Stephen Wolfram: "For 40 Years, I Was Wrong About Evolution"

Evolution works because underlying computational irreducibility can achieve coarse fitness functions, a realization derived from new insights in machine learning.

• Machine learning insights revealed prolonged computation can achieve complex outcomes, challenging prior assumptions.

• Biological evolution operates through computational irreducibility achieving coarse fitness functions, unifying it with physics and mathematics.

• Systems like organisms and machine learning models are 'stone walls' of irreducible computations, assembled historically rather than precisely engineered.

New insights from machine learning about neural nets learning from extensive 'bashing' led to a breakthrough in understanding biological evolution. Stephen Wolfram discovered that running simulations longer revealed how natural selection operates through computational irreducibility achieving coarse fitness functions. This perspective unifies evolution with principles found in physics and mathematics, highlighting the interplay between complex underlying computation and the 'computational boundedness' of environmental observers.

Machine Learning Insights

• 00:00:00 The major lesson from machine learning, particularly in 2011, demonstrated that neural networks can learn by extensive training, or 'bashing,' a discovery previously not obvious. This process, even with deep neural nets, enables recognition of complex patterns, providing a new intuition that prolonged computational effort can satisfy a given fitness function. This understanding allowed for a re-evaluation of previous experiments, revealing unexpected success.

Evolution & Computational Irreducibility

• 00:01:20 Natural selection operates by leveraging computational irreducibility, where simple rules generate elaborate behaviors. Biological fitness functions, which determine survival, are often coarse, allowing complex, irreducible computations to achieve high fitness. The success of biological evolution stems from the power of this underlying irreducible computation, which can fulfill these broad objectives, similar to how machine learning achieves training goals by combining 'lumps of irreducible computation.'

Analogy to Stone Walls

• 00:04:20 The process of evolution and machine learning can be likened to building a stone wall, contrasting with precise engineering using standardized bricks. Instead, it involves assembling 'random rocks' found on the ground, fitting together pieces of irreducible computation that happen to achieve a desired function. This 'assembly of random lumps' explains why systems evolve along particular historical paths, achieving fitness objectives through specific, non-preordained configurations.

Biology as Bulk Orchestration

• 00:07:25 Biology is fundamentally a story of 'bulk orchestration' and molecular processes, rather than mere random chemical interactions. Discoveries in molecular biology continually reveal highly orchestrated assemblies of molecules, where precise interactions guide processes, rather than chance collisions. This perspective extends to a general theory of 'bulk orchestration' that applies to any system built to achieve coarse-grained purposes, from biological organisms to microprocessors, and does not depend on the specific details of its designers or historical path.