spacy
@dosco
building ax, dspy in typescript. researching systems of llms and compilers + llms. proud vibe coder
we're on track to building the best dspy powered llm framework in the typescript world
Yes — @dosco’s Ax is pretty much the official DSPy in TS.
besides too much group think i feel the traditional startup grind is why we're getting so many people building the same stuff with nothing really shining at the ai app level.
in the age of ai, ideas are more valuable and coming up with good ideas needs space to think and engage a problem.
ok to get mipro working well on ax i'm gonna have to offload the tpe bayesian optimizer to a seperate python service no way around it.
read accelerando it's a crash course in posthuman economics. ai coporations, ai cats manipulating humans, mind uploads, best book i've read in a while
i've said this before you can achive this with a dspy program
Can we improve Llama 3’s reasoning abilities through post-training only? Introducing ASTRO, our new framework that teaches LLMs to perform in-context search and generate long CoT to solve math problems, via SFT and RL. Work done at @aiatmeta. 📄 Paper: arxiv.org/abs/2507.00417
when distilling from a teacher model like with dspy optimizers using a faster cheaper model can be better than using a big expensive one. you can even use a 7b to tune a 27b.
reasoning models like those used for math challenges are searching and verifying you could do the same with a big non reasoning model and a dspy program, it might even work better