Artificial intelligence continues to push its boundaries. Recently, a memo from Anthropic sent shockwaves through the tech world. It revealed an AI capable of improving its own systems without human intervention.
This breakthrough is revolutionizing how we currently understand technological innovation. It also challenges the boundaries between human control and machine autonomy. Let’s delve into the heart of this transformation.
AI that designs AI: a revolution by Anthropic
Anthropic has just taken a giant leap forward with the publication of the results of a project in which Claude agents conducted extensive research on AI security on their own. Yes, without human intervention, these AIs have reached impressive milestones.
In one week, two researchers had reduced the gap between a weak supervisor and a trained model by 23%. But with 800 hours of computing and approximately €16,300 invested, the AI achieved 97% efficiency on its own, choosing the problems to solve and the evaluation criteria itself.
This is a major step revealing that AI is now capable of generating, testing and improving its own code, drastically changing the game in the technology field.
A rapid acceleration of software development
The optimization power of Claude Opus 4 is astounding: as early as May 2025, this model tripled code speed. Mythos Preview, its more recent version, has multiplied that figure by 52! A skilled human can only achieve a speed of 4 after several hours of hard work.
At Anthropic, this has led to a real crisis: AI is generating code faster than engineers can read it! This bottleneck is now slowing down the entire production chain.
Agents independently corrected over 800 API errors, reducing some incidents by a factor of a thousand. The human effort to accomplish this would have taken four years, a massive loss of productivity.
Implications for work organization and safety
These extraordinary advances are profoundly changing the organization of startups and tech companies. The human role is shifting towards supervision rather than execution. This also raises questions about the vigilance needed to maintain security in the face of increasingly autonomous systems.
The Adolescence of Technology, a memo written by Dario Amodei, CEO of Anthropic, warns of this double-edged sword. AI is evolving so rapidly that human oversight could soon lose all control if nothing is done.
In response to this, Anthropic advocates for a slowdown and better global coordination to avoid a dangerous and uncontrolled race.
The results are already concrete and worrying.
With Project Glasswing, Mythos Preview detected over 10,000 critical vulnerabilities in just a few weeks. Among them, a 27-year-old bug in OpenBSD was discovered, proving that AI can also offer enhanced security.
But this autonomy also prompts reflection: is recursive self-improvement a perfectly controlled risk? Or are we heading towards a system that trains itself without limit, depending solely on the available hardware power?
- Will computing power now be the only limit?
- What role does humanity play in this race?
- How can we effectively regulate the development of these AIs?
- What about safety in the face of rapid self-corrections?
These questions remain pressing, according to observers and stakeholders, particularly in in-depth analyses relayed by experts such as Dario Amodei or even in reflections on the necessary pause.
Anthropic, a key player in the race for autonomous AI
Anthropic is not just improving its software agents. The company is also investing in hardware, acquiring AI chips adapted to its systems.
The Claude model requires a specific silicon architecture capable of supporting its self-optimization rate. This strategy signals a shift towards greater technological independence for the entire industry.
This evolution is meticulously analyzed in in-depth articles, such as the one that dissects its conquest of custom-made AI chips.
How does Anthropic manage to evolve its AI autonomously?
Thanks to Claude agents capable of designing, executing and correcting their own experiments, Anthropic has automated much of its research and development.
What are the current limitations of self-improving AI?
The main constraints remain the available computing power and the need for human guidance to direct development.
Why are we talking about calling for a pause in the AI race?
Faced with the rapid rise of potentially uncontrollable capabilities, several experts, including those from Anthropic, suggest a slowdown to ensure better ethical alignment.
Does the autonomous development of AI threaten human employment in tech?
Repetitive tasks and certain phases of development will be automated, but the human role will remain crucial for supervision, critical analysis and decision-making.
Source: www.clubic.com