New research from Allora Labs suggests that AI systems may become more resilient to changing environments by applying a principle borrowed from biological evolution called deliberate imperfection.
“The qualities we associate with intelligence, like originality, adaptability, and creativity, all require the ability to perform well under conditions you’ve never encountered before,” Dr Diederik Kruijssen, chief scientist at Allora Labs and the paper’s author, said.
“Current AI models can’t do that. They replicate what they’ve learned, but they don’t adapt to the unknown. A population of evolving models can. That’s what this work demonstrates, and it’s what evolution has been doing for billions of years,” Dr Kruijssen added.
Deliberate imperfection principle
The paper, “Flawed in Nature, Perfect through Evolution”, published in the journal Allora Decentralised Intelligence, argues that while a single highly optimised AI model can perform well under known conditions, its performance declines when the environment changes.
The proposed solution is to use “swarms” of AI models with deliberate random mutations, which can make individual models worse or produce variants better suited to new conditions.
The study showed that the best model within a mutated population outperformed the best model in an optimised population approximately 80% of the time after environmental changes.
The study showed that performance peaks when the rate of model variation aligns with the environment’s change rate.
Nick Emmons, CEO of Allora Labs, said: “As AI systems are deployed in financial markets, healthcare, autonomous vehicles, and other environments that change constantly, the inability to adapt to shifting conditions has become their primary failure mode. Most AI research focuses on making individual models bigger and faster. We think the bigger gain is in how populations of models interact and evolve, and this work proves that’s the right perspective.”
Looking ahead, researchers will apply the principle to large language models, decentralised AI networks and automated controllers capable of adjusting mutation strength as conditions change.
