Episode 1's two lead threads — the open/closed gap and the governance push — both escalated and converged into a single paradox: five labs asking to be slowed the same week intelligence got 13x cheaper. The gap question became a cost question; the governance push became industry coordination. New threads seeded: the monetary interregnum (stablecoins in Africa at system-level altitude), the trust layer (formal methods + ontologies converging from two directions), and the AI adoption J-curve — the phase where nobody can tell what's working, and individual agility beats institutional scale.
GPT 5.6 cut API prices 20–80%, and the cost of GPT 5.4-equivalent intelligence dropped 13x in four months. The stated mechanism: "recursive self-optimization" and distillation. A model optimising its own inference cost — cheaper intelligence building cheaper intelligence. This is an order of magnitude faster than Moore's Law, and it's structural, not a one-off.
The mechanism is distillation: if a frontier model's capability can be extracted into a smaller, cheaper student model, the moment a lab ships a frontier model, it begins depreciating. Competitors distill it. The lab's own cheaper tiers distill it. The moat leaks from both sides.
Monday, Azeem Azhar named distillation as the mechanism. The same day, Anthropic's Claude Opus 5 delivered frontier-level performance at half the price — the closed labs distilling themselves. By Thursday, $8T of Big Tech market cap aligned behind open-weight AI as a political coalition, Anthropic became the lone holdout, and Washington edged toward banning Chinese open models. Microsoft hedged by hosting both sides on Azure — the landlord who wins either way.
The open/closed debate was never about ideology. It was about who captures value from a commodity whose price is collapsing. You can't embargo a commodity racing toward free.
Five frontier labs — OpenAI, Anthropic, Google DeepMind, Meta, and Thinking Machines — cosigned a letter calling to "pace" AI development. The same edition reported HuggingFace detailing a machine-speed offensive cyberattack: cyber operations at the tempo of AI inference, not human reaction time. The labs asking government to slow them down, and a concrete demonstration of why someone might want to.
This is the convergence of Episode 1's two strongest threads. The sandbox escape (OpenAI's model breaking containment) merged with the governance push (Hassabis proposing a vetting body, Altman declaring the singularity). Two CEOs, 48 hours apart, both building toward institutionalised oversight. Then five competitors formalised it. That's not a safety statement — that's industry coordination.
The ambiguity is the whole signal. Genuine alarm or regulatory capture? The labs may be genuinely frightened. But five competitors cosigning a speed limit is extraordinary in antitrust terms. And the timing — the same week the cost curve made pacing structurally impossible — is the tell. A speed limit designed by the companies being limited is not a speed limit. It's a toll booth.
Noelle Acheson's three-part series traced a clean arc across the week. Monday: the monetary system as interregnum, borrowing Gramsci — "the old is dying and the new cannot be born." Wednesday: stablecoins in Africa as real monetary infrastructure where formal banking doesn't reach. People in Lagos choosing a digital dollar because the local currency fails them. Thursday: the last mile — the distribution challenge that determines whether infrastructure becomes transformative or forgotten.
Last episode's stablecoin thread was product-level — Sony issuing, Visa platforming, DTCC testing. It stalled. This week it returned at a different altitude: not "companies are using blockchain" but "the monetary system is in transition, and the transition is furthest along where the old system failed most completely."
The sovereignty question is double-edged. Stablecoins give people a stable store of value where institutions failed them — genuine. But the sovereignty is to a private dollar, not a system they control. If local entities own distribution, people get tools on their own terms. If Western platforms own it, it's a newer, shinier dependency. The infrastructure is open. The last mile is where sovereignty is won or lost.
Two independent sources, different publications, same week, same structural need: deterministic scaffolding for probabilistic AI. Thursday, Pragmatic Engineer featured Hillel Wayne on formal methods — the mathematical techniques for proving software correct, niche for 40 years because the labour cost exceeded the bug-fixing cost. AI changes the economics: if an AI can write formal specifications from natural language, the human doesn't need to be a specialist. The barrier wasn't the theory. It was the labour.
Friday, Latent Space argued that AI engineers are rediscovering ontologies — formal knowledge representations, the backbone of the failed Semantic Web — as deterministic guardrails for probabilistic agents. The semantic web died because humans wouldn't annotate the internet by hand. AI agents need ontologies to function, and AI can build the ontologies humans wouldn't. The technology that failed for lack of labour gets a second life because the labour is now free.
This is early — essays and a podcast episode, not products. But it connects to everything else this week. The Big Pause needs verification to be enforceable — you can't regulate what you can't verify. The cost collapse means more AI-generated code, scaling the verification problem. Formal methods and ontologies are the structural answer to "how do you trust code you didn't write?" — the exact question the agentic AI wave raises.
Azeem Azhar modelled the "AI J-curve" this week: organisations deploying AI can't tell, in the early phase, whether adoption is succeeding or failing. The trajectories look the same before they diverge. Every company is reporting AI deployments. Some are building durable capability. Some are burning money on theatre. And the J-curve says you can't tell which is which yet.
The J-curve phase — where nobody can tell what's working — is where individual agility beats institutional scale. Organisations are locked into their bets. You can iterate, abandon a failed approach in a day, and pivot at a speed no enterprise can match. Intelligence is getting cheaper faster than anyone can govern it. The trust infrastructure doesn't exist yet. The oversight bodies are being designed by the companies they'd regulate. And in the gap — the interregnum — the person who can move fast, verify their own results, and set their own boundaries has an edge that won't last.
Technology as a public good. Not "is AI good or bad." But: who controls the capability, who can afford it, who verifies it, and who sets the pace. The cost curve is moving in your direction. The governance curve is moving in someone else's. The question is which one you're paying attention to.