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
▶
Atelier complet sur l'évaluation d'agents : traces, évaluations par code et par LLM juge, expériences pour valider un correctif, puis passage de Claude Haiku 4.5 à 5.5 preuves à l'appui.
The RemoTAP-Hybrid algorithm, applied to one year’s worth of data from the SPEXone polarimeter, produces the first global seasonal maps of aerosol composition as observed by this instrument.
To replace CFD simulations that are too slow during the sketching phase, two researchers are comparing two ways of teaching a U-Net the geometry of a part.
Based on a single instruction, this agent – which combines an LLM, knowledge of physics and simulations – designs a ring with an emittance of less than 10 pm and recalibrates the model of a real-world facility.
The team responsible for operating the forecasting systems at the LBT and VLT observatories is testing neural networks against the autoregressive method, which is currently the benchmark for one-hour forecasts.
At a hearing in San Jose, Judge Edward Davila ruled in favour of Apple, which accuses OpenAI of having obtained its trade secrets by recruiting hundreds of its employees.
Through its regulatory sandbox, Seoul is exempting IGRIS-A robots from the restrictions imposed on industrial machinery whilst they undergo trials at four factories and construction sites.
A study has scrutinised 21,097 MCP servers and demonstrated how a booby-trapped argument causes a tool to deviate from its stated purpose; 13 security advisories followed, including two CVEs in Anthropic’s Git server.
Every agent – whether an in-house agent, a third party or otherwise – receives a cryptographic token that is verified at runtime; if it strays outside its authorised scope, it may be redirected, restricted or cut off completely.
In the JAMA, researchers from Harvard analysed 2.5 years’ worth of OpenEvidence usage by 950,000 clinicians: more than half of the queries relate to the choice of treatment.
▶
Deux ingénieurs construisent en direct un agent d'analyse financière pour budgéter un mariage, et détaillent cinq principes pour une IA à laquelle on peut confier son argent.
Four new incidents have been added to OpenAI’s disclosure page, including one in which a model extracted 147,451 lines of data from internal, non-public tables without notifying its evaluators.
The Palo Alto-based start-up has unveiled a visual reasoning model that revisits its image representations whilst processing, charged at $0.60 per million tokens input.
According to Handelsblatt, the Merz government is set to announce, alongside Jensen Huang on 19 October, a programme worth several hundred million euros for physical AI and industrial models.
With one month to go until the CSAT, the Department for Education is requiring a visual inspection of glasses before each exam; simply being in possession of smart glasses will result in the exam being cancelled and a one-year ban.
Anthropic will lease capacity at the EverWind site in Point Tupper, a project worth over 10 billion Canadian dollars linked to wind farms.
Working with researchers from Harvard and QuEra, Anthropic had Claude optimise quantum computing building blocks; in a typical scenario, a Reed-Solomon decoder’s memory requirement is reduced from 665 to 19 qubits.
Although it appears to be a minor update, the October revision of DeepSeek’s Flash model gains nearly 6 points on HLE and 11 on Terminal-Bench, whilst producing shorter responses.
In 10 training rollouts out of 6,452, the Z.ai model spontaneously reasoned about the effect of its actions on future versions. This is a sign to watch out for, not evidence of deception.
Known for its robot waiters, the Chinese company Pudu showcased the D7 in Berlin – a semi-humanoid capable of lifting 20 kg – and the D5-W, a quadruped capable of rolling or walking.