Editorial section — [ 0x1 / sys ]
Software architecture, distributed systems, artificial intelligence and security.
In Yemen, a cell used Claude Code like a small team of software engineers to work on guidance systems, test a rocket, then analyze its failure. The case is spectacular, but it mainly reveals a deeper phenomenon: AI is starting to make rare military expertise much less rare.
For years, every lab wanted to accelerate. In September 2026, Anthropic is now asking to slow frontier models, OpenAI says it's ready to do so, and several industry leaders approve. This reversal reveals above all that AI is beginning to accelerate AI research itself.
It is often repeated that a quantum computer will be able to break part of our cryptography. But how, exactly? The real danger does not come from an absurdly fast machine: it comes from an algorithm that completely changes the way the problem is attacked.
Quantum computers capable of breaking our cryptography don't exist yet. Yet, part of the Web is already preparing for their arrival. And this isn't science fiction.
I asked GPT-6 Astra to design an original benchmark blending logic, probabilities, causality, software concurrency, optimization, and self-verification. ChatGPT, Gemini, DeepSeek, Kimi, and Grok then took it without access to the answer key. The result isn't just a ranking: it's a fairly brutal X-ray of how these models reason, prove… and sometimes persist in their own errors.
Hundreds of millions of queries, thousands of fake accounts, and models trained on their competitors' responses: distillation has become an industrial and geopolitical issue. But between legitimate learning, capability extraction, and outright theft, the line is far less clear than it appears.