Conference Presentation, Keynote
Why the Brain Computes 1,000,000x More Efficiently Than A GPU: Unconventional AI's Naveen Rao
- The company name "Unconventional AI" is expected to eventually change to "conventional" as the market evolves.
- Tape-out timelines are projected to shrink from multi-year cycles to a matter of months, contrasting with traditional semiconductor manufacturers.
- An actual prototype chip is scheduled for completion this summer, approximately six months from the commencement of the talk, following a team expansion from January.
- Within two to four years, global energy availability is predicted to become insufficient for AI training and inference operations.
- Current compute efficiency gains are characterized as incremental, with no significant improvement in energy cost per floating-point operation involving memory access.
- Human-level intelligence regarding discovery is anticipated to be reached in a very short timeframe, albeit at a high energy cost.
- The industry is approaching a thermodynamic limit to intelligence defined by the Landauer principle, which serves as an uns surpassable asymptote.
- Biological systems are estimated to be approximately two orders of magnitude from the thermodynamic asymptote, whereas current silicon technology is three orders of magnitude distant.
- The organization plans to leverage nonlinear dynamics and physical substrate properties to achieve "galaxy brainness" while significantly increasing compute efficiency.
- Dynamic systems are expected to be trainable to steer toward arbitrary trajectories, including the generation of specific image classes such as cats or horses.