Remix.run Logo
▲ mbgerring a day ago

“Because I am impressed and mystified by its outputs, I believe that the next token prediction machine must be doing something more than that, regardless of the lack of any evidence”

AGI is when the output of LLMs is plausible enough to fool putatively intelligent people that they are alive

▲preg_match a day ago | parent | next [-]

LLMs are most definitely doing more than just predicting the next token. They can reason. This is emergent behavior. The idea is that, while one small part is doing something simple, the entire system together produces behavior that’s much more than that.

I kind of view it as how slime mold can solve a maze. Mold doesn’t think or see, but the system of a bunch of slime mold together has almost a rudimentary intelligence.

▲_justinfunk a day ago | parent | next [-]

I'm of the opinion that what is emerging is the human intelligence inherent in the training data.

▲mbgerring 21 hours ago | parent | prev | next [-]

Be specific. Describe what you think the computer is doing that is “more than” predicting the next token, and explain why you believe that.

▲solyin 20 hours ago | parent | next [-]

It generalizes reasoning. If we are to believe the link's linked paper, that's new. It didn't exist before. Yes a human helped. AI also played a part. Somewhere in the calculation of a pile of linear algebra, gigabytes, upon gigabytes of training data and RAM, whilslt still predicting the next token, the prediction that went into choosing that specific next token, reasoning occured.

Internal to the model, it can represent intermediate relationships, combine them, and then use those results to generate a unique answer.

▲soon_to_be_sb 18 hours ago | parent [-]

> into choosing that specific next token, reasoning occured.

This is where you fall.

Imagine when you ask an LLM, "can cats fly", it reply "No".

But then, you can feed it so much fake training data based on this specific question so that it will say "Yes" to this question, despite all the genuine training data it had before that point.

You won't be able to teach a reasoning entity such contradicting facts. There is no mechanism in the LLM to reject fake data that contradicts with its existing knowledge.

So an actual intelligent entity should curate the training data first, for the LLM to show the "appearance" of intelligence.

▲preg_match 18 hours ago | parent | next [-]

No, LLMs can reason. If you ask it to explain how and why cats can fly, it will show its reasoning. It’s not a secret. You can try to convince it otherwise, and of course if you poison the data that will affect it.

But then it will realize… well most mammals don’t fly. And cats don’t have wings. So no, cats cannot fly. That’s rudimentary reasoning that a child could probably do, but it’s reasoning. You can try the same experience with a human child. Tell them over and over if cats can fly. They might then say yes! Then ask them to explain how, and they’ll reason and decide no, they cannot.

Of course LLMs can do reasoning much more advanced than that, as evidenced by theorem proving.

▲Planktonne 9 hours ago | parent [-]

> No, LLMs can reason. If you ask it to explain how and why cats can fly, it will show its reasoning

You're eliding an important difference here.

> can reason...show its reasoning

The appearance of a thing is not necessarily the same as the thing.

▲lostmsu 5 hours ago | parent [-]

Yesterday GPT 6 wrote a new joke. Same day world famous comedian came up with a word for word the same joke independently. Except his joke was funny!

▲Planktonne 2 hours ago | parent [-]

The argument would be stronger if you had a real example rather than a made-up one.

▲solyin 13 hours ago | parent | prev | next [-]

> "On two occasions I have been asked, 'Pray, Mr. Babbage, if you put into the machine wrong figures, will the right answers come out?' In one case a member of the Upper, and in the other a member of the Lower, House put this question. I am not able rightly to apprehend the kind of confusion of ideas that could provoke such a question."

- Charles Babbage, 1864 memoir, Passages from the Life of a Philosopher

▲solyin 12 hours ago | parent | prev [-]

We tell children stories of Santa and they believe their parents. Are children merely the appearance of intelligence as well?

▲emtel 18 hours ago | parent | prev | next [-]

You first: explain how “predicting the next token” can be done without doing anything special.

“Predict the next token” isn’t a process or algorithm. If you traveled back in time and said “you can build AI by predicting the next token” this wouldn’t help them do anything they didn’t already know how to do.

▲preg_match 18 hours ago | parent | prev [-]

Nobody can do this, nobody knows exactly what it’s doing or why. The system is far, far too large to comprehend or understand, and it’s opaque.

At its core it’s just a bunch of weights, just floats. Put together though it can solve problems and reason, including problems which has not been solved before.

Similar to the slime mold. It’s just a single cellular organism with benign, and unimpressive, DNA programming. It doesn’t have any organs, just organelles, and certainly no nervous system.

But, put a bunch together, and it can solve a maze like a rat. Interesting.

Largely I think these “stochastic parrot” type takes are just perpetrated by people who do not use LLMs. If you use LLMs to program, as virtually every software dev working does, it’s indisputable they can reason. Not always well, but they can reason.

▲bigstrat2003 a day ago | parent | prev | next [-]

> LLMs are most definitely doing more than just predicting the next token. They can reason.

No, they cannot, and we have seen this proven time and time again.

▲matkoniecz 21 hours ago | parent [-]

> No, they cannot, and we have seen this proven time and time again.

how it was proved?

▲solyin 20 hours ago | parent [-]

Didn't you see on Twitter months ago? It couldn't count the number of "R"s in strawberry, so we should just call the whole thing off!

▲bothers a day ago | parent | prev | next [-]

[dead]

▲alpaca128 a day ago | parent | prev [-]

Someone on YouTube showed that the same maze can also be solved by plain water. I don't see how the medium being slime mold makes it more impressive or a sign of intelligence.

▲lostmsu 5 hours ago | parent [-]

Isn't water literally doing gradient descent here?

▲ashdksnndck a day ago | parent | prev [-]

It’s like saying humans brains are just a bag of subatomic particles mechanically following the physical laws that predict the next state. Vacuously true.