↓Skip to main content
Zeitgeist — a spike by Chris Gathercole
  1. Signals/
  2. Symptom Catalogue/

Update — 11 June 2026

Symptoms #

  • Microsoft blocks its own employees from Fable 5 via GitHub Copilot model picker due to 30-day data retention, while simultaneously offering Fable 5 to all external GitHub Copilot and Foundry customers — a vendor that won’t use its own product internally. [claude-integrations + claude-teams]
  • Fable 5 breaks Zero Data Retention (ZDR) — the first Claude model to require Anthropic data retention (30 days for safety classifiers; up to 2 years for flagged prompts); all previous Claude models support ZDR, which is the baseline enterprise compliance configuration. [claude-expertise + claude-teams]
  • MiniMax M3 displaces Kimi K2.6 as open-weight SWE-Bench Pro leader (59.0% vs 58.6%) within 7 days of Kimi’s crossing being reported — the benchmark leadership shelf life is now sub-week. Open vs Closed AI Ecosystems
  • Andrej Karpathy — who coined “vibe coding” in February 2025 — publicly declares in June 2026 that “this era is ending” and positions himself as a proponent of “agentic engineering” instead; 16 months from coinage to public self-distancing. Vibe Coding Approaches
  • 81% of enterprise technology leaders report production failures from AI-generated code; self-assessed AI readiness score is 83.6/100 — the gap between felt readiness and actual production outcomes is statistically large and directionally inverted. [claude-teams + vibe-coding-applications]
  • Anthropic’s own production codebase: 80% of code authored by Claude; Claude Code team: 90% of Claude Code itself is AI-written, engineers run 5 PRs/day, PR output +67% while team doubled — the internal adoption benchmark that every enterprise will now be measured against. Team & Org Use of Claude
  • Claude Cowork (research preview) is explicitly local-only: “Cowork Projects are local to your computer. Your colleague can’t access your projects.” — an AI product positioned as a “coworker” that colleagues cannot actually share. Team & Org Use of Claude
  • 15 state attorneys general coalition forms against GAAIA preemption clause, calling it “a gift to the AI industry” — the first organised multi-state opposition to a federal AI bill, and the first time state AGs have explicitly argued a federal AI floor would lower existing state protections. AI Impact on Society

Synthesis: What connects these? #

The second wave of today’s observations has a distinct character from the morning’s capability-and-governance cluster: these symptoms expose the gap between institutional claim and institutional reality. Microsoft claims to offer Fable 5 while blocking it internally. Enterprises self-assess 83.6/100 readiness while producing AI-code failures 81% of the time. Karpathy names something (“vibe coding”), builds a movement, then steps back. Claude Cowork is positioned as a shared coworker but physically cannot share. The readiness-reality inversion is not random — it’s systematic to the moment when adoption has outrun both institutional vocabulary and institutional infrastructure.

  • Team & Org Use of Claude The Anthropic 80%/90% production code figures and the local-only Cowork limitation are the central claude-teams symptoms.
  • Claude-Specific Expertise The ZDR break is the practitioner-critical Fable 5 fact not covered in this morning’s extraction.

Meta-observations #

  • Emerging pattern: Second-order enterprise friction is now visible — not “can we access AI tools” but “can we govern what we’ve already deployed, and does the tool do what it claims at team scale.”