06-reference

innermost loop singularity duopoly

2026-06-20·reference·source: Innermost Loop·by Alex Wissner-Gross
AI market structureAnthropicopen-source AIplatform dynamicsAI geopolitics

"Welcome to June 20, 2026" — @alexwg

Why this is in the vault

Alex Wissner-Gross's weekly curated digest maps the tightening AI duopoly around Anthropic and OpenAI while tracking the simultaneous rise of Chinese open-source models — directly relevant to RDCO's model-selection posture and platform-structure thinking.

The core argument

The issue opens with John Jumper (2024 Chemistry Nobel co-laureate for AlphaFold) leaving Google DeepMind for Anthropic, arriving at a moment when DeepMind insiders reportedly concede "the AGI race to Anthropic and OpenAI" as internal morale slides. Google's flagship model sits fifth on the intelligence index, behind Chinese competitors.

The insurgent story is Chinese open-source: Qwen, DeepSeek, Kimi, GLM, and MiniMax now command the majority of token use on OpenRouter's top ten, up from under 2% in late 2024. GLM-5.2 beat Claude Opus 4.8 on a deliberately AI-resistant take-home coding challenge — aided in part by Anthropic's Fable model rollback. A former DeepMind VP called GLM-5.2 "the first open model good enough to be a daily driver." The frontier labs' counter-move is going dark: Andrej Karpathy can no longer contribute to his own open-source repos after joining Anthropic.

Secondary clusters in the issue:

Mapping against Ray Data Co

Strong connection on several axes:

  1. Anthropic/Claude as primary platform bet: RDCO runs Claude Code and Claude-based agents. The Jumper hiring and the GLM-5.2 coding benchmark (which surfaced the Fable rollback as a concrete weakness) are directly actionable intelligence for model selection and hedging strategy.

  2. Duopoly thesis: The framing of Anthropic+OpenAI as the consolidating frontier, with Chinese open-source as insurgent, shapes how a phData Deal Solutions Architect should position AI-stack recommendations to clients — the "safe" commercial choices vs. cost-efficient open alternatives.

  3. AI market structure: The OpenRouter token-share data (Chinese models >50% of top-10 usage) is a leading indicator of commoditization pressure on proprietary APIs. Relevant to RDCO's harness-engineering investments — platform lock-in risk is real.

  4. Claude Code specific: The Linear A crack via Claude Code is a reference point for "agentic coding on novel problems" use cases — useful in client conversations about what autonomous coding agents can do.

Related