threads.net/@herval

The amount of Dunning-Kruger on people who use ChatGPT and think they know what they’re doing is *astonishing*

agora ele tem uma oportunidade de virar influencer de turismo com milhoes de followers

Function calling comparison Function calling comparison. GitHub Gist: instantly share code, notes, and snippets.

Sharing the code as well, in case anyone has suggestions on how to make this work!

gist.github.com/herval/e341d...

I actually tried the same thing with multiple frameworks (crewai, langchain & dspy), and the results were similar with them too (some of the frameworks, like crewai, try to inject chain-of-thoughts to make the models get out of their loop, but it doesn't seem to work)

The results:
- Claude got it right every time
- GPT got the response wrong
- tinyllama couldn't figure out how to call the function at all
- mistral, qwen2.5, llama3.3 and deepseek-r1 hallucinated functions that don't exist
- llama3.2 get in an infinite loop, calling the same function forever

I'm using litellm since it helps calling ollama + remote models with the same interface. It also automatically adds instructions around function calling to the system prompt.

Here's a fun benchmark showing the performance of 2 online models and 5 lines of local models, when trying to use function calling.

The test is simple:
- there's a function for listing files in a directory
- the question is simply how many files exist in the current folder + its parent

I don't know if this is just me, but the sheer distance between open source LLM model capabilities and GPT/Claude/Gemini is so ridiculous, I can't see how anyone can use any OSS models IRL. They don't seem to follow instructions, hallucinate wildly and get stuck in repeat loops repeat loops....

É preguiça demais..

bets sao o novo PIB do Brazil

* shocked Pikachu *

Nao necessariamente. Muitas vezes é só contrarianismo mesmo. E convenhamos, tem crença que é mais facil de aceitar que outras - homeopatia é bem menos absurdo que o velhinho de barba branca que mora no céu

This dude was never an entrepreneur in any capacity, shape or form…

Ateismo não é garantia de inteligência

brb asking an LLM to talk like a machine

Please post more here, I still get to twitter from time to time just to read a select few posts. It’s like sticking your hand in a cesspool to fish for a nugget of gold

Summarization and text edits on iOS work pretty well

Destruição presente, tem engarrafamento ate na calçada, nao cabe mais gente na cidade

Fuck Zucc buuuuut at least Threads isn’t a desert town. Follow me there frens?

Is we on the tape

…anyone there?

Sunday in the #Meatverse

Come try Clio! would love to hear what you think 👉 https://clio.so/maker

Applying effects to sketches with controlnet is ✨magical✨

Doomer guy IRL, made with clio.so/studio

It pretty much taught me Python by example

Hello world?