What If We Stopped Using GPUs? | YC Paper Club

Video summary

YC’s Alternative Compute Club explores optical diffusion models, brain-inspired chips, and living neurons playing Doom beyond GPUs.

At YC’s Alternative Compute Club, Focal Systems founder Francois Chaubard argues that AI hardware and learning methods have co-evolved around transformers, backpropagation, and GPUs, even as compute-efficiency gains slow. He presents zero-order optimization and SOMA, a proposed mixture of small, sharded models, as alternatives to gradient training. Ilker Oguz explains how light can perform massively parallel, low-energy computation and describes an EPFL-Google experiment using optical propagation to generate diffusion-model images; converting data between digital and optical systems, storing weights, and implementing nonlinear operations remain obstacles. Alok Vasudev says neuromorphic computing is still in R&D, though bringing memory closer to computation may be nearer-term. Finally, Sean describes training cultured brain cells to play Doom with reinforcement learning, learned stimulation and decoding, and feedback based on surprise. The speakers emphasize that practical alternatives will require matching algorithms to each substrate’s strengths and solving scale-up challenges.

Chapters

1. 0:00Francois Chaubard: Why Alternative Compute?—Transformer Memory Demands and Stalled Compute Efficiency
2. 3:38Beyond Backpropagation: The Brain’s 20-Watt Compute and Chaubard’s All-Light Alternative
3. 8:08Zero-Order Optimization and SOMA: SPSA and Sharded LSTM Experts
4. 15:06Ilker Oguz: Computing With Light—10,000× Lower Loss and Parallel Optical Beams
5. 19:27Why Optical Computing Isn’t Everywhere Yet: Conversion Costs and Nonlinear Activations
6. 22:23Building a Diffusion Model With Light: MNIST Tests and a Proposed Billion-Parameter Milestone
7. 30:03Optical Computing Q&A: Lightmatter’s HBM Pivot and Deep-Network Nonlinearity Challenges
8. 34:55Optical Computing Q&A: Wavelength Multiplexing, Optical Memory, and Polarization
9. 39:00Alok Vasudev: Neuromorphic Computing and the Brain’s 20-Watt, Adaptive Design
10. 45:23What the Brain Can Teach Us About Chips: Threshold Spikes and Hardware-Software Co-Design
11. 49:35Where Neuromorphic Computing Stands Today: 2026 R&D, D-Matrix, IBM, and Intel’s Drone
12. 54:24Neuromorphic Computing Q&A: D-Matrix, Coupled Oscillators, and Optics for Data Movement
13. 1:00:00Neuromorphic Computing Q&A: Diamond Substrates, Useful Noise, and SPSA Reproducibility
14. 1:04:24Sean Cole: Teaching Brain Cells to Play Doom With 59 Channels and 54 Actions
15. 1:07:48Training Biological Neurons With PPO: Stochastic Stimulation and Decoder Controls
16. 1:12:28How Brain Cells Learn From Feedback: Synchronous Signals and TD-Error Scaling
17. 1:15:40Biological Computing Q&A: Entropy, Divergent Intelligence, and Scaling to Billions