3 ms·
IME llama et all require LoRA or fine-tuning to be usable. That's their real value vs closed source massive models, and their small size makes this possible, ap
by yunohn 7mo ago
IME llama et all require LoRA or fine-tuning to be usable. That's their real value vs closed source massive models, and their small size makes this possible, appealing, and doable on a recurring basis as things evolve. Again, rendering ASICs useless.
- fxnn 7mo agoRead the blog post. It mentions that their chip has a small SRAM which can store LoRA.
- yunohn 7mo agoNeither the blog nor Taalas' original post specify what speed to expect when using the SRAM in conjunction with the baked-in weights? To be taken seriously, that is really necessary to explain in detail, than a passing mention.