Tulips were not one of the greatest scientific advances in human history. AI that can write and converse intelligently on virtually any subject is. When the chief complaint about AI is that it lies too much, that's a massive breakthrough.
We still don't understand how these models really work, at least not in a way that teaches us anything about language or about humanity.
I find that a strange take. Almost all scientific advancement ends in one thing. Another set of unanswered questions. Quite often we gain more knowledge, but know less because we realize there are thousands more questions we should have been asking.
For example, theory of mind being observed out of just processing language is a big arrow that we may have been thinking about many things wrong in the past and open new avenues for 'scientific' tests that give us answers.
Is there a "hacker news" of theory of mind researchers? Probably some interesting conversations happening over there.
We’ve effectively made like a general teenager that can be up-trained for specific tasks. It’s super useful if you know how to use it.
FTFY. At least any kind of actually notable scientific advancement like using metals, farming, steam power, electricity, cars, planes, computers, etc.
If a person can use a chatbot to speed up their work then that's a major advancement in productivity right there.
We understand how LLMs work conceptually. Human language, when parameterized as a sequence of tokens, has geometry and structure in a billion dimensional space. A priori, we didn't know that. So in a very real sense, a discovery was made.
The unexplained Chain-of-Thought emerges from LLM surely would an huge gain in "knowledge". In fact I'd say it potentially the most important meta-knowledge human race can gain, you know, to unlock the mystery of "conciousness" itself.
Ask for a biography of a random public figure without giving notes and there is 0 chance they will provide an accurate result. This may mean we need systems that are literally memories without the noise driven generalization, and that these need to be exposed to the said LLMs as a resource. It's not insurmountable and it certainly isn't tulips.
I feel like everybody commenting on this thread ought to have to write "it's text completion, not intelligence" on a chalkboard 500 times or so first.
There’s some great chess videos for example where the bots start by making decent moves until you quickly find out they not only don’t know how to play well they don’t even understand the rules or the board. https://youtu.be/rSCNW1OCk_M What’s fascinating is how easy it is to give something the benefit of the doubt because it’s using language, we just aren’t very suspicious of things that can talk.
Further if you watch the video it wasn’t moving the pieces incorrectly it was just creating them from thin air. This wasn’t like an 8 year old learning the game this was like your cat randomly hitting the keyboard.
It acts similar to a human that has a button for "recite chess rules" and presses it when you ask them to recite the chess rules. You'd then get a reply with the chess rules without the button-presser getting the same info, so they wouldn't gain any knowledge from it. I'm sure you can find a lot of people who could, if asked, read you the chess rules but even after reading wouldn't have understood them themselves. But are they not generally intelligent?
I get a feeling that people compare ML-models to a smart person that concentrates on what you ask them, and LLMs fail in comparison but do so confidently. If you instead compare them to an average person at their daily average focus level, the ML probably looks pretty smart.
I have a neighbor who is not very bright. He'll see some news on TV and will tell you about it, only he doesn't understand it and will just repeat words and half-sentences as best he can remember. He doesn't know he doesn't understand (or he does but doesn't care), so he'll confidently state utter nonsense. He's not smart and I wouldn't go to him for advice, but I think he qualifies as "intelligent" in the binary sense.
Even a dumb person is vastly smarter than an ant. So when I am saying ChatGPT is dumb I mean compared to having a conversation with a person which seems like the only reasonable scale to use with a chat bot.
ChatGPT’s model requires quite a of of processing power and sophistication to form its responses. It just doesn’t comprehend what it’s saying.
I'm sure you've talked to people before who used certain terms but it was obvious to you they didn't actually know what those terms meant. They're able to use them without understanding the concept. Just like ChatGPT, only that it's more likely to sound like ChatGPT understands the concept.
What I'm trying to say: humans have a very clear limit to what concepts they can grasp. They can punch above their weight by pretending to understand a concept, but the proof is in the pudding, so unless they're able to apply it to solve problems, they haven't understood it. And even if they do understand something, they'll make mistakes. If the test for intelligence is essentially 'can perform certain tasks with a maximum error rate of x, and can learn to perform new tasks', does 'understanding' really matter? And can you even really test for understanding, aren't you only testing for 'gives answers to questions that signal understanding'?
I don’t have specific level beyond that because it’s deficiencies don’t map to human condition.
As to a touring test given enough time and freedom in questions you can easily tell it’s a machine so it just fails. That’s the thing, the touring test isn’t can you tell if this is a machine in 7 seconds while talking about baseball it’s can to tell period.
The models are not operating at a lower or higher level along a scale of human intelligence, so they're not genius-like or child-like. They're not operating on the same scale at all.
They're text prediction models, and many seem to insist on anthropomorphizing them in a way that causes them to forget that every other second or so.