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

What Does It Mean for AI to Understand?

quantamagazine.org

Post reply on HN
Print view View original

91–100 of 134 posts

Re: What Does It Mean for AI to Understand?

#91

Post by justinpombrio »

jonplackett wrote: post #58
bpizzi wrote: post #34

Earlier quoted context omitted.

> I can say something like ‘a tiger is just a lion with stripes’ to a 3 year old and they now ‘understand’ what a tiger is almost as well as if they saw a picture of one. They could definitely identify one from a picture now. Assuming the 3 year old already knew what a lion looks like, and point at 'things with stripes' and 'things without stripes'. I think that a model that can already recognize separately lions and…

Maybe… but this is just one very easy example and also using something very obvious and visual. I could also say “a Cheetah is like a lion but it’s smaller and has spots and runs a lot faster. And a leopard is like a lion but smaller and can climb trees and has spots.” I could probably start with a house cat and describe an elephant if I wanted to and I’ll bet the kid would work it out. The ability to take apart and…

> I could also say “a Cheetah is like a lion but it’s smaller and has spots and runs a lot faster. And a leopard is like a lion but smaller and can climb trees and has spots.”

The OpenAI website is unresponsive at the moment, so I can't actually demonstrate this, but you could totally tell GPT-3 that, and it would then make basic inferences. For example, saying "four" when asked how many legs a cheetah has, or guessing a smaller weight for a cheetah than a lion when asked to guess a specific weight for both. Not perfectly, but a lot better than chance, for the basic inferences.

(You wouldn't actually tell it "a Cheetah is like a lion but..." because it already knows what a Cheetah is. Instead you'd say "a Whargib is like a lion but ...", and ask it basic questions about Whargibs.)

Re: What Does It Mean for AI to Understand?

#92

Post by jonplackett »

ShamelessC wrote: post #68
jonplackett wrote: post #60

Earlier quoted context omitted.

I’d be interested in where to look for more info on that. Do you think it could work on anything more complex? As I said in comment below I reckon I could make a much more elaborate explanation and still have the kid get it.

https://openai.com/blog/dall-e/ shows a decent ability to generalize to previously unseen concepts. You're correct to consider the complexity of the phrase and just how good humans are at this sort of thing without needing much "training". For now, concepts that aren't explicitly in the training set are effectively composed from those which are. This can lead to some bizarre and outright incorrect results, particular…

Thanks. this is interesting. In a way it's opposite thought-direction of what we are talking about.

eg. can it look at an avocado shaped chair and recognise it as a chair in the shape of an avocado - for me that would display a lot more understanding of the concept of 'chair' and 'avocado' than being able to produce an image of the phrase 'a chair shaped like an avocado' - but maybe the same process must be happening in there somewhere to make this possible? What do you think?

Re: What Does It Mean for AI to Understand?

#93

Post by chermi »

simplestats wrote: post #81
chermi wrote: post #53

Earlier quoted context omitted.

*Edited because I can't type* I do not understand this argument. I'm going use the word magic for whatever you're calling this non-"materialistic" spiritual explanation. I'm not trying to be a jackass, that's truly the best word I can find for it. I don't think magic is a bad thing, I just don't think you should summon magic until you need to. Magic is not an explanation, it's a non-explanation. It's non-falsifiable…

I don't think this line of dismissal necessarily fits. If we ended up having to expand physics to include some new (today seen as mystical) property of the universe, then you are essentially wrong here. With QM (which I doubt is related to consciousness) we had to add unmeasurable quantum phase to everything and accept certain previously-"nonphysical" new properties of systems. Similarly with other revolutionaly theo…

I don't think I follow? What am I dismissing?

I don't quite understand QM analogy. QM arose to explain new observation from more detailed measurements. Consciousness and intelligence and all things brain have been (to an extent) known 'properties' and difficult to explain (from first-principles) for quite some time. The more we probe, we answer some questions, and in the process reveal some more detail, leading to more detailed questions.

My qualm is that nowhere along this trajectory have I ever seen any need to invoke magic. It's a complex, many-body system with lots of noise, many relevant interacting length and time-scales that are difficult to cleanly separate, model, observe and probe. It's hard, but it seems we're making progress. Maybe it will take CERN, LIGO or ISS scale endeavors, I don't know.

It's not all similar, IMO, to how QM arose.

But I'm not a neuroscientist nor a historian of science.

Re: What Does It Mean for AI to Understand?

#94

Post by machiaweliczny »

KhoomeiK wrote: post #89

People outside the NLP research community need to understand that a language model does nothing more than calculate probabilties for the next word given some context. The way it learns that probabilistic function can be quite complex, involving billions of parameters, but it's still fundamentally the same. Most of the increase in performance in recent years has come from the ability to train LMs that better generaliz…

What if to predict next word you need to compress/understand universe in the end? AFAIK we have no clue how exactly NNs pick up/structure information.

For me backprop and learning is similar to evolution thus result might be simialar. My knowledge is very limited though but I am happy to read any proofs/insights.

Re: What Does It Mean for AI to Understand?

#95

Post by yboris »

jonplackett wrote: post #69
yboris wrote: post #22

Earlier quoted context omitted.

I'm curious, when you say "This kind of understanding won’t work with an AI", do you mean currently, or in principle, even in the future? note: children's brains come pre-loaded with so much stuff when we are born (we are not "blank slates").

What I mean is AIs aren’t really built with the goal of ‘understanding’ anything currently. They are awesome at individual tasks but they don’t have the kind of common sense a person can use to reason with and build up an understanding of how pieces of knowledge fit together. Eg. Driverless cars can identify a car, or a motorbike or a cyclist and maybe work out a trajectory for it. But they don’t understand that a bi…

Thank you for the response. A bit of a push back: On the driverless cars example you're right that AI models abstract away needless complexity, but so do we, as humans, can live without understanding that bodies are made up of cells (or how muscles move during a tennis match). If we want AI models to include a person's gaze in their calculations, we can make it happen. From an abstract-enough perspective, the AI model will do enough that whether it "understands" may become irrelevant.

Language understanding is much harder, because words are about stuff in the world. Just like when a toddler says "love" we know they don't fully understand what they mean, AI won't have the capacity to mean "love" unless it has a lot more it "understands" along the way. But it feels like it could in the near few decades "understand" enough about "duck" to mean it when it says "I see a duck".

Re: What Does It Mean for AI to Understand?

#97

Post by ShamelessC »

jonplackett wrote: post #92
ShamelessC wrote: post #68

Earlier quoted context omitted.

https://openai.com/blog/dall-e/ shows a decent ability to generalize to previously unseen concepts. You're correct to consider the complexity of the phrase and just how good humans are at this sort of thing without needing much "training". For now, concepts that aren't explicitly in the training set are effectively composed from those which are. This can lead to some bizarre and outright incorrect results, particular…

Thanks. this is interesting. In a way it's opposite thought-direction of what we are talking about. eg. can it look at an avocado shaped chair and recognise it as a chair in the shape of an avocado - for me that would display a lot more understanding of the concept of 'chair' and 'avocado' than being able to produce an image of the phrase 'a chair shaped like an avocado' - but maybe the same process must be happening…

https://openai.com/blog/clip/ CLIP is the corrolary model created purely for classification purposes rather than generative as in DALL-E and is quite impressive across a range of tasks. Give it an image and a caption, and in return you get a score (0.0 to 1.0) telling you how much they match.

I think it is more in line with your premise. Others have taken CLIP and combined it with frozen language models (GPT2) to create automatic captioning models that are very impressive.

edit:

To try to address your question about whether or not actual semantic composition occurs, I think the answer is "yes" but it would be challenging to convince you this is true without going into details of the "self attention" mechanism which allows both methods to work. The short version is that these networks are able to find meaning in extremely high-dimensional problems by having a mechanism specifically tasked with learning positional statistics of the training data. In language this refers to e.g. how often the word "pillow" is directly next to the word "fort". In vision, this similarly refers to how often e.g. trees are positioned next to gift-wrapped presents.

That's quite simplified but I hope that makes sense to a degree!

Re: What Does It Mean for AI to Understand?

#98

Post by machiaweliczny »

BTW humans are generally stupid. We are only smart in collective with all trial/error saved in books, so don't underestimate computers. Also most of progress happens by outliers in humanity so it can be after AIs start to collaborate. Distributed MoE architectures are step in this direction.

My prediction: codex and programming/operating APIs will give NNs ultra boost, then programming/math gets automated (hard to master, easy to verify) and rest/us will be history. John Conor is already alive. Mark my words.

Re: What Does It Mean for AI to Understand?

#99

Post by interstice »

AnIdiotOnTheNet wrote: post #90
interstice wrote: post #76

Earlier quoted context omitted.

Is it the case that playing the games are effectively a way of asking and answering questions, specifically ‘do I win if I do this’?

That wouldn't qualify as 'unexpected' in any way. The system's ability to express any questions it might have is severely curtailed by the limitations of its inputs and outputs.

And yet the go games had some very unexpected and creative plays, so if the inputs and outputs were language based are we simply moving the goalposts for what is considered 'unexpected'?

Re: What Does It Mean for AI to Understand?

#100

Post by PeterisP »

tested23 wrote: post #16

What does it mean for humans to understand? There are many times in the past where i thought i understood something and then i grow older and i see the holes.

One definition of "understand" would be "have or obtain an internal model of Thing-or-process-to-be-understood which is close enough to reality that it allows you to reasonably predict what will happen and make effective decisions regarding that thing". It does not have to be a perfect model - if it would, then I'll be the first to say that I don't understand anything according to that definition, but it's a bit more tricky than it sounds on the surface. For example, for a self-driving car, "understanding pedestrians" according to this definition does require an ability to predict how they will behave and thus "know" what factors affect that - that the likelihood of a kid suddenly springing towards the middle of the road is highly dependent on the presence of a ball or a pet in that direction; that certain wobbly and jagged movements are indicators that the person might behave in a less predictable manner than the average person, etc, etc; and if a system does have this practical knowledge (measured by how well it is able to effectively apply it for its goals) then I'd say that it does have some understanding.
Post reply on HN