Conference report

If this reaches you at a weird time compared to the usual it is because I am on San Francisco time. I am having a good time over here. I can't quite put into words how odd and contradictory my experience of San Francisco as a city is so far. So I will leave it at odd. I'm game for an odd experience though so it is fine by me.

This is the most flavorful and unusual venue I've been to for a Code Sync-organised conference. Pretty, old building. Lots of military memorabilia and walls to honor the fallen. It continues the theme of odd and I appreciate the weirdness. Coming from a culture that barely admits to having military this is quite different. As per usual things run smoothly when Code Sync organises, they are professionals and nice people to boot.

I felt my talk went well. It is not the most ambitious talk I've given considering I had about a week to pin down the demo and two days to craft a journey to take folks on. I'm happy with the laughs I got and the feedback has been good. If you want the repo for the demo it is here. No docs though. I believe they'll post recordings later.

The starting keynote was interesting. Sean Moriarty of Nx fame spoke on LLMs and some of my major takeaways were that the chatbox is probably not the interface. I am curious if his idea around DSLs being the answer will check out. But it did lead to a brainwave I have around NervesHub and Ash Framework.

From speaking to Zach Daniel (Ash creator) and Josh Price (Alembic, sponsoring the Ash work) I am fairly confident that the most practical and well built-out way of exposing Elixir application functionality to an AI model is by using Ash. We will not be whole-sale adopting Ash for NervesHub anytime soon. We don't have experience with it and we have a lot of existing code.

We do have incentives to explore AI possibilities for NervesHub. For one thing I'm sure it would help get funding for NervesCloud if we have an AI story. That's a non-zero factor. There is also some interest in exploring this from the SmartRent side. They would benefit a lot from anything that lowers support burden and helps discover problems for their many, many devices. So this is on my agenda.

I think the answer to any question about "Can we make this better with AI?" needs to be "Don't know". Most applications slap an LLM in, call it good and are pretty terrible. I've seen decent implementations that are helpful, going back to Sean's point. Ideally this is not a chat box. You need to try things and see what works, how it works and if it is useful. Investing deep in tooling up the data and the interface surface area for just one specific thing is a lot of risk and you'll immediately start feeling sunk costs if it doesn't seem to be working. I think that's how people end up shipping a writing assistant even if it sucks.

The fundamental part of being able to do anything with an LLM in a codebase is to have a schema that you can offer it to make tool calls from or a dataset to ask questions about. Ash can generate schemas from your existing Postgres database. I plan to use this for prototyping. I can derive Ash Resources from my Postgres database meaning I get a more detailed schema than Ecto can offer and I can patch in policies (probably using Igniter) for access control and such and this essentially becomes a compile step. I'll likely git ignore the entire Ash resources chunk of code while trying this. Just re-generate it when I need it.

Fundamentally I want to see what kind of questions I could make the model answer about the system and how I can ask those questions in a smart way that doesn't require the user to type a question. I don't know what's feasible. I don't know what would work well.  I'm not convinced we can make something that provides meaningful value. Other people want me to figure this out and I'd rather be able to make a lot of lightweight experiments with a little code than commit hard to something that turns out to not matter. Zach has explored the general parts with ash_ai already. It'll get MCP and support for model providers I'm sure.

I've been poking at that a little bit in the breaks and Zach has been Just-in-Time-fixing bugs I run into as I'm using it in an atypical way. Good fun :)

Digit did his presentation on Flutter running on top of Nerves and publicized nerves_flutter_support. Super fun talk :)

Alright, back to the people. Hope you have a lovely weekend.

Questions welcome. You can get to me on the Fediverse where I'm @lawik@fosstodon.org or by responding to this email to lars@underjord.io.

Thank you for reading. I appreciate you.

The Elixir shirt pre-orders will ship sooooooon oswag.org.