- cross-posted to:
- [email protected]
- [email protected]
- cross-posted to:
- [email protected]
- [email protected]
ChatGPT is losing some of its hype, as traffic falls for the third month in a row::August marked the third month in a row that the number of monthly visits to ChatGPT’s website worldwide was down, per data from Similarweb.
Meh.
That’s a very fallibilistic viewpoint. There are lots of certainties that can be answered correctly.
There are fields and fields in science that work on things that are “certainties”.
If you’re talking about simple stuff like “what is the first letter in the english alphabet”, then sure. But many people, even in this thread, say they use the engines for hours, to get answers, guide them, and discuss.
It is a parrot on steroids, but even a parrot has knowledge. LLMs have 0% knowledge.
Well, we are back at my earlier point. There is no need for knowledge if the statistical models are good enough.
A weather forecast does not have any knowledge whatsoever. It has data and statistical models. No one goes around dismissing them due to them not have any knowledge. Sure, we can be open to the fact that the statistical models are not perfect. But the models have gotten so good that they are used in people’s everyday life with rather high degree of certainty, they are used for hurricane warnings and whatnot, saving tens of thousands of life’s - if not more - yearly.
Your map app has no knowledge either. But it’s still amazing for knowing with a high degree of certainty how much time you’ll need from place A to B and which route will be shortest. Even taking live traffic into account. We could argue it’s just a parrot on steroid, that has been fed with billions of data points with some statistics on top, and say that it doesn’t know anything. But it’s such a useless point, because knowledge is not necessary if the data and statistical models are sound enough.
It is exactly my point.
None of the “predictive” apps pretend to have knowledge, to give you answers, to “think”, to “hallucinate”, to “give you wrong answers”.
Everybody knows the weather app is “ballpark predictions”, even though it’s based on physical events that are measurable and extrapolatable.
Same with maps. People who follow maps 100% end up in lakes. The predictions the maps give are based on real-life measured data, topical for that particular frame of time.
With LLMs, the input is language. The output is language. It wraps the generated text in pleasantries to imitate knowledge. Unless it’s fed 100% correct material (no such thing), the output is 100% bullshit that sounds about right; right enough to lure naive and, maybe, less IT-literate people to make them feel they’re getting “correct” information.
Statistical engine. No knowledge. Garbage input, garbage output. No sign of “intelligence” whatsoever.
“asking” it questions is not carring about the “information” it returns.
So you can feed a weather model weather data, but you cannot feed a language model, programming languages and get accurate predictions?
Basically no one is saying that “yeah I just go off the output, it’s perfect”. People use it to get a ballpark and then they work off that. Much like a meteorologist would do.
It’s not 100% or 0%. With imperfect data, we get imperfect responses. But that’s no difference from a weather model. We can still get results that are 50% or 80% accurate with less than 100% correct information. Given that a large enough amount of the data is correct.
Yeah, no difference between real-life physical measurements & data calculations made from proven formulas, and random shit collected of random places on the internet (even, possibly, random “LLM” generated sentences).
People do “just go off the output”. There are people like that in this very thread.
Statements like “no difference” are just idiotic.
Of course there is. But weather forecasting have also gotten ridiculously much more accurate with time. Better data, better models. We’ll get there with language models as well.
I’m not arguing language models of today are amazingly accurate, I’m arguing they can be. That they are statistical models is not the problem. That they are new statistical models are.
A broken clock is accurate twice in a day.
I’m arguing that they will never be accurate, because accuracy is not possible. I mean, look at Wikipedia. At least it’s written by people.
Full self driving next year, right?