3 ms·

Coming from a frontend background this reminds me a lot of the frontend situation some years ago when it was super common to npm install the stupidest pointless

Coming from a frontend background this reminds me a lot of the frontend situation some years ago when it was super common to npm install the stupidest pointless packages. Of course some people still do, but hard lessons were learned and all experienced people I know are a lot more mindful and cautious what dependencies they add. It seems to me that this space, and maybe data science more broadly, is currently in that situation. Maybe it's the lack of coding skills, maybe it's the transition from one-off research notebooks to production applications, or maybe it's just the norm to have big do it all libraries (like pandas, scikit etc), idk, but I expect the same lessons will be learned. For Langchain I wonder how they will keep up with changes in all the things they have their abstractions on. I guess having a huge community helps, but that doesn't help you with compatibility. It would not surprise me if that will get really ugly.

ShamelessC 3y ago
The problem is that coders are used to dealing with code. GPT-4 is a robust processor for semantics as conveyed by strings of characters. The whole point is that you don’t need code. You just ask it what you want. But programmers love to think they can still make improvements to such a system using code. In reality, any improvements that can be made are the responsibility of f-strings and/or a templating mechanism and even that may be overkill in many cases.
- codeptualize 3y ago
  Often there is still a lot of programming to do, as a lot of the use cases involve some sort of integration into a bigger systems, or you might have to process larger quantities of data which gets you into queues and for example managing rate limits. That said I 100% agree that often the basic tools are sufficient to solve these problems, and where it gets complex there are many battle tested solutions.

ShamelessC 3y ago

The problem is that coders are used to dealing with code. GPT-4 is a robust processor for semantics as conveyed by strings of characters. The whole point is that you don’t need code. You just ask it what you want. But programmers love to think they can still make improvements to such a system using code. In reality, any improvements that can be made are the responsibility of f-strings and/or a templating mechanism and even that may be overkill in many cases.

codeptualize 3y ago
Often there is still a lot of programming to do, as a lot of the use cases involve some sort of integration into a bigger systems, or you might have to process larger quantities of data which gets you into queues and for example managing rate limits. That said I 100% agree that often the basic tools are sufficient to solve these problems, and where it gets complex there are many battle tested solutions.