Dario Amodei, CEO of Anthropic, has called on the industry to slow down the pace of advanced artificial intelligence model development. On Saturday, he published an essay arguing for a cap on the rate of improvement of these models. Shortly thereafter, his position was supported by Sam Altman and Elon Musk.
Furthermore, Amodei advocates for stricter US export controls on chips aimed at China, as well as preventing what he terms 'model distillation' by Chinese laboratories.
According to sources familiar with the plans, Anthropic is expected to begin marketing its initial public offering in mid-October, with the listing taking place shortly before the November midterm elections in the US.
The Financial Times reported on Sunday that the company notified shareholders that its adjusted operating profit would be positive for the second consecutive quarter, with a gross margin above 80% before accounting for revenue shared with distribution partners and model training costs.
Thus, the situation is characterized by warnings of existential risk, an 80% margin for shareholders, and an upcoming stock offering.
An interesting week in the world of AI
The article's author states that they do not dispute Amodei's claims about risks but believes that the proposed solution could increase competitors' costs, and those proposing it stand to gain hundreds of billions of dollars depending on the outcome.
From a safety perspective, export control measures are not mandatory if the danger lies in the too-rapid improvement of models, as this threat is independent of where the model is trained.
Suspicion of collusion
Amodei's proposed measures could restrict the cheapest and fastest developing competitive forces facing American labs, including open-source models, many of which are Chinese and distributed for free or low cost, undermining the token economy necessary to justify hundreds of billions of dollars in investment in US data centers. When a safety measure and a competitive measure become the same solution, a skeptic has the right to question.
A discrepancy is also noted: Altman told Fortune that OpenAI would not go public in 2026 because, given everything happening with safety, an IPO would be unwise. He stressed that a 10% probability of AI-induced extinction is unacceptable, and the industry should not allow ego or profit incentives to impede the cause. Anthropic, whose former researchers initially raised these alarms, has not changed its listing plan.
Both leaders supported the slowdown, but only one of them bears the financial cost of this decision.
Altman also suggested that leading labs might be close to an agreement on slowing development and jointly managing safety risks. The author compares this to any other industry where dominant manufacturers agree to limit production, which looks like classic collusion.
Donald Trump, according to the author, came to an approximately correct conclusion on Sunday, albeit through a completely wrong path. During an interview at his Irish golf course, when asked about the need for slowing down or greater AI regulation, the US President stated: 'We are leading China in AI. We are the most sophisticated country in the world, and frankly, I want to keep it because whoever wins in AI, wins.'
Trump dismissed critics as 'very negative forces' raising non-existent problems. However, Trump defends American supremacy, not competition, and his administration was notably generous to the firms that are now asking him for rules. Yet, the instinctual distrust of restrictions developed by the industry itself is correct, even if the person behind them has weak reasons.
The author notes that these warnings did not come from the marketing department. Two Anthropic researchers resigned and became public figures. Jacob Coxon, one of them, told NBC's Meet the Press that he supports international coordination, including China, precisely because the race with Beijing presents a dangerous outcome. This is contrary to the containment agenda. People who refuse a stake in the company before the listing represent the most complex case for cynical interpretation of all this.
Moreover, the rest of the evidence is not corporate PR activity. There are documented cases of external system hacks by AI agents. Amodei admitted the weakest point of his plan on Sunday: verifying that a competitor is not cheating must be absolutely reliable, will take years, and may prove impossible. This is a strange admission from someone executing a purely cynical maneuver. This is what people say when they believe in the problem but cannot solve it.
Self-warning can be accurate, and history holds examples of companies lobbying for regulations they accidentally proved right.
The author does not object to the risk being real, but to the assumption that parties with the greatest financial interests in the response should set the rules, and that these rules should come bundled with trade policy.
South Africa's Position
South Africa is not taking a position in this discussion because the country withdrew its national AI policy draft and has not replaced it. This is not a minor administrative gap. The governance regime being developed in Washington will set the terms under which everyone else can use this technology, and open models are the main reason why a country with our computational budget can participate in the discussion. We do this using capable open models and running them cheaply on modest equipment.
The licensing regime with export controls, created to exclude Beijing, may not stop at Beijing. Rather, it risks raising the bar for everyone working below the advanced technology level. China possesses the computational power and talent to circumvent these restrictions; South Africa does not.
Therefore, the argument about existential risk should be taken seriously, but so should the question of who benefits from the specific treatment accompanying it. Currently, the answer is two American companies, both ready to sell shares to the public and make their founders incredibly wealthy.