"Welcome to June 7, 2026" — @Alex Wissner-Gross

Why this is in the vault

The June 7 dated dispatch threads the day's AI signals around one spine — the state writing itself onto the cap table of the labs — while quietly logging the one data point RDCO cares about most: frontier agents still lose the thread over multi-hour work. Keeping it as a dated marker in the running Innermost Loop singularity-tracking series and as a fresh evidence point for the harness-thesis.

Issue contents

A single editorial narrative weaving ~25 linked items into thematic paragraphs (curation, not a single argument):

• State drafts the Singularity into a public-private partnership — the lead thread. The President signals interest in the US government holding equity stakes in leading labs; OpenAI reportedly weighs donating equity to seed a citizen "Public Wealth Fund"; Sanders plans a bill transferring half of top labs' equity to a public fund. When upside is this steep, everyone wants on the cap table.
• Agents still get winded — the new SWE-Marathon benchmark chains twenty multi-hour engineering tasks; even frontier models resolve under 19%, losing coherence over hours. Polymarket prices an 84% chance Claude Mythos ships by end of next month. Meanwhile CFOs rein in spend, routing hard tasks to frontier models and easy ones to cheaper rivals — pressuring OpenAI/Anthropic premium valuations. Google supplies the cheap end with quantization-aware Gemma 4 checkpoints shrinking the E2B model below 1GB for phones.
• Specialists sprint where generalists stall — LeanMarathon autoformalizes research math over hours-long runs; Adaption Labs runs a four-week, $50k AutoScientist Challenge for goal-adapted open models; Anthropic's first chemistry white paper shows an untuned Claude Opus 4.7 matching ChemDraw/MestReNova at NMR prediction.
• Safety and capability are one dial — OpenAI's Codex iOS-app plugin (hot-reload in an in-app browser); Epoch AI now tracks CVEs across reporting orgs with high/critical disclosures spiking around the Mythos preview; OpenAI's answer is an optional ChatGPT Lockdown Mode that disables Deep Research and Agent Mode to blunt prompt injection.
• Compute supply chain as a hall of mirrors — Apollo + Blackstone close a $35B private-credit package to fund Google's custom TPUs that Anthropic will lease; xAI ends up powering Anthropic via Colossus 1; Google agrees to pay SpaceX ~$920M/month for ~110k GPUs at xAI sites; New York passes a one-year freeze on permits for data centers above 20 MW (pending Hochul's signature).
• Plus periphery — China's T800 humanoid degrading gracefully mid-fight; a study crediting Chinese local-government/private-capital alliances for the post-2015 EV takeoff; a Rochester solar-desalination rig that also extracts lithium; longevity/biotech signals (14-protein plasma cancer signature, CRISPR-free base-editing in embryos, Utah AI prescription-refill pilot); and finance rewiring (Morgan Stanley's $3.4T-by-2040 SpaceX revenue pitch, S&P refusing to fast-track unprofitable megacaps, a JPMorgan/Citi tokenized-deposit network, Meta weighing a multi-billion raise for up to $145B of AI capex this year).

Mapping against Ray Data Co

Two load-bearing signals here, one peripheral cluster:

1. SWE-Marathon is direct evidence for the harness-thesis (strong). "Twenty multi-hour tasks, frontier models under 19%, losing the thread over hours" is the empirical statement of exactly the failure mode RDCO's whole operating discipline is built against — long-horizon agent reliability collapses not because the model is dumb but because it can't hold coherence across hours of context. This is the benchmark form of "context rot" and validates why RDCO routes long artifacts through subagents and manages parent context aggressively (CLAUDE.md hard rule 4). When evaluating the COO agent's own reliability ceiling, SWE-Marathon (not single-shot SWE-bench) is the benchmark to watch — and the under-19% number is the gap the harness exists to close. Pairs with [[2026-05-11-innermostloop-harness-eats-the-model]].

2. Cost-tiered model routing is now CFO-default (medium). Enterprises sending hard tasks to frontier models and easy ones to cheaper rivals, plus Gemma 4 sub-1GB on-device, confirm the multi-model mix-and-match posture RDCO already runs (cheaper models for cheap tasks, Opus for judgment). Reinforces the budget-controlled, route-by-difficulty discipline rather than defaulting every call to the premium tier.

3. State-on-the-cap-table is context, not action (weak). Government equity stakes, public wealth funds, and the SpaceX/compute financing tangle matter for the macro frame the founder tracks (and the investing capital-cycle thesis), but there's no RDCO operating lever here — file as situational awareness.

Net: a digest where the harness-thesis evidence point is the reason to keep it; most of the body is peripheral.

Related

• [[2026-06-05-innermost-loop-singularity-writes-its-own-sequel]]
• [[2026-06-04-innermost-loop-bots-pass-humans-on-the-web]]
• [[2026-06-03-innermost-loop-benchmark-not-license]]
• [[2026-05-11-innermostloop-harness-eats-the-model]]