AI Leaders Question the Race to Go Faster
By Suad Seferi ·
The debate over artificial intelligence changed tone this week after several of the executives leading the development of frontier models backed calls for greater restraint in how quickly the most capable systems are developed. Anthropic chief executive Dario Amodei set out the clearest version of that argument in an essay titled We Must Pace the Frontier, published in September. He called for independent evaluation of advanced AI systems, stronger safety standards and international coordination as capabilities continue to improve. OpenAI's Sam Altman and xAI's Elon Musk subsequently expressed support for elements of the proposal, adding unusual weight to an argument that has often come from researchers and outside critics rather than the companies competing at the frontier. The significance is not that the AI industry has agreed to stop developing more powerful models. It has not. The companies involved remain in intense competition over performance, infrastructure, customers and talent. Anthropic itself is growing rapidly while preparing for a possible public listing, illustrating how commercial expansion and concerns about safety are unfolding at the same time. What has changed is that the pace of development itself is becoming part of the argument. That matters beyond the frontier laboratories because the rest of the economy is moving in almost the opposite direction. Businesses, governments and individual users are being encouraged to increase AI adoption, automate more processes and delegate a growing range of tasks to systems whose capabilities are developing faster than many organisations can evaluate them. The result is a widening gap between what AI can do and the ability of institutions and users to determine where it should be used. Capability is moving faster than the systems used to measure it Amodei's concern is rooted partly in the speed of technical progress. Stanford's 2026 AI Index found that frontier systems improved by around 30 percentage points in a single year on Humanity's Last Exam, a benchmark specifically designed to remain difficult for advanced models. Other evaluations are also becoming saturated quickly, while researchers have raised questions about the reliability of some widely used benchmarks. That creates a practical problem for developers and regulators. If evaluation methods become outdated within months, safety frameworks built around those evaluations can also struggle to keep pace. Transparency has moved in the opposite direction. Stanford found that industry produced more than 90% of notable AI models in 2025, while some of the most capable systems now disclose less information about their training data, parameter counts, computing requirements and development process. Amodei argues that this combination of rapidly improving capabilities and limited external visibility makes independent evaluation increasingly important. His proposal is aimed primarily at frontier systems, particularly as models improve in areas such as autonomous operation, cybersecurity and assistance with AI research itself. There is already political resistance to that argument. US President Donald Trump rejected calls for a significant slowdown, arguing that maintaining American leadership over China remains a strategic priority. The disagreement illustrates one of the central problems surrounding frontier AI governance: governments may recognise new risks while simultaneously fearing the consequences of allowing another country to advance faster. That tension is unlikely to disappear. Adoption is accelerating for a different reason Outside the frontier laboratories, the pressure is not primarily geopolitical. It is economic. Organisations increasingly view AI adoption as a competitiveness issue. Stanford estimates that 88% of surveyed organisations were using AI somewhere in their operations in 2025, while 70% reported using generative AI in at least one business function. Generative AI reached an estimated 53% population-level adoption within three years, considerably faster than the early adoption curves of the personal computer or the internet. Those figures show how quickly the technology has spread. They do not demonstrate that every use creates value. This distinction is becoming more important as organisations move from experimentation toward integration. Giving employees access to a chatbot is relatively easy. Redesigning a process around AI requires decisions about reliability, data access, security, supervision, accountability and what happens when the system produces an incorrect result. The adoption numbers also conceal large differences between applications. Using AI to transcribe a meeting creates a different risk profile from allowing an autonomous system to change production code. Generating a first draft of a marketing text is not comparable to using AI to support a decision about employment, credit, healthcare or public services. Treating all of these activities as a single category called "AI adoption" makes the overall statistics less useful for understanding whether organisations are actually becoming more capable. The more important measure is increasingly what has been delegated, under what conditions and with what evidence that the change improves the underlying process. Agents make the distinction harder to ignore The issue becomes more consequential as AI systems move beyond producing content and begin performing actions. Agentic systems can already navigate websites, interact with software, execute code and complete sequences of tasks with less direct human involvement. Stanford's data suggests that enterprise deployment of agents remains relatively limited, with adoption still in the single digits across most business functions, but their technical capabilities are advancing rapidly. This changes the nature of an AI error. An incorrect answer from a chatbot can be reviewed before it is used. An agent connected to operational systems may be able t…