08:19 <o0x1eef> OpenAI "tool calls" are not at all what I thought but also pretty cool. I can see how you could design agents with this API. The biggest surprise is that it is an API for calling code that's local to you rather than OpenAI itself. OpenAI just parses natural language into what is essentially a function call, and then says "hey, call your function X with args Y and Z".

11:58 <adam12> o0x1eef: I have a client right now using OpenAI tool calling to do some intersting things. I had never heard of it until I joined on with them.

12:44 <o0x1eef> adam12: It has got me really excited

12:46 <o0x1eef> This is what I built with it so far -- a tool that can run any shell command via natural language: https://github.com/llmrb/llm/commit/1fcabd201f357d055a35ecaf0e25ad6f03df9fb6

12:48 <o0x1eef> I think in theory you could probably fake it on Gemini with JSON Schema as well

13:35 <adam12> o0x1eef: Oh cool.

16:13 <o0x1eef> https://ollama.com/blog/tool-support

16:13 <o0x1eef> Nice

17:22 <havenwood> o0x1eef: LLM is looking nice! IMHO, switch from venture-backed Ollama to llama-cpp and llama-swap.

17:23 <havenwood> Ollama is a thin wrapper and adds little for me except a Golang runtime I'd just as well avoid. YMMV.

17:24 <o0x1eef> That's nice. I didn't know that. Agreed.

17:26 <o0x1eef> Wow, it even supports audio: https://github.com/mostlygeek/llama-swap

17:31 <havenwood> o0x1eef: I haven't missed a thing from Ollama and have liked a lot of odds and ends from both llama-cpp and llama-swap.

17:31 <havenwood> Also upstream and open source, so seems like wins all around.

17:32 <o0x1eef> Yeah, I am definitely going to look into this more. I'll add an issue to my local gitea instance. It is not entirely clear what changes I will need to make, sounds like I might need separate provider(s) but since there's some OpenAI compatibility it might be relatively easy