Anticipating AI like defense is vital: its capabilities are progressing very quickly, could overtake experts before 2030 and already pose a control challenge, with a risk of runaway in the short term.
On March 2, at Île-Longue, the President of the Republic unveiled the name of the Invincible, the first third-generation nuclear missile submarine, designed to sail until 2090. On March 18, at Indret, he named the France Libre, a new generation aircraft carrier scheduled to enter service in 2038. These two programs illustrate the ability to commit considerable resources, over several decades, to guarantee a sovereignty which only has value if it precedes the threat.
You don’t design a response the day you need it. We size it before, because afterward, it is too late. It is this same anticipation that must now be applied to artificial intelligence, and more precisely to the problem of controlling general-purpose AI systems.
A systematic analysis of 60 standardized benchmarks covering the main cognitive capabilities of advanced AI systems, reasoning, mathematics, programming, scientific knowledge and autonomous operational capabilities, indicates, based on a hierarchical Bayesian model, that 98% of these assessments could be saturated before 2030.
On benchmarks related to offensive cybersecurity, automation of AI R&D, and dual-use biological and chemical knowledge, saturation is projected before 2028. Cutting-edge models are already outperforming experts with doctorates in physics, chemistry, and biology on knowledge tests. They solve research-level math problems. They identify new software vulnerabilities. And these capacities are increasing by tens of percentage points per year, faster than experts had anticipated over the last decade. To give an idea of this dynamic: on GPQA Diamond, a benchmark of doctoral-level scientific questions, researchers specializing in the disciplines evaluated only obtained a score of 69.7%. Current models exceed 90%. In 2022, experts estimated that an AI would win a gold medal in the International Mathematics Olympiad around 2030, with an 8.6% probability before 2025. It was obtained that year. In cybersecurity, during DARPA’s AI Cyber Challenge in August 2025, AI systems discovered 18 real-world zero-day vulnerabilities in open-source software.
No one is asking policymakers to decide when artificial intelligence will be trained to surpass the full capabilities of human experts in all cognitive tasks. The exact when is a difficult, perhaps impossible, bet. On the other hand, determining whether the control problem requires an answer now is an easy gamble. Current alignment techniques are insufficient. There is no method to ensure that a system beyond the capabilities of human experts will remain controllable. However, capacity trajectories are not slowing down.
Above all, there is a threshold effect that planners in headquarters must keep in mind. If AI systems become capable of contributing autonomously to AI research, a feedback loop is triggered and preparation time is suddenly compressed. Autonomous software engineering benchmarks indicate that this threshold could be crossed in the next two to three years. Whatever scenario is chosen for the future, the time available to act is shrinking.
The French Libre will be operational a decade before the withdrawal of the Charles de Gaulle. The Invincible will patrol long before the threat it must deter materializes. Because we anticipated. For artificial intelligence, it is the same requirement, with an additional constraint: the capacity schedule is not in our hands. If we still want to make a difference and not be downgraded or exposed to security problems, it’s today. Maybe tomorrow. But after that… who knows?
