Haven't seen this posted on here, no idea why
So Taalus are working to create LLMs that run directly on chips. Basically, the chip IS the LLM
The website (chat Jimmy) is using llama 3.5 so it's an old model, out of date already, but it's a proof of concept. It's lightning fast (try it out, all it to make a poem or something) and Taalus claim it's 20x cheaper compute than Nvidian B300 chips. Also the chips themselves are cheaper but that's besides the point, the cost of running AI is in the compute
Taalus have raised about $200m, they're private but I expect whoever buys them to be a big winner in the AI arms race. Thoughts?