Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Position: LLMs Can't Jump

openreview.net

Post reply on HN
Print view View original

181–190 of 233 posts

Re: Position: LLMs Can't Jump

#181

Post by AnthonyMouse »

walrus01 wrote: post #172
gabbagool wrote: post #143

Probably too late for this, but I have argued before that language is a fundamentally lossy encoding of the human experience. We do our best to describe what we're seeing and experiencing using language, which is fantastically expressive, but it has its limits. I think we see glimpses of this when we find ourselves saying things such as, "it's impossible to put it into words" or we overload certain words when we mean…

> Probably too late for this, but I have argued before that language is a fundamentally lossy encoding of the human experience. We do our best to describe what we're seeing and experiencing using language, which is fantastically expressive, but it has its limits. Yes, and sometimes this is very intentional . Take for example a short poem which if you sit and really think about it for a long time, you could go off on…

> Being terse and allowing for human interpretation is kind of the entire point of something being written like this.

It's also to a certain extent why LLMs work.

When IBM Watson was playing Jeopardy, one of the game prompts was:

> It was the anatomical oddity of U.S. gymnast George Eyser, who won a gold medal on the parallel bars in 1904

The man was missing a leg and used a prosthetic. Watson's output was, "What is leg?"

At first it was regarded as correct. If a human said that you could conclude that they knew the answer. But then the judges decided not to give Watson the point because its output didn't provide enough specificity to prove that it understood the context.

If you ask an LLM what kinds of things taste sweet it can give you examples like cotton candy or strawberries, but it has never actually tasted anything. All it knows is that the training data contains the association between those tokens. But the human reading the output knows what strawberries are, which is what allows the output to be meaningful.

Re: Position: LLMs Can't Jump

#183

Post by cwmoore »

lukifer wrote: post #147
gabbagool wrote: post #143

Probably too late for this, but I have argued before that language is a fundamentally lossy encoding of the human experience. We do our best to describe what we're seeing and experiencing using language, which is fantastically expressive, but it has its limits. I think we see glimpses of this when we find ourselves saying things such as, "it's impossible to put it into words" or we overload certain words when we mean…

The lossy compression of language is why we should find it unsurprising that LLMs tend to perform better at code, than at human language tasks or reasoning. While there can be subtle semantic differences in real codebases (using "null" to mean "unknown" in one context, versus "intentionally blank" in another), there is a much tighter coupling of semantics to meaning (low ambiguity) compared to "love" in English (let…

I have come across a concept that given a file of compressed text files, adding a new text to it expands it more or less depending on how different the new text is from the compressed.

The compression series sounds interesting!

I think the deltas are growing at different rates.

Re: Position: LLMs Can't Jump

#184

Post by cwmoore »

AnthonyMouse wrote: post #181
walrus01 wrote: post #172

Earlier quoted context omitted.

> Probably too late for this, but I have argued before that language is a fundamentally lossy encoding of the human experience. We do our best to describe what we're seeing and experiencing using language, which is fantastically expressive, but it has its limits. Yes, and sometimes this is very intentional . Take for example a short poem which if you sit and really think about it for a long time, you could go off on…

> Being terse and allowing for human interpretation is kind of the entire point of something being written like this. It's also to a certain extent why LLMs work. When IBM Watson was playing Jeopardy, one of the game prompts was: > It was the anatomical oddity of U.S. gymnast George Eyser, who won a gold medal on the parallel bars in 1904 The man was missing a leg and used a prosthetic. Watson's output was, "What is…

Human language requires a human receiver, like art requires an audience.

Re: Position: LLMs Can't Jump

#185

Post by jhbadger »

throwaway0123_5 wrote: post #154
Dylan16807 wrote: post #150

Earlier quoted context omitted.

I think I could get very close via video. But that depends on a lot of my non-canyon experience moving around the world, and LLMs are very flawed in their ability to input video, so I think they would not get nearly as close.

I've been to a few natural wonders (including the Grand Canyon) that I saw in advance on video. At least for me it isn't close at all. Even if audiovisual elements could be near-perfectly reproduced by video (imo not even close with modern tech, no screen is capturing the brilliance of sunlight), you aren't capturing the temperature, the feel of wind or rain, the smell of the plants around you, etc.

I don't know. I think some "wonders" like the Grand Canyon and Niagara Falls are more impressive in aerial videos and are kind of a let down in person -- you can't really see the whole thing when you are next to it.

Re: Position: LLMs Can't Jump

#186

Post by AnthonyMouse »

cwmoore wrote: post #184
AnthonyMouse wrote: post #181

Earlier quoted context omitted.

> Being terse and allowing for human interpretation is kind of the entire point of something being written like this. It's also to a certain extent why LLMs work. When IBM Watson was playing Jeopardy, one of the game prompts was: > It was the anatomical oddity of U.S. gymnast George Eyser, who won a gold medal on the parallel bars in 1904 The man was missing a leg and used a prosthetic. Watson's output was, "What is…

Human language requires a human receiver, like art requires an audience.

https://people.math.harvard.edu/~ctm/home/text/others/shanno...

> The fundamental problem of communication is that of reproducing at one point either exactly or approximately a message selected at another point. Frequently the messages have meaning; that is they refer to or are correlated according to some system with certain physical or conceptual entities. These semantic aspects of communication are irrelevant to the engineering problem.

Re: Position: LLMs Can't Jump

#187

Post by kkukshtel »

gabbagool wrote: post #143

Probably too late for this, but I have argued before that language is a fundamentally lossy encoding of the human experience. We do our best to describe what we're seeing and experiencing using language, which is fantastically expressive, but it has its limits. I think we see glimpses of this when we find ourselves saying things such as, "it's impossible to put it into words" or we overload certain words when we mean…

Noah Smith had an interesting idea related to this in a recent newsletter:

https://www.noahpinion.blog/p/what-will-more-intelligence-ac...

> Another way of saying this is that there may be laws of the universe that humans can’t understand but AI can. I call these “cloud laws” — causal regularities that can be exploited by technology, but which are too diffuse and complex for an individual human being to either intuit or communicate. Human language seems to obey cloud laws, so why not other phenomena too? Perhaps social sciences like economics, sociology, and political science obey similarly complex regularities, and AI can help us find them. Perhaps there are physical processes — plasma, or topological materials, or aerial turbulence, etc. — that obey cloud laws instead of chaos?

Re: Position: LLMs Can't Jump

#188

Post by NegativeLatency »

throwaway0123_5 wrote: post #154
Dylan16807 wrote: post #150

Earlier quoted context omitted.

I think I could get very close via video. But that depends on a lot of my non-canyon experience moving around the world, and LLMs are very flawed in their ability to input video, so I think they would not get nearly as close.

I've been to a few natural wonders (including the Grand Canyon) that I saw in advance on video. At least for me it isn't close at all. Even if audiovisual elements could be near-perfectly reproduced by video (imo not even close with modern tech, no screen is capturing the brilliance of sunlight), you aren't capturing the temperature, the feel of wind or rain, the smell of the plants around you, etc.

Mt Rushmore is better and more impressive in media than IRL. Saw North by Northwest a few times and went later in my life just didn’t compare.

Re: Position: LLMs Can't Jump

#189

Post by Mikhail_Edoshin »

True understanding requires not knowing, and LLMs cannot "not know". LLMs have to come up with an answer, this is their nature. They are search engines. We do have a similar mechanism; one can notice it by reflecting. The mechanism is an opposite of true thinking, as it merely looks up what is already "known". We "jump" when we temporarily turn this mechanism off.

That said, here's an experiment conducted by some Soviet psychologist, I forgot the name. The man wanted to study intuition. So he invented an experiment that was supposed to trigger it in laboratory conditions. (Take a moment to marvel at that; how would you approach such a task?) He gave people a few puzzles. One was to place some sticks according to some rules. Yet another was to find a path in a maze. The secret was that the path in the maze was the same figure as the solution to the stick puzzle.

And he observed interesting results. People who solved the maze after the sticks found the path much faster than the control group. If a subject was asked to comment how he was solving the maze, at the start or halfway through, the speed dropped to typical. Subjects normally didn't notice the similarities.

So there is something to study here, although it is obviously a case of pattern matching, only subconscious. This is a jump of sorts, but not the one I mean. What I mean is a Zen jump.

Re: Position: LLMs Can't Jump

#190

Post by handoflixue »

jvanderbot wrote: post #2

Came for: "A computer once beat me at chess, but it was no match for me at kick boxing." TFA was actually about leaps of intuition, sadly. One of the experiments I've heard proposed around here is to somehow create an LLM from all text up to 1980 or 1990 and see if it can get back to making itself.

They did that with "Talkie", a model trained on 1930 and before. It had the ability to assemble crude Python programs, but there were definitely some leaks in the training data (it knew about stuff like World War 2) so not 100% perfect. Still seems like a pretty reasonable "proof of concept"

https://didof.dev/blog/talkie-1930-llm-reasoning/ seems like a decent overview

Post reply on HN