AI is Advancing: Act Before It's Too Late
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisis ton rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
The Urgency of Anticipation in the Field of AI
The rapid advancement of artificial intelligence (AI) capabilities necessitates a level of anticipation similar to that required in defense. Experts estimate that by 2030, AI systems could surpass human capabilities in many areas, presenting a significant control challenge. This situation demands immediate reflection to prevent an uncontrollable technological escalation.
Military Projects Illustrating Anticipation
On March 2, the President of the Republic unveiled the name of the Invincible, a third-generation nuclear submarine designed to operate until 2090. A few weeks later, on March 18, he named the France Libre, a next-generation aircraft carrier expected to enter service by 2038. These projects demonstrate France's ability to commit substantial resources over several decades to ensure its sovereignty. This proactive approach is essential for anticipating future threats.
The Importance of Preemptive Preparation
It is crucial to size a response before the need arises. Once the threat is present, it is often too late to react effectively. This necessity for anticipation also applies to artificial intelligence, particularly concerning the control of general-purpose AI systems.
Revealing Benchmarks of AI Progress
A thorough analysis of 60 standardized benchmarks, covering key cognitive capabilities such as reasoning and mathematics, shows that 98% of these assessments could be saturated before 2030. Advanced AI systems are already outperforming human experts in fields such as physics, chemistry, and biology, progressing at a faster rate than expected.
Impressive Performances
To illustrate this dynamic, consider the GPQA Diamond benchmark, which evaluates doctoral-level scientific questions. Specialized researchers achieved a score of 69.7%, while current AI models exceed 90%. In 2022, it was estimated that an AI could win a gold medal at the International Mathematical Olympiad around 2030, with an 8.6% probability before 2025. This performance was achieved as early as this year.
Cybersecurity Challenges
In the field of cybersecurity, during the AI Cyber Challenge organized by DARPA in August 2025, AI systems discovered 18 zero-day vulnerabilities in open-source software. These advancements underscore the need to develop robust control methods for these powerful technologies.
The Challenge of AI Control
It is difficult to predict when an AI will surpass all human capabilities, but it is clear that the control issue requires an immediate response. Current alignment techniques are insufficient to ensure that these systems remain under control. AI capabilities continue to progress without slowing down.
An Imminent Threshold Effect
Planners must be aware of the potential threshold effect. If AI systems become capable of autonomously contributing to AI research, a feedback loop could be triggered, drastically reducing preparation time. Benchmarks for autonomous software engineering suggest that this threshold could be reached in the next two to three years.
The Necessity for Immediate Action
The France Libre will be operational well before the retirement of the Charles de Gaulle, and the Invincible will patrol before the threat it is meant to deter materializes. This anticipation is crucial for AI, with an additional constraint: the timeline for capabilities is not under our control. To avoid being outpaced or exposed to security issues, it is imperative to act now.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.