Debate over large language models: Are LLMs becoming more versatile - or more specialized?

Andrew Ho left OpenAI after eight months, citing poor generalization and uneven capabilities of large language models even in well-funded areas like programming. He believes the problem lies in insufficient training data, as most economically relevant skills are barely represented in existing data offerings. Ho estimates that AI labs will need to spend over 100 billion dollars on targeted data acquisition because scaling alone is not sufficient. He is also skeptical of the high valuations of frontier labs like OpenAI and Anthropic, noting they are chronically unprofitable and must invest constantly to compete with cheaper rivals such as Qwen or Kimi. His first products focus on datasets for bioinformatics and everyday laboratory work, where current models like GPT-5.6 Sol achieve only about 30 percent success rate. Adam Hunt from Cambridge supports this view, observing that the latest models become more specialized, with growth in coding and complex math while other areas stagnate. Tom Zahavy from Google DeepMind provides a structural explanation, stating that language models master deduction and induction but fail at creative abduction. He suggests action-controllable world models as a possible way out.