| LLMs and neural nets from first principles: https://arxiv.org/pdf/2404.17625 (pdf) https://news.ycombinator.com/item?id=40408880 (llama3 implementation) https://news.ycombinator.com/item?id=40417568 (my comment on llama3 with breadcrumbs) Admittedly, I'm way behind on how this translates to software on the newest video cards. Part of that is that I don't like the emphasis on GPUs. We're only seeing the SIMD side of deep learning with large matrices and tensors. But there are at least a dozen machine learning approaches that are being neglected, mainly genetic algorithms. Which means that we're perhaps focused too much on implementations and not on core algorithms. It would be like trying to study physics without change of coordinates, Lorentz transformations or calculus. Lots of trees but no forest. To get back to rapid application development in machine learning, I'd like to see a 1000+ core, 1+ GHz CPU with 16+ GBs of core-local ram for under $1000 so that we don't have to manually transpile our algorithms to GPU code. That should have arrived around 2010 but the mobile bubble derailed desktop computing. Today it should be more like 10,000+ cores for that price at current transistor counts, increasing by a factor of about 100 each decade by what's left of Moore's law. We also need better languages. Something like a hybrid of Erlang and Go with always-on auto-parallelization to run our human-readable but embarrassingly parallel code. Short of that, there might be an opportunity to write a transpiler that converts C-style imperative or functional code to existing GPU code like CUDA (MIMD -> SIMD). Julia is the only language I know of even trying to do this. Those are the areas where real work is needed to democratize AI, that SWEs like us may never be able to work on while we're too busy making rent. And the big players like OpenAI and Nvidia have no incentive to pursue them and disrupt themselves. Maybe someone can find a challenging profit where I only see disillusionment, and finally deliver UBI or at least stuff like 3D printed robots that can deliver the resources we need outside of a rigged economy. |