1/ TIL @Microsoft has been building 1 Bit LLMs. I know right! WTF are 1 bit LLMS!!! React or comment if you didn't know the following! I was like what are 1 Bit LLMs??? basically all weights in models are numbers (floats) can be 16 bits or 32 bits (or 8, or 4 etc) essentially everything in AI can be 'quantized' down to these smaller bits so computation in easier and storage is cheaper this was known and is supported by large scale AI solutions like vector databases but I didn't realize that when you do this to 1 bit the math just becomes addition meaning we don't need all these fancy FLOPS and dRAM hacks that @nvidia @GroqInc are pioneering we just need chips that can add 1's and 0's fast... once we figure out these 1 bit LLMs either GPUs are gonna be worthless and the market will crash or simpler chips will be made and AI is going to be more efficient than ever HUGE UNLOCKS with contemporary tech it started with AirLLM, now bitnet.cpp, it's gonna be crazy Just bought a @Apple Macbook pro M4Max 48gb unified I hope its both.