The Lone Banana Problem in AI • Hacker News Board

Re: The Lone Banana Problem in AI

#21

Post by bluenose69 » Tue, Jul 04, 2023, 10:03 AM UTC

kgwgk wrote: post #7

tudorw wrote: post #6

I've seen lots of long prompt suggestions, however single words are fascinating, some words are extremely heavy, for example, we might assume 'topic' and 'subject' are somewhat interchangeable, but when I'm talking to an LLM swapping those words cause a huge change. Same with names, if you ask it to act like Dave, it's not some random personality, it's a distillation of every Dave it has encountered! I don't know how…

These are the Daves I know: https://m.youtube.com/watch?v=8nvzEqsZIGo

Somehow, I needed that blast from the past!

Re: The Lone Banana Problem in AI

#22

Post by bunga-bunga » Tue, Jul 04, 2023, 10:03 AM UTC

rishabhjain1198 wrote: post #9

I wish this article was just 3 paragraphs. The verbose writing style was a little tiring, I found myself scrolling impatiently to find what the actual "Lone Banana Problem" was.

Someone needs to create a browser that parses and displays every website in a format that I request.

My preferred format would probably be:

- answer to the title / synopsis

- the article, without ads, using the site’s style

Re: The Lone Banana Problem in AI

#23

Post by llogiq » Tue, Jul 04, 2023, 10:14 AM UTC

Again, there is a lot of words to describe the fact that machine learning is just lossy compression for a bunch of data with the possibility to interpolate between data points and get somewhat plausible results. This means data points may get lost during compression/training, and certain things will look off, whether it be a preference for banana pairs, even numbers of fingers or certain weasel words in verbiage.

Re: The Lone Banana Problem in AI

#24

Post by sebzim4500 » Tue, Jul 04, 2023, 10:17 AM UTC

tudorw wrote: post #6

I've seen lots of long prompt suggestions, however single words are fascinating, some words are extremely heavy, for example, we might assume 'topic' and 'subject' are somewhat interchangeable, but when I'm talking to an LLM swapping those words cause a huge change. Same with names, if you ask it to act like Dave, it's not some random personality, it's a distillation of every Dave it has encountered! I don't know how…

I wonder how much of this is due to mode collapse. With RLHF, there isn't really anything stopping it deciding that all Daves have a certain personality.

Re: The Lone Banana Problem in AI

#25

Post by anotheryou » Tue, Jul 04, 2023, 10:22 AM UTC

Midjourney solution: "--no bunch, two, multiple": https://i.imgur.com/QnJmLr0.jpeg

These models have a tendency to move towards the average, especially if unprompted. As we see here, sometimes even if prompted otherwise.

They just have to get better or need a more precise interface like the —no :). We also couldn't have "a man crawling" before and now we can: https://i.imgur.com/ycVpk3i.jpeg

Re: The Lone Banana Problem in AI

#26

Post by brazzy » Tue, Jul 04, 2023, 10:27 AM UTC

thriftwy wrote: post #11

Humans are in the business of consuming bananas, whereas neural nets are in the business of peddling bananas. They don't get to actually use these bananas so they can't gain deeper insight into what's they for. This is the classic lamb vs. mutton issue. Wealthy land owners who use one set of idioms vs. servants who use a different one. Happens to neural nets on human chassis as well.

That touches on something I've pondered for a while:

Maybe the popular sci-fi image of "artificial brains" with humanoid bodies actually turns out to be more accurate than the technical view that robotics and artificial intelligence are mostly separate fields!

What if, instead of machine learning being a source of features we can maybe later build into robots, it's the other way round? What if direct interaction with the real world through robotic limbs and sensors turns out to be what it takes to get AI to stop hallucinating?

What if, furthermore, having a body turns out to be an essential requirement for AI to develop a sense of self? Not just because of sensory feedback, but simply because it provides an anchor point for identity?

Re: The Lone Banana Problem in AI

#27

Post by bippihippi1 » Tue, Jul 04, 2023, 10:31 AM UTC

llogiq wrote: post #23

Again, there is a lot of words to describe the fact that machine learning is just lossy compression for a bunch of data with the possibility to interpolate between data points and get somewhat plausible results. This means data points may get lost during compression/training, and certain things will look off, whether it be a preference for banana pairs, even numbers of fingers or certain weasel words in verbiage.

all gradient learning creates kernel machines

Re: The Lone Banana Problem in AI

#28

Post by lproven » Tue, Jul 04, 2023, 10:33 AM UTC

rishabhjain1198 wrote: post #9

I wish this article was just 3 paragraphs. The verbose writing style was a little tiring, I found myself scrolling impatiently to find what the actual "Lone Banana Problem" was.

The pictures did not immediately tell you?

They told me what the problem was in about 2 seconds, and frankly, the title and the first sentence alone were.

All right, I concede, a blind reader would not know, but most people are not blind.

Re: The Lone Banana Problem in AI

#29

Post by tsukikage » Tue, Jul 04, 2023, 10:35 AM UTC

"Three cats in a trenchcoat" is a good one. Three cats in a trenchcoat standing on each other's shoulders, pretending to be a human, Vincent Adultman style.

You can get them standing side by side wearing trenchcoats. You can get a cat pyramid or a cat totem or a stack or tower of cats (though often the ability to count to three is then lost). You can't get them to share the trenchcoat. Nothing like that occurred in any training set, and the AIs do not understand spatial relationships between objects ("X is on top of Y, inside Z") so you cannot describe how to arrange the things it does know about in the scene. Dall-E won't do it. Midjourney won't do it. Stable Diffusion won't do it.

Eventually, enough images will be seeded into the training sets for this to stop being a useful test. But right now, it gives one a fascinating window on what happens when you try to extrapolate outside the cloud of thingspace described by the training data, rather than just interpolating within it.

https://www.toothycat.net/~sham/trenchcoats

Re: The Lone Banana Problem in AI

#30

Post by lionkor » Tue, Jul 04, 2023, 10:38 AM UTC

> Then think of all the wealth disparity that has been introduced into our world. Think of the social anxiety of always being online. Think of the undermining of our democratic institutions.

How is this at all whatsoever needed