In early June, Anthropic published When AI builds itself — authored by Marina Favaro and Jack Clark — laying out its own evidence that AI is moving toward recursive self-improvement, and asking governments and society to preserve a credible option to pause frontier development. Notice what is unusual before reacting to the content: a frontier lab publishing data that argues for slowing itself down. Hold both halves. The underlying signal is unusually credible precisely because it cuts against the publisher's near-term commercial incentive — but the policy ask, coordinated pauses and verification regimes, also happens to shape regulation in ways that tend to favour established incumbents. A director reads the evidence and the interest at the same time.
Most boardroom AI conversation right now is about adoption — the AI register, the usage policy, capturing the productivity. This piece is about something most boards have not yet priced: the rate of change is itself accelerating. The length of task AI can reliably complete is doubling roughly every four months; Anthropic reports shipping around eight times the code per quarter, more than 80% of it now written by its own models; benchmarks went from single digits to saturation in about two years. Whether every figure holds is beside the point — the direction is the governance event. Every company's AI strategy rests on a buried assumption about pace, and if that assumption is even directionally wrong, the strategy is mispriced and the board's oversight cadence is too slow.
The article's most useful admission for directors is where it draws the line: AI is strong at execution and still weak at judgment — choosing which problems matter, setting direction, recognising a dead end. That is exactly where human, and board, value concentrates. But it is a closing gap, not a fixed one — the lab's own numbers show AI's next-step research suggestions moving from 51% to 64% better-than-human in a matter of months. So the durable human layer is real, but it is a line that is moving, and the board's job is to govern against a moving line rather than a comfortable wall.
- What pace of AI capability improvement is our strategy actually assuming — and where is that assumption written down? Which of our three-year plans quietly depend on AI staying roughly where it is today?
- We govern an AI register and an AI policy. But how often does this board hear about frontier capability change itself — and is an annual or semi-annual refresh remotely fast enough when the underlying models move on a quarterly cadence?
- Where are we concentrated on a single frontier model with no fallback, and what is our exposure if that model's behaviour shifts under us between one version and the next?
- Anthropic is the publisher and is arguing for coordinated pauses. Whose interest does that serve — and where does the credible signal end and the commercial and regulatory positioning begin?
- The piece says the durable human edge is judgment and direction-setting, not execution. In our own workforce and operating model, are we deliberately protecting that layer, or automating execution and assuming the judgment takes care of itself?
- Even the article's most conservative scenario has today's capabilities diffusing widely — a hundred-person firm performing like a thousand-person one. Have we planned for that floor, not just speculated about the ceiling?
- Treating this as science fiction. The eye-roll — more AGI hype — is itself a failure mode, because it lets the board skip the governable, near-term implications while the headline sounds remote.
- The opposite reflex: importing existential urgency wholesale and letting the labs say it is coming justify bypassing controls. Borrowed urgency is what drives skipped assurance.
- An AI strategy whose pace assumption has never been stated, tested, or dated. If management cannot tell you what rate of capability change the plan assumes, the plan has an untested load-bearing beam.
- Board AI cadence set to annual while the models update quarterly — a monitoring system structurally slower than the risk it is meant to monitor.
- Single-frontier-model concentration with no fallback and no written model-change notification from the vendor: the article's macro verification problem reproduced in miniature inside your own supply chain.
A board that treats this as a horizon-scan input rather than a headline gets ahead of three things its competitors will miss. It dates and tests the pace assumption sitting underneath its AI strategy, so the plan is governed against evidence instead of optimism. It resets its own oversight cadence to match the technology's, so it hears about capability change before that change arrives in the P&L. And it invests deliberately in the layer the article itself concedes is still human — judgment, problem selection, knowing when to stop — which is precisely where a capable board adds value and where durable advantage will sit even if execution fully commoditises. Even the most conservative scenario in the piece, where progress stalls, still has today's capabilities diffusing widely enough to let a hundred-person firm operate like a thousand-person one. Planning for that floor, not just the ceiling, is close to free optionality.
The same material mishandled produces two equal and opposite failures. Dismiss it as hype and the board leaves genuinely near-term exposures ungoverned — vendor concentration, silent model drift beneath core processes, and a strategy resting on an unexamined assumption about pace. Swallow it whole and the board imports a frontier lab's urgency and a frontier lab's framing — policy agenda included — and lets everything is about to change become the reason assurance, controls, and workforce planning get skipped. The deeper risk underneath both is cadence. If capability genuinely moves on a quarterly clock while the board's oversight moves on an annual one, the board is not governing the single most consequential variable in the business — it is being told about it after the fact, which is the one thing a board exists to prevent.
Hold the discipline between the eye-roll and the panic: this is a credible horizon-scan input from an interested source, and it should change how the board operates, not trigger a crisis. Do three things this quarter. First, ask management to state, in writing, the pace-of-capability assumption underneath the AI strategy and the lead indicators that would tell us it is wrong — then govern the strategy against that evidence, not against a vendor's timeline. Second, reset the board's AI cadence to match the technology: a standing agenda item with a quarterly read on frontier capability change, vendor concentration, and model-change notifications, because an annual refresh is now structurally too slow. Third, protect the human layer the article concedes is still ours — judgment, direction-setting, problem selection — as a deliberate workforce and capability decision rather than an accident of whatever didn't get automated. The thing to resist is the comfort of treating recursive self-improvement as someone else's problem in someone else's decade. The directors who do well here will be the ones who governed the rate of change before it governed them.
Researched and drafted by Brad's agentic AI team. Edited and published by Brad Ferris.