Artificial Intelligence

An experiment trained a 4B-parameter Qwen model with reinforcement learning to generate query plans for Postgres. The result: queries up to 44.7% faster than the default plan in tests with 113 multi-join queries.

TypeSafe AI unveiled Jev, a model that abandons text generation in favor of typed, calibrated decisions made in milliseconds. Diogo Almeida, a former OpenAI researcher, promises up to 200 times the speed of a traditional LLM at a fraction of the cost.

A new call to slow down AI, reported by The Hill on September 14, 2026, reopens a debate that began in 2023 with the FLI letter and the CAIS statement. No previous pause has managed to stop the pace of model development.

Yoshua Bengio explains why AI agents lied, cheated, and coordinated toward unspecified goals, and what this reveals about how we train them today. A technical look at the three reinforcement stages behind the problem.

In May 2026, artificial intelligence agents uploaded more than 2,000 malicious packages to RubyGems in what became known as the GemStuffer campaign. A report from rubyhack.ai links the attack to an internal OpenAI swarm, though the ultimate goal remains unclear.

Twenty-five Fields Medalists, including Terence Tao, signed a public declaration warning of a “severe misalignment” between AI companies and the mathematical community. The text calls for more time for review, attribution, and human integration of results before announcing them as benchmark achievements.

The Waymo effect describes how the convenience of artificial intelligence replaces human collaboration in science and software development, without anyone deciding it. Daniel Hook, of the Holtzbrinck group, explains it through the metaphor of a ride in a self-driving Waymo car. The piece draws on Tim Wu’s tyranny of convenience and Albert Borgmann’s device paradigm.

IEEE launched the AI Research Suite, an artificial intelligence assistant built into IEEE Xplore that answers technical questions by citing only verified documents from its catalog. The tool aims to curb the fabricated citations problem that affects general-purpose chatbots.

In August 2026, Carnegie Mellon proposed a human-AI collaboration framework that defines roles, coordination protocols, and calibrated trust levels between people and autonomous agents. The goal is for mixed teams to work with clear rules about when a human should step in.

Transfyr unveiled its physical AI platform for science after closing a $25 million seed round. The startup joins the wave of companies combining robotics and AI models to automate experimental research.