Former researcher at OpenAI launches training data startup, argues that scaling AI alone is not enough

Andrew Ho left OpenAI after eight months convinced that large language models generalize poorly and perform inconsistently even in well-covered areas such as programming. In his view, the main cause is a lack of training data for economically valuable skills, which are only marginally represented in existing datasets. Ho argues that scaling alone will not solve this problem and estimates that AI labs will have to invest over 100 billion dollars in targeted data collection in the coming years. He also says he is skeptical of the high valuations of leading labs such as OpenAI or Anthropic, which he describes as chronically unprofitable because they need to continually invest in new models to maintain their lead over cheaper competitors such as Qwen or Kimi.

The first products from his startup target two areas: datasets for complex analyses in bioinformatics, where, according to the figures cited, even current models such as GPT-5.6 Sol achieve a success rate of only around 30 percent, and datasets for routine laboratory work, such as AI models evaluating photographs of experiments. According to the company, chemistry, materials science, healthcare, and more general knowledge work are set to follow.

Adam Hunt, a researcher at University of Cambridge, shares a similarly skeptical view. He argues that the latest models are becoming more specialized rather than more versatile: they are improving in programming and complex mathematics, while language quality and basic logic are stagnating or deteriorating. According to Hunt, this is because reinforcement learning works well in areas with clear rewards and complete training data, such as code, while such data is lacking elsewhere. However, he puts his own confidence in this pessimistic thesis at only around 40 percent. In the position paper “LLMs can't jump”, Tom Zahavy from Google DeepMind offers a structural explanation: models are good at deduction and induction but fail at creative abduction, meaning the ability to come up with a cause for which no linguistic precedent yet exists. According to the source, the debate over whether models are capable of true generalization beyond their training data remains unresolved.