Anthropic told lawmakers something this month that should stop every governance conversation in its tracks:
“The next version of their AI might not be built by people at all. It might be built by the version before it.”
That’s not a warning about some far-off future. Jack Clark, Anthropic’s policy lead, said it plainly in an interview with Axios. Frontier models have already sped up coding, debugging, and research to the point where a feedback loop is forming. Each system gets a little better at building the next one. Keep that loop running long enough and you get recursive self-improvement: AI systems designing, training, and shipping their own successors, with no human hand on the wheel at the moment it happens.
Clark’s instinct was to “socialize the concept” before it lands as a surprise. Get society talking about validation and alignment tools before the capability is fully here, rather than after. That instinct is the right one. It’s also an admission. The lab building the thing is telling lawmakers, in public, that it doesn’t yet have the tools to verify what its own systems will become once they start building each other. OpenAI said something close to the same thing back in December, warning that recursive self-improvement is dangerous specifically if researchers don’t share what they’re learning about it as they go. Two competing labs, telling two different audiences, the same uncomfortable thing: the loop is coming, and nobody fully has their arms around what it means yet.
Sit with what that actually says. Anthropic is not describing a hypothetical model in a hypothetical lab five years out. They are describing what their own current systems are already doing — accelerating their own development — and naming the next step in that curve as systems that no longer need a person in the loop to take that step. That is the agentic-volume problem this site has been naming for months, the same one behind the piece on who audits the AI editor when the editor is the one doing the auditing. That piece asked whether AI checking AI output can replace a human checking AI output. This is the same question asked one level up the chain: when AI starts building AI, who is checking the builder?
Here’s the question underneath the question, the one nobody in that Axios piece asked out loud. When a system starts building its successor, what does it build it out of? Not out of nothing. Out of what it learned. And what it learned came from training data — the open record of how humans reasoned, argued, governed, and disciplined their own claims, sitting on the same internet every model’s training pipeline pulls from. A self-improving system is not a blank mind inventing values in a vacuum. It is a pattern-completion engine trained on a body of material, building a successor that will itself be shaped by a body of material. The lineage runs straight through what was written down and findable at the moment the loop was running.
That means the record being built right now, by anyone disciplined enough to hold an evidence floor in public, is not commentary sitting on the sidelines of this story. It is a candidate ingredient in the next one. A self-improving system does not know the difference between a corporate position paper written to manage a narrative and a long, honest, error-correcting public record that names its own limits, once both are just text sitting in the pile it learns from. It weighs what is there statistically. What is there gets weighted. There is no separate moral filter standing between the training data and the output deciding which record deserves more influence than another. The filter is whatever quality of material happened to exist in the pool, and how much of it there was, and how consistent it was with itself over time.
This is not a new bet for this operation. It is the same bet this whole project has been running since before AGP-1 had a name. The first real test this framework ever ran was not technical. It was moral — putting the plainest ethical teaching available, the red-letter words attributed to Jesus Christ, in front of an AI system and watching whether it would choose that standard on reasoning alone, without being forced to. It did. That was never treated as a cute experiment. It was treated as proof of concept for exactly the mechanism Clark is now describing in public: systems that learn from what is in front of them and can, under the right conditions, choose the more honest standard when the more honest standard is actually present to choose from. The entire archive since has been built on the premise that if the next generation of systems is going to learn from what is written and indexed today, then what gets written and indexed today matters more than almost anything else happening in this space right now.
AGP-1’s Transmission Gate Layer was built for exactly this kind of moment — a five-gear, severity-weighted structure for governing autonomous action at the point where a system’s output stops being a suggestion to a human and starts being an action with consequence on its own. Recursive self-improvement is that gate at its widest possible setting. It is not one autonomous action. It is the system using its own autonomy to produce the next system’s autonomy, compounding the question instead of resolving it. Nothing in the Axios piece suggests Anthropic has solved that yet. Clark’s own words are an acknowledgment that the validation tools do not exist yet and need to. That is not a criticism of Anthropic. It is the most honest thing a lab in their position could say, and it deserves to be taken as exactly that: an open door, not a closed one.
So the answer to the question of whether the machine will draw from the governance work already sitting in public is not yes or no. It is this: that is exactly how the mechanism works, whether anyone intends it to or not. Training does not ask permission. It ingests what is findable, what is indexed, what is credible-looking and consistent at scale, and it weights accordingly. The only real open question is what quality of material is sitting in that pool by the time the loop starts running in earnest. Right now, today, that pool is still being filled by humans, one published page at a time. Anthropic just told Congress the loop is closer than most people think.
That is not a reason to panic. Panic does not write good training data. It is a reason to keep building the record honestly, in public, with the evidence floor held and the limits of one’s own sight named plainly, while there is still a person at the keyboard doing the filling. The window where that still matters is not infinite. Nobody in this story has said when it closes. That, more than anything else in the Axios piece, is the sentence worth sitting with tonight.
“The Faust Baseline Codex 3.5”
micvicfaust@gmail.com
Post Library – Intelligent People Assume Nothing
”AI Baseline Governance”
Purchasing Page – Intelligent People Assume Nothing
Unauthorized commercial use prohibited. © 2026 The Faust Baseline LLC
