06-reference

innermost loop singularity as market

2026-06-19·reference·source: The Innermost Loop·by Alex Wissner-Gross
ai-futuressingularityfrontier-modelsai-policyhardwareenergyrobotics

"Welcome to June 19, 2026" — @alexwg

Why this is in the vault

A weekly pulse-check across AI capability, policy, hardware, and energy infrastructure — the exact stack RDCO's thesis rides on.

The core argument

The hook — "The Singularity is a market, not a monarch" — sets the tone: frontier AI evolution is driven by competitive dynamics, not a single tipping point. The edition covers:

Model frontier: Artificial Analysis ranks Fable 5 first, Opus 4.8 second, and open-weight GLM-5.2 third on multi-week knowledge work. GPT-5.6 rumored imminently. Unsloth compressed the 744B GLM model from 1.51TB to 217GB for local deployment. MIT's Self-CTRL work trains models toward faithful self-description; OpenAI finds rewarding honesty and humility produces alignment that holds under adversarial fine-tuning.

Policy/geopolitics: Anthropic and the White House reportedly shifted from standoff to co-authoring AI security grading rules. Project Glasswing enterprise testers retained access to Mythos Preview even after a US export order pulled public Fable 5. SK Telecom named as the firm whose Mythos access was revoked over alleged China ties. Microsoft quietly resells mostly-OpenAI models to Chinese tech giants via Azure — ByteDance alone exceeds $1B/year.

Product layer: OpenAI added Record & Replay to Codex (demo a task, it becomes a repeatable skill). AWS shipped Continuum (machine-speed code repair) and Context (knowledge graph for enterprise data).

Hardware/memory crunch: CCS Insight forecasts ~15% smartphone shipment drop as memory flows to server chips. ASML/EUV export concerns resurface. Amazon in talks to sell Trainium chips into rival data centers. Intel jumped 10% after Apple announced plans to help build chips domestically.

Energy/infrastructure: Meta signed 1.6 GW from Crusoe across Texas and Missouri. First Street flagged 79% of data-center capacity sits in climate-exposed markets. Rolls-Royce SMR won three-reactor deal in Sweden; Switzerland moving to lift nuclear ban.

Robotics/drones: Anthropic's Frontier Red Team clocked Claude Opus 4.7 on robodog tasks running ~20x faster than last year's best humans. Drone regulation tightening in Beijing; Orlando launched first Drone as First Responder program.

Medicine: OpenAI o3 Deep Research surfaced leads yielding 18 new diagnoses across 376 unsolved rare-disease cases. Midjourney pivoted into healthcare with a 60-second whole-body imaging scanner.

Labor/capital/citizenship: Bezos predicted AI causes labor shortages, not redundancy — against a backdrop of 97,000 May layoffs, 40% AI-linked. Sanders proposed a $7T public AI fund. Argentina's Milei argued for legal personhood for AI firms. Pew: Americans using more chatbots while their views of AI keep souring.

Mapping against Ray Data Co

Investing thesis (chip-fab/memory capital cycle): The memory demand signal is explicit — smartphone shipments declining ~15% as memory flows to server chips, Amazon selling Trainium into rival data centers, Intel/Apple domestic chip news. This confirms Phase 2 of the Markov capital-cycle thesis (supply racing to meet AI compute demand). The ASML/EUV export concern is a tail risk to monitor for the fab supply chain.

Model landscape / phData DSA role: Fable 5 ranking first on multi-week knowledge work, Opus 4.8 second — directly relevant for the Snowflake GenAI Specialty and Claude Certified Architect cert paths. The Anthropic-White House co-authoring of security grading rules is context for the enterprise sales conversations Ray will be in at phData.

Agent positioning: OpenAI Codex's Record & Replay (demo → repeatable skill) maps directly onto the agent-harness loop RDCO runs. AWS Continuum at machine-speed code repair is a competitive signal for the agentic dev-tools space.

Energy infrastructure bet: The Meta/Crusoe 1.6 GW deal and the nuclear SMR developments reinforce that power availability is an upstream constraint on AI buildout — relevant for any infrastructure-adjacent position in the portfolio.

Policy risk surface: The Microsoft/Azure resale to Chinese firms and the Mythos export revocation confirm that enterprise AI distribution is being used as a geopolitical lever. phData client engagements that touch export-sensitive industries need this framing.

Related