Sensory-First Intelligence: An Agent-Driven Approach to Brain-Inspired Neural Architectures

The dominant approach in artificial intelligence today is scaling — ever-larger models trained on ever-more data. While this has delivered impressive results, it is becoming unsustainable. Training frontier models now costs tens of millions of dollars and consumes vast amounts of energy. We are reaching hard economic and environmental limits. There is a more promising path forward — one that closely mirrors human cognitive development. The human brain achieves intelligence on roughly twenty watts of power through plasticity, sparse connectivity, and elegant local rules. Current transformer models, by contrast, are rigid and extremely compute-hungry. This essay proposes a fundamental shift: building sensory-first intelligence using teams of AI agents that collaboratively evolve new neural architectures with three equally important goals — strong task performance, mathematical simplicity, and radical compute efficiency. The Foundation: Sensory-First Development Rather than beginning with language as current large language models do, intelligence should be built in the same order nature uses. Sensory systems must come first. The architecture should begin by creating tokens exclusively from human-scale sensory data — vision, sound, touch, proprioception, balance, and thermal sensations — all limited to normal human perceptual ranges. Only after a rich, multi-sensory foundation has been established should language be introduced. In this design, language becomes a high-level compression layer grounded in deep sensory understanding, rather than the foundation itself. This sensory-first principle is essential. Without it, we risk building systems that are fluent with words but lack genuine understanding or empathy. Agent-Driven Architecture Evolution To move beyond current limitations, we can create a collaborative team of specialised AI agents that work together iteratively. One agent proposes changes to the neural architecture — refining how information flows, introducing new forms of sparsity, or improving cross-modal integration. A second agent implements and tests these ideas on small-scale models. A third agent observes, evaluates, and judges whether each change improves performance, mathematical simplicity, and compute efficiency. Successful architectures are gradually scaled up in stages, with the roles rotating among the agents. This evolutionary loop treats neural architecture not as a fixed design, but as a living, adaptive system — much like the brain’s own plasticity. Why This Direction Matters By grounding intelligence in sensory experience first and then using agents to actively evolve more efficient mathematical structures, we can develop systems that are not only more capable, but dramatically more efficient. The focus on mathematical simplicity and low compute cost, inspired by the brain’s twenty-watt intelligence, offers a genuine alternative to today’s brute-force scaling race. This approach also opens the door to a more democratised research ecosystem. Much of the early exploration can be done on modest hardware using lightweight agent teams, allowing independent researchers and smaller labs to contribute meaningfully. Conclusion The future of artificial intelligence should not be determined solely by who has the most compute. It should be shaped by who discovers the most elegant and efficient paths to intelligence. By placing sensory experience before language and using collaborative agents to evolve better architectures, we can move towards systems that are both more powerful and far closer to the remarkable efficiency of biological intelligence. The tools to begin this work are largely available today. The real question is whether we are ready to move beyond scaling and start building intelligence the way nature always intended