A new book by the co-founder and CEO of You.com and Recursive Superintelligence goes onto the ins, outs, and what-have-yous of how science may work in a new era of automation and agentic superintelligence.

Socher, from the beginning, suggests that neural networks plus big data will lead to a scientific renaissance, and that what he describes as “full-stack AI” will revolutionize our journey toward new heights of discovery.

In chapter one, Socher looks back at technologies like Watson and Deep Blue, describing the development of two “tribes” in human analysis: connectionists, and proponents of symbolic reasoning systems. Using the example of a game called Pachinko, Socher explains how AI self-corrects over time, while citing Moravec’s paradox: AI is good at rote knowledge, but not intuition. Mentioning classifiers like AlexNet in the early days of AI, Socher notes:

“Language is full of subtle shades of meaning. Each dimension captures a faint trace of a pattern, maybe something about the size, or living versus non-living, or emotional tone. Alone, each dimension says a little, but together they allow the system to capture the nuanced meaning of a word.”

I loved this example he gives, of a molecular construct we call Oxygen:

“You can put the word oxygen into a large language model and access vast amounts of written information,” Socher writes. “But any description of oxygen in a sentence leaves out far more than it captures. Much of what we know about the molecule exists in other forms, mathematical equations, spectral data, quantum wave functions. Our deepest understanding of oxygen is encoded in data that cannot be fully conveyed in words.”

Moving on, Socher continues to describe the meteoric rise of the technologies in question, explaining benchmark saturation and the ferocity of competitors in this area.

“AI research has taken on the atmosphere of a global sport, complete with scoreboards, spectators, and bragging rights,” he writes.

Asking “what is left to know?” Socher describes a paucity of actual new rules and revelations, which is strange, given investments in systems like the Large Hadron Collider. He suggests AI will reverse the trend, though.

Citing the work of Thomas Kuhn, he points to a kind of conformity and “gap-filling” in new research that lacks the eccentricities needed to innovate.

Another of Socher’s major thesis statements is the compelling idea that while scientists of the twentieth century were focused on breaking down larger systems into smaller ones, “teasing apart” the universe, today’s work is perhaps best supported by starting to “weave the world back together,” and AI will help us to see the big picture. It sounds like an argument for unified field theories. Suggesting that science, as a discipline, is “alien” to human societies, he notes that it might take “alien intelligence” like AI to boost its potential.

“Crucially, abstraction would occur at each layer, condensing complexity into representations,” he writes. “To show essential patterns without reproducing every fine-grained interaction. This layered abstraction would make the model computation feasible, also retaining protective power and biological coherence.”

There’s a lot more in here exemplifying Socher’s major arguments: he talks about staging drug discovery, new potential for humane testing, and democratizing knowledge systems, as well as that monolithic premise in AI research: digital twinning, the simulation of a complex system in digital representations that the neural net can go to work on.

And we’re close: Socher describes cancer cures, pollution remediation, and that old sci-fi dream of organ libraries.

“In these virtual worlds, we can find a deeper understanding of how organisms adapt, evolve and respond to shifting conditions,” he writes. “We will be able to see the big picture and the small, at scales ranging from the microbial to the biosphere. In time, this knowledge may lead us to a deeper grasp of life-organizing principles – the hidden symmetry and emergent properties that allow molecules to grow into organisms as complex as ourselves.”

Applications to Economics

For so long, Socher notes in a subsequent chapter of “The Eureka Machine,” economics was just what we could prove mathematically. Again, big-picture work facilitated by AI has the power to affect change, uncovering behaviors and tracking the invisible hand of the market.

And then he brings it back to the Sims.

“In the future, these sim-citizen populations will increase exponentially,” Socher writes. “Researchers will likely be able to dispense with the time-consuming process of individual testing, because data describing our preferences and economic behaviors already exists in numerous forms.”

All of this, he suggests, can lead into policy-making, for big wins.

Referencing a timeline including, for example, Hippocrates and the four humours, Socher talks about something he calls “mechanomorphism,” where we tend to think of the human body in terms of the technology we have. Today, that would be the “brain as a computer,” an idea popularized by pioneers like Turing and Von Neumann.

Suggesting that we may some day be “decoding human thoughts,” Socher predicts that brain computer interactions will “generate useful surprises.”

“Our brains managed to run a distributed parallel processing system of roughly a trillion neurons using about 20 watts,” he writes. “Understanding this remarkable efficiency and applying it to AI systems would be transformative.”

Turning to cosmology, Socher describes “beautiful math and messy data” citing systems like the SKA in Australia, exploring the astronomical applications of recurrent neural networks. Current technologies, he notes, are often relatively narrow in scope. That could soon change.

“What we have seen so far is not the main act, but an overture,” Socher concludes, “an introduction to the themes and motifs that will define the future.”

With that, he delineates four essential pillars for making this type of research reality.

Pillar 1 is the Foundational Model of Human Knowledge, basically, the underlying data to be used. Socher also enumerates the Unified Model of Reality, which I understood as the tools to move around in a model world, and the Simulated World, or the model world itself.

The fourth pillar, the Lab in the Loop, involves experimentation.

Noting that there’s “no substitute for reality,” Socher describes how real-world trials bring everything full circle, delving into agentic AI, proof of concept, and recursive self-improvement.

I’ll leave it to the reader to go over Socher’s Ten Spaces of Intelligence, where he illuminates more of the architecture that can lead us toward our goals.

Overall, he calls for collaboration and partnerships, as we strive to use AI well in aid of robust discovery, writing:

“This Eureka machine represents a monumental challenge that no single university or company can meet alone.”

I found this book to be quite useful in pondering the big questions about our shared frontier.