I’ve been showing clients how to use ChatGPT to help write code (for Anvil, and other systems). Here’s a good article which explains how the author integrates Copilot and other LLMs into workflows to improve productivity:
Is it within the scope of anvil.works’ resources to train and fine tune an LLM (likely one of the forks of LLaMa) on a broad dataset of Anvil code? I expect many users would find an integrated ‘AnvilPilot AI’ a genuine productivity booster - perhaps one which would make Anvil much more appealing to new users in the future. Is this something that would be a good community project to get started?
While I agree that it will be really cool to see in the future, the model can only be trained on publicly available code (official Tutorials and clone links posted in the forum) and that will not be a very large dataset.
Although it may reduce posts asking help on something that is already there in the forum.
I’d think it’d need to be trained on a set of “best practices” examples.
Anvil changes frequently enough, I wonder how much of a hassle it’d be to retrain on major changes (e.g. accelerated tables changed best practices for efficiency)?
This feature is already integrated in FlutterFlow: FlutterFlow AI Gen. My hope is that Anvil continues to display a competitive position to new users, when compared against products such as Flutterflow, Microsoft Power Apps, etc.
Using AI to write code is a current buzz, and I expect public interest in that topic to continue growing, to the point that not having integrated tools might be a potential stopper for many new users who’d otherwise get started with the Anvil platform (I expect in a few years that any platform without AI support may begin to appear like a dinosaur).
In the near future, it may be difficult to compete with code generation advances by OpenAI, so perhaps it would be worth considering the possibility of a ChatGPT plugin (with usage fees charged to the user…).
Of course, anyone can just interact with ChatGPT, so maybe the idea of creating an integrated solution is simply not important. I’m just thinking about the power of the marketing affect and the appearance of usability when compared to other tools (of course AI can actually be helpful in writing code, but at this point, my thoughts are all about hoping that the Anvil community continues to grow and that the company’s financial position continues to gain strength).
Perhaps a series of blog posts about using ChatGPT to help create Anvil apps could produce the same marketing effect, and appeal to new users. Perhaps creating content like that will naturally evolve as a community driven effort.
One other angle, hopefully not too far off topic: perhaps demonstrating Anvil’s capability to create apps that involve some AI toolchain is something that could be written about more. There’s a lot of buzz about that in the Streamlit (and of course the general Python) community, for example:
I’ll settle for Copilot X in VSCode if it is reasonably priced. I know some people have hacked together plugins for ChatGPT in their IDE’s but there’s something to be said for the experience of Copilot.
The most feasible way to implement this into the Anvil Editor is to work with GitHub/Microsoft directly to build an integration, as it doesn’t look like Copilot has an open platform for integrating with IDEs. Otherwise every semi-popular IDE would already have the integration. I imagine competitors to Copilot will spring up that are more “open”.
I do like that firenvim thing, that’s clever!
A note about training: the beauty of LLMs like this is that they’re trained on so much Python code that they only need a few Anvil app examples to become useful.
ChatGPT isn’t nearly as good with Anvil yet as it is with, for example, Flask (or Bottle) + SQLalchemy (or PyDAL). The tools for building self hosted LLM document seach are improving quickly:
