ai tech is advancing so rapidly, i just looked into the latest tech and there is lora+peft for easy fine tuning of larger models, image to video models, chatgpt quality local coding (deepseek). You blink once and you miss tons of stuff.
Login to reply
Replies (7)
I spent yesterday replacing all my regex output parsing with forced Json output
Basically a probability mask to force tokens to json
Sometimes model would screw me by ignoring my <parsing>tags</parsing>
Small models like Llama3.2-3B + Zed.dev editor work like a charm!
Agreed, I’m so into this technology that sometimes I forget how I used to do things before AI 😂
everybody is looking for the next smartest model. not many talking about curating the datasets. the smartest is qwen 2.5 right now. but it is terrible in terms of human alignment. best alignment: llama 3.1 after of course ostrich 😆
The /lmg/ thread on 4chan's /g/ board will keep you up to date.
heard good things about qwen, going to try it out
The agenda is to replace humans. Too many humans aren’t paying enough attention.