Ask HN: Are all these new LLM/AI features just bloat?
6 comments
Some implementations are bloat (especially LangChain), but in general the "AI can use tools" approach is not bloat.
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’.)
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.
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.)
(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.)
No, its not bloat.
A lot of it is experimentation that will be dead ends as useful techniques get proven and less useful ones filtered out, but that’s normal with new technology and how the best approaches grt established.
A lot of it is experimentation that will be dead ends as useful techniques get proven and less useful ones filtered out, but that’s normal with new technology and how the best approaches grt established.
It's too early to call it bloat. I think what's happening is that people are experimenting in all kinds of directions right now and only some of them will end up being useful long term.
I use OpenAI's official chat & API - but nothing else. To those that are making use of new features in existing software, or picking up new tools: what are you using, and how has it affected your day-to-day?