Evolution and AI

What can 540 million years of evolution teach us about AI? Quite a lot [1].

Intelligence did not evolve simply because nature built a better brain. Evolution repeatedly improved the system around the brain.

During the Cambrian explosion, organisms did not become more capable because of a sudden, isolated leap in cognitive architecture [1]. Instead, they developed complex, image-forming eyes, allowing them to actively sense and navigate their environments [4]. While evolutionary biologists debate whether this sensory leap was the primary catalyst for the Cambrian diversification—a concept known as Andrew Parker’s "Light Switch" hypothesis—or whether geochemical changes like ocean oxygenation and genetic innovations were the true drivers, the systemic impact of vision remains undisputed [1, 4]. Sensing was only the beginning.

This biological leap is unfolding in modern technology. Early large language models were effectively blind, trained on static, historical text corpuses. The modern AI "Cambrian explosion" is not driven by making the core model larger, but by equipping it with sensory systems: multimodal ingestion, live search, and real-time retrieval-augmented generation (RAG) [4]. The model is no longer isolated; it can see its environment.

As biological intelligence grew more sophisticated, energy became the ultimate constraint. The human brain represents only about 2% of body weight but consumes roughly 20% of the body's resting energy [2]. To make this high metabolic cost sustainable, nature had to develop complex energy-management systems [2]. Without these biological guardrails, advanced cognition would have been an evolutionary dead end.

Enterprise AI faces an identical metabolic wall. Brute-force scaling of model parameters is hit by a massive power bottleneck, threatening local energy grids. To resolve this, the system around the model must evolve. Just as biological systems manage energy through localized blood flow and specialized glial cells, enterprise architectures must rely on sparse architectures like Mixture of Experts (MoE), parameter-efficient fine-tuning (PEFT), and localized edge deployment to keep AI sustainable.

After metabolism came the need for faster networking. In the nervous system, this was solved by the evolution of myelin—a protective sheath that wraps around nerve fibers [3]. While myelin is famous for accelerating signaling speed by up to 100-fold, biophysical modeling reveals that its original evolutionary purpose was actually energy conservation [3]. It prevented signal decay and saved precious metabolic energy along long pathways [3]. Myelin solved the physical bottleneck of long-distance communication in larger bodies [3].

Yet, neurobiology reveals a fascinating nuance: the most complex, highly evolved regions of the mammalian brain, such as the cerebral cortex, actually show intermittent or reduced myelination [3]. This structural trade-off sacrifices uniform transmission speed to preserve synaptic plasticity, allowing the brain to remain adaptive and learn [3].

This is the blueprint for enterprise AI orchestration. The critical latency bottleneck in the enterprise is rarely the token generation speed of the model itself. The delay occurs in the "uninsulated" pathways surrounding the model: heavy middleware, slow vector database queries, and clumsy API calls. To scale, we must "myelinate" these pathways with high-bandwidth memory, optical interconnects, and distilled routing layers. Yet, we must also embrace the lesson of intermittent myelination. We do not need rigid, hyper-optimized speed across every workflow; instead, we must design adaptive systems that balance rapid routing with the computational flexibility of multi-agent orchestration.

I have been thinking about what all of this means for enterprise AI.

Frontier models already possess extraordinary cognitive capabilities, and they continue to improve. But the difficult work begins after a model is selected: connecting it to proprietary data, defining the boundaries within which it can act, integrating it into existing workflows, and measuring whether it produces reliable economic outcomes.

In other words, the challenge is no longer just building a better AI brain. It is building the system around it. That system must also include the human-AI interface, because AI without human integration represents capability without judgment.

This is the last mile of enterprise AI: crossing the chasm from a powerful model to an intelligent enterprise system. The next decade of enterprise AI will belong to the organizations that learn how to build that system—reliably, repeatedly, and at scale.

In my latest Forbes Technology Council article, I explore what biology can teach us about building enterprise AI—and why intelligence has always been a systems problem.

The brain matters. But what intelligence can accomplish has always depended on the system around it.

Evolution and AI

References

Comments