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
A team from MIT, Carnegie Mellon, NYU and Stanford has beaten the best human Stratego players by focusing on planning at the moment of play rather than on vast amounts of self-learning.
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.
In theory N=4 super Yang-Mills, the direct computation had been capped at eight loops since the work of Lance Dixon. Fable 5.1 went to nine, for 96 processors for a week and a few thousand dollars.
Researchers at NYU publish an open-source tool that determines the stable shape of tautomers; out of 5,075 reference complexes, it reports 126 structures that are probably poorly recorded.
Eric Lu at Cognition factored the 862-bit number into 4,900 GPU-days using an algebraic screen implementation developed and piloted with Devin, at a cost ten times lower than public methods.
The Biological Computing Co. is marketing with AWS an optimization based on measurements on cultured neurons, which weighs less than 0.1% of the model it accelerates.
Facility-0 gives models control over 40 real instruments; on the SciUniverse bench, Claude reaches 45.3% compared to 32.5% for GPT-6.
By replaying the history of its explorations instead of rerunning them, the method reduces the number of calls required by a Scientific Discovery Agent by up to 162 times.
A method presented at EMNLP 2026 reduces the size of the evaluation sets by running only a small number of reference models, with mean absolute errors between 0.002 and 0.02.
A vision model and a reinforcement agent pushed molecules on copper to form a honeycomb network, in 900 iterations and 25 hours, without breaking the tool.
The start-up claims that an AI system can design an inference chip on its own, with 95% code coverage without a verification engineer. Silicon, on the other hand, remains to be engraved.
Justin Fanelli, chief technology officer of the US Navy, details to investment funds the five areas where his $150 billion annual will go to find start-ups.
Helmholtz Munich and TUM have assembled natural protein bricks and structures invented by generative AI. The best of the hundred variants tested delivers more RNA with less dose.
The company confirms that it is conducting physical experiments with its models, on basic biology rather than on drug discovery.
Fermilab says it has reduced the development of a fault-tolerant architecture from five months to three weeks thanks to this open source extension of the CUDA-Q platform.
Trained on proprietary data, a modified OpenFold3 accurately predicts more than half of the protein-ligand complexes tested, compared to a third for the public version.
Google Research builds valid tool sequences before writing related queries, to produce training data more efficiently.
DeepMind provides researchers with an atlas of possible changes to a letter in human DNA, along with a score to prioritize variants.
Google describes an experimentation loop in which an agent adjusts the settings, runs the training and retains the measured improvements.
An MIT method filters compositions at an early stage to increase the proportion of stable structures proposed by generative models.