Tilde Lowengrimm on Nostr: If a transformer model could take a picture of the schedule I was given as a handout, ...

If a transformer model could take a picture of the schedule I was given as a handout, or the conference web page with all the talks on it and convert that into a collection of events on my calendar, I think that would also be a reasonable use of the thing. But they can’t, at least not reliably and generally. If I set a generative model to do this, I’d want to check its work before relying on it, and that massively detracts from its usefulness. I have no doubt that someone could tweak a transformer to be good at this, but I’d still have less confidence in its accuracy than a piece of regular software that someone spent the same effort writing for the same task.

And that’s basically the thing about these models: they can kinda perform any task, but so unreliably that you’d be better off doing it yourself. And they can be carefully tweaked to do one thing reasonably well. And simply use three orders of magnitude more compute to do the same thing you could probably have achieved more reliably by writing a piece of software you actually understand and can run on an old cellphone.