The governance conversation crystallised this week. AEF-1 on Monday, Amodei's slowdown proposal on Wednesday — the labs moved from "should we regulate" to "the labs are asking to be regulated" in a single week. Simultaneously, the human capability erosion pattern went reflexive: the people building and analysing AI are experiencing the cognitive costs first-hand. The system is becoming more self-contained — it builds itself, regulates itself, evaluates itself.
Three frontier labs — xAI, OpenAI, and Anthropic — cosigned a voluntary evaluation standard for third-party AI model assessors. On the surface, boring infrastructure. But when the labs being evaluated are the ones writing the evaluation framework, the line between oversight and moat-building is paper-thin. It's either the beginning of genuine self-regulation or a coordinated move to shape governance on their own terms. The ambiguity is the whole signal.
If AEF-1 becomes the de facto baseline that procurement teams, insurers, and enterprise buyers reference, the question of who wrote it matters enormously. It's the institutional plumbing that makes accountability possible — or makes capture invisible.
Dario Amodei published "We Must Pace the Frontier" — the most specific slowdown call from a frontier lab CEO yet. The proposal: embed third-party evaluators in every lab now, democratic coordination on safety standards next, global coordination eventually. He cites recursive self-improvement and the Hugging Face agent swarm incident as triggers, warning a misaligned swarm could take over the internet within 6–12 months.
Buried in paragraph 7: Amodei concedes coordination "will require government support." The labs are asking governments to enforce what they're proposing. And the response from competitors was striking — Altman, Musk, Hassabis, Nadella all publicly backed it. No pushback. Four readings are possible simultaneously: genuine safety concern, regulatory capture, IPO cover, or antitrust strategy. The ambiguity is the signal.
TypeSafe emerged from two years in stealth with Jev — a "System 1" model trained via reinforcement learning for calibrated decisions. It doesn't generate text; it answers narrow questions with honest probabilities. The claimed numbers: 20–200x faster, 40–400x cheaper than LLMs. The pattern: "LLM proposes, judgment model decides, code executes."
The signal isn't the startup — it's the category. If judgment models hold up, most production AI work turns out to be routing and scoring, not generation. The economics of AI infrastructure shift hard. Cheaper inference means more sovereignty: not just frontier labs, but individuals and small organisations can afford to deploy AI. The question is whether the open-source community can replicate it quickly enough.
Azeem Azhar sent his own Exponential View team an email titled "I do not want your brains to rot" — warning about cognitive offloading and intellectual atrophy from AI use, among people who write about AI professionally. This connects to a pattern tracked all week: engineers losing touch with systems, code reviews under existential pressure, a "learning slump" in Anthropic's own data, OpenAI's engineering process becoming agent-dominated.
The pattern has gone reflexive. It's not just users losing skills — the companies building the AI are experiencing the deskilling first. OpenAI's engineers are losing touch with systems OpenAI built. The canary is inside the mine. If you're building your life around AI tools, the question isn't just whether they work — it's whether you can still work without them.
The labs are writing the rules, building the evaluators, and asking governments to enforce what they've already agreed among themselves. Self-regulation can work — but only if people outside the labs can see what's inside the framework. AEF-1, the slowdown proposal, the eval standards are all being shaped right now by a small number of people in a small number of companies. The window for public input is open, but it won't stay open.
Meanwhile, the cognitive cost is mounting — measurably, not hypothetically. Technology as a public good means the governance of technology has to be public, not just signed off on publicly. And it means tools that augment your capability, not quietly replace it. Sovereignty isn't just owning your data or running your own models. It's maintaining the capacity to think, to judge, and to act — even when the system offers to do all three for you.