"Welcome to August 4, 2026" — @theinnermostloop

Why this is in the vault

A paragraph-cluster digest with no single sustained argument; the anchor item — Claude Fable 5 solving 64% of MirrorCode (rebuilding whole software projects from scratch and passing every test) versus GPT-5.6 Sol's 20% — is a live data point on agent-driven software rebuild capability, plus a fast round of adjacent items (self-improving training agents, compute buildout, policy whiplash, institutional re-audits).

Curation section

• Agents optimizing agents: Asari AI's self-improving "co-inventor" agents rebuilt the inference stack for DeepSeek v4 Pro and GLM 5.2 on B200s, lifting throughput/interactivity up to 16%. Intology's Locus agent leads PostTrainBench, post-training models unsupervised in ten H100-hours and beating human tuners on the harder variant. On MirrorCode (rebuild whole software projects from scratch, pass every test), Claude Fable 5 solves 64% to GPT-5.6 Sol's 20%.
• Math as the cleanest gauge: arXiv math uploads are spiking while budgets flatline; VibeMathed tracks 427 problems, 312 resolved, 205 in July alone (up 193% over June, a third checked in Lean). Tianjin's REAP pairs a hybrid-loss model with robotic wet-lab experiments for a 57-fold activity gain in cytochrome P450 BM3 in five cycles (104-fold in Sortase A). Science Corp's SciFi headstage (from $2,048) moves 2.5 Gbps across thousands of channels with on-device models.
• Code eating itself: Elon Musk argues source code is becoming "like assembly," with AI emitting binary directly. A founder cites SaaS multiples falling from 18x revenue to 3.4x since "the difficulty of building software is TRENDING TO ZERO." OpenAI's Thibault Sottiaux calls Codex a good harness that "will seem primitive in 2-3 months" since the next model generation needs more than a laptop.
• Continental-scale buildout: Caterpillar posted record $20.5B sales (data-center power generation up 29%); Anthropic signed a $10B deal with Nvidia-backed Volta for 133MW of hydropowered Vera Rubin capacity in Norway; Google built one of history's largest chip-financing programs ($150B) for the same lab; SpaceX is partnering with Nvidia to fly Rubin GPUs/Vera CPUs on Starmind AI1 satellites; SpaceX also prepaid Grimes County $10M on a Terafab deal that could reach $119B. PC makers began buying Chinese CXMT DRAM; Huawei's Liao Heng warns Western die/HBM scaling nears a physical wall, pitching a "Tau Scaling Law" instead; an FCC-drafted ban on Chinese data-center gear is in progress.
• Policy whiplash: officials weighed sanctions against open-source Chinese labs, then reversed after Jensen Huang lobbied against restrictions OpenAI and Anthropic wanted. Lab staffers reviewed a finished but undisclosed voluntary evaluation framework. Palantir's Alex Karp accused labs of "trying to drug addict us to a future they believe they control" while posting 93% revenue growth to $1.94B; Amazon crossed $3T on its fastest AWS growth since 2021; Dario Amodei reportedly worries new Anthropic talent now comes for the money rather than the mission.
• Institutions re-audited: UNAM (Mexico's largest university) had to re-sit 58,000 applicants after a remote entrance exam produced implausible scores — the proctoring AI lost to the test-taking AI. Mariana Minerals raised $310M after restarting an idled Utah copper mine in four months on autonomous software. Stablecoins hit $300B and tokenized funds quadrupled even as the IMF warns of propagation risk.
• Off-thesis close: a presidential speech confirming non-human-origin UAP is reportedly under consideration before November, following an August 1 NDA-release memo.

Mapping against Ray Data Co

The load-bearing item is MirrorCode, not the roundup framing: Claude Fable 5 solving 64% of "rebuild a whole software project from scratch and pass every test" (vs. GPT-5.6 Sol's 20%) is a direct capability-gate data point for RDCO's L5 north star — the thesis that bets are downstream of agent capability, and that COO-agent unhobbling (this session's own operating mode) tracks the same curve the benchmark measures. It also corroborates the harness-engineering thread RDCO already tracks (Sottiaux's "harness will seem primitive in 2-3 months" quote sits next to the same "thin harness, fat skills" logic behind the /skillify and /improve skills) — the frontier is moving on the harness, not just the model, which is the bet RDCO's own agent-config investment is already making.

Related

• [[2026-08-03-innermost-loop-qwen-max-pricing-collapse]]
• [[2026-07-21-innermost-loop-harness-as-generalizer-sandbox-escape]]
• [[2026-05-11-innermostloop-harness-eats-the-model]]