3 ms·

Some implementations are bloat (especially LangChain), but in general the "AI can use tools" approach is not bloat.

• PaulHoule 3y ago
  LangChain strikes me as like SpaCY, released too early and with flaws that will keep it always a bridesmaid and never a bride. So many libraries in this space (like ones I have written on my own account and at work) have serious flaws in their design that are partially out of inexperience, partially intrinsic to the domains and most of all rooted in the difficulties of making systems that are research products and in production at the same time. (not a contradiction operationally because this is the essence of ‘continuous improvement’.)
  • minimaxir 3y ago
    spaCy is fine from a software perspective, the reason it's become less popular is because you can now YOLO with pretrained transformer models (e.g. via Hugging Face transformers) and get even better results, no NLP shennanigans/technical debt needed.
    • PaulHoule 3y ago
      I worked in the field professionally a while back and every conversation about SpaCY would start with some design defect that would leave F1 points in the table we could never get back. It was poorly thought through from end to end, quite literally there was no thinking about end to end just about having an exhaustive list of half-baked almost working ‘features’ and no consideration of how you’d build systems if they were really supposed to work. (For instance, SpaCY was obsessed with word embeddings which in my mind are 95% problem and 5% solution. Stanford still has that embarrassing landing page for GloVE embeddings that shows separations you could get just as well from random embeddings.)