Analysis of AI, automation, cloud and software. No hype: what is changing, what works, and what to verify before taking a promise at face value.
Latest article Science and mathematics
OpenAI has published 722 manuscripts with results on open problems in mathematics. What they contain, which ones would really matter, what a machine has checked and what is still pending.
From the imitation game of 1950 to experiments with current models: what the Turing test measures and its place in AI research.
AI-assisted mods reopen a question troubling artists and mathematicians: why keep going when a machine can do it? Chess offers a clue.
We tried Cloudflare OS at Codin: an environment for creating small applications with AI, adapting them and sharing them. Our CRM is a first example.
OpenAI says it has solved more than a hundred open problems and part of the mathematical community responds with a letter. What is being debated, what is being asked and what we think at Codin.
A steadily growing underlying capability, tasks that chain many steps and labs that train what they can verify: why AI advances in jumps on specific tasks.
The tests highlighted in the launches of GPT-2, GPT-4, o1 and Grok show how our expectations of their capabilities have evolved.
What the Great Filter is and why AI could be part of it: technological power, fragile institutions, and a capacity for control that falls behind.
Six capabilities for understanding future LLM risks: the dangers they could enable, the evidence available and what is still missing from the most serious scenarios.
A fly connectome opens a still speculative debate: whether an emulation could feel, and what obligations we would have if it did.
OpenAI claims to have solved Navier–Stokes. Beyond the authorship dispute, its construction with external forcing would leave a fundamental question open.
Cycling pelicans, balls inside a hexagon and Blender scenes: viral tests show how expectations of AI have risen and what we can infer from its demos.
The important threshold for an AI agent may not be consciousness or superintelligence, but the ability to obtain resources, remain active and reconstruct itself without a clear owner.
What local inference can contribute to a process involving sensitive data, its real limits, and what it takes to operate it in production.
AI is accelerating robotics, but not all its bottlenecks are moving at the same pace. The plumber test separates intelligence, dexterity, reliability and cost.
Single-use apps propose generating software for a specific need and discarding it afterwards. The interesting idea is not producing more applications, but deciding which parts of software need to last.
The human form is not optimal for every task. Its advantage appears when a general-purpose robot must use our tools and operate in spaces that already exist.
AI systems are beginning to produce mathematical research results. The challenge may shift from proving theorems to verifying, understanding, and selecting them.
A study of 26,811 students found better homework in less time but worse exams. The issue is not simply using AI, but which effort it replaces.
