YC Paper Club: what comes after LLM + GPU? (#AI)
An interesting video for anyone wondering what comes after LLM + GPU :) YC Paper Club brought together a discussion of optical computing, neuromorphic chips, and living neurons. The four participants offer a shared perspective: the next leap in AI efficiency may require different hardware, a different model architecture, and a different way of learning at the same time.
1️⃣ François Chaubard, the host from Focal Systems, begins with the observation that GPUs, backpropagation, and modern models evolved together. An algorithm that looks poor on a GPU might therefore become interesting when operations have a different cost.
He studies learning without a backward pass: slightly change weights in a random direction, compare results, and decide where to move. Large models require too many such trials. His SOMA approach splits the model into small experts to reduce estimation noise. The next bet is that if forward passes become extremely cheap, as they might in optics, the economics of this training will change. Chaubard’s prediction that backpropagation will be abandoned in ten years should be understood as a research bet.
2️⃣ Ilker Oguz, an optical computing researcher from EPFL who is now a Stanford postdoc, shows where this idea meets physics. Light passes through optical elements and performs the required transformation. Around it remain memory, calibration, nonlinear operations, and conversions between electrical and optical signals. These can consume the benefit.
In his image diffusion experiment, the algorithm was adapted to optics: the same weights are reused over successive steps. A device limitation thus becomes an algorithm design principle. The demonstration is currently limited to MNIST and Fashion-MNIST. Its energy savings estimate assumes passive, fixed elements, while the prototype uses a programmable modulator.
3️⃣ Alok Vasudev, an engineer by training and a partner at the venture firm Standard, adds useful skepticism about attempts to copy the brain. Its structure and information processing evolved together. Moving an individual mechanism onto silicon does not automatically transfer its efficiency.
In his assessment, approaches that bring memory and computation closer together, reducing data movement, are nearer to practical use. More literal reproduction of brain function remains a research direction. A good question for any neuromorphic project is which specific limitation biological similarity helps overcome.
4️⃣ Sean Cole, who developed algorithms for the experiment with Cortical Labs, takes the discussion to living cells controlling Doom. A digital encoder turns game state into electrical stimulation; the cells respond with spikes; a decoder translates those into actions.
Here is the most interesting detail. According to Cole, an early decoder was so powerful that it could play even with cell signals zeroed out. The team reduced it to a simple linear readout and ran checks to establish that neuronal signals were actually needed. Otherwise, one can produce an impressive demonstration in which an ordinary computer does all the useful work.
Overall, this is an interesting video about research outside today’s mainstream that has intriguing prospects.
#AI #Hardware #Research #Engineering