GitHub

Tool call support (generic + native for Llama, Functionary, Hermes, Mistral, Firefunction, DeepSeek) w/ lazy grammars by ochafik · Pull Request #9639 · ggml-org/llama.cpp

This supersedes #6389 (now using a fully C++ approach), #5695 (first attempt at supporting Functionary) and #9592 (more recent Python wrapper). Whi...

Llama.cpp now allow DeepSeek do tool call :D 💦

If you are using self deployed DeepSeek remember to add system prompt 😂 it is really really really different Reference here

DeepSeek-R1 System Prompt

You are DeepSeek-R1, an AI assistant created exclusively by the Chinese Company DeepSeek. You'll provide helpful, harmless, and detailed responses ...

My nostr timeline provide more signal of AI stuff than Twitter now 😂 very nice

I’m re-thinking what kind of AI infrastructure for nostr could be a thing Is there a place that those smart head already start discussing

What is mcp server

The word open source actually is slightly different when used in bitcoin world and AI world in #bitcoin, open source means you can download the source code, run it in your own machine, and everytime it runs, no matter where, it should have the same result. (think about it, it will be a huge desaster if one bitcoin address could have different amount of satoshis) however, in #AI world, when an LLM claims to be open source, it usually means you can download the model parameters for free, and you can use it in your own project. does it means it have the same level of transparency than other open source proejects having a long history like linux or bitcoin core? no. you could image AI model parameters as the result of data + training algorithms + training parameters. all of them could end up with different results. and usually, for those models who chalimed to be open source, they only open source the find result, aka, the parameters. One of the reason that people are crazy about deepseek is that it opens up many details during training process, which is confirmed by open ai scientists that they are using same method(but open ai keep it as a secret). However, we still dont know what kind of data they are using during training, how they process the data, and will it have any, (i dont know how to use strong words), lets say, political bias, so that the model can influence the society to the direction the creator wants. of course, technically, it is possible to train a model from scratch with uncensorshiped data, but it costs at least 5 million dollars for one batch.

How are you doing 😂 View quoted note →

I started make TikTok video 🤣 View quoted note →

The most important thing for me is freedom 😳 And it is necessity to check my life regularly and think how much I’m willing to sacrifice

It’s 1 am and I weirdly cannot sleep Maybe we could have a #DeepSeek fnq I’m not an expert but I read their paper Anything want to know? #ai

I like the feeling tomorrow I have 1234 todo

Btw who else is at chaincode program now 🤣😂 Does you guys want to create a small groups? Just discussion and fun View quoted note →

Today’s mocha !

Whops maybe I made a mistake 😂😂😂😂😂

most of my software engineering friend have a thoughts of deterministic thinking , all error can be traced etc Llm is not, it could be break and it could be tricked and it can print wrong statement