A former OpenAI employee who left the lab after just eight months is betting that the future of artificial intelligence lies not in building bigger models, but in collecting better data. Andrew Ho, who worked on bioinformatics at OpenAI, has founded a startup to produce high-quality training datasets, arguing that the industry's current scaling approach is failing to deliver genuine generalization.
Ho predicts that AI labs will spend more than $100 billion on targeted data collection in the years ahead. He is skeptical that simply scaling large language models (LLMs) will solve their fundamental weaknesses, even in well-funded areas like programming. "Most work is highly contextual and not easily encoded into a gradable environment; even if we can observe a 'golden path' taken by a human which we believe to be good, it's challenging to understand whether alternate, counterfactual paths produce good or bad outcomes," Ho said.
His skepticism extends to the business models of frontier labs. He described OpenAI and Anthropic as chronically unprofitable because they must keep pouring growing sums into new models to stay ahead of cheaper rivals like Qwen or Kimi. Ho's first products target datasets for complex scientific analyses in bioinformatics and datasets for everyday lab work, such as researchers submitting photos of experiments for AI evaluation. Current models like OpenAI's GPT-5.6 Sol hit only about a 30% success rate on bioinformatics tasks. Planned future areas include chemistry, materials science, healthcare, and broader knowledge work.
The Specialization Ceiling
Ho's views are shared by Adam Hunt, a researcher at the University of Cambridge, who described his own perspective on LLMs shifting from optimistic to increasingly pessimistic. Hunt argues that the latest models are becoming more specialized, not more versatile. While programming and complex math capabilities keep improving, language quality and simple logic are stagnating or getting worse.
Hunt said reinforcement learning works well in code because clear reward signals and complete training data exist. In other areas, that data isn't available. He believes the early progress of LLMs and their apparent ability to generalize through scale and reasoning was a side effect of training on a broad text corpus, not a sign of true general understanding. Hunt put his confidence in his own view at only about 40% and acknowledged that technical advances could prove him wrong, for instance by combining specialized models into something resembling general intelligence.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
A Structural Explanation for Gaps
Tom Zahavy, a researcher at Google DeepMind, authored a position paper titled "LLMs can't jump" that offers a structural explanation for the uneven performance of these models. Zahavy said LLMs are good at deduction and induction but fail at creative abduction, the ability to invent a cause with no linguistic precedent.
Zahavy suggests action-controllable world models that allow counterfactual experiments as a possible fix, noting that LLMs could still play a key role in such advanced systems. Researchers at Cambridge and Google DeepMind back the perspective that current AI systems trend toward greater specialization rather than versatility and have hit a ceiling on creative problem-solving.
The Broader Debate
The debate centers on whether LLMs can develop capabilities beyond reproducing or recombining training data, especially in areas without automatic verification. Scientists do not agree on the generalization question. Two paths of AI development are described in the discussion: gradual broad improvement versus extreme specialization, where a few capabilities spike while core skills stagnate or shrink.
Ho's departure from OpenAI and his new venture represent a bet that the industry's current trajectory is unsustainable. By focusing on targeted data collection for economically valuable but poorly represented skills, he aims to address what he sees as the root cause of LLM failures. His prediction of over $100 billion in spending on such data collection signals that he believes the market will eventually recognize the same limitations he does.