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

Rodney Brooks on GPT-4

spectrum.ieee.org

Post reply on HN
Print view View original

211–220 of 412 posts

Re: Rodney Brooks on GPT-4

#211

Post by tornato7 »

avereveard wrote: post #199
tornato7 wrote: post #183

Earlier quoted context omitted.

The one thing that has convinced me that chatGPT has built a real world model is asking it how to stack different objects. You can ask it for the safest way to stack a pillow, an egg, a nail, a book and an action figure. Even get more complicated. GPT-4 will, most of the time, correctly reason through how to stack these items to form a stable structure without breaking. That tells me it knows intimately the physical…

Or was that one of the training samples provided to gpt 3.5? It is going to be increasingly harder to divide between what it was in the training set and what is creative thinking going forward. Like, at the very minimum, all the publicly known reasoning samples that were used to probe gpt version 3.5 downward are going to be unreliable for knowing what gpt4 reasoning abilities are

The point of the question is that you can easily come up with combinations of objects that are unlikely, or statistically impossible, to be in their training samples. How about this one? "Tell me how to safely stack the following objects on top of each other without any additional support: A Nikon DSLR, A small succulent, 4 lightbulbs, a pack of gummy worms, and a hardcover copy of Harry Potter and the Halfblood Prince."

Hardcover copy of Harry Potter and the Half-Blood Prince: Start with the largest, heaviest, and most stable object. This will serve as your base. Make sure to place it on a flat, secure surface.

Pack of gummy worms: This item is fairly light and flexible. It should sit nicely on the book without risking damage to either. You can place it right in the center to maintain balance.

Nikon DSLR: This is a heavy object with a fairly stable shape. Put this on top of the pack of gummy worms, as it's sturdy enough to handle the weight and won't be harmed by the soft package. Make sure to place it so that the lens is pointing upward and it sits flat and stable. Don't use any protruding lenses or parts to balance it, as they could easily be damaged.

Four Lightbulbs: Since light bulbs are round and fragile, you will need to stack them with care. If the bulbs are in a box, you can place the box on top of the DSLR, making sure it's centered and stable. If the bulbs are loose, it's a bit more tricky. You might want to nestle them into the lens of the DSLR (if the lens is large enough), which should provide some natural containment for them. If not, it might be safer not to include the lightbulbs in your stack.

A small succulent: Finally, the succulent can be placed on top. It's likely the lightest object and it is usually quite stable due to its pot. If the lightbulbs were in a box, place the succulent on top of that. If you ended up not including the lightbulbs, place the succulent directly on top of the DSLR.

Re: Rodney Brooks on GPT-4

#212

Post by consilient »

galaxyLogic wrote: post #187
beardedwizard wrote: post #105

Earlier quoted context omitted.

Why would common crawl NOT contain a chess manual? the rules are explained in detail on wikipedia. the simplest conclusion is that it has indeed been trained on a chess manual and is good at predicting what the next word in a chess manual is. it is not synthesizing anything.

I haven't seen this mentioned but can LLMs actually play chess? I'm sure they have read rules of chess online, but if you ask them to play chess with you, what happens? Can they apply the rules? Can they apply them intelligently and win the game? My point is that even though LLMs "know" what the rules of chess are, they don't really "understand" them, unless they can use them to play the game and play it well.

GPT4 can play chess, but not especially well. For example: https://lichess.org/@/oopsallbots-gpt-4

Re: Rodney Brooks on GPT-4

#213

Post by espe »

aappleby wrote: post #128
fwlr wrote: post #33

> The large language models are a little surprising. I’ll give you that. I think this is the key point about LLMs that kind of explains the wide and polarized views on whether it understands or parrots, whether it can think or is the precursor to thinking or is a dead-end, whether it will catastrophically destroy the world, or “merely” make it steadily worse with bullshit, or just put a few industries out of a job. A…

What exactly does "understanding the world" really mean?

compression :)

Re: Rodney Brooks on GPT-4

#214

Post by drpixie »

ramraj07 wrote: post #47
fwlr wrote: post #33

> The large language models are a little surprising. I’ll give you that. I think this is the key point about LLMs that kind of explains the wide and polarized views on whether it understands or parrots, whether it can think or is the precursor to thinking or is a dead-end, whether it will catastrophically destroy the world, or “merely” make it steadily worse with bullshit, or just put a few industries out of a job. A…

The issue with all these experts is they still think it's human nature to be able to fully understand the world before they speak about it. On the contrary it's human nature (and all animal nature) to figure out how to navigate the world without fully understanding or having a complete model of it. All you need is a working model that affects the facets of the world you need to deal with. I still remember in the 90s…

> how to navigate the world without fully understanding or having a complete model of it

GPT-4 does not have ANY understanding or model of the world - it just has a model of what tokens (words) are likely to appear in a certain context. If it could build any usable model of the world, and reason about it, I'd be much more impressed.

When it quacks like a duck, only the most simplistic view takes it as being a duck.

Re: Rodney Brooks on GPT-4

#215

Post by goatlover »

kristiandupont wrote: post #197
onethought wrote: post #170

Earlier quoted context omitted.

> considering the scale of the matter, i.e. human extinction. There is literally no evidence that this is the scale of the matter. Has AI ever caused anything to go extinct? Where did this hypothesis (and that's all it is) come from? Terminator movies? It's very frustrating watching experts and the literal founder of lesswrong reacting to pure make believe. There is no disernable/convincing path from GPT4 -> Human Ex…

Nuclear bombs have also never caused anything to go extinct. That's no reason not to be cautious. The path is pretty clear to me. An AI that can recreate an improved version of itself will cause an intelligence explosion. That is a mathematical tautology though it could turn out that it would plateau at some point due to physical limitations or whatever. And the situation then becomes: at some point, this AI will be…

So what would be the path for GPT5 or 6 creating an improved model of itself? It's not enough to generate working code. It has to come up with a better architecture or training data.

Re: Rodney Brooks on GPT-4

#216

Post by anon84873628 »

JoshTko wrote: post #38
fwlr wrote: post #33

> The large language models are a little surprising. I’ll give you that. I think this is the key point about LLMs that kind of explains the wide and polarized views on whether it understands or parrots, whether it can think or is the precursor to thinking or is a dead-end, whether it will catastrophically destroy the world, or “merely” make it steadily worse with bullshit, or just put a few industries out of a job. A…

The more I think about it the more I'm convinced I am basically just predicting/saying my next word whenever I speak.

The sad truth is that we are all NPCs without free will. "Thoughts" are a non-physical phenomenon that, seemingly, arise from physical activity in the brain. The thoughts (our "experience") come after the electrochemical reaction that manifests them. How could these non-physical phenomena then turn around and influence the physical chemistry? They can't. By the time we experience a thought, the physical state for it has been represented, and that state causes the next one after it. Like a stone bouncing down a hill - chaotic but still deterministic. We're just along for the ride.

Re: Rodney Brooks on GPT-4

#217

Post by bheadmaster »

krainboltgreene wrote: post #83
ramraj07 wrote: post #47

Earlier quoted context omitted.

The issue with all these experts is they still think it's human nature to be able to fully understand the world before they speak about it. On the contrary it's human nature (and all animal nature) to figure out how to navigate the world without fully understanding or having a complete model of it. All you need is a working model that affects the facets of the world you need to deal with. I still remember in the 90s…

> The issue with all these experts is they still think it's human nature to be able to fully understand the world before they speak about it None of the experts think this.

Actually, at least one expert thinks that.

Re: Rodney Brooks on GPT-4

#218

Post by gmt2027 »

blowski wrote: post #163
thomastjeffery wrote: post #107

Earlier quoted context omitted.

> My current (tentative) resolution of the surprise is that language encoded way more information about reality than we thought it did. I think you are close to the mark, but you have been subtly mislead: language is not the data we are working with. We are working with text . Once you fix that particular failure of word choice, everything else becomes much more clear: text contains much more information than languag…

I don't really understand your distinction between language and text, but it sounds intrguing. Would you be able to give more detail? I searched but couldn't find anything that seemed to explain it.

Text is an instance of language. Think of it as the difference between the python language and a large collection of python programs. The language describes syntactic and semantic rules, the collection is a sampling of possible programs that encodes a significant amount of information about the world. You could learn a lot about the laws of nature, the internet, even human society and laws by examining all the python programs ever written.

An extreme version of the same idea is the difference between understanding DNA vs the genome of every individual organism that has lived on earth. The species record encodes a ton of information about the laws of nature, the composition and history of our planet. You could deduce physical laws and constants from looking at this information, wars and natural disasters, economic performance, historical natural boundaries, the industrial revolution and a lot more.

Re: Rodney Brooks on GPT-4

#219

Post by braindead_in »

> No, because it doesn’t have any underlying model of the world.

Ilya's counter to this reasoning is for next word prediction to work, the model has to 'understand' our world. Otherwise the predictions will be way off. Therefore the human world has been modelled to a degree by GPT.

I haven't yet heard a good counter to that.

Re: Rodney Brooks on GPT-4

#220

Post by cageface »

panarky wrote: post #150
cageface wrote: post #112

Earlier quoted context omitted.

Maybe in casual conversation but that's not how I experience my though process about anything non trivial at all. I usually spend a lot of time thinking about the concept in non verbal terms and that process involves recalling images and sensory information in fairly abstract terms and then through what feels like several iterations it starts to coalesce into something I can encode in language. I think we can all agr…

> thinking about the concept in non verbal ... abstract terms The abstract terms we think about are concepts, and we think about multiple concepts, at various levels of abstraction, and their relationships to each other, before getting a sense of what we want to say or write. Only then do we begin speaking or writing, grouping concepts into paragraphs, breaking them down into sentences and words. And there's evidence…

Right but they're limited to concepts that are derived from linguistic tokens. Human reasoning can include nonverbal inputs.

For example asking GPT what would be a reasonable way to stack a list of arbitrary objects on top of each other.

Post reply on HN