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

Summary

This All-In interview has Naveen Rao present Unconventional AI’s case that energy and data movement, rather than transistor count alone, are becoming binding limits on AI growth. The company proposes physical dynamical computing built from coupled oscillators, using three-dimensional integration and time-varying state as a so-called 4D substrate in which computing elements also retain state.

The source’s concrete proof point is UNO, first simulated as a trainable image-generation model and then implemented in an early physical prototype. Rao reports roughly 500 nanojoules per generated image and targets a 1,000-fold gain in intelligence per watt, but the episode supplies no independent benchmark, matched output-quality comparison, full-system measurement, production reliability evidence, or manufacturing economics.

Key Claims

• Rao argues that AI expansion is approaching an energy wall because model demand grows faster than deliverable electricity and conventional computers spend substantial energy moving data between memory and compute.
• Biological brains show that useful intelligence can operate at much lower power, but the proposed response is semiconductor architecture inspired by physical dynamics rather than a literal biological processor.
• Physical Dynamical Computing seeks useful computation in the time evolution of coupled oscillators instead of implementing every neural operation as conventional digital arithmetic.
• UNO is presented as evidence that a sparse oscillator system can be trained in simulation, generate recognizable images, and transfer into a fabricated physical prototype.
• Rao calls the architecture “4D computing” because die stacking supplies three physical dimensions while time is an intrinsic computational dimension.
• The first prototype reportedly generated an image with about 500 nanojoules, but the episode does not establish an independently comparable end-to-end efficiency result.
• Existing models would need to be ported through a new programming layer, so compatibility, engineering effort, reliability, and workload fit remain commercialization gates.
• Rao predicts that much cheaper computation could expand total AI use through Jevons paradox, including distributed infrastructure and energy-constrained robotics.

Key Quotes

“intelligence per watt” - Rao’s optimization target for the architecture.

“4D computing” - the company’s label for three-dimensional physical integration plus temporal dynamics.

Connections

• All-In - show context for the interview.
• Naveen Rao, Unconventional AI, and UNO - founder, company, and model/prototype at the center of the episode.
• Physical Dynamical Computing, AI Data Movement Energy Cost, and Intelligence Per Watt - core architecture and evaluation concepts introduced by the source.
• Data Center Power Bottleneck, AI Energy Bottleneck, and Jevons Paradox In AI - infrastructure and rebound-demand context.
• Memory Wall, In-Memory Computing For Edge AI, and Semiconductor 3D Stacking - adjacent approaches to memory movement, locality, and vertical integration.
• Biological Processor Energy Efficiency - biological-efficiency benchmark that motivates the architecture without validating its performance.
• Intel, Databricks, and GPU - Rao’s prior company-building and conventional infrastructure context.

Contradictions

• No settled contradiction with existing wiki content was found.
• Rao’s reported 500-nanojoule result and 1,000-fold target are company claims without a disclosed standardized benchmark, matched quality threshold, independent replication, or full rack-level accounting.
• The episode’s token-volume, joules-per-token, data-center-capacity, brain-bandwidth, and thermodynamic-limit figures remain source-scoped estimates rather than independently verified current facts.
• The two-year product plan and longer-term distributed-AI and robotics forecasts depend on model portability, software tooling, fabrication yield, reliability, cooling, system integration, customer adoption, and economics not demonstrated by the prototype.