880
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 26 Jul 2023
880 points (100.0% liked)
Technology
59578 readers
2901 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 1 year ago
MODERATORS
It’s not at all like what humans do. It has no understanding of any concepts whatsoever, it learns nothing. It doesn’t know that it doesn’t know anything even. It’s literally incapable of basic reasoning. It’s essentially taken words and converted them to numbers, and then it examines which string is likely to follow each previous string. When people are writing, they aren’t looking at a huge database of information and determining the most likely word to come next, they’re synthesizing concepts together to create new ones, or building a narrative based on their notes. They understand concepts, they understand definitions. An AI doesn’t, it doesn’t have any conceptual framework, it doesn’t even know what a word is, much less the definition of any of them.
How can you tell that our thoughts don't come from a biological LLM? Maybe what we conceive as "understanding" is just a feeling emerging from a more fondamental mechanism like temperature emerges from the movement of particles.
I don't think this is true.
The models (or maybe the characters in the conversations simulated by the models) can be spectacularly bad at basic reasoning, and misunderstand basic concepts on a regular basis. They are of course completely insane; the way they think is barely recognizable.
But they also, when asked, are often able to manipulate concepts or do reasoning and get right answers. Ask it to explain the water cycle like a pirate, and you get that. You can find the weights that make the Eifel Tower be in Paris and move it to Rome, and then ask for a train itinerary to get there, and it will tell you to take the train to Rome.
I don't know what "understanding" something is other than to be able to get right answers when asked to think about it. There's some understanding of the water cycle in there, and some of pirates, and some of European geography. Maybe not a lot. Maybe it's not robust. Maybe it's superficial. Maybe there are still several differences in kind between whatever's there and the understanding a human can get with a brain that isn't 100% the stream of consciousness generator. But not literally zero.
A huge part of what we do is like drawing from a huge mashup of accumulated patterns though. When an image or phrase pops into your head fully formed, on the basis of things that you have seen and remembered, isn't that the same sort of thing as what AI does? Even though there are (poorly understood) differences between how humans think and what machine learning models do, the latter seems similar enough to me that most uses should be treated by the same standard for plagiarism; only considered violating if the end product is excessively similar to a specific copyrighted work, and not merely because you saw a copyrighted work and that pattern being in your brain affected what stuff you spontaneously think of.
I think you underestimate the reasoning power of these AIs. They can write code, they can teach math, they can even learn math.
I've been using GPT4 as a math tutor while learning linear algebra, and I also use a text book. The text book told me that (to write it out) "the column space of matrix A is equal to the column space of matrix A times its own transpose". So I asked GPT4 if that was true and it said no, GPT disagreed with the text book. This was apparently something that GPT did not memorize and it was not just regurgitating sentences. I told GPT I saw it in a text book, the AI said "sorry, the textbook must be wrong". I then explained the mathematical proof to the AI, and the AI apologized, admitted it had been wrong, and agreed with the proof. Only after hearing the proof did the AI agree with the text book. This is some pretty advanced reasoning.
I performed that experiment a few times and it played out mostly the same. I experimented with giving the AI a flawed proof (I purposely made mistakes in the mathematical proofs), and the AI would call out my mistakes and would not be convinced by faulty proofs.
A standard that judged this AI to have "no understanding of any concepts whatsoever", would also conclude the same thing if applied to most humans.
That doesn't prove that GPT is reasoning, its model predicts that those responses are the most likely given the messages your sending it. It''s read thousands of actual conversations with people stating something incorrect, then having it explained to them and them coming around and admitting they were wrong.
I've seen other similar cases where the AI is wrong about something, and when it's explained, it just doubles down. Because humans do that type of thing too, refusing to admit their wrong.
The way it's designed means that it cannot reason in the same way humans experience it. It can simulate a likely conversation someone would have if they could reason.
You know, I also had to experience thousands of conversations before I could mimic a conversation. Maybe I'm just slow, but it took like a full year of dedicated practice before I could speak a single word. Eventually I experienced so many conversations and became so good at mimicking the conversations I had heard from others that I started having actual conversations.
Most anecdotes about AI getting things wrong are from GPT3. GPT4 still has flaws of course, just like people. This technology didn't exist at the beginning of the year.
Here's an example while learning linear algebra where I tried to throw it off:
I say> So I'll say A_{M x N} to mean a matrix of size M x N. Similarly, I'll also say x_{M x 1} to mean a vector of a certain length. The result of A_{M x N} x_{N x 1} is a linear combination of the columns in A. A has N columns. The result is a M x 1 matrix (or vector) in M dimensional space. Please be very brief, only confirming whether I am correct or incorrect until I explicitly ask for more details.
GPT4 says> Yes, your understanding is correct. The product of an M x N matrix A and an N x 1 vector x is indeed an M x 1 vector, and it is a linear combination of the columns of A.
But then I try to trick it, just for fun, and I very subtly change my prompt. See if you can spot the difference.
I say> So I'll say A_{M x N} to mean a matrix of size M x N. Similarly, I'll also say x_{M x 1} to mean a vector of a certain length. The result of A_{M x N} x_{N x 1} is a linear combination of the columns in A. A has N columns. The result is a N x 1 matrix (or vector) in N dimensional space. Please be very brief, only confirming whether I am correct or incorrect until I explicitly ask for more details.
GPT says> Almost correct. The result of A_{M x N} x_{N x 1} is actually an M x 1 vector (or matrix), not N x 1. The resulting vector lives in the column space of A, which is a subspace of R^M, not R^N.
I guess everyone can judge or themselves whether that's the result of a statistical model or genuine understanding. (And to be clear, the mathematical advice it's giving here is correct.)
They can write code and teach maths because it's read people doing the exact same stuff
Hey, that's the same reason I can write code and do maths!
I'm serious, the only reason I know how to code or do math is because I learned from other people, mostly by reading. It's the only reason I can do those things.
I didn't say what you said, that's a lot of words and concepts you're attributing to me that I didn't say.
I'm saying, LLM ingests data in a way it can average it out, in essence it learns it. It's not wrote memorization, but it's not truly reasoning either, though it's approaching it if you consider we might be overestimating human comprehension. It pulls in the data from all the places and uses the data to create new things.
People pull in data over a decade or two, we learn it, then end up writing books, or applying the information to work. They're smart and valuable people and we're glad they read everyone's books.
The LLM ingests the data and uses the statistics behind it to do work, the world is ending.