Unconventional AI is building a novel physical computing substrate in silicon, designed from first principles to reduce energy consumption for AI workloads to biology-scale efficiency. The approach rests on hardware–software co-design: custom mixed-signal and analog circuits that exploit oscillatory dynamics and processes resembling neural computation, paired with corresponding software interfaces and probabilistic models.
The technical scope spans analog circuit design, silicon implementation, AI systems, computing theory, and neuroscience. Rather than optimizing within conventional digital architectures, the work translates insights from biological systems to fundamentally alter the substrate itself - addressing the energy-latency-throughput trade-off at the physics level rather than through algorithmic or architectural tweaks to existing von Neumann designs.
This is substrate-level work: the bottleneck is not software optimization on fixed hardware, but redesigning what computation looks like when energy density, heat dissipation, and latency are redefined by the underlying medium. Success requires integration of analog reliability and circuit-level robustness with AI inference performance and software ergonomics - a coupling that introduces operational complexity and evaluation challenges distinct from digital accelerators.