What If We Stopped Using GPUs? | YC Paper Club

Video summary

YC’s Alternative Compute Club explored optical AI, brain-inspired chips, and living neurons playing Doom as alternatives to GPU-based learning.

At YC’s Alternative Compute Club, Focal Systems founder Francois Chaubard argued that AI hardware and algorithms have co-evolved around transformers, backpropagation, and GPUs, even as compute efficiency gains slow. He highlighted the human brain’s roughly 20-watt power use and described SPSA zero-order optimization and SOMA, a proposal to train many small model experts in parallel. Ilker Oguz presented optical computing, where photons offer about 10,000 times lower loss and higher bandwidth than electrons, and showed a diffusion-based image-generation experiment with an estimated sevenfold energy advantage over a comparable GPU model; conversion, memory, and nonlinear activation remain obstacles. Alok Vasudev said neuromorphic computing is still in R&D, though memory-compute integration is nearer-term. Finally, Sean described training cultured brain cells to play Doom using learned stimulation and decoding, while emphasizing that scaling to widely served intelligence remains an open challenge.

Chapters

1. 0:00Francois Chaubard: Nvidia’s Shift from CNN FLOPs to Transformer Memory
2. 3:38Beyond Backpropagation: A 20-Watt Brain and Alternative Compute
3. 8:08Zero-Order Optimization and SOMA: SPSA, Sharded Experts, and 64 Perturbations
4. 15:06Ilker Oguz: Photons Offer 10,000× Lower Loss for Parallel Computing
5. 19:27Why Optical Computing Isn’t Everywhere Yet: Conversion Costs and Nonlinearities
6. 22:23Building a Diffusion Model With Light: MNIST Tests and a Billion-Parameter Goal
7. 30:03Optical Computing Q&A: Lightmatter’s HBM Pivot, Digital Backpropagation, and Scaling Limits
8. 34:55Optical Computing Q&A: Wavelength Multiplexing, Optical Memory, and Polarization
9. 39:00Alok Vasudev: Neuromorphic Computing and the Brain’s 20-Watt, Spike-Driven Design
10. 45:23What the Brain Can Teach Us About Chips: Threshold Spikes and Hardware–Software Co-Design
11. 49:35Where Neuromorphic Computing Stands in 2026: D-Matrix, IBM, Intel, and R&D Challenges
12. 54:24Neuromorphic Q&A: D-Matrix, CMOS Coupled Oscillators, and Optics for Interconnects
13. 1:00:00Neuromorphic Computing Q&A: Diamond Substrates, Thermodynamic Computing, and Reproducible Noise
14. 1:04:24Sean Cole: Cortical Labs Uses 59 Channels to Teach Brain Cells to Play Doom
15. 1:07:48Training Brain Cells With PPO: Stochastic Stimulation and a Constrained Decoder
16. 1:12:28How Brain Cells Learn From Feedback: Synchronous Signals and Surprise-Scaled TD Error
17. 1:15:40Biological Computing Q&A: Entropy, Organoids, and the Challenge of Scaling to Billions

What the full analysis includes

17 chapters · 94 key points · 17 section summaries · mind map

The remaining 91 key points, section summaries and mind map are in the full analysis. Analyze this video to see everything →

This video is 1.4 hours. The Free plan analyzes videos up to 30 min; plans from Plus cover videos up to 2 hours.