06-reference

innermost loop image model mirror blink

Wed Apr 22 2026 20:00:00 GMT-0400 (Eastern Daylight Time) ·reference ·source: The Innermost Loop ·by Alex Wissner-Gross

“Welcome to April 23, 2026” — @theinnermostloop

Why this is in the vault

Weekly Wissner-Gross compressed scan of the AI/tech buildout. This issue’s hook: image models are now generating photorealistic screenshots of themselves, METR Time-Horizon forecasts have Mythos at a full 40-hour human work week (Opus 4.7 at 19), and the coding-IDE wars went from competitive to feudal in a single news cycle. Useful as the top-of-stack situational frame for the rest of the week’s deep reads.

The core argument

“When the image model can screenshot itself, the mirror has finally learned to blink.” The throughline is recursion at every layer: image models auditing their own outputs, AI writing 75% of Google’s new code, agents colonizing the OS, and capex flows that look more like industrial policy than venture investing.

Stack-by-stack snapshot:

Mapping against Ray Data Co

Strong. Three load-bearing items for RDCO this week:

  1. METR Time Horizon: Mythos at 40h, Opus 4.7 at 19h — this is the operative number for the COO-as-Claude positioning. RDCO’s whole pitch is that an Opus-level agent can hold a multi-day workstream coherently. Confirmation that the public benchmark community now prices Opus 4.7 at roughly half a human work week (and Mythos at a full week) is direct corroboration of the durable-context thesis. Worth pulling into the ../03-positioning/coo-as-claude-positioning doc as a citation.
  2. GitHub Copilot pauses signups, token billing, weekly cost doubled since January — the “agent economics get harder before they get easier” signal RDCO clients need to hear. If Copilot — the cleanest consumer agent product line — is rationing, then RDCO’s “we own the agent and the budget” framing becomes more valuable, not less. Pricing/packaging implication for the Sanity Check pipeline.
  3. OpenAI Chronicle (background agents from Codex screen captures) + workspace agents in ChatGPT — competitive watch on the “always-on agent that builds memory from your work” surface. That’s exactly the COO-as-Claude moat. Chronicle is closer to it than anything OpenAI has shipped before. Adds urgency to documenting RDCO’s distinct angle (Claude-resident, Mac-mini-hosted, founder’s vault rather than screen captures).

Secondary:

Curation section — notes

Pure-curation issue: every claim is a substack-redirect-wrapped link to an external source (OpenAI announcements, Steve Yegge post, vendor PR, IEA report, etc.). No links back to other Innermost Loop posts — no self-cross-promo detected. No sponsor block. Wissner-Gross’s value is the connective tissue (“substrate is densifying” / “agents are colonizing”) not the link selection. Did not deep-fetch any links — every item passes a one-line headline test from the body summary; the deeper read for any specific item (e.g. Yegge’s post on the Antigravity strike team) belongs in a dedicated note if it crosses a threshold on its own.

Quotes ≤15 words, paraphrase otherwise. Source: Innermost Loop, Apr 23 2026 — view at https://theinnermostloop.substack.com/p/welcome-to-april-23-2026