[Even the smartest large models need practice: Former OpenAI researcher starts a data-focused venture] According to monitoring by Beating, former OpenAI researcher and co-author of GeneBench-Pro, Andrew Ho, has left the company to start his own venture. The new company will produce high-quality reinforcement learning data for large models (training tasks that allow models to practice and receive scores based on results). The first batch of products will focus on biology and statistical reasoning, though the company name has not yet been disclosed. Ho believes that large models still have significant weaknesses. Even in heavily invested fields like programming, while models can solve problems and revise code, real-world delivery often requires manual finishing touches. For many practical tasks, there aren’t even suitable training problems available. These tasks depend on specific contexts and are difficult to evaluate automatically for correctness. His approach involves generating data and problems that closely resemble real scientific research, while retaining known answers. This allows models to explore freely, and the training system can clearly determine right from wrong. GeneBench-Pro employs this method, featuring 129 computational biology tasks, where GPT-5.6 Sol Pro achieved only a 31.5% success rate. The new company also plans to create training data for experimental images such as petri dishes and Western blots, with future expansions into chemistry, materials science, healthcare, and office tasks. Ho predicts that cutting-edge AI labs will invest over $100 billion in precise training data in the future. [Original article link]