[giancarlostoro]

This is the kind of project I really want to make, especially now that LLMs are very capable, even the small ones. Just need a good computer to run the model off of. I wonder what happened to Mycroft? I looked forward to seeing more of it.

Mycroft was an open source alternative to Alexa.

https://github.com/mycroftai

[giancarlostoro]

I cannot edit my original comment, I posted a thread on HN apparently Mycroft AI was killed by a patent troll that drained their funds with legal fees. Patent trolling should really have legal consequences for suing people while not actually providing value to society.

https://news.ycombinator.com/item?id=47678354

[stavros]

This is why we can't have nice things.

[juancn]

You don't need that much for simpler models. Some can even run on a PI.

Qwen3 and Gemma models are fairly capable, they are slow-ish (a few tokens per second) but will run.

You can start building with cheap hardware and simple models, and use something more capable once you're more confident on the use case.

[kukanani]

Local inference is definitely a good way to go here. Latency when talking to an embodied robot is extremely noticeable though, and pauses during voice chats are way worse than during text chats.

It’s something I’m exploring - stay tuned :)