Akarsh Kumar and Tim Scarfe distinguish artificial life from ordinary task optimization: it studies possible worlds and the processes that produce complex behavior. Akarsh Kumar argues that evolutionary histories and curricula can discover structures that a fixed objective or gradient-based training may miss, while acknowledging that these ideas do not establish a route to general intelligence.
Akarsh Kumar explains the fractured entangled representation hypothesis and contrasts it with organized representations found in some evolutionary systems. He describes experiments across cellular automata, Lenia, Boids and Particle Life, using foundation-modelA foundation model is a broadly trained AI model that can be adapted to many downstream tasks and applications. embeddingsAn embedding is a numerical representation that lets an AI system compare information by learned patterns or meaning. to search for target behaviors, diversity and open-ended change. These measures are useful proxies with limitations, not definitive tests of life or intelligence.
Akarsh Kumar and Tim Scarfe discuss ASAL's exploration of hundreds of thousands of Game of Life rules and the Digital Red Queen project, which uses language modelsA language model learns patterns in sequences so it can interpret context and generate likely continuations such as text, code or structured responses. to mutate competing Core War programs. Akarsh Kumar reports strong results against a set of human-written programs and examines repeated adversarial evolution, robustness and behavioral variety. The conclusions remain bounded by the selected simulators, evaluation measures and opponents.
Watch on YouTube
