"Welcome to June 3, 2026" — Alex Wissner-Gross

Why this is in the vault

Daily frontier-AI digest written as woven prose rather than bullets, organized around a single thesis: US policy has decided to measure the Singularity instead of permitting it. The lead frame — "governments would rather benchmark it than license it" — is a clean read on the regulatory posture RDCO is building its COO agent under, and the issue's silicon-and-debt thread (Broadcom backstopping a $36B private-credit TPU deal for Anthropic, CoreWeave junk notes, Majorana 2) is live datapoint material for the chip-fab/memory capital-cycle investing thesis. Secondary threads on labor displacement, Claude Code per-seat cost caps, and Windows-level agent sandboxes are adjacent to RDCO's own harness-engineering practice.

The core argument

A new White House executive order ("Promoting Advanced AI Innovation and Security") directs agencies to build a classified test of AI cyber capabilities and invites labs to voluntarily share frontier models for up to 30 days pre-release — while explicitly forbidding any mandatory licensing regime. Politico read it as the industry dodging heavier federal oversight again. Wissner-Gross's frame: regulation has shifted from gatekeeping (license before release) to instrumentation (benchmark after release), and freed from preclearance, the labs are racing to feed the machines.

Supporting movements woven through the issue:

• Knowledge work catching up to math. Sixteen mathematicians (backed by the IMU) published the Leiden Declaration asking the field to disclose AI use and keep humans accountable, weeks after a model disproved an 80-year-old Erdős conjecture. Codex ships "Sites" (a Lovable competitor deploying apps from a prompt); a Stanford blind study found law professors preferred AI legal answers ~75% of 3,000 comparisons, flagged harmful a third as often as human ones. ChatGPT hit 1B MAU fastest ever; Claude's smaller base compounding ~640%/yr; Codex past 5M weekly users with six role-specific plugins. Microsoft answered with Scout (always-on Outlook/Teams assistant), Project Solara, and Execution Containers (Windows-level sandbox already adopted by OpenAI, Nvidia, Manus, Nous). Anthropic expanded Project Glasswing / Claude Mythos Preview to ~150 critical-infra orgs.
• Borrowed silicon, borrowed money (load-bearing). Broadcom pledged to backstop a record $36B private-credit deal structured to buy Google TPUs and lease them to Anthropic — senior tranche compressed to ~5.75%, unbacked slice 8–9%. A CoreWeave-linked data center raised $900M in junk notes at 7.5%, part of $27B+ borrowed this year. Microsoft unveiled Majorana 2, a topological quantum chip designed with its own agentic AI, claiming 1000x qubit reliability and pulling its scalable-quantum target to 2029.
• Bodies and orbits. Barclays sees humanoid robots a $200B market within a decade; SpaceX won FAA approval for Starfall orbital-manufacturing reentry capsules (commentators flagged the obvious "rods from God" dual use).
• Who actually did the work. NY Fed researchers found remote work, not AI explains ~two-thirds of rising young-graduate unemployment (employers stopped hiring un-mentorable juniors). Where AI is the cause, bills arrive fast: Uber capped engineers at $1,500/mo per coding tool like Claude Code after burning a year's budget in four months; Thrive Holdings is betting $1B buying accounting firms to automate them. Meta rolled back a keystroke/screen-logging surveillance tool after worker pushback. A NYT audit of 600+ Musk claims found he hit 75% of his 2015 goals on time.

Closing line reframes Gibson: "The future is already here, the last 25% just isn't evenly distributed yet."

Mapping against Ray Data Co

Strong. Two threads land directly on RDCO's active work.

(1) Governance posture. The benchmark-not-license framing is the single most useful policy datapoint for an agent-deployer building toward L5. A voluntary 30-day pre-release sharing window with no mandatory licensing means the operating environment for an autonomous COO agent stays permissive — the constraint on RDCO's harness ambition is capability and trust, not a federal gate. It also validates the "control/harness is the real frontier" read: if the state has given up on licensing model capability and is instead instrumenting model behavior, then the durable governance surface moves downstream to deployment harnesses (exactly where RDCO's own implementation-notes/verification discipline lives), echoing [[2026-05-11-innermostloop-harness-eats-the-model]].

(2) Capital cycle. The silicon-and-debt movement is on-thesis corroboration for the chip-fab/memory capital cycle the founder places RDCO in Phase 2 of. The Broadcom-backstopped $36B TPU-lease structure for Anthropic, $27B+ in data-center junk issuance this year, and Majorana 2's 2029 pull-forward are the demand-and-financing signals of a maturing capex up-cycle — and the credit-spread detail (5.75% senior vs 8–9% unbacked) is the kind of risk-pricing texture worth watching for a cycle-top tell. Caveat per usual: this is a digest, not primary data — every number (the $36B, the spreads, the 640% Claude growth, the Stanford 75%) needs source verification before it feeds a thesis doc. Treat as lead-generation, not evidence.

The Uber/Claude-Code per-seat cost cap is a small but pointed operational datapoint: agentic coding spend is real and lumpy enough that a sophisticated buyer hit a $1,500/seat/mo ceiling after blowing a year's budget in four months — relevant to how RDCO reasons about its own API-cost budgeting posture.

Related

• [[2026-06-02-innermost-loop-june-2-singularity-leaderboard]] — prior day's issue; Singularity-as-leaderboard frame, same author cadence
• [[2026-06-01-innermost-loop-memory-worth-more-than-oil]] — capital-cycle + Anthropic-financing thread this issue extends
• [[2026-05-11-innermostloop-harness-eats-the-model]] — the "control/harness is the real frontier" thesis the benchmark-not-license framing reinforces
• [[ai-governance-posture-for-agent-deployers]] — concept stub (does not exist yet): how US benchmark-not-license policy shapes RDCO's L5 build envelope