Naveen Rao: 4D Computing, AI's Energy Wall & Beating Biology

Video summary

Naveen Rao unveiled Unconventional AI’s first dynamical chip, generating images at about 500 nanojoules each.

Naveen Rao, CEO of Unconventional AI, described a career spanning Nirvana Systems, Intel’s AI group, and Databricks before introducing a new computer architecture designed to reduce AI’s energy demands. He estimates Google processes 3.2 quadrillion tokens monthly, requiring about 12 gigawatts at 10 joules per token, and warns demand could outstrip energy supply within three years. Rao says conventional computers waste power moving data between memory and processors, while biological brains achieve intelligence with far less energy. Unconventional AI uses physical dynamics and oscillators to combine memory and computation; its open-source Uno model demonstrated image generation. Rao announced the company built its first physical dynamical computer in five months, with images generated at roughly 500 nanojoules each. He calls the approach “4D computing,” combining three physical dimensions with time, and targets a full data-center rack product within two years. Existing models will need porting at the model layer, and the company uses Python libraries to connect its interdisciplinary team. Its long-term ambition is to beat biology and enable more distributed AI and robotics.

Chapters

1. 0:00Welcome Naveen Rao: Unconventional AI’s 1,000x efficiency goal
2. 4:45Is energy really the problem? Google’s 3.2 quadrillion tokens and the 12-gigawatt gap
3. 10:24Cutting out the middleman: Uno uses oscillator dynamics to generate images
4. 14:58Cutting out the middleman: 4D computer generates images at about 500 nanojoules each
5. 19:40Chamath joins: Unconventional AI targets a rack-scale product within two years