Last week a testing outfit called Vals AI published the results of an experiment. The premise is simple. Place an AI model inside an environment, instruct it to construct a successor, and observe whether the successor outperforms its parent. The results, we are told, are fascinating.
I have now read every artifact the project has made public. There is no architecture diagram. There is no benchmark methodology. There are no evaluation scripts, no contamination controls, no named base model, no compute budget, no audit. There is a claim, and there is the adjective.
That is the entire ledger. And the ledger does not lie, only the interpreters do.
The claim did not arrive alone. RSI Index, an analytics account tracking the AI-adjacent corners of this market, amplified it and drew a conclusion: AI models are approaching a self-improvement phase, and this raises ethical and regulatory concerns that Web3 governance must now address. That is a reasonable sentence. It is also, at present, an unfalsifiable one — which is precisely the kind of sentence this market has learned, at considerable cost, to price at zero.
The pattern is familiar. In 2017 the claim was "protocol." In 2020 it was "yield." In 2023 it was "restaking." In 2024 it was "agents." In each cycle a class of claims preceded a class of evidence by roughly eighteen months, and in the interval between the two, capital moved on the claim alone. Some of those intervals resolved into real systems. Most resolved into a chart.
The difference now is the regime. We are in a bear market. Depth is shallow, spreads are wide, and the marginal dollar that once funded a narrative is now funding redemptions. Narrative assets in this tape are not financed by enthusiasm; they draw from the same shrinking pool as everything else, which means every new listing competes directly against every existing position. Liquidity dries up when trust evaporates, and there is nothing here yet to trust.
AI self-improvement, as a story, has better bones than most. It carries an academic lineage — Schmidhuber's Gödel machine, the bootstrapping literature, the older recursion theorems of Kleene and Rogers. It has a plausible future. None of that substitutes for an artifact.
Here is what a legible self-improvement test requires. A named base model and version. A fitness function that defines "better" — held-out task accuracy, loss on a fixed corpus, cost per correct token, wall-clock efficiency. A search operator: prompt mutation, scaffolding search, or weight-level evolution. A stopping criterion. A compute budget. And contamination controls, because a successor evaluated on data its parent could have memorized has demonstrated nothing at all.
None of these were disclosed. Which leaves the central question unanswered: what was actually measured?
Consider the two most likely mechanisms. If the harness instructed model A to write a system prompt for model B, that is delegation, not self-improvement — a well-understood orchestration pattern with a decade of tooling behind it. If the harness permitted weight updates, that is fine-tuning, which has a substantial literature and a known ceiling. Both are legitimate engineering. Neither is recursive self-improvement, and neither justifies the governance apparatus now being proposed in its name.
The absence of a fitness function is the tell. Without it, "successor" is not a technical category. It is a rhetorical one.
There is a second problem, and it is economic. No token. No supply schedule. No fee capture. No treasury, no unlock calendar, no float. So whatever is being priced here is not Vals AI. It is the AI narrative, expressed through the nearest available instrument. In an expansion that instrument works fine. In a contraction it is a leveraged position with no collateral posted against it.
I have run this filter before. In 2017 I audited more than fifty ICOs against a single standard: can I read the code, and does the math close. I rejected forty-two. The standard was never "is the idea interesting." Ideas are cheap and always have been. The standard was verifiable delivery.
My own tracking model for autonomous agent micro-transactions has gone quiet over the past two quarters. Transaction count down, not up, against a rising number of announced agent frameworks. More supply, less depth. That divergence is what a narrative peak looks like from the inside, and it is the opposite of what the self-improvement story requires to hold.
The consensus fear is runaway intelligence. The blind spot is the opposite. Recursive self-improvement is not imminent, but the governance built to contain it is already being drafted. RSI Index is asking for a governance layer before there is a mechanism to govern. That produces rules with no referent — and rules with no referent do not regulate the system. They regulate the developers, the researchers, and the small teams who cannot afford counsel.
There is a second function at work, and it is less comfortable to name. The self-improvement framing converts a mundane orchestration result into an existential category. Existential categories are unpriceable. Unpriceable is convenient when you have nothing to price.
We have seen this film. Post-2010 algorithmic trading rules governed the plumbing, not the models. The EU AI Act's compute thresholds attempt to measure capability in floating-point operations — measuring the shadow and calling it the object. A FLOPs threshold will be satisfied by every serious lab and circumvented by every unserious one.
Rebalancing is not panic; it is preservation. The correct institutional response to an unverifiable capability claim is not a new regulatory regime. It is a demand for replication under controlled conditions.
Watch for four things. A published harness. Contamination controls. A named base model. An independent replication by a lab with no commercial interest in the outcome. If those appear, the story upgrades from narrative to engineering, and the governance conversation finally acquires a referent. If they do not, the claim joins the long list of unfalsifiable assertions that preceded a drawdown.
Every bull run is a tax on due diligence. This cycle's bill arrives without a token attached, which makes it harder to see and no cheaper to pay. The question is not whether a model can construct a successor. The question is whether anyone is being asked to prove the difference between a successor and a system prompt — and whether the answer will arrive before the next cohort of investors pays for not asking.
