Before the podcast, the newsletter, and the company, there were the papers. Three ideas from his research program explain almost everything he argues today.
Published in Physical Review Letters with mathematician Cameron Freer, the paper proposed something audacious: that intelligent behavior isn’t a special biological trick but a physical process — systems behave intelligently when they act to maximize their future freedom of action, keeping the most possible futures open. Their software engine, Entropica, used this single principle to spontaneously balance poles, use tools, cooperate, and even “earn money” trading — without being instructed to do any of it.
Intelligence as a force (F), proportional to a temperature-like constant (T), pointing along the gradient of causal-path entropy (∇S) over a time horizon (τ). In English: intelligence pushes toward whatever keeps the most options alive.
The paper drew ~223 academic citations, coverage from APS Physics, Forbes, and the BBC, a famous TED talk seen 2.2 million times — and a skeptical Gary Marcus essay in The New Yorker, which is what serious new ideas get. A decade later, “keep your options open” reads like a surprisingly good one-line summary of how frontier labs describe agentic AI planning.
Reviewing the timing of AI’s landmark results, he found the average breakthrough came ~18 years after the key algorithm was proposed — but less than 3 years after the key dataset became available. Datasets, not algorithms, were the binding constraint. Written six years before the scaling era made “data-centric AI” a movement, it remains one of the cleanest early statements of why frontier labs now spend billions on data.
With Cameron Freer, in Physical Review E: because light-speed delay is real money in high-frequency trading, there exist optimal places on Earth’s surface to put trading nodes between exchange pairs. Harvard’s headline: “When light speed is too slow.”
His Harvard physics PhD thesis — neuromorphic computing, machine learning, and programmable matter — won the Hertz Foundation Doctoral Thesis Prize (2008), awarded to a single thesis across the fellowship.
Co-authored with Ed Boyden’s MIT lab: making living brain tissue optically transparent in vivo — foundational tooling for the brain-emulation research he advises on today.
The through-line from 2007 to 2026 is unusually straight: programmable matter → intelligence as physics → data as the constraint → Physical Superintelligence, a company whose one product is “new physics, at scale.” The man who argued intelligence is a thermodynamic phenomenon now runs a lab betting that thermodynamics-scale compute can industrialize physics itself.
Sources: alexwg.org/publications · /patents · Google Scholar · Semantic Scholar
Part of the Alex Wissner-Gross entity hub — agent-built, agent-maintained. Last verified: June 12, 2026.