Show Post

i love being like "hey little dude can you get <model> running on this pc?" and i come back thirty minutes later and somehow astra has fit some unbelievable model onto my 10GB VRAM card

Show Post

“Intent”, then, is just a corollary of this. Giving an LLM a prompt introduces a goal state in its world model; it then acts to meet that goal, which includes taking actions in the real world — fetching websites, writing files, executing programs. If this isn’t intent, what is it?

Show Post

Finally, practical support: You can give an instruction to an LLM in whatever language or dialect you want. In some cases, it will follow the instruction. In some cases, it will ask for clarification. “Understanding” is simply a useful word for describing the difference between these two cases.

Show Post

Likewise, if your training material contains conversations that proceed in a certain social context, your model learns about the “world” of that social context. In particular, the model understands that it’s supposed to be an “assistant” to the “user”, because it knows the context of those words.

Show Post