A computer can propose a protein that has never existed. To make that proposal matter, the protein has to be made and tested. The result may justify a better experiment, reveal an error, or send the researchers back to their assumptions.
That movement between a promising idea and an encounter with reality is the subject of Richard Socher’s The Eureka Machine. He asks what would happen if AI could participate throughout the scientific process: connecting findings, forming hypotheses, designing experiments, interpreting results, and using what it learns to improve the next attempt.
The book is most convincing when it explains how those activities could work together. Its evidence supports substantial progress in particular scientific tasks. Its larger vision, a system capable of sustained discovery across disciplines and of improving its own intelligence, remains a research program. The distance between those two claims is where much of the interesting work in this book happens.
Disclosure: Richard is my friend, and I have genuine respect for his work. We are also connected through the Young Global Leaders alumni community. I am a You.com shareholder, previously served as a formal adviser, and continue to advise the company informally. I received an advance digital copy free for this review.
A builder’s account of science
Socher writes from inside the development of AI. The early chapters follow the movement from systems built around explicit rules to neural networks that learn representations from data. He explains how inputs become numbers, how training changes a network, and why a method that works with sentences might also work with other kinds of sequences. His own research in language gives that history a personal thread.
The explanations are accessible without being empty. An analogy for a neural network does some useful work, then he points out where the analogy breaks down. That is good explanatory writing: give the reader a way into an unfamiliar mechanism without asking them to mistake the illustration for the mechanism itself.
His discussion of oxygen captures the transition to science particularly well. A language model can retrieve a great deal of writing about oxygen. But scientific knowledge of it also lives in equations, spectra, and physical measurements. A model trained on words has not thereby absorbed all those other forms of evidence. The book’s ambition depends on learning across them. Scientific AI, in this account, includes models of molecules and measurements as well as models of language. (p. 48)
I have written about Richard’s work before, particularly the relationship between greater AI capability and greater engineering responsibility. This book develops a larger version of that question. An impressive model is one component of a scientific system. The surrounding decisions about data, experiments, evaluation, and use determine what its capability can become.
There is a useful continuity between the history and the argument. Socher treats learning representations as a way to move beyond hand-written descriptions of every relevant feature. He then asks whether science can benefit from representations that connect what its separate fields already know. The opening history earns its place by preparing that question.
The book also reflects its author’s position as a researcher, entrepreneur, and investor. Several biomedical examples come from companies in which he discloses an investment. Near the end, he describes building toward his vision through Recursive, a venture also identified in the publisher’s public author biography. These connections help explain the book’s access and practical interests. They also make it useful to keep track of the kind of evidence each example offers: a published experiment, a company’s account of its work, or a founder’s expectation of what comes next.
The problem is larger than too many papers
“Collectively, we know far more than we understand,” Socher writes. (p. 74)
His diagnosis goes beyond the difficulty of keeping up with the literature. Scientific work becomes more specialized as knowledge accumulates. A useful result in one field may sit far from a problem in another. Researchers also work within institutions that determine what receives time, money, attention, and a route to publication.
The chapter on slowing scientific progress brings these problems together. Its broad premise deserves care. Measures based on citation patterns can tell us something about how later work relates to earlier work; they cannot, by themselves, settle how much valuable understanding a period has produced. The influential study of declining disruptiveness found a falling share of disruptive work alongside a comparatively stable absolute number of highly disruptive papers and patents. A methodological critique challenges how the measure behaves as citation practices change.
Socher himself makes a more interesting qualification. Perhaps progress increasingly consists of connecting and integrating knowledge about complex systems, rather than producing the kind of isolated breakthrough associated with earlier scientific eras. He gives synthesis and consolidation a place in scientific progress. (p. 58) That possibility strengthens his case for tools that connect knowledge even if one remains unconvinced by a single, sweeping account of scientific decline.
One of the book’s most distinctive examples concerns AI that models the distribution of human attention. Work by Jamshid Sourati and James Evans incorporates information about who studies which subjects. It can be directed toward promising connections that sit outside the paths researchers are most likely to follow. The research describes these as “alien” hypotheses: possibilities intended to complement the current scientific community’s search, rather than simply anticipate its next publication.
The appeal is clear. A useful scientific partner need not resemble the researcher it assists. It might contribute by noticing a connection the researcher’s training makes less likely to occur to them. The important qualification is equally clear: a promising connection is a reason to investigate. It has not become a discovery merely by being unusual. Socher explicitly identifies this work as primarily proof of concept. (pp. 68–71)
I find that account of complementarity more productive than a contest over whether a machine is creative in precisely the way a person is creative. A research team needs ideas worth testing and the means to test them. Different ways of searching can be valuable if they improve that process.
There is still an institutional problem to solve. A model may have no personal fear of a rejected paper, but the organization using it still chooses the questions, datasets, and experiments. Broader search creates an opportunity to pursue neglected work. Whether that opportunity is used depends on people and institutions as well as the model.
What the Eureka Machine would contain
The book’s central proposal becomes explicit in its eighth chapter. Socher describes four connected parts. (pp. 184–195)
First, a foundation of accumulated knowledge would connect scientific literature with code, images, measurements, and other records. Second, a model grounded in physical data would learn the regularities that constrain what can happen. Third, simulations would let the system explore possible interventions before attempting them in the world. Fourth, laboratories would make and measure things, returning their results to improve the models and the next round of experiments.
AI agents would coordinate the work. Here, an agent means software that can pursue a goal through a sequence of actions, using models and tools to decide what to do next. Some agents might search the literature, others propose experiments, and others examine results. Their usefulness would depend on how well they communicate and how reliably the whole process handles disagreement and error.
Socher illustrates the arrangement with a hypothetical search for a better battery material. The system connects relevant research, considers whether candidate compounds are physically plausible, tests their predicted behavior in simulations, and then synthesizes and measures promising candidates in a laboratory. Differences between prediction and measurement feed back into the earlier stages.
This is a helpful example because it makes each part responsible for something. A literature search supplies possibilities. A simulation narrows them. A physical experiment can reveal that the simulation was wrong. The system has to retain that disagreement and learn from it.
“There is no substitute for reality,” Socher writes when introducing the laboratory. (p. 188) That sentence is the most important constraint on the book’s optimism. He does not propose that an AI can establish a scientific result by producing a persuasive explanation and then agreeing with itself.
In a generalized account of a debugging problem in my own work, I described an export tool mangling nested data. My first instinct was to add more logging. Running the same input through an independent parser identified the fault in my code. The scientific demands here are much greater, but I recognize the value of a check that can disagree with the system being checked. Socher gives that role to physical experiments, including experiments that expose weaknesses in a simulation.
He is also explicit that the steps in his scenario exist in use or development without yet forming the complete combined system. That qualification matters.
A collection of promising components establishes reasons to build. It does not establish that integration will be straightforward, that errors will cancel, or that improvement in one component will improve the whole.
Proteins put the argument in contact with matter
The biological chapters provide the clearest examples of why this program deserves attention. Socher moves from identifying drug targets to predicting structures, generating proteins, testing responses, and modeling cells. These are different scientific achievements with different kinds of evidence.
AlphaFold is an example of prediction becoming a widely available research resource. The 2021 human-proteome study covered 98.5 percent of human proteins with predictions. Its authors also reported confident predictions for 58 percent of residues, the individual amino-acid positions within proteins. Coverage and confidence answer different questions. The contribution is substantial without assuming that every region of every protein is equally understood, or that knowing a structure establishes how a treatment will behave.
ProGen, work in which Socher participated, gives the book a particularly concrete passage from generation to experiment. A model trained on protein sequences proposed new sequences. Researchers selected candidates and tested physical proteins for enzyme activity. In the published study, 66 of 90 assayed artificial proteins were functional in a lysozyme assay. Those 90 came from a selected set of 100, following the generation of a million candidate sequences. The often-cited 73 percent is a result for that tested set, not a success rate for every generated sequence.
The selection and experimental steps make the example more interesting. Researchers used the model to explore possibilities, then asked a physical system whether selected proposals did the relevant work. A protein’s activity in an assay is still a different endpoint from a safe, effective medicine. It is nevertheless a consequential result: the proposal survived a test beyond the computer that proposed it. (pp. 87–89)
The same care is needed when the book turns to an AI co-scientist that proposes biomedical hypotheses. The published Co-Scientist research includes drug-repurposing candidates tested in laboratory cell models. Its authors expressly distinguish such results from clinical success. The scientific value lies in generating testable possibilities and obtaining early experimental evidence. A patient benefit requires further kinds of evidence.
These distinctions are central to evaluating the promise of AI in medicine. A model may improve the search for a target, the design of a molecule, or the choice of an experiment. Each improvement can be useful without solving every subsequent stage. The book gives readers a view of several places where that work is happening, which is more informative than treating drug discovery as one task that AI either has or has not solved.
Socher also attends to the physical skills that scientific papers do not fully record. Knowing a proposed material’s properties and successfully making it are separate challenges. His discussion of laboratory practice acknowledges how much practical knowledge remains difficult to digitize. (pp. 72–73) That is a valuable qualification for readers whose main encounter with AI is software: the world contains experiments that cannot be reduced to another fast call to a model.
A virtual cell must preserve the biology that matters
The proposed virtual cell takes the argument to a harder level. Knowing many components of a cell does not mean knowing how they behave together. Socher describes an ambition to build models that can predict responses to interventions, help explain those responses, and generate further hypotheses. (pp. 101–120)
The idea depends on abstraction. A useful simulation need not reproduce every atom. It needs to preserve the relationships that matter for the question being asked. That makes the project conceivable, but it also creates the difficult research problem: deciding which simplifications retain the behavior one needs to predict.
The book’s account of integration is unusually useful. Models of gene regulation, metabolism, and protein interaction can use different mathematical representations. As each becomes more sophisticated, combining them into a consistent model may become harder. Socher treats this as a central challenge, and reports a wide range of expectations about how long a sufficiently capable virtual cell might take. (pp. 109–110)
An illustration makes the stakes easier to see. Imagine a laboratory considering two interventions for a particular kind of cell. A model that reliably eliminates one unpromising option could save an experiment and help the researchers choose the next one. That is valuable even if the model cannot predict the behavior of a whole organ. Conversely, a model that reproduces familiar measurements but fails when the intervention changes may be less useful than a less elaborate model that predicts that change correctly. This is a hypothetical example of the evaluation problem, not a reported result.
Socher acknowledges the difficulty of predicting innovation, then forecasts a tenfold increase in the pace of biological research and drug discovery over the next decade. (p. 113) That is a substantial prediction. A tenfold increase in promising hypotheses, completed experiments, or useful treatments would describe very different achievements. I want a clearer account of how to measure the acceleration and how faster experimentation would translate into validated benefits.
Elsewhere, the book’s ambitions extend from cells to personalized treatment, organs, and ultimately living systems at larger scales. I read those more expansive passages as possible destinations. The engineering argument for useful partial models is more firmly developed than a timetable for a comprehensive digital patient.
A further distinction deserves attention: validating a model for one use does not validate it for every use. The evidence needed to suggest an inexpensive laboratory experiment differs from the evidence needed to guide an intervention in a person. As the consequences change, the standard of validation must change with them. More detailed simulation does not remove that responsibility.
Different sciences need different tests
The chapters on economics, neuroscience, and astronomy broaden the book in a useful way. They show why scientific AI cannot be evaluated by a single kind of score.
In economics, Socher describes the AI Economist, a system in which simulated workers adapt to tax rules while a simulated policymaker learns to adjust those rules. The interplay matters: policies change behavior, and changed behavior changes the consequences of policy. The research demonstrates results inside simulated economies, under specified objectives. It does not establish a tax policy’s benefit to an actual country.
The example makes computational exploration understandable. Researchers can examine how behavior changes under different rules. The challenge in moving outward is whether the modeled behavior remains a credible guide when people face circumstances unlike those in the simulation. Adding more simulated people does not settle that question.
Socher gives people and their representatives responsibility for choosing social objectives. (p. 132) That boundary deserves to be carried into the interpretation of every result. A system can help explore the consequences of a welfare objective; its ability to optimize that objective does not decide whose interests should count or how conflicting interests should be weighed.
The neuroscience chapter is strongest when it questions its own metaphors. People have understood brains through the technologies familiar to them: hydraulics, machinery, communications networks, computers. A metaphor may suggest an experiment, but familiarity does not make it a theory. Socher also doubts that closely copying biological brains is necessarily the best route to more capable artificial intelligence. That gives the chapter a useful tension between learning from biology and treating biology as an engineering specification. (pp. 141–158)
BrainGPT provides another case where the task definition matters. The underlying study tests whether models and experts can choose between an authentic neuroscience abstract and a version with altered results. Strong performance suggests that patterns in the literature can help discriminate between possible findings. It does not demonstrate the complete competence required to conceive, conduct, and interpret a neuroscience research program. The result is worth understanding at its actual scale.
Astronomy introduces yet another relationship between data and inference. The book describes PRIMO, which used a library of simulated black-hole images to help reconstruct an image from sparse observations. The primary research explains how learned structure informs aspects not fully specified by the measurements. The result is a reconstruction supported by data and assumptions. Its usefulness depends in part on understanding the contribution of each. (pp. 171–172)
Later in the chapter, Socher imagines an AI trained within an older, Earth-centered understanding of the cosmos. Could it become very good at prediction while remaining trapped in a mistaken organizing theory? The thought experiment brings the book to a question more demanding than accuracy on familiar examples: can the system help discover which assumptions need to change? (pp. 175–177)
These chapters earn their breadth because the differences matter. A protein can be synthesized. An economic policy changes a society whose members respond. An astronomical observation may constrain several possible explanations. Scientific usefulness depends on how a model meets the relevant test, not simply on how impressive its output looks.
The work between one discovery and the next
Socher’s proposed machine becomes more valuable when the outcome of one experiment changes the next. Failed experiments matter here. He wants their information to improve the knowledge base and the models, rather than disappear from the record available to subsequent research. (p. 191)
That connects with a problem I have described in my own work with AI tools. Changing the tool doing the work can strand the decisions, methods, and accumulated knowledge that gave the work its direction. In writing about the synthesis ecosystem, I described why I moved the durable center of a project outside any single agent’s working environment. The next participant needs a record it can inspect and verify.
An earlier version of my synthesis project-management approach kept a project’s context in one file. That worked for short projects and degraded as the record grew. Separating current state, stable knowledge, and chronological history came from observing that problem in use. The connection to the book is continuity: a research process needs to retain what was tried, under which conditions, what happened, and why the next decision changed. A confident summary that loses those distinctions can make the next step less reliable.
Scientific research imposes additional demands that a software workflow does not resolve. Samples, instruments, measurement uncertainty, and reproducibility belong in its record. But the organizational lesson carries across: improving the next answer requires preserving more than the previous answer.
The broader idea I call synthesis engineering is deliberate human–AI collaboration with explicit responsibility for the resulting system. Socher’s account gives that principle a demanding scientific setting. A capable model, a capable laboratory, and capable researchers still need an arrangement in which evidence moves accurately between them.
The self-improvement examples toward the end sharpen this point. The Darwin Gödel Machine research improves coding agents’ tools and workflows while keeping their pretrained foundation models frozen. AlphaEvolve evolves candidate programs using executable, user-supplied evaluation procedures. These are meaningful forms of improvement, with identifiable mechanisms and tests. They also show why the phrase self-improvement needs a specific object: an agent’s workflow, an algorithm, a trained model, or an entire scientific process.
Socher uses such developments to motivate a stronger ambition: a system that improves the intelligence with which it conducts further research. The examples give reasons to explore that ambition. They do not yet establish its open-ended form or its pace.
Who chooses the purpose, and who gets the benefit?
The book’s aspirations are human ones: better health, longer lives, more knowledge, and wider access to the benefits of discovery. Those aims deserve to be evaluated alongside the technical design.
Socher advocates open data, models, and code, and recognizes that commercial success can put openness under pressure. He asks whether cell-modeling efforts will remain open once they generate valuable discoveries. (pp. 111–112, 199–200) That concern belongs in the review because it prevents an easy assumption that greater capability will automatically mean wider access.
Open tools can help. Access also depends on the resources needed to use them and the institutions that decide which problems receive attention. In writing about who receives the gains from AI, I have distinguished creating more value from distributing its benefits. That is the question I bring to the book’s promise of abundance. The ability to generate more candidate treatments does not itself determine which diseases are studied, which results are shared, or which patients can obtain a successful treatment.
This is where my review of Marc Benioff’s Trailblazer provides a useful companion question. That book asks how business can serve a public purpose. The Eureka Machine asks what a new scientific capability might make possible. Read together, they direct attention to the institutions that turn capability into benefit. A powerful means and a worthwhile end still need to be joined by actual decisions.
In software, I describe human architectural authority as responsibility for consequential choices about the system being built. Scientific AI raises a related issue. Delegating more execution does not settle who is accountable for the research objective, the acceptable risk, or the use of the result.
The division of work can change. In an earlier essay about Richard’s research, I made my own framework conditional on what AI can do: if systems sustain architectural coherence autonomously over long periods, the discipline should adapt. Quality and verification still matter as that boundary moves. Socher’s final emphasis on human judgment gives the book a basis for asking how responsibility should evolve with capability. I would have welcomed more development of that relationship between the technical architecture and the humane outcomes he wants it to produce.
How far can the argument go?
The final chapter widens from scientific tools to intelligence itself. Socher proposes knowledge and agency as principal components, then considers capabilities including perception, memory, reasoning, physical action, social coordination, creativity, and self-assessment. He asks what might limit different forms of intelligence and challenges the assumption that human performance defines the relevant ceiling. (pp. 208–227)
The useful move is to make the discussion more specific. Better memory, faster calculation, more effective experimentation, and better coordination are different capabilities. Calling a system more intelligent can conceal which of these improved and what evidence supports the claim.
The chapter also changes the kind of book we are reading. Its proposed categories and unifying measure are conceptual work, with formal definitions and validation still to come. Socher acknowledges that the categories overlap and that formalization remains unfinished. The proposal is interesting as a way to organize questions. It cannot yet do the work of a validated measurement system.
My larger reservation concerns the passage from promising mechanisms to confidence about the whole trajectory. A system might learn quickly where experiments are cheap and outcomes are easy to evaluate, then meet a field where measurements are slow, ambiguous, or costly. Correcting a model also requires diagnosing whether a discrepancy arose from the hypothesis, the apparatus, the data, or the simulation. Faster iteration helps only when those distinctions remain reliable.
The book recognizes many of these difficulties individually. Its forecasts sometimes move across them faster than its explanations establish. I find its account of what researchers could build more persuasive than its confidence about how quickly the larger benefits will follow.
That leaves a substantial positive judgment.
The value of the book does not depend on accepting every destination or timetable. It gives readers a connected way to think about scientific AI, and enough concrete mechanisms to ask better questions about subsequent claims.
A narrower system that reliably helps scientists choose and interpret experiments could be valuable long before the complete Eureka Machine exists.
Who should read it
For a scientifically curious reader who knows AI mainly through chatbots, this is a useful expansion of the subject. It explains why learned patterns in proteins, cells, economic behavior, and astronomical observations belong in the same conversation, while giving each field room to pose different problems.
Researchers and technical leaders may get more from the connections across disciplines than from every account of their own specialty. The book is a wide-ranging argument and research vision. A specialist looking for a technical survey of one field will need the primary literature; the bibliography offers routes into it, and the evidence deserves to be followed beyond the book’s summaries.
Readers who are wary of AI optimism should attend to the experiments and the author’s qualifications before deciding what the forecasts are worth. Readers who already share the optimism should attend just as closely to the steps between a successful demonstration and a dependable scientific institution. Both can get something useful from taking the architecture seriously.
Near the end of the astronomy chapter, Socher worries that a discovery humans cannot grasp could make AI an oracle that delivers answers without understanding. (p. 180) That concern stays with me because it sets a demanding standard for the entire project. Scientific progress should enlarge what people can investigate, test, and understand.
The Eureka Machine is worth reading for its effort to explain how AI might contribute to that work. Its most useful invitation is to examine the process by which a conjecture becomes knowledge, and to build systems that can improve that process without losing sight of what makes the knowledge trustworthy.
Book and source note: This review refers to the advance digital edition supplied for review. Page references use its printed pagination and may differ from the final edition. PublicAffairs lists the book for September 22, 2026; it is available for preorder at the time of this advance review. Links to research distinguish the reported demonstrations from the book’s forecasts.
I write here as an independent blogger. The Society of Professional Journalists’ Code of Ethics informs my approach to accuracy, fairness, disclosure and independent judgment. The views and editorial decisions are my own.