Do we need to integrate AI into Lazarus? lazarus / 21-General ---------------------------------------------------------------------- Marion · 2023-06-19 17:29 Will we need to do this to compete with other development systems? ---------------------------------------------------------------------- Warfley · 2023-06-19 18:22 Tensorflow or Torch or pretty much most ML libraries are just normal dynamic libraries you can use with any language, and there are already existing wrapper for using them in Pascal. That said, personally I don't see any reason to do so, most ML models are conceptionally extremely easy, and basically the code you need to realize them is just some glue code. For example an autoencoder/decoder model (basically the thing that powers all the new generative models like dall-e and the other stuff that is currently so much hyped) is less than 100 lines of python code. It will be much more complex in pascal, because pascal is much more rigid and requires much more boilerplate. For projects this small, it is simply not worth it to use anything more complex than a scripting language. You must put in much more work without gaining anything. Especially as you can simply embed these python scripts into your Pascal program and then call them from there. ---------------------------------------------------------------------- Marion · 2023-06-19 18:26 I was referring to integrating AI into the IDE like Github Copilot to assist in the writing of programs. ---------------------------------------------------------------------- Warfley · 2023-06-19 20:42 I think you already can use GitHub copilot with Lazarus source if you use GitHub. That said, as there is currently lawsuits ongoing about the legality of GitHub copilot, better to stay away from int until this is settled ---------------------------------------------------------------------- Marion · 2023-06-20 22:05 Quote from: Warfley on June 19, 2023, 10:42:22 pmI think you already can use GitHub copilot with Lazarus source if you use GitHub. That said, as there is currently lawsuits ongoing about the legality of GitHub copilot, better to stay away from int until this is settled I forgot about the lawsuit. ---------------------------------------------------------------------- SymbolicFrank · 2023-06-21 08:30 The nodes and layers are actually simple, and we already have an extensive library for those, written in FPC, here. The thing you would want to integrate into an IDE are the Big Data tools needed. For an AI, you need to make a suite of predefined "unit tests", to be able to train it. Which require "mappable" inputs. Or, put differently: they need to be linear graphs. As an example, let's say you want to use an AI to help your customers to find the right pair of shoes for them. You ask the customer their gender, the length of their feet and the intended use: running, walking or general use. Then again, short people tend to prefer thick soles / high heels. Ask for their height. And the shoe size has an implicit parameter: the width. While the shoe size has a direct relation to the length of your feet, the width isn't as defined. Most feet have the form of an average shoe, instead of the other way around. If you get older, or generally wear slippers, the form of your feet changes. If you get older, you get flatter feet, which are wider in the middle. If you wear slippers, your toes expand. So, all of a sudden, that single parameter, "shoe size", has two additional components, which aren't clearly noted on the product specification. And which is why many people have to try the fit of their shoes physically. So, the success rate of your AI is determined by two "unknowns". Anyway, a successful AI IDE would allow you to extract the required data from many sources, do that data analysis and tabulate all the parameters you need in a linear fashion. Even when processing images, you first have to pre-process them like that: (https://www.simplilearn.com/ice9/free_resources_article_thumb/CNN_rec ognizes_a_bird1.png) And lastly, define the kind of layers and interconnects to turn it into a usable AI.