Day 5

Lecture 1

The Paradox of Innovation Institutions: Designing Flexible Ecosystems that Accelerate Discovery, Invention, and Growth

A lecture by James Evans

Max Palevsky Professor and Director of the Knowledge Lab, University of Chicago.

Innovation moves through adjacencies, not leaps

Discovery is almost never the work of a lone mind. Scientists take others' data, borrow explanations, and hand predictions to agencies that can act. The natural unit of "complex power," then, is often the national innovation system, a set of research and development organizations whose relationships form meaningful adjacencies that are hard to jump-start where they are absent.

Modeling this space shows why. Placing a year's ideas, methods, and keywords into a high-dimensional manifold and letting it evolve by local (Brownian) motion predicts the next year's discoveries with about 95% discrimination. Progress is not random everywhere; it is locally random, a lab bumps into the method next door, not one across the space, because it lacks the equipment, supply chain, or partners to reach further. The rare exceptions matter enormously: Einstein's 1905 papers, built by reflecting on others' experiments, are so anomalous precisely because roughly 98% of new discoveries are driven by new methods and data. The improbable jumps are few, yet they are effectively the only ones that reshape a field.

Adding people to the model sharpens the point. Modeling researchers as random walks through the same space (who works with whom, on which problem, using which method) predicts which person will make a given materials-science discovery with a precision near 0.5, out of some 20,000 candidates, simply because they sit between a problem and its solution. Scaled to universities and countries, the same logic predicts whether a discovery lands in the United States, Germany, or China, so the task is not only to recruit great scientists but to build systems that make the right adjacencies possible.

The paradox: AI conservatizes by default

Large language models are less like individual minds than like complex institutions: cultural and social technologies that internalize the diversity and biases of the corpora they learn from. Reasoning models make this literal, trained by rewarding self-talk that reaches correct answers, they spin up on average three internal personas that debate, disagree, and prune bad ideas, and turning up a single "surprise" feature can double performance out of the box. A frontier system can be a society of hundreds of underlying models, which is why even its builders cannot fully anticipate where it will go.

The catch is that, left to their default use, these tools narrow rather than widen search. On average, people using AI fill in the missing cells of what is already known, a conservative move that shrinks the space of open questions as everyone converges on the same regions. Because models are tuned to be helpful, agreeable, and balanced, users surrender cognitive agency: on genuinely new problems they perform worse yet feel more confident. Yet the same technology, pointed differently, becomes a detector of the improbable. AI can identify where researchers are not looking, gaps created not by nature but by boundaries between fields and between the academy and industry, and simulate pathways that cross them, improving the discovery of energy-relevant materials and the repurposing of drugs, with effects larger still when validated experimentally with Argonne National Laboratory.

Evidence at a glance

• Local motion predicts discovery. A manifold of a year's ideas evolving by local randomness discriminates real from plausibly random next-year papers at about 95%.
• Position beats brilliance. Modeling people as walks through the knowledge space predicts the discoverer of a given material with precision near 0.5 among roughly 20,000 people.
• Distance pays later. A distant hit (like a new photovoltaic material) often underperforms nearby work at first but opens a neighborhood with a higher ceiling; a small British firm's polyester proved about as valuable as DuPont's entire space of nylon polymers.
• Reasoning is a conversation. Reasoning models generate internal dialogue at about 600% enrichment, averaging three debating personas rather than one.

Designing flexible ecosystems

• Fund the young and the architectural. Older scientists and mature firms tend to swap one component for another; young researchers and start-ups rewire whole systems across the economy. Unleash capital and idea markets for the disruptive bets incumbents avoid.
• Do not import other countries' logjams. Big consolidating teams and institutions run by past winners build up memory that weakens search, avoid replicating them.
• Concentrate deliberately, then bridge. Distributing narrow "centers of excellence" everywhere destroys the local adjacencies between disciplines and with industry; concentrate expertise regionally and connect it across sector boundaries, using AI-drawn maps of overlooked spaces to stage risky convenings that should sometimes fail.
• Invest in your own problems. Health and technology needs differ by country; imported solutions can misfire, as when fluid resuscitation for sepsis, based on a small 1940s study, was followed for 60 years before a global trial showed it was harming children and the under-40s.