Eric Schmidt, the former CEO of Google and founder of Schmidt Sciences, contends that the next wave of AI-driven scientific discovery depends less on massive datasets and more on AI systems that can reason through complex problems iteratively. This perspective challenges the dominant model of recent AI breakthroughs and points toward a fundamentally different approach to accelerating research across multiple fields.

Why Did AlphaFold's Success Create a False Template?

In 2024, Google DeepMind scientists won the Nobel Prize in Chemistry for AlphaFold, a neural network that predicts protein structures with remarkable accuracy. The achievement seemed to validate a clear formula for AI-driven science: feed the system enormous amounts of validated data, and it will solve hard problems. But Schmidt and Suhas Mahesh, who leads AI for science work at the AI Center of Schmidt Sciences, argue this model may not be replicable across most scientific domains.

AlphaFold's success rested on a specific advantage: researchers had assembled roughly 170,000 experimentally validated protein structures over 53 years, representing approximately $21 billion in experimental work. That dataset was the foundation of the breakthrough. But comparable datasets will be difficult or impossible to create in many fields, from climate science to materials discovery to drug development. The protein-folding problem was, in some sense, an outlier.

What Are AI Agents, and How Do They Differ From AlphaFold?

Rather than relying on massive pre-existing datasets, Schmidt and Mahesh propose a different approach: AI agents that model the iterative, contingent process of actual scientific research. Unlike AlphaFold, which applies a powerful technique to a single, well-defined question, agents are generalists. They don't represent a new way to do science; instead, they digitally simulate how human researchers actually discover things.

This distinction matters because real science is messy. A researcher forms a hypothesis, runs an experiment, observes unexpected results, adjusts the hypothesis, and tries again. Each step depends on the previous one in ways that are hard to predict. An AI agent can model this contingent, iterative process rather than trying to solve a static problem with a fixed dataset.

How Is Schmidt Sciences Supporting This Research Direction?

Schmidt Sciences, funded by Eric and Wendy Schmidt, operates the AI2050 program, an initiative that supports academics whose work involves AI. The program convenes some of the most accomplished and promising AI researchers in the world, and fellows receive funding they can use to purchase GPUs (graphics processing units), the specialized computing hardware required to train and run advanced AI models.

This support addresses a critical bottleneck in academic AI research. Universities simply cannot afford the computing power required to train frontier AI models, and companies like Anthropic and OpenAI keep the inner workings of Claude and ChatGPT proprietary. Academic researchers can study how these models behave from the outside, but they cannot conduct detailed research on their design and training, nor can they steer that development themselves.

What Challenges Do Academic AI Researchers Face Today?

The landscape of university-based AI research has shifted dramatically in the past four years. The cutting edge of AI development has moved from academic institutions to private companies, leaving university researchers in a precarious position. Several key obstacles now shape their work:

• Computing Power Scarcity: Universities lack the GPUs and infrastructure needed to train large language models, forcing researchers to either work with smaller models or rely on expensive API calls to commercial services.
• Proprietary Model Access: Anthropic and OpenAI do not allow external researchers to examine the design and training details of Claude or ChatGPT, limiting the scope of academic inquiry.
• Funding Constraints: Federal scientific funding in the United States has declined, making it harder for researchers to purchase computing resources or conduct large-scale experiments.
• Research Relevance Gaps: Many academics deliberately avoid working on problems they believe tech companies will solve, since companies prioritize profitable research over questions that might reveal unflattering findings.

Nika Haghtalab, a computer science professor at UC Berkeley, captured the frustration in a striking analogy. She said that being an AI academic today was like being a biologist in a world in which private companies had exclusive control over CRISPR, the gene-editing tool. Researchers can observe the outcomes, but they cannot access the mechanism.

Are Non-LLM AI Researchers Facing Different Pressures?

A large group of AI academics work with specialized models rather than large language models (LLMs). These researchers build AI systems that analyze data, make predictions, or simulate physical systems. They are not directly competing with frontier labs like Google DeepMind, yet they face their own challenges.At the AI2050 convening, several researchers voiced concerns about how widespread ignorance of non-LLM AI was affecting their work. Those building specialized AI tools to address climate change, for example, sometimes struggle to advocate for their research when so many people believe that "AI" means energy-intensive large language models. This perception gap can make it harder to secure funding and institutional support for work that may be equally important but less visible.

The situation has prompted some prominent academics to take leave from universities to join frontier labs, and many AI2050 fellows now hold industry positions alongside their academic roles. Additionally, recent breakthroughs in AI-driven mathematics have sparked concerns among some researchers about the future of human mathematicians, with one fellow expressing worry about the mental health of her mathematician peers.

Is There Hope for Academic AI Research?

Despite these headwinds, several factors suggest that academic researchers may find new pathways forward. Empirical science, which requires collecting real-world data, may prove much harder to automate than pure mathematics, since data collection is an intrinsically slow process. Some researchers view AI mathematicians and scientists not as threats but as tools that could make human scientists far more efficient, freeing them to pursue ambitious ideas they might otherwise never explore.

The resource constraints that prevent academics from training frontier models also push them to discover new ways to make models smaller, faster, and more efficient, or to explore entirely new architectures. If the next major AI breakthrough comes not from a major company but from a scrappy academic lab, it would not be surprising.

Schmidt's emphasis on reasoning-based AI agents for science reflects this broader shift. By funding researchers who are exploring alternatives to the data-hungry AlphaFold model, Schmidt Sciences is betting that the future of AI-driven discovery lies not in scaling up existing approaches, but in rethinking how AI can model the creative, iterative process of human research itself.

" }