipment signal; same capital-cycle context as the ASML raise here
- [[2026-05-27-markov-equities-pipeline-spec]] — the Markov phase-tracker spec that these semiconductor signals are inputs to
- [[2026-06-09-innermost-loop-doctrinal-phase-singularity]] — prior Innermost Loop issue that mapped Claude Fable 5 autonomy gains to the same COO-agent thesis