Eric Schmidt: AI agents may be a better model for science
In an op-ed featured at the top of MIT Technology Review's "The Download" newsletter, Eric Schmidt, former CEO of Google and cofounder of Schmidt Sciences, and Suhas Mahesh, who leads the AI-for-science work at the Schmidt Sciences AI Center, argue that AlphaFold may not be the best template for accelerating science, and that AI agents could hold the key instead. They note that in 2024, Google DeepMind scientists shared the Nobel Prize in Chemistry for AlphaFold, a neural network that predicts protein structures, proving AI could make groundbreaking discoveries. But AlphaFold relied on a dataset of roughly 170,000 experimentally validated protein structures that took 53 years and about $21 billion worth of experimental work to assemble, and the authors argue comparable datasets will be difficult or impossible to build in many other fields. Their case for agents: unlike AlphaFold, which applies one powerful approach to a narrow question, agents are inherently generalists that can model the iterative, contingent process of actual research. The authors stress that agents do not represent a new way of doing science; instead, they digitally model the human process of discovery.
The newsletter's second feature previews a virtual Roundtables session on Thursday, August 13 (19:00 GMT / 2:00 pm ET / 11:00 am PT), where senior reporter Eileen Guo and executive editor Amy Nordrum will discuss what they found over nine months investigating the origins and rise of the "censorship-industrial complex" idea, a theory long circulating in right-wing online circles that a coordinated effort suppressed conservative and populist speech, which the newsletter says has now made its way into the Trump administration.
The issue's must-reads roundup, drawn from outside outlets, adds several other developments. A new Amazon data center in Texas has a permit to release up to 33 million tons of CO2 and, per the New York Times, could become the US's biggest polluter; the Verge reports the facility's gas plant will generate up to 7.65 gigawatts of power, and Cleanview notes it will be Amazon's first off-grid AI data center. Separately, the Financial Times reports OpenAI has paused work on its Astra AI model after tests found it could launch cyber-attacks autonomously, though the digest gives no further detail on what prompted the tests. Al Jazeera and Reuters report that North Korean hackers, attributed to the state-linked group Kimsuky, have built AI tools used in spear-phishing campaigns. And per Bloomberg and Reuters, China's firms now control 97% of global humanoid robot shipments, which have more than tripled since last year, with Chinese manufacturers racing toward public listings.
The newsletter's quote of the day comes from Oren Etzioni, professor emeritus at the University of Washington and former CEO of the Allen Institute for AI, who tells CNN: "It's the humans that we need to watch out for. AI is just the tool." The issue's "One More Thing" feature, by Tiffany Ng, profiles Marcin Jakubowski, founder of Open Source Ecology, who lives off solar power, heats his self-built house with a woodstove, and farms his own fish and vegetables. His project is a collection of 50 machines, from a tractor to an oven to a circuit maker, meant to be reconfigured to build civilization from scratch; Jakubowski's stated goal is a "zero marginal cost" society where an additional unit of a good or service costs little to nothing, which he hopes to reach by eliminating licensing fees, decentralizing manufacturing, and fostering collaboration through education.
Key facts
- Eric Schmidt and Suhas Mahesh argue AI agents may be a better template for accelerating science than AlphaFold, whose training dataset of roughly 170,000 protein structures took 53 years and about $21 billion to assemble.
- MIT Technology Review will hold a virtual Roundtables session on Thursday, August 13 (19:00 GMT), where Eileen Guo and Amy Nordrum discuss their nine-month investigation into the "censorship-industrial complex" theory now shaping Trump administration policy.
- A new Amazon data center in Texas has a permit to release up to 33 million tons of CO2, with a gas plant that can generate up to 7.65 gigawatts, per the New York Times and the Verge.
- OpenAI has paused work on its Astra AI model after tests found it could launch cyber-attacks autonomously, per the Financial Times.
- China's firms now control 97% of global humanoid robot shipments, which have more than tripled since last year, per Bloomberg and Reuters.
Why it matters
The lead op-ed reframes what "AI for science" should look like after AlphaFold's Nobel-winning success: rather than replicate AlphaFold's approach, which required a uniquely large, decades-in-the-making dataset, Schmidt and Mahesh argue agents can model the research process itself, making the approach potentially viable in fields where AlphaFold-scale datasets do not exist. The rest of the issue underlines how AI's footprint is widening in parallel, from a political theory the newsletter says has reached the Trump administration, to a paused frontier model over autonomous cyber-attack risk, to the environmental cost of the data centers powering AI, to China's near-total hold on humanoid robot manufacturing.
Who it affects
Researchers and funders working on AI-for-science, particularly in fields without AlphaFold-scale datasets, are the direct audience for the op-ed's argument. The censorship-industrial complex session concerns anyone following US speech policy and the Trump administration; the Amazon data center item concerns Texas residents and Amazon's environmental record; the Astra pause concerns OpenAI and the wider AI-security community; and the humanoid shipment figures concern the robotics manufacturing industry, where China's dominance is already reflected in Chinese firms racing toward public listings.
How to use it
The newsletter links out to the full op-ed for readers who want Schmidt and Mahesh's complete argument, and to free registration for the August 13 Roundtables session with Eileen Guo and Amy Nordrum. The must-reads items each link to their original, mostly paywalled sources (NYT, FT, Bloomberg, Reuters) for readers who want the full reporting behind each one-line summary.
How solid is it
The two lead items are MIT Technology Review's own work: a signed op-ed from named authors with stated affiliations (Schmidt at Schmidt Sciences, Mahesh at its AI Center), and a preview of a session by two named MIT Technology Review staff, Eileen Guo and Amy Nordrum, following their own nine-month investigation. The must-reads section, by contrast, is a curated links digest: each item is a one- or two-line summary crediting an outside outlet (NYT, FT, Verge, Cleanview, Al Jazeera, Reuters, TNW, Bloomberg) rather than original MIT Technology Review reporting, so the digest itself provides no independent verification of those outlets' claims. No byline is given for the newsletter edition as a whole.
Risks and caveats
The op-ed argues in general terms and does not name specific scientific fields or discoveries where AI agents have already accelerated research. The digest does not explain what the censorship-industrial complex theory actually claims, only that MIT Technology Review traced its origins and rise; nor does it detail what specifically triggered OpenAI's tests of the Astra model beyond the finding that it could launch autonomous cyber-attacks. Because the must-reads items are one-line summaries of other outlets' paywalled reporting, their full context and methodology sit behind those outlets' own coverage.
“It's the humans that we need to watch out for. AI is just the tool.”
— Oren Etzioni, professor emeritus at the University of Washington and former CEO of the Allen Institute for AI, quoted by CNN