Keeping up is exhausting. The constant flow of WTF stories out of the global AI technopolitical ecosystem has rendered both the commentators and the curious (including this writer) a little punch-drunk.
There was Bernie’s Bill (I wrote about it here), Navier–Stokes (here), the Hugging Face hack (here), the data centre counter-revolution (here) the astonishing new DeepSeek 4.1 Flash and other models from China, the unprecedented swathes of capital moving hither and thither, and of course the revelation of a widely held belief within the AI labs that there is a better than 10 per cent chance of an AI-catalysed human extinction scenario.
And then came Saturday.
Dario Amodei of Anthropic published a 3,800-word essay titled “We Must Pace the Frontier”. It was built around a sentence few people expected from the boss of a company preparing a stock-market listing at a valuation of around two trillion dollars – “We must slow the pace at which we improve the capabilities of AI models.”
Whether Dario wrote this in response to growing anxiety on the streets, or the threats of politicians, or the shock of the Hugging Face hack, or whether he had been planning this for a while was not fully disclosed. What was even more astonishing was that Sam Altman and Elon Musk agreed. Remember, while alliances constantly shift under the pressures of coopetition, these three basically hate each other, even when they act otherwise.
It is easy to be cynical about this announcement, and to automatically assume that Amodei and his team had a good, disingenuous cackle when they pressed the send button, but that would be too easy. It is worth the time to dig deeper to see what was proposed, and to ask where these aspirations may stumble on the jagged rocks of both politics and technology.
Let’s parse the phrase “slow down”. It sounds like a policy. It functions rhetorically like a policy. But it resists definition.
Amodei himself was careful to clarify what he was not proposing. “Pacing does not mean halting model training or technical progress,” he wrote. In interviews, he emphasised that Anthropic was not going to stop training Claude or withdraw from model competition. The goal was to find “a new balance of speed between capabilities and safety”.
This is not a slowdown in any conventional sense. It is a slowdown relative to an unstated counterfactual – a world in which Anthropic and its competitors push capabilities as fast as physically possible, with no safety evaluations, no third-party oversight, no deliberation. The “slowdown” is a moderation of the acceleration, not a change in direction.
The problem is that no one has defined the baseline. How fast were we going before? How fast are we going now? How fast would we be going in the world where Amodei’s proposal is rejected? Without answers to these questions, “slow down” is more of a vibe than a policy prescription.
Amodei proposed a three-point framework – independent third-party evaluators with employee-like access embedded within frontier labs, coordination among AI companies in democratic countries on safety standards and limits, and global coordination, “to the extent this is possible,” between democratic and authoritarian governments. Anthropic committed unilaterally to the first step.
This tells us little. What, really, and actually and in fine-print detail, should be slowed?
Five things:
Slow the growth of compute. Slow the acquisition of dangerous capabilities (like how to make anthrax). Slow AI-assisted AI evolution (otherwise known as recursive self-improvement). Slow the deployment of models. Or slow everything, through an internationally agreed ceiling.
Only the first is relatively easy to measure (but the wrong metric, because each increment in the underlying software’s capabilities makes compute speed that much less important). The others require judgements about intelligence, danger and acceptable risk that nobody yet knows how to make reliably. And all ultimately collide with the same strategic problem – a unilateral speed limit is rather less attractive when you believe somebody on the other side of the world has kept his foot on the accelerator.
Which brings us to the deepest problem with “slow down” – it collides with a hard wall of geopolitics.
Amodei acknowledged this in his essay, noting that completely halting development “deprives humanity of benefits or simply places AI in the hands of authoritarian powers”. President Trump has been explicit in rejecting a slow-down is an all-caps post – “WHOEVER WINS AI, WINS!”. And Chinese mouthpiece Global Times reacted negatively to the whole idea, comparing it to the playbook of the Cold War. The framing of AI as a national security race, with the United States and China as the primary competitors, makes any unilateral slowdown politically untenable.
Amodei’s proposed solution – coordination between democratic and authoritarian governments – is aspirational at best. There is no mechanism for enforcing a global compute cap, no institution with the authority to verify compliance, no precedent for great powers agreeing to slow down a technology they believe is strategically decisive. The history of nuclear arms control offers a partial model, but nuclear weapons were never commercial products developed by private companies competing for market share.
This is why “slow down” may be structurally cosmetic (or even self-serving). It allows AI executives to signal concern, to position themselves as responsible stewards, to pre-empt regulation by offering voluntary measures – without actually changing the trajectory of capability development. The geopolitical competition ensures that someone, somewhere, will push forward. If not Anthropic, then OpenAI. If not OpenAI, then DeepSeek. The race continues regardless.
There is a danger that “pacing” becomes the perfect Silicon Valley word – sufficiently conservative to demonstrate responsibility, sufficiently vague to avoid specifying what anyone must actually stop doing.
So, is this all pointless, this “slow down” talk? No, because there is one genuinely good idea that has come out of it. I heard it on this week’s All-In podcast, on which Elon Musk was a featured guest. Whether this idea originated from him or not I do not know, but he described it succinctly and in plain language.
He said – “Have the AIs test each other’s models. So instead of grading your own homework, you would have competitors grading your homework and raising the alarms if they see concerns.”
He went on to point out that everyone would be able to view the security testing logs of everyone else, so it would not be possible to steal IP without immediately being seen, named and shamed. The cost of malfeasance would be extremely high in terms of reputational damage. And in his version, testing would be tightly constrained – bioweapons, nuclear weapons, deliberate deception, hacking
“I think that if you have the sum of all your competitor’s tests and you have all these heterogenous models, then you are not grading your own homework, and there is a reason you don’t grade your own homework”.
Would the Chinese agree? Elon thinks so – there is little downside for them, and no IP theft concerns because of the public logs. And the regulators? They would probably agree that competitors would have the sharpest teeth when publicly adjudicating each other’s threats to safety.
This, it seems to me, is the only smart idea to come out of this whole knotty matter, and it is much more efficient than Amodei’s “third party evaluator” suggestion. It costs little, can be implemented immediately without the sludge of government regulation and (more importantly) might even work.
Steven Boykey Sidley is a professor of practice at (ex-JBS, University of Johannesburg) and a partner at Bridge Capital and a columnist-at-large at Daily Maverick, Daily Friend and Financial Mail. His new book “It’s Mine: How the Crypto Industry is Redefining Ownership” is published by Maverick451 in SA and Legend Times Group in UK/EU, available now.
[image: Anisa Gauri for Unsplash]
The views of the writers are not necessarily the views of the Daily Friend or the IRR.
If you like what you have just read, support the Daily Friend