The SimpleTES framework decomposes trial‑and‑error into three scalable dimensions—Concurrency, Length, and Candidates—enabling test‑time scaling that lets open‑source models outperform closed‑source rivals across 21 diverse scientific benchmarks, from LASSO regression to quantum circuit compilation.
In just an hour, Claude Opus 4.6 cracked a 30‑year‑old combinatorial problem posed by Donald Knuth, showcasing a leap from pattern‑recognition to symbolic logical reasoning and suggesting that AI may become a core driver of fundamental scientific discovery rather than merely a supporting tool.
The article examines the Test‑Time Training to Discover (TTT‑Discover) approach, which applies reinforcement learning during inference to let large language models continuously improve on single test problems, and reports strong results across mathematics, GPU kernel optimization, algorithm design, and biology.
The article analyses the emergence of AI scientists—systems that can generate hypotheses, design and run experiments, and write papers—examining their classifications, speed and scale advantages, interdisciplinary breakthroughs, as well as challenges such as black‑box opacity, reliability gaps, ethical concerns, and the evolving talent landscape.
Google unveiled an AI co‑scientist built on Gemini 2.0 that uses multiple specialized agents to generate, evaluate, and refine research hypotheses, demonstrating promising results in drug repurposing, liver fibrosis target discovery, and antibiotic resistance, while also highlighting current limitations and community reactions.
OpenAI’s Sam Altman and chief scientist Ilya Sutskever discuss the company’s leadership in generative AI, highlighting its focus, talent density, and innovative culture, while examining AI’s economic, security, and existential risks, the need for regulation, and its transformative potential for scientific discovery and climate solutions.
