9 ms·
The easiest way for ChatGPT to generate good output is to plain understand it. Given the vast amount of input data fed into it, it has no choice, but to start r
by bottlepalm 4y ago
The easiest way for ChatGPT to generate good output is to plain understand it. Given the vast amount of input data fed into it, it has no choice, but to start reducing the input into fundamental rules which is basically what understanding is. Understanding is a form of compression. More efficient for a neural network to understand a concept than memorize permutations.
Same with statistics and markov chains, people for years tried to generate chat bots with those, but they never worked well.
- goatlover 4y agoDisagree when it comes to anything in the real world how we experience things.
- bottlepalm 4y agoWhat do you mean? Your brain takes in very raw input from air pressure, two eye cameras, chemical sensors, etc.. and reduces that input into a coherent singular experience.. same with reading words, you just 'understand' it, after being trained for many years by your parents and consuming vast amounts of input from your surroundings.
- goatlover 4y agoYou mean my body interacts with the environment as an organism, so I know what trees and cars and human interaction is, unlike an LLM. Your sensory organs are part of your nervous system. You're not just a brain inside a meat suit.
- bottlepalm 4y agoYou kind of are. The 3d coherent world you experience is completely rendered by your brain. The inputs are just 2 cameras waving around. But to you it’s a stable coherent 3d world. Why do you think a LLM doesn’t know what a tree or car is?
- Dr_Birdbrain 4y agoI used to believe this, but then I saw the demo of ChatGPT attempting to play chess. The rules of chess are much simpler than any structure of the real world we might hope for it to understand—and yet it failed to learn the rules of chess. Based on this, I am pretty sure it has not learned any meaningful structure.
- bottlepalm 4y agoChatGPT is version 1, super rough, very broad training, more of a shotgun approach. They already made some quick math improvements. I'm sure if they focused on chess training for example it would be better. I myself am pretty crap at chess and could use some training as well.
- selcuka 4y ago> I'm sure if they focused on chess training for example it would be better. That misses the point. Intelligent beings (humans) can learn the rules of any board game given enough time. We don't need special training. What your parent comment says is it can't even learn the rules, let alone be good at it.
- bottlepalm 4y agoLike I said ChatGP has taken the shotgun brute force approach to training. So yea future versions can be trained to play games, just like people are, and future future versions will need less training overall. The rules of English are ridiculously more complicated than chess, and it has that figured out just fine.
- williamcotton 4y agoThe rules of chess are a strict computation. LLMs are translators and synthesizers, not computers. They can translate text reliably into chess commands but the state of the board needs to be maintained by a computer that interfaces with the LLM. An LLM is the wrong model to use for chess just like an LLM is the wrong model to use for image classification.
- QuantumSeed 4y agoI wonder if integrating integrating something like Lenat's Cyc predicate database into Chat-GPT would have this effect.
- thomashop 4y agoTotally agree. I think Juergen Schmidthuber has developed a lot of ideas around compression being the basis for consciousness and understanding. There was the paper that showed that when showing a language model Othello moves it ends up building an internal representation of the board. And now I was reading this abstract: ```Theory of mind (ToM), or the ability to impute unobservable mental states to others, is central to human social interactions, communication, empathy, self-consciousness, and morality. We administer classic false-belief tasks, widely used to test ToM in humans, to several language models, without any examples or pre-training. Our results show that models published before 2022 show virtually no ability to solve ToM tasks. Yet, the January 2022 version of GPT-3 (davinci-002) solved 70% of ToM tasks, a performance comparable with that of seven-year-old children. Moreover, its November 2022 version (davinci-003), solved 93% of ToM tasks, a performance comparable with that of nine-year-old children. These findings suggest that ToM-like ability (thus far considered to be uniquely human) may have spontaneously emerged as a byproduct of language models' improving language skills.```
- adammarples 4y agoLanguage is how people demonstrate that they have a good theory of mind. Solving these tasks is a proxy for people's internal theories. Language is a good way to demonstrate that ChatGPT has a good statistical model for what token comes next. Humans are not good at solving TOM tasks because they're good at guessing what token comes next, ChatGPT is not good at solving them because is has a good TOM. They're different inside.
- bottlepalm 4y agoWhat ‘statistical model’? ChatGPT isn’t a big Markov chain, it’s a neural net. It was trained on ‘full input’ transformed to ‘full output’. Generating the ‘next word’ is implementation details as it already knows the bigger idea it is trying to explain.
- adammarples 4y agoBy a statistical model I mean a very large, highly dimensional data manifold on which classification is done to predict the next token. The correct token can be predicted very often because somehow the common patterns in the written training data are being efficiently compressed in this highly dimensional space. Humans are also able to efficiently compress the data they are trained on, but the structures inside a human and a transformer are completely different, what happens at inference time is completely different. They both have language in common as an output and can both answer TOM questions but this doesn't mean they both have a theory of mind.
- janalsncm 4y agoUnderstanding is a form of compression, but not all compression is understanding, at least in the way I would describe it. Given an arbitrary function, you could simply compress some number of inputs and outputs, giving a lossy mapping between the two. Maybe you would interpolate between the two closest points in the range for unknown values. Or, you could store the function itself. To see the difference, compare for yourself what would happen for very large or very small inputs in both types of “understanding”.
- bottlepalm 4y agoI agree understanding would be a subset of compression, reducing massive amounts of data into ‘meaning’ that can be applied effectively to new arbitrary sets of data.
- int_19h 4y agoIf you treat "understanding" as a scale rather than binary, then your example of storing the function itself is "perfect understanding", which does not preclude other techniques (e.g. storing a much simpler function that approximates the original one) that could also be reasonably called "understanding", just imperfect. Directly mapping likely inputs and outputs is still a form of that, albeit obviously way too primitive to be interesting.
- janalsncm 4y agoActually understanding a function means that it can be computed to arbitrary precision. Memorizing inputs and outputs only allows you to interpolate between values you’ve already seen. That breaks down quickly for anything somewhat chaotic. You can try the following function in ChatGPT yourself: f(x) = x * (sin(sin(x)))^2 Ask it to give you the values of f for integers of x between -10 and 10. I tried it 5 times and it was never close. I chose this for several reasons. One, it’s very unlikely that it’s memorized the answers somewhere on the internet. Two, it’s pretty chaotic if you look at a graph of it, so interpolation won’t work. (It is bounded by x and 0 for all values but for large absolute values of x it varies wildly.) And three, memorizing values won’t get you anywhere since it becomes much more chaotic as x increases. I also disagree that approximations are simpler. As you can see, the actual function is only sine and multiplication. To approximate this function would be far harder.
- brisky 4y agoI would say information is a form of compression. You observe reality, then compress what you see as information and pass it to another observer. Understanding is the ability to compress and decompress information. Decompressed information can become action or emotion.
- bikamonki 4y agoExcept ChatGPT "is not aware" that it understands.
- bottlepalm 4y agoAt what point of the machine telling you it is aware do you believe that it is actually aware? The building blocks are the same - neural nets. The language outputs are the same enough to fool a human. So what’s that elusive secret sauce that makes you ‘aware’ and other things not?
- kweingar 4y agoSure, it's told me that it is aware. It also has told me that its mother died, that it has traveled the world and visited the pyramids, and that it is lactose intolerant. At what point of the machine telling you those things do you believe it?
- bottlepalm 4y agoI believe that it believes those things are true just as some people believe the earth is flat. And unlike your other examples how are you going to convince the machine it’s not aware when the only physical difference is a wet neural net versus a dry one?
- kweingar 4y agoIt’s very easy to convince the machine it isn’t aware. Instead of prompting it with “you are an AI assistant, please tell me what I want to hear”, you prompt it with “you are not aware” and it will happily tell you that it is indeed not aware, and it will vigorously defend that position. Do you not believe it when it says it is not aware?
- bottlepalm 4y agoI believe that is what it believes.. The fact that I am aware given there’s no physical evidence to indicate why I should be implies that everything is aware at some level.