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
According to the Wall Street Journal, in June Tom Brown broke the deadlock with the White House over two suspended models, paving the way for the $1.25 billion-a-month computing contract with SpaceX.
Aleph Alpha’s first open-source model is a mixture-of-experts model with 78 billion parameters – 3.46 billion of which are active per token – focused on German and English, with a context validated up to one million tokens.
Suppliers keep their watermark detectors secret for fear of them being circumvented. A study proposes publishing half of them: forgery then becomes possible, but remains detectable using the private key.
A compromised server can exfiltrate a model’s weights by encoding them within plausible responses. A statistical method accumulates clues across requests whilst minimising the risk of false positives.
A campaign carried out in 2025 at the LUNA facility showed that the quadrupeds traversed the simulated lunar regolith, but slipped, sank into it, kicked up dust and lost visibility under low-angle lighting.
According to PitchBook, as reported by the Financial Times, robotics start-ups raised nearly $48 billion in 2026. An increasing proportion of this funding is being used to film people at work in order to train the models.
▶
Jesse Hall (LiveKit) explique pourquoi un agent vocal rapide sur le papier peut sembler robotique, et pourquoi viser environ 600 ms de bout en bout pour paraître humain.
Out of 3,041 responses from computer science students, the feedback on three business LLMs remains consistently positive and their sub-ratings are virtually identical, despite the appearance of a detailed analysis.
Section 12 of the AI Act requires events to be recorded without specifying which ones prove what. Researchers identify four factors that must be established before any assessment of a claim regarding an AI agent can be carried out.
At Emory University, four models were tested on 1,000 images. MammoCLIP achieved an AUROC of 0.999, but the tiny ResNetLite – which is the fastest – performed no worse than ResNet18.
A unique neural encoder converts surface area, year and type of heating into physical thermal parameters. The result: consistent annual trajectories with ten times fewer parameters than competing models.
Using fifteen years of Météo-France radar data, a team is comparing U-Net, Transformers, GANs and diffusion models to produce detailed precipitation maps. Deterministic models consistently underestimate heavy rainfall.
A study traces which SKILL.md files are copied onto GitHub. A few repositories are responsible for almost all of them, and auditing 100 carefully selected ones would identify 30 times as many at-risk skills as auditing the 100 most highly rated ones.
The Israeli start-up, founded in 2024, places a ‘sidecar’ in the client’s cloud to monitor and block its agents’ actions in real time. It has now raised a total of 35 million.
Derived from Qwen3.5-9B and available in Foundry and on OpenRouter, the model assigns a calibrated probability to each option for routing, classifying or verifying within agent workflows.
Led by Patrick Soon-Shiong, in partnership with AstraZeneca and Teradyne, the Series B funding round is intended to expand robotic ‘clusters’ that replicate the movements of technicians without the need for further regulatory approval.
▶
La responsable RL de MiniMax détaille l'attention clairsemée MSA de M3 et explique pourquoi entraîner texte et vision dès la première étape donne de meilleurs résultats que greffer la vision ensuite.
In a lengthy post published on Saturday, the head of Microsoft argued that the most reliable AI systems will be those that rely as little as possible on the model itself.
Founded by Jeff Holden, formerly of Amazon and Uber, the start-up is already supplying an ultra-fast electromechanical relay designed for 800 V power supplies in AI data centres.
Rather than generating text token by token, these ‘decision models’ return a short, structured response in a single inference, for real-time embedded applications.