They told me it would be able to ‘write Shakespeare’ all it needed was more training data, it would be ‘keeping old people company.’ All use cases (which are real) I listed are still the only ones. But I got the onslaught for merely pointing out they don’t know why it hallucinates or how to stop it.
I think it's moving asymptotically towards some maximum accuracy, but its boosters imagined that would naturally be 100%, whereas empirical evidence shows it to be more like 95% - just wrong enough to be more of an impediment than a help.
But sure—there are developments that are impressive in certain ways but for very discrete tasks. The failure rate is not even biggest problem with it—it’s just glaringly revealing of an absurd gap between rhetoric and reality and the culture of faith tech companies managed to create.
I think it's also revealing giant (and fascinating!) chasms in our understanding of what "intelligence" and "thought" are. To say nothing of "memory" and "consciousness."
At least I believe in the AI research realm researchers were aware these are all very complex things, and there is much we still need to know about them? Maybe not!
I think the researchers are well aware of the gaps. Less so the VCs, and I think there are a lot of non-technical tech industry people who have managed to convince themselves that they are actually inventing God, and so pouring money into this parlor truck is the Meaning they've been seeking.
They’re just going to declare it GOD like people used to (kind of) do in the past about certain natural objects.
And certain manmade ones!! Golden calves and such...
Oh, right. Golden Calf. I wonder how people thought about such objects…whether it itself is imbued with cosmic properties or if it is imbued by a spirit. A worship of objects is curious. You can smash an object. You can turn off the AI God. Seems like it to work there must be a ‘spiritual element’
I suspect a lot of it has to do with the inevitable caste of priests who attach themselves to the procedural worship of the object. You can't (easily) smash an object that has a high status caste dedicated to teaching people all the reasons to never smash it lol
This is why I keep coming back to the idea of a Savonarola as the only figure who can set us right - the mystical hater, the guy willing to devote body and soul to there being No More Fun of Any Kind and self-annihilating in the process.
I think a lot of the time, the spiritual element is the human imagination and mental investment we put into it. But it throws it into overdrive when the object can talk back very convincingly. The place of "truth" in our lives is about to undergo a radical shift.
I know. 😭 I prefer Classic Truth. I do not want New Truth.
Holy smokes—yes. That’s very striking now. But in the research realm, it seemed like people understood this? Then it jumped to the corporate realm and it’s like ‘thinking is outputs. Toaster thinks bread.’
It’s hard to know how Gen AI will go exactly. When people claim ‘there’s this thing that will happen’ and their ideas on how to make it happen have failed multiple times and there’s no defined path, it’s reasonable to be skeptical about optimistic predictions. There’s evidence in both directions.
The only thing it is 100% reliable AND useful for is converting text into different formats, like for R vectors.
All these things started out as translation engines to begin with. They're really only useful as translation and summarization engines, and you still have to be careful to ensure that it doesn't create or eliminate things that it shouldn't.
For me, at least
The scam part and insanity arose because this is a method for becoming a billionaire. So a thing which would normally be a discrete computing tool with costs built in got re-packaged as a society-altering commercial product (of no use) where HUGE costs are externalized but a few can get rich off.
Exactly. It should have remained a fun graduate student thesis for another 20 years. We didn’t need it stuffed into every crevice of our lives in such an imperfect state.
It’s a useful tool for certain things. We can admit that. If everything weren’t subject to the stranglehold of financialization on everything it would have stayed a research tool, then slowly designed to be available to a small number of people for discrete tasks, and would work for those.