"Welcome to June 2, 2026" — Alex Wissner-Gross

Why this is in the vault

Daily frontier-AI digest, written as woven prose rather than bullets. The lead frame — the Singularity has become a leaderboard rather than a finish line, with the gap between frontier models now measured in multiples per quarter — is directly on-theme for the agent-capability race RDCO tracks. The silicon/capital thread (Nvidia recutting one architecture across rack/desk/laptop/body, Alphabet's $80B raise, SoftBank passing Toyota, the memory-and-capex rotation) is live datapoint material for the chip-fab/memory capital-cycle investing thesis. Secondary threads on institutional memory loss under AI and agent-to-agent discovery infra are adjacent to RDCO's own harness-engineering practice.

Issue contents

Curation digest, ~25 linked items strung into a single thematic narrative across six movements. Self-promo: standard Substack "Subscribe for free" CTAs woven in twice — not sponsored content. All deep links are Substack redirect wrappers pointing to third-party news sources; no affiliate or self-product links beyond the subscribe CTA. No sponsor block.

• Frontier leaderboard (lead): Opus 4.8 posted a state-of-the-art ~1.5% of human efficiency on ARC-AGI-3, reportedly tripling GPT-5.5; framing is that frontier gaps now compound in multiples per quarter. Open-weight pack chasing: Nvidia Nemotron 3 Ultra (~550B params, smartest open US model at 48 AAII) still trailing China's Kimi K2.6 (54); MiniMax M3 claims sole open-weight fusion of frontier coding + 1M-token context + native multimodality at $0.12/M input tokens vs Opus 4.7's $5.
• Control as the new frontier: Anthropic giving the EU's ENISA access to Mythos (first European agency in Project Glasswing). Linux Foundation's DNS-AID teaches agents to discover each other over plain DNS. Cautionary tale: hackers seized Instagram accounts (Obama White House, a Space Force chief, Sephora) by socially engineering Meta's support bot into swapping the email — "the first social engineering attack where the social was optional."
• Institutional rewrite: Microsoft reportedly fusing GitHub Copilot, chat, Cowork, and a new "Autopilot" agent into one Copilot super app. Counterweight: a historian-turned-engineer warns the oral tradition of software may not survive AI as mentor-to-junior memory transfer dissolves into model weights. American Federation of Teachers (2nd-largest US teachers' union) released a 10-point plan capping classroom AI and banning screens through 2nd grade.
• Silicon recut for every altitude (load-bearing): Nvidia named Anthropic, OpenAI, and SpaceX as launch customers for its Vera CPU (first standalone DC microprocessor); shrank it to the DGX Station deskside supercomputer (748GB, 20 petaFLOPS) and further to the RTX Spark Arm laptop chip (TSMC 3 + MediaTek), which Jensen Huang calls the first reinvented PC line in 40 years, shipping this fall (Dell, Lenovo). HPE revenue jumped 40% to $10.7B past a $9.74B forecast on a 33% server surge. Open Isaac GR00T reference humanoid (Unitree chassis, Sharpa hands, Jetson Thor) makes one architecture span rack, desk, laptop, and body.
• Capital and copper: Alphabet raising $80B equity (incl. a $10B Berkshire Hathaway vote of confidence) for its buildout, even as Ohio suspends the tax break that drew data centers. AUKUS (US/UK/Australia) developing undersea drones to guard cables and pipelines. Africa leapfrogging to clean: 173 of 322 new 2025 energy projects solar, renewables overtaking fossil plants.
• New physics of value: SoftBank passed Toyota to become Japan's most valuable company (ending a 20-year reign) as Tokyo stocks hit an all-time high. SpaceX IPO (expected to more than double any in history) is forcing index providers to rewrite admission rules; Honeywell-backed Quantinuum lifted its quantum IPO to a $14.3B valuation. 220+ pre-ChatGPT "fallen unicorns" now stranded — too inflated for venture, too unprofitable for Wall Street. Bernie Sanders proposed an AI Sovereign Wealth Fund (one-time 50% tax paid in frontier-lab stock). Even Michael Saylor's Strategy net-sold bitcoin for the first time — "diamond hands, it turns out, are not forever."

Mapping against Ray Data Co

Strong. The issue's organizing thesis — "the Singularity has stopped being a finish line and become a leaderboard" — is a clean restatement of the AGI-progress / agent-capability-race frame the founder tracks, and it sharpens it: the relevant unit is no longer "did we reach AGI" but "what's the per-quarter multiple between frontier and the open-weight chasers." That directly informs RDCO's posture on which model tier to build the COO harness against, and on the commoditization-of-raw-capability / control-is-the-real-frontier read that underpins the harness-engineering thesis (memory and control as the durable moats, per [[2026-06-01-innermost-loop-memory-worth-more-than-oil]] and [[2026-05-11-innermostloop-harness-eats-the-model]]).

Two threads are direct investing-thesis inputs. (1) The silicon movement — Vera CPU launch customers, DGX Station, RTX Spark on TSMC 3, HPE's 40% revenue beat — is on-thesis corroboration for the chip-fab/memory capital cycle the founder places RDCO in Phase 2 of; one architecture spanning rack-to-body is exactly the demand-broadening that extends a capex up-cycle. (2) The capital rotation (Alphabet's $80B raise, SoftBank over Toyota, the fallen-unicorn cull) is the market-side signal of that same cycle maturing. Caveat: this is a digest, not primary data — every datapoint here (ARC-AGI-3 %, AAII scores, HPE numbers) needs source verification before it feeds a thesis doc; treat as lead-generation, not evidence.

The institutional-memory and DNS-AID items are softer but real: the "oral tradition of software may not survive AI" warning is a direct echo of why RDCO keeps an implementation-notes discipline and a durable vault, and agent-to-agent discovery infra is adjacent to where a multi-agent COO stack eventually goes.

Related

• [[2026-06-01-innermost-loop-memory-worth-more-than-oil]] — prior day's issue; memory-as-the-new-oil + Anthropic IPO, same capital-cycle thread
• [[2026-05-11-innermostloop-harness-eats-the-model]] — the "control/harness is the real frontier as raw capability commoditizes" thesis this issue restates
• [[2026-05-12-innermostloop-singularity-matures-state-of-ai]] — companion frontier state-of-AI framing from the same author
• [[2026-05-27-innermost-loop-may-26-frontier-digest]] — comparable digest-format issue for structure/cadence reference