Stephen Wolfram discusses his philosophy and methodology for doing good science, drawing on decades of work across computer science, fundamental physics, biology, economics, and the foundations of mathematics. The central thread is the computational paradigm — the idea that specifying simple rules and letting them run, rather than trying to solve everything from the top down, unlocks progress in fields that have been stuck for roughly a century.
How Wolfram Does Science
Tool-building and science feed each other. Wolfram alternates between developing technology (like Wolfram Language) and doing science. Having first dibs on his own tools lets him explore new methodological territory before others.
The computational paradigm as a unifying move. Many fields ground things down to elementary primitives in the late 1800s and early 1900s (atoms, formal axioms, discrete space) and then got stuck. Wolfram argues this was because they had implicitly reached computational irreducibility — the phenomenon where even though you know the rules, you cannot shortcut to the outcome; you must run the computation step by step. Gödel’s theorem is another reflection of this.
Three-part method: philosophy, computation, and homework. He forms a conceptual or philosophical hypothesis, translates it into a program and runs the experiment (removing human calculation error), then does historical homework to see how the results relate to what others have found.
Let the chips fall where they fall. A recurring theme: you must resist the psychological pressure to tweak or ignore experimental results that violate your theory. Some of his most important discoveries came from results that violated his own prejudices — and only years later did he realize the results actually confirmed a deeper version of his original intuition.
Computational Irreducibility and the Observer
Computational irreducibility is the key mechanism. It explains why simple rules can produce complex-looking behavior and why observers (whether physicists, mathematicians, or biological fitness functions) cannot simply deduce outcomes without running the computation.
The second law of thermodynamics as a case study. The reason molecules appear to get more random is not a magic source of randomness but the interplay between the computational irreducibility of molecular dynamics and the computational limitations of human observers, who cannot decrypt the detailed computation and so perceive it as random.
The observer is more central than he expected. Wolfram’s early work was “deeply dehumanized” — the universe as a giant hypergraph (the “Ruliad”), a completely abstract object. But he eventually realized that the nature of the observer is critical: the laws of physics we perceive depend on our nature as observers embedded within the Ruliad. Humans turned out to be more central to physics than he had imagined.
Multiway systems and quantum mechanics. He resisted his own invention (multiway systems, where the universe follows many branching paths of history) for nearly 30 years because he didn’t want the experienced world to be just one arbitrary branch. The 2019 physics project revealed the correct picture: observers are themselves branching and extended across branches in “branchial space,” so experience spans many branches rather than following one.
Physics: Discrete Space, Spacetime, and Feynman Diagrams
Why physics abandoned discrete space. In the early 1900s, most physicists expected space to be discrete. Heisenberg tried to build a discrete cellular model of space around 1930 but couldn’t make it work with relativity, so he pivoted to the S-matrix approach (only describing inputs and outputs). The idea of discrete space was dropped for a century — until Wolfram showed how to make it work computationally.
Spacetime was a foundational mistake. Wolfram argues that Minkowski’s 1908 insight — bundling time and space together because they appear in the same quadratic form — misled physics. Time (a computational process) and space (an extent of a data structure like a hypergraph) are fundamentally different in nature, even though relativity links them mathematically.
A deeper theory of Feynman diagrams. Wolfram is developing a new understanding of Feynman diagrams not as particle tracks but as representations of causality — events and their causal connections. This connects to “multi-causal graphs” from his physics project and may allow more effective calculations than the current k-factorial-to-the-fifth-power complexity.
Biology: Natural Selection and Computational Irreducibility
Biology has had no real theory. The closest thing is natural selection (1859), but biology textbooks lack a theoretical foundation comparable to physics. Wolfram argues this is because biology’s complexity is a consequence of computational irreducibility.
A 40-year failed experiment that finally worked. In 1985, Wolfram tried to evolve cellular automaton rules via mutation and selection, mimicking natural selection. It produced nothing interesting. In 2024, inspired by the lesson from deep learning that “if you bash a neural net hard enough, it will learn,” he tried running the experiment much longer — and it worked. Simple rules, under coarse fitness functions (e.g., how wide does the pattern get?), evolve elaborate structures by assembling pieces of irreducible computation.
Why biological evolution works. The fitness functions in biology are coarse (survive, reproduce), not impossibly specific. Computational irreducibility means the space of possible behaviors is rich enough that evolution can find lumps of irreducible computation that satisfy these coarse objectives. This is analogous to how neural networks work: they assemble “random rocks” of computation into a functional “stone wall.”
Toward a theory of bulk orchestration. Wolfram is working on a general theory of systems that are “bulk orchestrated” — where detailed assemblies of components achieve coarse-grained purposes. This would apply to biology, microprocessors, and any adaptively evolved system, independent of the specific historical path taken.
Economics and the Nature of Value
A proto-theory of economic value. Wolfram hypothesizes that the ultimate source of value is computational reducibility — pockets where you can shortcut the irreducible computation of the world. Because humans are mortal and time is scarce, anything that saves time (i.e., exploits reducibility) is valuable. A smartphone is valuable because using it avoids having to recompute everything from scratch.
Value in biology and invention. The same logic applies: when evolution or invention finds a piece of reducibility (a mechanism), it gets reused. Economics currently accounts for scarcity but not the value of the reducibility itself.
Scientific Methodology: Visualization, History, and Foundations
Visualization is critical. The most common mistake in science is to look at only a small slice of what’s happening (e.g., one summary curve). Wolfram insists on the most faithful, detailed visualizations possible, because our visual system is the highest-bandwidth way to get data into our brains. He was recently “bitten” by this in a current project when a better visualization revealed something he’d missed.
Do your homework on history. Understanding why people believed what they believed — and where they got stuck — is essential for knowing you’re actually making progress. Wolfram spent considerable effort untangling the history of the second law of thermodynamics to confirm his computational explanation fit with existing knowledge.
Foundational questions are often the most progressable. The foundations of a field are frequently “unprotected” — nobody has looked at them for decades, and the ambient methodologies have changed completely. The computational paradigm is the biggest such change. Progress on foundations tends to open up new questions rather than answering the old ones.
Beware of experimental prejudice. A formative experience: a QCD calculation Wolfram did in the 1970s predicted a particle interaction rate that contradicted an experiment. He trusted the theory; the experiment turned out to be wrong. The lesson: experiments are hard, and prejudice about expected outcomes can cause people to miss real phenomena. Computer experiments avoid many of these issues because the rules are exact and the results are irrefutable (though interpretation can still be wrong).
Ruleology: How Anyone Can Contribute to Science
Ruleology as an accessible frontier. The study of simple rules and what they do (cellular automata, Turing machines, lambda calculus, hypergraph rewriting) is a vast area where amateurs and students can make permanent, foundational contributions. The rules are abstract and timeless — a discovery about a particular rule will never become obsolete.
Low barrier to entry. You don’t need years of mathematical physics. Wolfram’s summer programs have high school students producing publishable results in two weeks using Wolfram Language tools. The key is being organized, doing the simplest possible experiment, and drilling down to foundational questions.
The computational universe is inexhaustible. Pick a random number as a rule — it has almost certainly never been studied before, and it will have richness. This is unlike fields like traditional physics, where there is a tall tower of existing knowledge you must climb before contributing.
Reproducibility and understandability. Everything Wolfram publishes includes runnable Wolfram Language code so anyone can reproduce the results. He also insists on writing explanations that anyone can understand, which forces him to actually understand what he’s talking about and prevents “floating over the formalism.”
What Is Science?
Science as a bridge. Science is the effort to produce a human-comprehensible narrative about what actually happens in the world. Traditional science tries to “wrap arms around the whole thing”; the computational paradigm offers a different kind of science where you understand the primitives but only have a meta-understanding of the whole arc.
Good science is high-leverage and foundational. It deals with simple, clean things that show up over and over, not three-page descriptions of narrow details. And it must be understandable — science nobody can understand fails its core purpose.
Bad science often comes from prejudice and muddiness. The most common cause of error is expecting a particular result and then having an experiment too complicated to reveal when you’re wrong. Computer experiments minimize this by being minimal and transparent.
