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
Out of 1,322 questions from 1,005 extracts of online tutoring, the best of the 11 models tested is 61.3%, compared to 91% for humans.
According to one study, self-improvement from user feedback increases the success rate of attacks from 14 to nearly 49 points according to the learning method.
This neurosymbolic approach predicts the actions to come from a short video, and corrects its predictions by editing the knowledge base, without re-training.
Variation autoencoders on diffuse background spectra show that six latent variables are sufficient, as well as in the standard cosmological model.
Tested on a FastNet configuration producing 1.5 km forecasts on the UK, a wavelet loss function gives more realistic precipitation than quadratic error.
▶
Sur les sites de documentation Mintlify, le trafic d'agents est passé d'environ 10 % à 80 % en un an ; Coinbase expose toute son activité par API et mise sur le paiement par agents.
The Shanghai start-up, which decodes movement and speech from flexible electrode implants, finances a certification trial in nearly 20 hospitals and a factory.
The voice recognition specialist offers himself the publisher of Pipecat and promises to keep the framework open and neutral vis-à-vis model suppliers.
The insurance software publisher relies on "agentic subscription": AI agents who evaluate and rate risks alongside policyholders and actuaries.
Built on DeepSeek Harness, this MIT-licensed search prototype listens to GitHub and Feishu, supports the team's requests and responds with sweaters requests.
Apple opens a beta developers on November 4th: in China, Alibaba's Qwen model will feed Siri, writing tools and visual intelligence.
Saturday, the owner of OpenAI exchanged with the executives of Samsung's electronics, finance and biotech branches, against a background of memory shortage for AI.
The Council for State Affairs sets out six areas: job creation potential, employment-friendly adoption, training, social protection and early warning.
Simulated environmental reinforcement learning doubles the clinical score of the model and reduces the underestimated suicidal risks by more than half.
On Llama, Mistral, Qwen, Gemma and Phi, the obfuscation averaged 54.3% of attacks; the model size does not predict its robustness.
Where BurnP3+ requires days or weeks per region, the network estimates in seconds the likelihood of burning, intensity and rate of spread.
28,200 simulated demonstrations on 282 drawing tasks allow to measure how these models understand the set, compose the space and generalize.
Compressed in 2 and 4 bits without retrain, the Kroko transducer processes 1.28 s of audio in 0.90 s on a chip without a neural accelerator.
Walls, doors, windows, beams and poles are reconstructed as IFC-compatible parametric parallelepipedes, which can be corrected by hand.
▶
Chez Upstart, des milliers d'appels de prêt par jour : comment l'équipe a fait passer l'accord entre juge LLM et évaluateurs humains de 33 % à 92 %.