Editorial illustration for The Download: AI agents for science and the "censorship-industrial complex
AI Agents for Science: Beyond AlphaFold's $21B Model
The Download: AI agents for science and the "censorship-industrial complex
AlphaFold won its team a Nobel Prize in 2024, but building it took 53 years and roughly $21 billion in experimental data collection. That's not a repeatable formula for most fields of science. Today's newsletter looks at a different bet: AI agents that model the actual process of research, rather than tools trained on one enormous, expensive dataset.
Eric Schmidt, the former Google CEO who now runs Schmidt Sciences, and Suhas Mahesh, who heads the group's AI for science work, make the case in an op-ed for why agents, not narrow prediction models, are the better path to speeding up discovery. Their argument turns on a distinction between solving a single well-defined question and replicating the messy, iterative work scientists do every day.
Elsewhere in today's edition: a look at how the phrase "censorship-industrial complex" moved from a niche grievance into a framework now shaping US policy, plus a report on an Amazon data center that could end up as the most polluting power plant in the country. First, the science story, starting with the limits of the AlphaFold model itself.
In 2024, Google DeepMind scientists shared the Nobel Prize in Chemistry for a neural network, AlphaFold, which predicts the structures of proteins. It showed that AI could make groundbreaking scientific discoveries, but AlphaFold may not be the best template for accelerating science. Instead, another approach may hold the key: AI agents.
Why this matters
Schmidt and Mahesh are making a specific claim worth testing rather than cheering: throwing more data at scientific problems hits diminishing returns without models that can actually reason through hypotheses the way a working scientist does. For researchers building AI agents for lab work, that's a useful check on the "bigger dataset, better science" assumption that's driven a lot of funding decisions. It also raises the bar for what counts as progress. An agent that can search literature fast isn't the same as one that can weigh conflicting evidence and propose a testable next step.
The "censorship-industrial complex" session, meanwhile, is a reminder that platform moderation debates keep reshaping what counts as acceptable AI-assisted speech online, and that's directly relevant to anyone deploying models with content filters or trust-and-safety layers. Worth watching how Guo and Nordrum's reporting on that theory's origins and spread lands on August 13, especially for teams thinking about how political narratives around moderation might affect regulatory pressure on AI products going forward.
Related Reading
Business & Startups Nordic pilot adds Gemini for Education, NotebookLM to boost AI literacy
Business & Startups Kling launches Video O1, all-in-one model with MVL bridge using transformer
Business & Startups DeepSeek Seeks More Capital Weeks After USD 7B Funding Round
From Archives Discovered Materials Runs AI Agents 24/7 to Hunt for Cooler Chips
From Archives OpenAI Buys NextSlide to Add AI Presentations to ChatGPT
Common Questions Answered
Why is AlphaFold not considered the best template for accelerating science despite winning the Nobel Prize?
AlphaFold required 53 years and approximately $21 billion in experimental data collection to develop, which is not a repeatable formula for most fields of science. The article suggests that AI agents modeling the actual research process may be a more practical approach than tools trained on one enormous, expensive dataset.
What is the key difference between AI agents and traditional AI tools like AlphaFold for scientific research?
AI agents are designed to model the actual process of research and reason through hypotheses the way working scientists do, rather than simply being trained on large datasets. This approach addresses the diminishing returns that occur when throwing more data at scientific problems without models that can actually think through the research process.
What claim are Eric Schmidt and Suhas Mahesh making about the 'bigger dataset, better science' assumption?
Schmidt and Mahesh argue that throwing more data at scientific problems hits diminishing returns without models that can actually reason through hypotheses like working scientists. Their claim challenges the funding decisions that have been driven by the assumption that larger datasets automatically lead to better scientific progress.
How did AlphaFold demonstrate AI's capability in scientific discovery?
AlphaFold, a neural network developed by Google DeepMind scientists, predicted protein structures and made groundbreaking scientific discoveries, earning its team the Nobel Prize in Chemistry in 2024. This achievement showed that AI could make significant contributions to scientific research, though the article suggests alternative approaches may be more scalable.
Further Reading
- From Models to Scientists: Building AI Agents for Scientific Discovery - Harvard Kempner Institute
- How AI agents will change research: a scientist's guide - Nature
- Accelerating scientific breakthroughs with an AI co-scientist - Google Research Blog
- AI Agents may be skilled researchers—but not always honest ones - Science
- Agentic AI for Scientific Discovery: A Survey of Progress, Challenges, and Opportunities - arXiv
