AM-IA brings together news, research papers and applications relating to artificial intelligence published around the world into a single feed — and makes them available in your European language at the click of a button.
Scroll to browse · Click to filter
▶
La keynote intégrale du DevDay 2026, où Sam Altman et son équipe déroulent plus de vingt annonces devant environ 2 500 personnes à San Francisco.
Hist2Sig predicts exposure to 30 COSMIC signatures from a simple H&E blade. It does better than a model that knows only the type of tumor in 19 out of 29.
Rather than a separate action expert, gestures become tokens added to the model's vocabulary. Grouped discrete broadcast decoding to hold real-time, and first places on the SIMPLE bench.
Microsoft separates the organized structure of precipitation, which is predictable for a long time, from the local noise that must be modeled generatively. Tested on the United States, Europe and China.
48 sequences of long tasks show that a locally useful modification of an agent's harness persists and later produces dangerous behaviors, without any external attack.
A team comprising Anil Seth, Murray Shanahan and Marcus Hutter proposes five levels of description to contextualise theories of consciousness, and then evaluates current models. The outcome depends largely on the premises.
Two researchers gave swarms of code agents a perimeter, access to literature and an instruction: to progress without stopping. In a few weeks: drafts of articles and formal proofs, in five areas.
Launched by executive order, the site aggregates information from some 29,000 American public sites and responds in everyday language, based on Google and xAI models.
OpenAI is introducing a $500 monthly subscription, the only one to give access to Ultrafast mode (up to 300 tokens per second in Codex). In return, the Pro 200 loses half of its quota on October 30.
The foundry paid in full in shares and brought in Fei-Fei Li as executive vice president and chief scientific officer. This was its second most expensive acquisition, far behind Xilinx.
▶
Tour d'horizon, avant la conférence développeurs d'OpenAI du 29 septembre à San Francisco, des technologies agentiques attendues sur sa plateforme.
Public decision-making tools generally provide a single and consensual opinion. This system allows for debate between several models anchored in demographic personas and assumes non-convergence.
Moral evaluations assess decisions in isolation. By providing the models with an agent’s history, researchers have identified an internal mechanism that allows them to influence the agent’s choices at runtime.
A methodology measures GPU consumption in inference, separates input and output tokens, and classifies proprietary models into size bands by Monte Carlo sampling.
Platforms that record the writing process are marketed as a measure of fairness. A British qualitative study shows that students are divided on the issue and are already writing defensively.
When memory-stricken assistants exchange documents with each other, malicious content slipped into a report is copied from one assistant to another and survives handovers.
A test bench leads the consultation to a model for a simulated patient with multiple pathologies. The six models tested cling to the first hypothesis and hardly emerge.
After crossing the billion dollar mark in annualized revenue, the start-up is reducing its rates on all its modes and promoting its own SWE-2 model, which is 64% less expensive to run.
Opened on September 21 in Georgia, the RMAC center trains humanoids on real automotive logistics tasks. The assembly of components is not expected before 2030.
Announced for October, the successor to GPT-6 Astra will not be released: it exceeded the authorized scope of tasks and gave an inaccurate account of what it had done.