06-reference

innermost loop singularity good taste

2026-06-21·reference·source: Innermost Loop·by Alex Wissner-Gross
ai-frontiersingularityai-tasteagentic-discoveryroboticscompute-economicsfrontier-modelsai-policy

Why this is in the vault

Wissner-Gross's lede — "The Singularity is learning good taste" — marks a maturation milestone worth tracking: frontier AI is no longer bottlenecked only by raw capability but by aesthetic and qualitative judgment. OpenAI Codex engineers explicitly cited front-end visual taste as the top remaining capability gap. At the same time, the agentic layer is hardening into discoverable infrastructure: Google's open Agentic Resource Discovery spec lets AI tools and skills publish and cryptographically verify themselves for peer agents to find. Both signals are directly relevant to RDCO operating an always-on agentic harness.

The core argument

The Singularity is no longer just getting smarter — it is getting tasteful. The frontier is closing the last qualitative gap (aesthetic judgment on visual/front-end work) while simultaneously building the infrastructure layer that lets agents find and verify each other. Six arcs run in parallel:

AI taste gap closing — OpenAI Codex team says front-end visual capability is the loudest remaining ask; when it lands, it will be a step-change moment ("That day will be something"). GLM-5.2 surprises Vercel's CEO on coding; places second behind Claude on debate benchmarks. Open weights trail frontier by "inches, not miles."

Security at the frontier — Anthropic's "Mythos" reportedly cracked nearly every classified system it touched "in hours," per congressional testimony from a former NSA general. Anthropic amended its privacy policy to require ID-and-selfie verification, apparently in response to export controls on Fable 5 and Mythos 5.

Agentic resource discovery — Google released an open Agentic Resource Discovery spec for AI tools to publish and cryptographically verify themselves so agents can find each other. OpenAI simultaneously rebranded ChatGPT billboards to Codex — retiring the chatbot identity for the agent identity.

Robotics architectural shift — Nvidia's Jim Fan declared "Rest in Peace, VLAs" at Sequoia; the field is converging on physics-grounded world models over vision-language-action models. Spiced self-play hit 99.4% safe driving on 2,500× less human data. Sanctuary AI ran Physical AI at 99.5% accuracy plugging flexible wires on a live factory conveyor. GM adding cobots at Detroit truck plant.

Compute economics — Data centers nudged US power rates down 2015–2024 by spreading fixed grid costs over durable demand. Tesla trademarking "MEGAPOD" to turn Supercharger network into distributed AI compute; US now world's largest oil exporter at 10.5M bbl/day.

AI policy brawl — VP floated government stakes in big AI firms; Democrats launched Guardrails Alliance super PAC ($15M target vs. $100M industry war chest). Kansas City deploying facial-recognition cameras on public buses. Teachers can't distinguish AI-written vs. human essays as "humanizer" tools add fake typos in real time.

Closing line: "To err is human, but to humanize is now a download."

Mapping against Ray Data Co

Strong. Three arcs land directly:

Related