There's this Google internal document, for example, that points out FLOSS community is close to eating Google's and OpenAI's cake:
ttps://www.semianalysis.com/p/google-we-have-no-moat-and-neither
So here is my question to you:
What are the best examples of useful, small, on-device models already out there?
Context: I am writing something about #AI and I want to show examples of useful small models in order to help non-techies understand why they are important.
@rysiek Yess, this so much! There's actually use in AI (despite me hating almost any AI tool out there). There's a huge perspective around personal empowerment and independence. Incidentally, those tools are usually also more ethical, IMHO, when it comes to how they're built or used (self promo https://ljrk.codeberg.page/ethical-ai.html)
Unless the training data are also free and the production models used to perform the work are substantially reproducible, then those models and the code to deploy them propagate power imbalances as fraught as proprietary object code: The model acts as a form of unauditable, AI-authored code.
Some folks in Debian did some work years ago, now, to try to parse out these issues, but scant notice has been afforded these challenges otherwise.
As a practical matter, the ability to bring enough computing to bear on running the training data through the training software offers a similar, but more quantitative rather than qualitative challenge to the abilities to make forks asserting typical software freedoms.
to be clear, by "scant notice" I mean in a FOSS-forward framing.
There is of course vigorous criticism of the misuses of AI. I wish not to contribute to how badly these have been disregarded more broadly, including but also beyond that frame.
Now I am trying to find examples of smaller models that do as well or better than LLMs from Big Tech, to demonstrate the point that we might be able to do just fine without LLMs, regardless of what Big Tech is trying to convince us of.
(incidentally, speaking of language models, free or not, DeepL did much better rendering your abstract into English, as accessed through the F-Droid-provided free-frontend-to-proprietary-backend than when I pasted it into LibreText's web form.)
@rysiek It's pretty neat how much can be done in WASM. Like, lichess.org embeds very strong in-browser chess position analysis with Stockfish.js, which can do both classical evaluation and evaluation with an NNUE.
@rysiek Apple's photo recognition too. It's not great, but it's steadily getting better — it can now do an OK job of recognizing birds and hedgehogs, for example.
@rysiek Are we only talking about the large language models as "AI" or are we following the trend of rebranding everything that was "ML" last year to be included in this year's new "AI" hype category?
@rysiek As far as I can tell Facebook's "LLaMA" is the biggest open-sourced language model that people are running various variants of on their home GPUs.
They remain a black-box of course, nobody has any idea what's going on inside any of those multi-billion parameter confusions of virtual wires and matrices.
But the model is freely redistrible and you can read the value of every node at every microsecond if you want.
Just that nobody knows what any of them mean.
Sad that it's Facebook, but FB are pretty good at Open-Source software libraries.
Add comment