JetBrains researchers turned the decision-making logic of a Maven-style build tool into a fixed weight matrix and used it to build an actual Java library. The result was identical, byte for byte, to what Apache Maven produces. A clean build with tests ran 1.49 times slower than the same logic implemented in Python.
In experiments, a few AI-edited photos and a fake AI summary of an accident video led hundreds of volunteers to remember invented details. In one test, participants answered wrongly even though they knew the text had been machine-generated.
Clubic examines how chatbot suggestions might steer users away from some ideas even as they help find others. As these tools enter daily life, the diversity of how people think, create and see the world could be affected.
Microsoft has proposed an AI method that corrects the errors of existing weather models for forecasts several weeks out, without heavy computing requirements. Results are encouraging, though the approach still cannot predict the weather with certainty.
Taste-Bench, a new benchmark, measures how well AI agents make strategic choices over long-running tasks. Even the top performer, GPT-5.6 Sol, picks the right path on just under 60% of simple binary questions, and letting models think longer does not help.
Google Research released Harness2 under the Apache 2.0 license, a framework that rewrites an agent's harness at test time. A single refinement step raised GLM-5's score on WorkBuddy-Bench by 11.5 points without touching the model's weights.