Stanford researchers built a virtual biotech run entirely by AI agents, which combed public data on lung cancer and proposed targeting a protein called B7-H3 with an antibody-drug conjugate. A major pharmaceutical company reached the identical strategy independently and later won FDA breakthrough therapy status for it, a real-world check on whether the AI's reasoning actually held up.
Source: med.stanford.edu →MIT's new HardFlow method lets an already-trained generative AI model satisfy strict, non-negotiable rules, like a collision-free robot path, without retraining it from scratch. Across tests in robotics, physical-process control, and computer vision, it consistently met the constraints while finding better solutions than existing techniques.
Source: news.mit.edu →UC Berkeley's GPN-Star, published in Nature, learns which DNA changes are likely to cause disease by studying how genomes have shifted, or stayed put, across hundreds of millions of years of evolution in other species. It outperformed rival models built with far more parameters, while training in days on just eight processors.
Source: news.berkeley.edu →