One good thing I have got out of my daily writing and social activity recently is an increased sense of continuity, thinking more about how events connect to each other, finding things worth following up on and refering back to. This whole recent AI/large language model hoopla is a great example. Today, we released an episode about Kodsnack in large part about what "AI" tools might and might not do for us. It was recorded two weeks ago, and inbetween recording and release Microsofts' new Bing chat bot exploded all over the news, and Google apparently felt pressured to announce their own competitor as well.
All of this of course felt like it connected to the episode, and to the increasing flow of AI related skepticism in my Mastodon feed. Thanks to writing about all of this a couple of times, I got an addendum to the episode recorded where I could comment a bit on recent events - mostly by asking more questions - and the whole thing was much more coherent than if I had just walked around with vague ideas in my head for two weeks.
My own trajectory right now is definitely one of decreasing excitement for positive new tools being created, and decreasing worry about the world being flooded in generated information, the world needing dramatic adjustments, and so on. And today, I came across (again through Mastodon) another great piece to connect to the others. I came across a video presentation of a paper - "On the dangers of stochastic parrots: can language models be too big?" - which very concisely listed problems and limitations of our current "AI" tools. It not only very clearly talked about things like bias and complete lack of actual knowledge, but also truly interesting questions such as power consumption, unequal benefits, and if it is actually even worth investing so much effort in the current approaches considering their limitations.
It is so refreshing to hear someone plainly say that perhaps it is worth considering if we could reach a similar or better result by doing something else entirely.
Now, here is the real kicker: That presentation sounded as if it was created today, reacting to today's exact problems. But that is of course not how academic paper writing works, these things take time. No, the presentation was from May 5th, 2021! Which of course means that the paper was actually being written long before that, and that everyone who is building our current models could and should have been aware of the whole set of problems long ago.
Some probably were. Those people were probably not all that much in power.
And so here we are, a whole world discovering that, indeed, AI models do not actually know anything, and that when we see intelligence or creativity, it truly is us reading those traits into mindless text generated for us.
The whole presentation is just 21 minutes and well worth watching all the way through, but the last six minutes contain the very juciest summary of problems and questions.
(As a side note, it is fun and annoying to see Youtube's automatic subtitles completely fall over itself pretty much every time a person is named. I think one paper was credited to "Two women, at all" in the subtitles.)
(Stochastic parrot, now there is an expression I want to use as an insult.)
Voquest - the reasonable approach
Reading about my voquest, a friend went and threw reason into my ranting: "Why not just ask Focusrite?"
Well damn, I should try that. Updating my list.