In flight:

• ElixirConf EU talk
  • Final revisions are in. Very final. Surely.
  • Rehearsing, rehearsing, rehearsing.
• ElixirConf trip:
  • Team swag arrived. I have some extras. Need to decide how much we can bring.
  • Flying out sunday morning.
  • If you are going and are a reader, let me know :) I plan on hitting the Meetup a few days before the conf. Find me there or at the conference and just go up and demand stickers. I have many stickers! I humbly request interactions with y'all.
  • I will only film and interview you if you want to :D

Constant changes

I have been exploring a fair bit recently. Eventually I will exploit what I learn for work purposes, right now it mostly gives me more answers when stakeholders ask about the current hype topics.

We can start with ML. I think I've mentioned that I've poked around with Stable Diffusion, Whisper and some other Elixir-friendly utility models. I recently also have tried to run some of the more accessible LLM type deals. A downscaled Cerebras and Flan T5 (which is in Bumblebee) were some I got up and running. I tried some Llama variant and it did language well but was weird to deal with. Overall I've compared this to what I can get out of ChatGPT and so far that's night and day. ChatGPT comes off as immensely competent compared to most others I've tried. Llama seems like the competitor and whoever can pull of a license-unencumbered Llama variant will likely take the crown as the text generation model for self-hosting and retraining.

ChatGPT is neat. I'm very skeptical about how it has been made, I don't trust it. I'm well aware of the problems and criticisms. It is also impressive tech. Those things are not in opposition, they are in tension. I don't trust OpenAI further than I can throw them but they've built something interesting and compelling with many potential uses and a lot of weird risks.

Still on the waiting list to access GPT-4 via API.

I would happily help a client explore it because it is kind of fun and the APIs are very reasonable. For my own stuff, for anything serious, I'd definitely go with whatever I can self-host. Honestly, business-wise, I'm thinking about whether this is a time to gear up to sell shovels in what is definitely a gold rush. That's not strictly cynicism. I think people will find useful and profitable ways of using LLM and diffusers. I'm not very interested in building those companies myself but I don't mind working with it. Maybe I'll learn math!

I'm learning Rust for some of the work I do with Electric SQL. Looks like I'll be implementing some aspects of CRDT logic and making their tooling more cross-platform. Rust is interesting and a good complement to Elixir. CRDTs are interesting and a good complement to .. LiveView, I think. Not these. These are for SQL and probably quite specific for what this system wants to do. CRDTs in general are interesting though.

I've poked around with Automerge 2.x which was implemented in Rust to be useful across languages. I hope to eventually produce a Rustler NIF library for it. I'm curious if that plus the WASM version for the browser could be used to do some interesting local-first-ish work with LiveView. I like LiveView a lot but I also appreciate the goals of local-first a lot and those are currently not compatible. That's fine but it makes me think.

Automerge also touts supposedly good performance for text collaboration which is very interesting to me and should put it on many people's radar. OT (Operational Transformation) which always seems like the "simple solution" for collaboration tends to get quite painful every time I try to use it. I might just understand it poorly or the libraries just never fit my brain. People tell me OT is corner-case city as well which is less than appealing.

Additionally I've teed up Mimic3 from Mimicry as something I want to try for text to speech (TTS). Whisper solves STT well enough for my needs. With TTS I could do a talking assistant kind of deal which would be very fun to DIY. I'd like a decent self-hosted LLM to try and make it turn natural language text into a set of available commands. I could very likely get that with the ChatGPT API but that's what I don't want.

I've also read up on what the options are for face tracking. This is so that I could experiment with image and video visuals in some fun ways. Deepface (multiple models) and mediapipe face mesh both seem to be python-centric and I'd likely need to shell out for the immediate future. I could see Bumblebee getting something like Facenet (one of the models used in deepface) eventually but someone has to spend the time making it accessible in that way.

A bunch of the thinking that leads to me exploring these options will be in my talk. It is all about exploring and experimenting. I consider it a key part in how I build skill and knowledge.

I get that not everyone has time for experimentation. So the question is, what are you experimenting with currently OR what would you want to experiment with if you could? I appreciate getting your thoughts in email or federated at me via lars@underjord.io or as @lawik@fosstodon.org.

Thanks for reading. I appreciate it.