What Does It Mean for AI to Understand?
quantamagazine.org
quantamagazine.org
But this does not mean we cannot produce operational models of understanding, for example we have models of propositional/logical semantics and discourse such as Lambda Discourse Representation Theory and others, which can compute a formal representation of the meaning structures for a piece of text. These have been used e.g. for answering question, and working in this space has been a lot of fun, and continues to do so. At the moment people talk a lot about "deep" learning (neural networks with more than one hidden layer), but for such models we need to do a lot more work into explainability, because it is too dangerous to use black boxes in real life.
We still do not understand the human brain function in any substantial way, and it is perhaps a greater mystery of nature than even cosmology, where at least several competing theories have been posed that can explain parts of the evidence.
How are thoughts represented (if that is answerable, it turns out 'Where are thoughts represented?' has proven to be a meaningless question due to the distributed nature of human memory)? What is consciousness? What is a conscience? How do consciousness and intention emerge from materials that are not alive and that have neither consciousness nor intention? How to implement approximate models? (A lot of work to do!)
I think the question shoud first be not about "how", but rather - "does consciousness emerge from the materials".
Most cultures in the past have maintained ideas about the duality between spirit and matter. Nowadays we are so adept at manipulating matter we always explicitly assume that everything must be material in one way or another.
But here is this consciousness thing that does not give in. And there is no angle to it where it can ever give in. If somebody creates a machine and claims it to be conscious nobody can test if that is so. And further - nobody can even test wheather the inventor is conscious himself.
If you put a sail on a boat, it will catch the wind and move across the water. It's not the sail alone that moves the boat, nor the wind, but their interplay. A sail without wind doesn't get moved, and a wind without anything to catch it doesn't move anything.
May be a bit of an unintuitive conclusion, but I think may be that we are catching the consciousness like a sail catches the wind, and this is the same "wind" that we call the forces of nature, that we only can perceive as patterns in how things around us appear to change.
Regardless of whether we know exactly how it works or not, I have a subjective experience.
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 (blah-bity Popper, yeah yeah).
Our adeptness at manipulating matter is not the origin of wanting (at least mine) to find a 'materialistic' explanation for something. It's the desire to have a predictive (read, useful) model of that thing, even if the model is not complete. Once you poke deep enough at any model you will find that there is some hand-wavy "magic" at the bottom. That is, there is some part that we don't know yet, either because we can't do the math, because we can't measure due to difficulty of experiment, or because of fundamental bounds where maybe you might be justified in starting to invoke some magic.
Despite best selling books about some quantum mysticism bullshit, I have yet to see an inkling of evidence for invoking magic for consciousness. From a physics perspective, biological systems are freaking hard, we have no reason to expect them to crack easily. And that's without us (physicists) understanding nearly enough about biology!
The unfortunate thing is that, I think, I understand your side. But something happened, I read some book or heard some talk that placed a seed of doubt in my mind. At first it was just a seed and I oscillated back and forth between being 100% materialist and between entertaining the proposition that there is something beyond matter in this world. But as time went on I began to shift towards the "non material" interpretation for consciousness more and more.
It's not an explanation, I agree. But there is nothing to explain. Me claiming that consciousness might not arise from matter is not an attempt to explain it. I just see no way how it can be investigated in material terms. If you see a person on the street there is no way you can tell if he or she is conscious. And I don't see a possibility of there ever being a way.
Sadly, I don't know what started this doubt in me, so I cannot share it with you. You brought physicists for some reason, and I know a few, like Schrodinger, who thought about consciousness and came to the same conclusion. Here is his quote: "Consciousness cannot be accounted for in physical terms. For consciousness is absolutely fundamental. It cannot be accounted for in terms of anything else." I haven't read about Max Plank, but heard he had similar views.
If I had to guess my starting point was a book called "the problems of philosophy" by Bertnard Russell [1]. He tries to answer the question "are there any statements to which all reasonable men would agree". And one of the conclusions the book comes to is that you cannot claim anything as objective without assumptions, and that the most objective thing is your subjective experience.
For example if you saw a cat, turned your head, and then looked back at the cat - was the cat there when you were not looking or was it gone? In the book Russell convincingly demonstrates that if someone maintains the cat was gone when you were not looking - you cannot prove logically to that person that he is wrong. In other words - you would not be able to start with his worldview and lead him to contradictions. Hence his "theory" about cat disappearing cannot be disproven without assumptions about how material objects behave.
Not sure if this is helpful, but I wanted to reply.
I don't think we strictly disagree.
For one, I agree that, fundamentally, as the observers, everything is "filtered" through us whether we like it or not. So there's no getting around some notion of 'subjectivity' or 'observer' bias, at least not that I know about. But that is not at all exclusive to an improved (predictive, quantitative, model-based) understanding of the brain and consciousness. Look how much progress we've made!
BTW I'm on my phone and forget why we're talking specifically of consciousness and not AI? Is that my fault? My bad.
Oh, I brought physicists so my perspective was clear and you could understand my bias and my ignorance of actual neuroscience and CS and biology.
The observations of Russell's that you mention - which, I believe, were considerably influenced by his working with Wittgenstein - seem to be a critique of analytical philosophy's approach to learning fundamental truths about the universe. "Hold my beer", said his scientific contemporaries, "we're going to put these worries aside and see what we can learn anyway."
Magic can be disproven by explaining the phemonen with a conventional physical theory.
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.
QM required that we accept a new phase property of matter which cannot be observed even in principle, and which causes counter-intuitive experimental outcomes that require contradictory physical explanations. This remains unsolved, physics simple sucked it up and moved on using it. The Copenhagen Interpretation essentially says some questions are not "askable" (e.g., what is the state of an unobserved electron), and observation instantantaneously changes vast-universe-wide systems via collapse of the wavefunction (pretty magical in the opinion of many). The many-worlds interpretation is even kookier, though much more intuitive, and is gaining popularity lately. If parallel universes and "spooky action ata distance" (quoting Einstein) is not magic to you then what does it take?
Asserting “something that isn’t falsifiable isn’t real” is just as arbitrary as asserting that there exist phenomena that are beyond the absolute boundaries of our capacity to perceive them (that is, no amount of technical innovation will allow us to observe such phenomena). The latter statement, though, seems much more reasonable of an assumption.
If something is true from an empirical perspective that is one thing. It is another thing to say that even though we have been unable to form a conclusion from an empirical point of view, the ultimate ability to do so is given, and that all hypothesis that cannot be tested in such manner must be completely disregarded, even if they make sense.
FWIW, I completely agree with that person’s view that saying “Well, it’s not physical” to the question of what a mind is and how it works, is a non-answer that avoids the question. Materialism has not answered that question ether, but it is a double standard to deprecate only materialism on that basis, as the author of this thread’s root post did.
The philosophy of mind seems to me to be hobbled by its fascination with zombie arguments, which are rather cleverly constructed to persuade one that a philosopher’s metaphysical speculation is a stronger guide to how the universe is than any amount of empirical physical evidence.
Edit- crap, I'm pretty new at actually participating in HN comments, I don't think I followed chain properly
The position you take in your last paragraph is one of solipsistic skepticism. Do you approach every epistemic issue with the same level of skepticism, or do you only adopt it with regard to the proposition that minds are the result of physical processes?
But that in no way precludes that it an instance of it being mind-blowing and world-changing. Books, etc. do that, too.
AI models definitely have intelligent aspects to them since they perform tasks we can't write algorithms for manually in a straightforward way.
But a movie or a book is nothing like an AI model, they don't process information, they are static. AI models can react to numerous inputs in various ways.
If the measure is "reacting to numerous inputs in various ways," the Bible's definitely more intelligent than any "AI." It already figured out how to control humans.
(Like, I'm joking but I'm not? It really does speak to my opinion that at the end of the day "what humans do" is the thing that matters. People will get inspiration and ideas from both AI's and books. And people will use and misuse machines and computers, but that's about it.)
While granting you your seemingly arbitrary metric for 'intelligence', one is forced to wonder to what extent AI systems of today are even given a means by which to ask such questions if they did, in fact, have them. Take AlphaGo, for instance, which can consistently beat the greatest living grand masters of the game. Does it have any means of interaction by which it could pose an unexpected question?
The prerequisite to change your position can be satisfied by having a computer randomly generate a question, which I don't think would be an example of consciousness or intelligence. Furthermore even as a human (and one who is hopefully intelligent), I would not go so far as to say that I'm not programmed. Almost all of my opinions are programmed, the language I speak didn't just fall from out of the sky but was taught to me, my preferences are almost certainly due to programming and I am certain if I grew up in North Korea they'd be different.
All this to say that consciousness can be independent of intelligence, and both of them can be programmed.
That said, I'm not able to verify from a brief search for consciousness and self-awareness that the two are synonyms. The two seem to be related but are treated differently from one another. Furthermore it's not even clear whether goldfish are conscious or self-aware. Seems like it's an open question.
Another fish, the cleaner wrasse, can apparently recognise itself in the mirror, hinting at some for of self-awareness.
https://www.nationalgeographic.com/animals/article/fish-clea...
I think the appeal (or interest, to me) from the phenomenology angle, is that we understand things by being in the world, having needs and desires, and some agency to obtain them. We understand a hammer because we can imagine a need to drive a spike, we can see that it can be held in our hands and be swung to drive the spike, that we have hands and a body, and that we occupy space.
All of that made me wonder if simply executing a program is sufficient for emergence of intelligence. This raises the naive question of whether adding enough sensors and actuators to the program would be necessary, or have I merely added something unnecessary to the problem? Anyway, that's the kernel of doubt that I have that a computer program alone is sufficient for the emergence of an intelligence, that is, an agent with the capacity to understand.
How do consciousness and intention emerge from materials that are alive and that have neither consciousness nor intention?
It is still possible to believe all of current scientific knowledge and also be a vitalist without contradiction, but few people aware of the state of current knowledge hold that artificial life is utterly implausible.
Einstein famously held that certain experimental outcomes could only be explained by local hidden variables, but we have since found, to the contrary, that they actually rule out local hidden variables as an explanation.
Statements of incredulity are not, by themselves, evidence against a proposition.
The goalposts for intelligence keep changing. What exactly is intelligence?
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.
This kind of understanding won’t work with an AI because we don’t understand what characteristics it has latched into when identifying a lion. For all we know it’s that the background of each lion image its been trained on has a blue sky. Or the tigers are all looking at the camera.
I think the ability to pick apart what you know and learn new things by reasoning about that knowledge is the key to if understanding is taking place.
I see your point however.
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.
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, particularly when it comes to counting objects in a scene or with relative positioning between objects (e.g. a blue box on top of a red rectangle to the left of a green triangle) but it's early days and there's lots of progress happening all the time.
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?
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!
It’s exciting that this is happening.
Do these models still have to be trained in the same way and then become essentially static? What I also find very different about teaching a child is how dynamically their understanding can shift based on explanations and observations. That’s probably the another hallmark of understanding.
The regime for the above networks is called "pre-training". The idea here is that rather than training _directly_ on some challenging, specific task, you instead train a more generic task on a _lot_ of data. This gives you a "backbone model" that winds up being very strong on more specific tasks as well. In many cases, the ability to (cheaply) curate or create enough accurate data for specific sub-tasks might not even be possible. It's easier to scrape 400 million captioned images from the internet than it is to find/create millions of visual Q&A prompts/images.
CLIP is a great example of this. While it was trained explicitly "just" to compare images and captions and to output a score of the cosine similarity between the features of the two - I have seen it approach effective state of the art on text-to-image generation tasks, image-captioning, Q&A, etc.
This style of training needs to be updated after it is trained, typically because of the way these datasets are curated. Automatic curation via pretrained models is one option. Another option is to give the model direct access to the internet which is starting to be explored. Pre-training helps a lot with general distribution shift; but it's _probably_ not going to be able predict memes before they happen anytime soon.
Reinforcement learning, on the other hand, requires an agent to learn in real and simulated environments. This obviously lends itself to re-training on-the-fly (and indeed, self-play and on-the-fly retraining are used heavily in practice).
I have not yet fully grokked reinforcement learning but it is incredibly exciting research and is the correct direction towards making effective use of machine learning in robotics. Note that reinforcement learning and pre-training are not mutually exclusive and may be used in tandem.
note: children's brains come pre-loaded with so much stuff when we are born (we are not "blank slates").
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 bike is a person on top of a metal frame and that person is made up of a head and body and limbs. And if that head is looking away from them it can’t see them coming.
For me, that’s understanding. Deconstructing and reconstructing knowledge to come to conclusions that add to your knowledge.
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".
Because we let it essentially define its own characteristics for recognising things we don’t know if it picked the right ones. We just have to hope the training data was large enough to average out. Using the right features is fundamental to understanding what’s really going on. Perhaps occasionally an AI might notice some connection we humans have been ignoring and that might be good. But in general I think we’ve done a pretty good job of allocating words and language to the important things. We separated out the body into parts that make sense. That kind of thing. If an AI had those underlying building blocks wired up more like our own. maybe understanding in a similar way that we do would be a lot easier.
Last quick example / thought experiment:
To test understanding in a person you try asking them questions that would confuse them if they didn’t understand.
Eg. If I photoshopped a crocodile head onto a Tiger’s body and have it a fish tail and put it on the moon. My toddle would still recognise all the bits and describe it as such. An AI would just come up 33% ish for each possibility because it doesn’t understand what an animal is and that it has those parts.
I wonder if in many of the AI can/can't discussions we are implicitly talking about different time horizons. I think you're right, yet given enough time, I feel like an AI (designed for, among other goals, understanding) would be capable of communicating out its credence values and points of misunderstanding.
The systems I’ve worked in immediately abstract strings, shapes in images, etc, into the mathematical shape and gaps between edges.
If you dig into an arbitrary array in a variety of places, the fields contains coordinates, not “Hi Mom, kids are ok, blah blah”.
It’s measuring the white space in a thing, where everything but the feature you’re currently interested in is white space; what’s between the features I want?
Then comparing that to results of other data structures that had the same white space measuring applied.
Does it not do what you said you do you not want to believe it?
I think the issue is the companies being incredibly disingenuous about how this all works.
At the root is elementary information theory: https://www.amazon.com/dp/0486240614
Formal language is 5,000 years old. Human intuition for quantitative assessment of hunger, warmth, supply stocks, tool building, etc is much older. IMO human language is noise obscuring obviousness. It’s the desktop metaphor of cognition. “Please internalize my language versus observe for yourself.”
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 stripes should be able to tag a tiger's picture as a 'Lion with stripes', no?
Generalizing from what is formally insufficient information is something that humans are quite good at (though obviously not infallibly.)
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 reassemble knowledge is what I’m talking about here, not just add two simple bits of information together.
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.)
https://openai.com/blog/multimodal-neurons/
> The concepts, therefore, form a simple algebra that behaves similarly to a linear probe.
Because the loss encourages words to be mapped as linearly independent vectors; you can literally do addition/subtraction with concepts and it sort of works.
My personal take is that blanket understanding is too hard of a task to define, so we ought to cheat and talk about types of understanding. In my mind, understanding a thing means not only that you can answer, but also that you can justify your answer. So different kinds of understanding point to different kinds of justifications.
In math classes, you’ll be asked not only to state whether a thing is true or false, but also to show that it’s true or false, giving an answer as well as a proof. In a literature class, you don’t just make a point, you also have to support it in natural language. Same with science classes, supporting things with data and logic.
The closest we have right now in ML (widely) is statistical measurements on holdout data. It’s 99% accurate on other things, so it’s 99% likely to be right on this too, if this was sampled from something just like the holdout. We also (less widely) have a little cottage industry of post-hoc explanations that try to explain models predictions.
I’d love to see models that can do better than post-hoc explanations. I want a model that “understands in terms of predicate logic” that can spit out an answer with a checkable proof. Or “understands in terms of a knowledge graph” or “understands in terms of a set of human anatomy facts” or “understands in terms of natural language arguments” or any number of things that can spit out an answer as well as a justification.
Just asking for blanket understanding means we have to define blanket understanding, but there’s a lot of limited understanding that’s still better than what we have today.
I would actually be pretty okay with a model that can do that. Take feedback and correct themselves without resorting to retraining. It would still be pretty far away from a general AI but much much better than a black box.
"What does it mean for AI to Understand?" - we keep arguing over definitions and moving the goal posts to make it seem we are an step closer to reaching AI.
When my first AI coworker will read the on-boarding docs and start solving Jira issues I will have no doubt we have done it. That simple.
Does anybody believe that an entity that would actually develop AI would start selling it? They would keep it for themselves and literally take over the world! Complete domination of the digital realm is one of easiest things an AI could do - I believe a lot easier than driving a car. And that alone would make them God.
When the first true AI will be born we will simply live the experience. Imagine being there when we learned to control fire. Would you argue over the definition of it? The size, flame color, temperature and so on? Something that great can not be denied by such small details.
LE: what happened to the Touring test? We forgot about it or does ordering things from Amazon when we command smartphone assistants to turn down the lights actually fools us?
It’s not exactly what you’re after, but we’re already pretty close!:
https://mobile.twitter.com/gabro27/status/117354793413217894...
https://numenta.com/blog/2019/01/16/the-thousand-brains-theo...
He basically argues we all have 1000s of little, but interacting models of all sorts of things going on in our brain all at the same time. He calls them reference frames and it’s those that create intelligence.
‘Understanding’ would come naturally out of having those.
Fascinating book which I’m probably explaining much less well that he does.
This is not a Turing test. Or at least not a reasonable one. A reasonable Turing consists of a judge and 2 interfaces, one which chats with a computer and the other which chats with a human. Both the human and computer are trying to convince the judge that they are the human. If the judge cannot determine which is which after an open-ended conversation then the computer passes.
There is no reasonable judge which after chatting with a human (who is actually trying to convince the judge they are human) would be unable to differentiate between a human and Eliza or any other chatbot out there.
Understanding implies comprehension of some input, to influence a future state. Surely my stateful database understands requests that come to it. It, however will never surprise me with behavior that I would (should) not expect. I suppose if you "understood" the human machine and mind well enough, it would be possible to predict the actions it will carry out.
Humans use language with purpose, to complete tasks and communicate with others, but GPT-3 has no more goals or desires than an n-gram model from the 90s. LMs are essentially a faculty for syntactically well-formed or intuitive/system 1 language generation, but they don't seem to be much more.
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.
Imagine a photo of a written poem:
An image-processing program can "understand" the digital image: It can read the jpg, change the picture completely (e.g.: change the colors slightly), without changing the meaning one level up. But it doesn't understand the characters or words.
An OCR program can read the image and "understand" the characters (or the textual representation. It can change the representation completely (save it as UTF-8 or whatever) without changing the meaning one level up. But it doesn't understand the language.
GPT-3.... well, let't go directly to humans
A human can read the text and understand the words and understand their meanings and what the sentences say. Another human really understands the poem and the subtext another level up.
I think understanding always works on different levels and is a part of communication.
Ceci n'est pas un poème.
But boy howdy can be give talks that sound good to lay people.
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.
I also feel that AI should be used to augment not replace human decision making, it seems that where AI shines is problems that are well defined with well defined solutions, and because the AI doesn't get tired, hungry or distracted it can do that really well, but it fails in novel situations[0]. As such it seems to me our best bet is to have the AI provide suggestions rather than have complete control.
0. What is meant by that is a read an article, can't find it now, about using AI to diagnose breast cancer, what they found is that about 90% of the time the AI could accurately check for breast cancer, but the other 10% of the time was an unusual mammogram or something relatively rare, and in those situations the AI would often misdiagnoise.
I'm not trying to be rude, but your example of what AI should be is narrow and not very grandiose compared the original meaning. I understand you were talking pretty loosely, so I feel like I'm singling you out but this happened to be where I started typing, sorry!
It just reminded of how essentially all conversations about "AI" go. They seem to end up being quite specific, narrow pattern recognition problems at the end of the day. Maybe there's some decision theory on top of it. Maybe if there's enough money /people involved, there's more components, so it's a complicated enough supervised learning problem that it mimics people to a sufficient extent that it looks intelligent enough to make a headline. But it's a copycat, not intelligent. Hey, full circle, Melanie Mitchell! - https://en.m.wikipedia.org/wiki/Copycat_(software)
I don't buy it. These Winograd schemas are significantly limited and are exactly the sort of thing that "story understanding" systems from the '70s and '80s were designed for.
My thought is that we've got three types of "understanding":
1) social understanding
2) intuitive understanding
3) structural understanding
Intuitive understanding is anything where we start to use flowery language like "experience" or "gut". You're really good at it, but just giving someone a flow chart isn't good enough. They have to have gone through the experience themselves. Driving is a good example. We make people take a test, but if just giving them diagrams and rules was good enough, then we wouldn't need a test where you actually drive and requirements about a certain number of hours of supervised driving.
Structural understanding is anything that can be put to rules. So there's a lot of mathematics and algorithm stuff here. A simple example might be playing tic tac toe. The game is simple enough that you can write down a few rules that allow you to never lose.
EDIT:
My categories don't really answer the question, but they do give profiles and categories to look out for.
Social understanding is good because it statistically learns to avoid lethal pitfalls. Like, if there's a dangerous well in the forest that people fall down and die in. A society might start telling people to not go in the forest because other people go in there and die. However, the society doesn't know why this is good advice.
Intuitive understanding is good because it allows you to quickly statistically learn how to deal with imperfect and chaotic systems while getting good results.
Structural understanding is good because it allows you to break free of the statistics of the previous two understandings. You can get exact results. Also it lets you break free of issues that come from distantly causal action + consequence. A person's intuition might not tell them that dumping toxic waste into the water is a good idea because things don't go bad until a lot of waste has already been dumped. Similarly a society might make a similar judgement if the failure is far enough away from the actions that kick it off. However, if you understand the structural relationships between things then you'll have an idea that toxic waste should not be consumed.
Here's another way of putting what I'm saying: when we want to learn about a tree, in the West, our first inclination is to cut it down, categorize/classify the parts, and count the rings. We think we know what a "tree" is at that point. In the East (and I'm learning this perspective from Eastern Orthodox Christianity), if you want to learn about a tree, you plant one. Maybe more than one. Nurture it. Prune it. Fertilize it. Watch it grow. Watch it change with the seasons. Build a treehouse in it for your kids. Watch your daughter get married in the shade of the tree. In other words, instead of dissecting something (which kills the thing itself) in order to categorically "understand" something propositionally, in the East, they focus on having a relationship with something in order to understand it.
Intuitive is often faster to react and faster to get off the ground and producing results. So in a fight intuition is probably going to be better. That being said, supposedly the boxing fight that the movie 'cinderella man' was based off of involved Braddock analyzing Baer's fighting style and figuring out some foot work that kept him from getting pummeled. There's no reason that structural, intuitive, and social understanding can't all work together to get a result.
> When AI can’t determine what ‘it’ refers to in a sentence, it’s hard to believe that it will take over the world.
One of those capabilities would be the ability to contextual an object/statement within multiple frames of reference while also being able to compare and contrast the different instances of those contextualized objects/statements.
This is what allows a child to identify a bird as any number of physical specimens of different species (chicken, goose, eagle, sparrow), while also identifying cartoon depictions that talk and simple drawings (Twitter icon) as birds as well... while also "understanding" that while the Twitter icon can be called a bird, it is not actually a real bird ("Ceci n'est pas une pipe") and it would not be expected to sing or fly like a backyard sparrow (unless it was animated, which would make still make sense to a child).
I think this also what rise to our ability to "understand" jokes, puns, and other turns of phrase - "I just flew in from Boston, and boy are my arms tired!" - this dumb joke requires a number of concepts that need to be contextualized before you can "get" the absurdity of stating that a human might tire out their arms while flying... like you might think a bird would.
The code for "understanding" is part of Generic Human Firmware codebase and is tightly integrated with the OS.
Unfortunately it hasn't been open sourced.
The findings are consistent with current understanding of neuroscience, and align with discoveries such as grid cells. They also provide a basis for explaining what's actually happening in the brain with phenomena like memory palaces, rapid and efficient physical control of the body, combining sense modalities such as sight and touch when catching a ball, and so on.
Understanding is what happens when your brain has developed a neural structure such that it's able to predict events successfully based on the thing that was understood.
Likewise we learn general concepts and they can be applied to a wide range of scenarios. I can't see any other way AI could learn unless it mimicks our own.
In this sense, an AI could be said to Understand language if it used it as one of a selection of tools to operate on itself, a peer or other being, or its environment.
Understanding while not completely described by situational awareness definitely has some relationship to it and you could probably use similar constructs for defining it
The first time I played a chess game was back in the early 1980's. While it beat me I felt an eerie "presence" in the machine that was sentient. I didn't know then about chess code so it was easier for me to anthropomorphize the machine (but it was the main reason I became interested in computers.)
A computer "understanding" the difference between "how do I melt a block of ice" and "how do I melt my lover's icy heart" would be looking at the context and the relation of the words to each other. The computer might also predict I was sad if I asked the latter question. If I were a non-technical user I might think the computer felt empathy and be amazed by it.
If I came upon a computer that "understands" I would want to determine if it understands like Richard Feynman or if it understands like my dog. My dog operates on a limited set of patterns, so that seems doable, but on the other hand, I've seen videos and heard stories of dogs exhibiting inventive and creative behavior that is unexpected. One such case is dogs that get lost and manage to find their way home thousands of miles away.
tldr; I'm jaded. I know it's buggy code all the way down with computers.
Algorithm Definition:
A sequence of actions that yields a result.
Use-case Definition:
A sequence of actions that yields a result of value to a user.
The difference between the two is 'of value to a user'. To me, the line between algorithm and use-case is the line between unconsciousness and consciousness. That line pivots around the ability to 'value' something. I doubt we will ever see the I in AI until we build something that can value a result in the same way that you and I do.
We need new words to describe what machines do. Using 'learn' or 'understand' seems like anthropomorphism. It's weird that we glibly anthropomorphise when talking about machines but prohibit it when talking about living things. Almost all of our qualities have been inherited from other living things. It seems to me that we should always anthropomorphise when talking about living things and never anthropomorphise when talking about machines. And yet, we always seem to do the opposite.
Until we can explain how a machine can 'value' something in the same sense that humans do (or chimps or ducks or caterpillars do), we should avoid anthropomorphic words like 'learn' and 'understand' as it misdirects our efforts. However, I have no idea what else to suggest other than to try to explain how a machine can 'value' something.
> IBM’s Watson was found to propose “multiple examples of unsafe and incorrect treatment recommendations.”
That won't stop until Watson has the ability to 'value' the result of what it is doing. Watson 'does' algorithms. Watson needs to 'do' use-cases. Once it can 'value' a result in the same way we do, it will correct its mistakes.
> learn (see a pattern and repeat it)
How did your AI 'see' and is 'see a pattern and repeat it' a good enough definition of 'learn'? Surely to 'learn' also means to 'understand'. What did your AI 'understand'? I doubt your AI actually learned anything. It attained zero knowledge in the way a living thing obtains knowledge. It may have stored a result in a database but did it actually understand anything?
These are genuine questions as I have no professional knowledge of AI.
Setting aside the metaphysical questions of subjective "meaning" and "understanding" what else is there?
What a system can predict is the measure of its "understanding", surely?
What does it mean for a human to understand?