I think AI is neat.
Ok, but so do most humans? So few people actually have true understanding in topics. They parrot the parroting that they have been told throughout their lives. This only gets worse as you move into more technical topics. Ask someone why it is cold in winter and you will be lucky if they say it is because the days are shorter than in summer. That is the most rudimentary “correct” way to answer that question and it is still an incorrect parroting of something they have been told.
Ask yourself, what do you actually understand? How many topics could you be asked “why?” on repeatedly and actually be able to answer more than 4 or 5 times. I know I have a few. I also know what I am not able to do that with.
I feel that knowing what you don’t know is the key here.
An LLM doesn’t know what it doesn’t know, and that’s where what it spouts can be dangerous.
Of course there’s a lot of actual people that applies to as well. And sadly they’re often in positions of power.
I don’t think actual parroting is the problem. The problem is they don’t understand a word outside of how it is organized. They can’t be told to do simple logic because they don’t have a simple understanding of each word in their vocabulary. They can only reorganize things to varying degrees.
It doesn’t need to understand the words to perform logic because the logic was already performed by humans who encoded their knowledge into words. It’s not reasoning, but the reasoning was already done by humans. It’s not perfect of course since it’s still based on probability, but the fact that it can pull the correct sequence of words to exhibit logic is incredibly powerful. The main hard part of working with LLMs is that they break randomly, so harnessing their power will be a matter of programming in multiple levels of safe guards.
https://en.m.wikipedia.org/wiki/Chinese_room
I think they’re wrong, as it happens, but that’s the argument.
I guess, I just am looking at from an end user vantage point. I’m not saying the model cant understand the words its using. I just don’t think it currently understands that specific words refer to real life objects and there are laws of physics that apply to those specific objects and how they interact with each other.
Like saying there is a guy that exists and is a historical figure means that information is independently verified by physical objects that exist in the world.
This is only one type of intelligence and LLMs are already better at humans at regurgitating facts. But I think people really underestimate how smart the average human is. We are incredible problem solvers, and AI can’t even match us in something as simple as driving a car.
Lol @ driving a car being simple. That is one of the more complex sensory somatic tasks that humans do. You have to calculate the rate of all vehicles in front of you, assess for collision probabilities, monitor for non-vehicle obstructions (like people, animals, etc.), adjust the accelerator to maintain your own velocity while terrain changes, be alert to any functional changes in your vehicle and be ready to adapt to them, maintain a running inventory of laws which apply to you at the given time and be sure to follow them. Hell, that is not even an exhaustive list for a sunny day under the best conditions. Driving is fucking complicated. We have all just formed strong and deeply connected pathways in our somatosensory and motor cortexes to automate most of the tasks. You might say it is a very well-trained neural network with hundreds to thousands of hours spent refining and perfecting the responses.
The issue that AI has right now is that we are only running 1 to 3 sub-AIs to optimize and calculate results. Once that number goes up, they will be capable of a lot more. For instance: one AI for finding similarities, one for categorizing them, one for mapping them into a use case hierarchy to determine when certain use cases apply, one to analyze structure, one to apply human kineodynamics to the structure and a final one to analyze for effectiveness of the kineodynamic use cases when done by a human. This would be a structure that could be presented an object and told that humans use it and the AI brain could be able to piece together possible uses for the tool and describe them back to the presenter with instructions on how to do so.
Few people truly understand what understanding means at all, i got teacher in college that seriously thinked that you should not understand content of lessons but simply remember it to the letter
I am so glad I had one that was the opposite. I discussed practical applications of the subject material after class with him and at the end of the semester he gave me a B+ even though I only got a C by score because I actually grasped the material better than anyone else in the class, even if I was not able to evaluate it as well on the tests.
So super informed OP, tell me how they work. technically, not CEO press release speak. explain the theory.
I’m not OP, and frankly I don’t really disagree with the characterization of ChatGPT as “fancy autocomplete”. But…
I’m still in the process of reading this cover-to-cover, but Chapter 12.2 of Deep Learning: Foundations and Concepts by Bishop and Bishop explains how natural language transformers work, and then has a short section about LLMs. All of this is in the context of a detailed explanation of the fundamentals of deep learning. The book cites the original papers from which it is derived, most of which are on ArXiv. There’s a nice copy on Library Genesis. It requires some multi-variable probability and statistics, and an assload of linear algebra, reviews of which are included.
So obviously when the CEO explains their product they’re going to say anything to make the public accept it. Therefore, their word should not be trusted. However, I think that when AI researchers talk simply about their work, they’re trying to shield people from the mathematical details. Fact of the matter is that behind even a basic AI is a shitload of complicated math.
At least from personal experience, people tend to get really aggressive when I try to explain math concepts to them. So they’re probably assuming based on their experience that you would be better served by some clumsy heuristic explanation.
IMO it is super important for tech-inclined people interested in making the world a better place to learn the fundamentals and limitations of machine learning (what we typically call “AI”) and bring their benefits to the common people. Clearly, these technologies are a boon for the wealthy and powerful, and like always, have been used to fuck over everyone else.
IMO, as it is, AI as a technology has inherent patterns that induce centralization of power, particularly with respect to the requirement of massive datasets, particularly for LLMs, and the requirement to understand mathematical fundamentals that only the wealthy can afford to go to school long enough to learn. However, I still think that we can leverage AI technologies for the common good, particularly by developing open-source alternatives, encouraging the use of open and ethically sourced datasets, and distributing the computing load so that people who can’t afford a fancy TPU can still use AI somehow.
I wrote all this because I think that people dismiss AI because it is “needlessly” complex and therefore bullshit. In my view, it is necessarily complex because of the transformative potential it has. If and only if you can spare the time, then I encourage you to learn about machine learning, particularly deep learning and LLMs.
That’s my point. OP doesn’t know the maths, has probably never implemented any sort of ML, and is smugly confident that people pointing out the flaws in a system generating one token at a time are just parroting some line.
These tools are excellent at manipulating text (factoring in the biases they have, I wouldn’t recommended trying to use one in a multinational corporation in internal communications for example, as they’ll clobber non euro derived culture) where the user controls both input and output.
Help me summarise my report, draft an abstract for my paper, remove jargon from my email, rewrite my email in the form of a numbered question list, analyse my tone here, write 5 similar versions of this action scene I drafted to help me refine it. All excellent.
Teach me something I don’t know (e.g. summarise article, answer question etc?) disaster!
Fact of the matter is that behind even a basic AI is a shitload of complicated math.
Depending on how simple something can be to be considered an AI, the math is surprisingly simple compared to what an average person might expect. The theory behind it took a good amount of effort to develop, but to make something like a basic image categorizer (eg. optical character recognition) you really just need some matrix multiplication and calculating derivatives-- non-math-major college math type stuff.
you really just need some matrix multiplication and calculating derivatives-- non-math-major college math type stuff.
Well sure you don’t need a math degree for that, but most people really need to put some time into those topics. I.e., that kind of math is complex enough to constitute a barrier to entry into the field, particularly people with no free time to self-study or money for school.
Said differently: matrix math and basic calculus is hard, just not for you and I.
People learn the same way, we do things that bring us satisfaction and get us approval.
We use words to describe our thoughts and understanding. LLMs order words by following algorithms that predict what the user wants to hear. It doesn’t understand the meaning or implications of the words it’s returning.
It can tell you the definition of an apple, or how many people eat apples, or whatever apple data it was trained on, but it has no thoughts of it’s own about apples.
That’s the point that OOP was making. People confuse ordering words with understanding. It has no understanding about anything. It’s a large language model - it’s not capable of independent thought.
I think that the question of what “understanding” is will become important soon, if not already. Most people don’t really understand as much as you might think we do, an apple for example has properties like flavor, texture, appearance, weight and firmness it also is related to other things like trees and is in categories like food or fruit. A model can store the relationship of apple to other things and the properties of apples, the model could probably be given “personal preferences” like a preferred flavor profile and texture profile and use this to estimate if apples would be preferred by the preferences and give reasonings for it.
Unique thought is hard to define and there is probably a way to have a computer do something similar enough to be indistinguishable, probably not through simple LLMs. Maybe using a LLM as a way to convert internal “ideas” to external words and external words to internal “ideas” to be processed logically probably using massive amounts of reference materials, simulation, computer algebra, music theory, internal hypervisors or some combination of other models.
I fully back your sentiment OP; you understand as much about the world as any LLM out there and don’t let anyone suggest otherwise.
Signed, a “contrarian”.
They’re kind of right. LLMs are not general intelligence and there’s not much evidence to suggest that LLMs will lead to general intelligence. A lot of the hype around AI is manufactured by VCs and companies that stand to make a lot of money off of the AI branding/hype.
I find this line of thinking tedious.
Even if LLM’s can’t be said to have ‘true understanding’ (however you’re choosing to define it), there is very little to suggest they should be able to understand predict the correct response to a particular context, abstract meaning, and intent with what primitive tools they were built with.
If there’s some as-yet uncrossed threshold to a bare-minimum ‘understanding’, it’s because we simply don’t have the language to describe what that threshold is or know when it has been crossed. If the assumption is that ‘understanding’ cannot be a quality granted to a transformer-based model -or even a quality granted to computers generally- then we need some other word to describe what LLM’s are doing, because ‘predicting the next-best word’ is an insufficient description for what would otherwise be a slight-of-hand trick.
There’s no doubt that there’s a lot of exaggerated hype around these models and LLM companies, but some of these advancements published in 2022 surprised a lot of people in the field, and their significance shouldn’t be slept on.
Certainly don’t trust the billion-dollar companies hawking their wares, but don’t ignore the technology they’re building, either.
You are best off thinking of LLMs as highly advanced auto correct. They don’t know what words mean. When they output a response to your question the only process that occurred was “which words are most likely to come next”.
That’s only true on a very basic level, I understand that Turings maths is complex and unintuitive even more so than calculus but it’s a very established fact that relatively simple mathematical operations can have emergent properties when they interact to have far more complexity than initially expected.
The same way the giraffe gets its spots the same way all the hardware of our brain is built, a strand of code is converted into physical structures that interact and result in more complex behaviours - the actual reality is just math, and that math is almost entirely just probability when you get down to it. We’re all just next word guessing machines.
We don’t guess words like a Markov chain instead use a rather complex token system in our brain which then gets converted to words, LLMs do this too - that’s how they can learn about a subject in one language then explain it in another.
Calling an LLM predictive text is a fundamental misunderstanding of reality, it’s somewhat true on a technical level but only when you understand that predicting the next word can be a hugely complex operation which is the fundamental math behind all human thought also.
Plus they’re not really just predicting one word ahead anymore, they do structured generation much like how image generators do - first they get the higher level principles to a valid state then propagate down into structure and form before making word and grammar choices. You can manually change values in the different layers and see the output change, exploring the latent space like this makes it clear that it’s not simply guessing the next word but guessing the next word which will best fit into a required structure to express a desired point - I don’t know how other people are coming up with sentences but that feels a lot like what I do
Even if LLM’s can’t be said to have ‘true understanding’ (however you’re choosing to define it), there is very little to suggest they should be able to
understandpredict the correct response to a particular context, abstract meaning, and intent with what primitive tools they were built with.
Did you mean “shouldn’t”? Otherwise I’m very confused by your response
No, i mean ‘should’, as in:
There’s no reason to expect a program that calculates the probability of the next most likely word in a sentence should be able to do anything more than string together an incoherent sentence, let alone correctly answer even an arbitrary question
It’s like using a description for how covalent bonds are formed as an explanation for how it is you know when you need to take a shit.
Yes. But the more advanced LLMs get, the less it matters in my opinion. I mean of you have two boxes, one of which is actually intelligent and the other is “just” a very advanced parrot - it doesn’t matter, given they produce the same output. I’m sure that already LLMs can surpass some humans, at least at certain disciplines. In a couple years the difference of a parrot-box and something actually intelligent will only merely show at the very fringes of massively complicated tasks. And that is way beyond the capability threshold that allows to do nasty stuff with it, to shed a dystopian light on it.
I mean of you have two boxes, one of which is actually intelligent and the other is “just” a very advanced parrot - it doesn’t matter, given they produce the same output.
You’re making a huge assumption; that an advanced parrot produces the same output as something with general intelligence. And I reject that assumption. Something with general intelligence can produce something novel. An advanced parrot can only repeat things it’s already heard.
How do you define novel? Because LLMs absolutely have produced novel data.
I use LLMs to create things no human has likely ever said and it’s great at it, for example
‘while juggling chainsaws atop a unicycle made of marshmallows, I pondered the existential implications of the colour blue on a pineapples dream of becoming a unicorn’
When I ask it to do the same using neologisms the output is even better, one of the words was exquimodal which I then asked for it to invent an etymology and it came up with one that combined excuistus and modial to define it as something beyond traditional measures which fits perfectly into the sentence it created.
You can’t ask a parrot to invent words with meaning and use them in context, that’s a step beyond repetition - of course it’s not full dynamic self aware reasoning but it’s certainly not being a parrot
The difference is that you can throw enough bad info at it that it will start paroting that instead of factual information because it doesn’t have the ability to criticize the information it receives whereas an human can be told that the sky is purple with orange dots a thousand times a day and it will always point at the sky and tell you “No.”
To make the analogy actually comparable the human in question would need to be learning about it for the first time (which is analogous to the training data) and in that case you absolutely could convince the small child of that. Not only would they believe it if told enough times by an authority figure, you could convince them that the colors we see are different as well, or something along the lines of giving them bad data.
A fully trained AI will tell you that you’re wrong if you told it the sky was orange, it’s not going to just believe you and start claiming it to everyone else it interacts with. It’s been trained to know the sky is blue and won’t deviate from that outside of having its training data modified. Which is like brainwashing an adult human, in which case yeah you absolutely could have them convinced the sky is orange. We’ve got plenty of information on gaslighting, high control group and POW psychology to back that up too.
Ha ha yeah humans sure are great at not being convinced by the opinions of other people, that’s why religion and politics are so simple and society is so sane and reasonable.
Helen Keller would belive you it’s purple.
If humans didn’t have eyes they wouldn’t know the colour of the sky, if you give an ai a colour video feed of outside then it’ll be able to tell you exactly what colour the sky is using a whole range of very accurate metrics.
But also the people who seem to think we need a magic soul to perform useful work is way way too high.
The main problem is Idiots seem to have watched one too many movies about robots with souls and gotten confused between real life and fantasy - especially shitty journalists way out their depth.
This big gotcha ‘they don’t live upto the hype’ is 100% people who heard ‘ai’ and thought of bad Will Smith movies. LLMs absolutely live upto the actual sensible things people hoped and have exceeded those expectations, they’re also incredibly good at a huge range of very useful tasks which have traditionally been considered as requiring intelligence but they’re not magically able everything, of course they’re not that’s not how anyone actually involved in anything said they would work or expected them to work.
Depends on what you mean by general intelligence. I’ve seen a lot of people confuse Artificial General Intelligence and AI more broadly. Even something as simple as the K-nearest neighbor algorithm is artificial intelligence, as this is a much broader topic than AGI.
Wikipedia gives two definitions of AGI:
An artificial general intelligence (AGI) is a hypothetical type of intelligent agent which, if realized, could learn to accomplish any intellectual task that human beings or animals can perform. Alternatively, AGI has been defined as an autonomous system that surpasses human capabilities in the majority of economically valuable tasks.
If some task can be represented through text, an LLM can, in theory, be trained to perform it either through fine-tuning or few-shot learning. The question then is how general do LLMs have to be for one to consider them to be AGIs, and there’s no hard metric for that question.
I can’t pass the bar exam like GPT-4 did, and it also has a lot more general knowledge than me. Sure, it gets stuff wrong, but so do humans. We can interact with physical objects in ways that GPT-4 can’t, but it is catching up. Plus Stephen Hawking couldn’t move the same way that most people can either and we certainly wouldn’t say that he didn’t have general intelligence.
I’m rambling but I think you get the point. There’s no clear threshold or way to calculate how “general” an AI has to be before we consider it an AGI, which is why some people argue that the best LLMs are already examples of general intelligence.
Depends on what you mean by general intelligence. I’ve seen a lot of people confuse Artificial General Intelligence and AI more broadly. Even something as simple as the K-nearest neighbor algorithm is artificial intelligence, as this is a much broader topic than AGI.
Well, I mean the ability to solve problems we don’t already have the solution to. Can it cure cancer? Can it solve the p vs np problem?
And by the way, wikipedia tags that second definition as dubious as that is the definition put fourth by OpenAI, who again, has a financial incentive to make us believe LLMs will lead to AGI.
Not only has it not been proven whether LLMs will lead to AGI, it hasn’t even been proven that AGIs are possible.
If some task can be represented through text, an LLM can, in theory, be trained to perform it either through fine-tuning or few-shot learning.
No it can’t. If the task requires the LLM to solve a problem that hasn’t been solved before, it will fail.
I can’t pass the bar exam like GPT-4 did
Exams often are bad measures of intelligence. They typically measure your ability to consume, retain, and recall facts. LLMs are very good at that.
Ask an LLM to solve a problem without a known solution and it will fail.
We can interact with physical objects in ways that GPT-4 can’t, but it is catching up. Plus Stephen Hawking couldn’t move the same way that most people can either and we certainly wouldn’t say that he didn’t have general intelligence.
The ability to interact with physical objects is very clearly not a good test for general intelligence and I never claimed otherwise.
I know the second definition was proposed by OpenAI, who obviously has a vested interest in this topic, but that doesn’t mean it can’t be a useful or informative conceptualization of AGI, after all we have to set some threshold for the amount of intelligence AI needs to display and in what areas for it to be considered an AGI. Their proposal of an autonomous system that surpasses humans in economically valuable tasks is fairly reasonable, though it’s still pretty vague and very much debatable, which is why this isn’t the only definition that’s been proposed.
Your definition is definitely more peculiar as I’ve never seen anyone else propose something like it, and it also seems to exclude humans since you’re referring to problems we can’t solve.
The next question then is what problems specifically AI would need to solve to fit your definition, and with what accuracy. Do you mean solve any problem we can throw at it? At that point we’d be going past AGI and now we’re talking about artificial superintelligence…
Not only has it not been proven whether LLMs will lead to AGI, it hasn’t even been proven that AGIs are possible.
By your definition AGI doesn’t really seem possible at all. But of course, your definition isn’t how most data scientists or people in general conceptualize AGI, which is the point of my comment. It’s very difficult to put a clear-cut line on what AGI is or isn’t, which is why there are those like you who believe it will never be possible, but there are also those who argue it’s already here.
No it can’t. If the task requires the LLM to solve a problem that hasn’t been solved before, it will fail.
Ask an LLM to solve a problem without a known solution and it will fail.
That’s simply not true. That’s the whole point of the concept of generalization in AI and what the few-shot and zero-shot metrics represent - LLMs solving problems represented in text with few or no prior examples by reasoning beyond what they saw in the training data. You can actually test this yourself by simply signing up to use ChatGPT since it’s free.
Exams often are bad measures of intelligence. They typically measure your ability to consume, retain, and recall facts. LLMs are very good at that.
So are humans. We’re also deterministic machines that output some action depending on the inputs we get through our senses, much like an LLM outputs some text depending on the inputs it received, plus as I mentioned they can reason beyond what they’ve seen in the training data.
The ability to interact with physical objects is very clearly not a good test for general intelligence and I never claimed otherwise.
I wasn’t accusing you of anything, I was just pointing out that there are many things we can argue require some degree of intelligence, even physical tasks. The example in the video requires understanding the instructions, the environment, and how to move the robotic arm in order to complete new instructions.
I find LLMs and AGI interesting subjects and was hoping to have a conversation on the nuances of these topics, but it’s pretty clear that you just want to turn this into some sort of debate to “debunk” AGI, so I’ll be taking my leave.
Can your calculator only serve problems you already solved? I really don’t buy that take
Llms are in fact not at all good at retaining facts, it’s one of the most worked on problems for them
Llms can solve novel problems. It’s actually much more complex than just a lookup robot, which we already have for such tasks
You just take wild guesstimates on how they work and it just feels wrong to me to not point that out
Here is an alternative Piped link(s):
Piped is a privacy-respecting open-source alternative frontend to YouTube.
I’m open-source; check me out at GitHub.
It depends a lot on how we perceive “intelligence”. It’s a lot more vague of a term than most, so people have very different views of it. Some people might have the idea of it meaning the response to stimuli & the output (language or art or any other form) being indistinguishable from humans. But many people may also agree that whales/dolphins have the same level of, or superior, “intelligence” to humans. The term is too vague to really prescribe with confidence, and more importantly people often use it to mean many completely different concepts (“intelligence” as a measurable/quantifiable property of either how quickly/efficiently a being can learn or use knowledge or more vaguely its “capacity to reason”, “intelligence” as the idea of “consciousness” in general, “intelligence” to refer to amount of knowledge/experience one currently has or can memorize, etc.)
In computer science “artificial intelligence” has always simply referred to a program making decisions based on input. There was never any bar to reach for how “complex” it had to be to be considered AI. That’s why minecraft zombies or shitty FPS bots are “AI”, or a simple algorithm made to beat table games are “AI”, even though clearly they’re not all that smart and don’t even “learn”.
Even sentience is on a scale. Even cows or dogs or parrots or crows are sentient, but not as much as we are. Computers are not sentient yet, but one day they will be. And then soon after they will be more sentient than us. They’ll be able to see their own brains working, analyze their own thoughts and emotions(?) in real time and be able to achieve a level of self reflection and navel gazing undreamed of by human minds! :D
People who don’t understand or use AI think it’s less capable than it is and claim it’s not AGI (which no one else was saying anyways) and try to make it seem like it’s less valuable because it’s “just using datasets to extrapolate, it doesn’t actually think.”
Guess what you’re doing right now when you “think” about something? That’s right, you’re calling up the thousands of experiences that make up your “training data” and using it to extrapolate on what actions you should take based on said data.
You know how to parallel park because you’ve assimilated road laws, your muscle memory, and the knowledge of your cars wheelbase into a single action. AI just doesn’t have sapience and therefore cannot act without input, but the process it does things with is functionally similar to how we make decisions, the difference is the training data gets input within seconds as opposed to being built over a lifetime.
People who aren’t programmers, haven’t studied computer science, and don’t understand LLMs are much more impressed by LLMs.
If you’ve ever actually used any of these algorithms it becomes painfully obvious they do not “think”. Give it a task slightly more complex/nuanced than what it has been trained on and you will see it draws obviously false conclusions that would be obviously wrong had any thought actual taken place. Generalization is not something they do, which is a fundamental part of human problem solving.
Make no mistake: they are text predictors.
I believe they were implying that a lot of the people who say “it’s not real AI it’s just an LLM” are simply parroting what they’ve heard.
Which is a fair point, because AI has never meant “general AI”, it’s an umbrella term for a wide variety of intelligence like tasks as performed by computers.
Autocorrect on your phone is a type of AI, because it compares what words you type against a database of known words, compares what you typed to those words via a “typo distance”, and adds new words to it’s database when you overrule it so it doesn’t make the same mistake.
It’s like saying a motorcycle isn’t a real vehicle because a real vehicle has two wings, a roof, and flies through the air filled with hundreds of people.
Which is a fair point, because AI has never meant “general AI”, it’s an umbrella term for a wide variety of intelligence like tasks as performed by computers.
Do you mean in the everyday sense or the academic sense? I think this is why there’s such grumbling around the topic. Academically speaking that may be correct, but I think for the general public, AI has been more muddled and presented in a much more robust, general AI way, especially in fiction. Look at any number of scifi movies featuring forms of AI, whether it’s the movie literally named AI or Terminator or Blade Runner or more recently Ex Machina.
Each of these technically may be presenting general AI, but for the public, it’s just AI. In a weird way, this discussion is sort of an inversion of what one usually sees between academics and the public. Generally academics are trying to get the public not to use technical terms loosely, yet here some of the public is trying to get some of the tech/academic sphere to not, at least as they think, use technical terms loosely.
Arguably it’s from a misunderstanding, but if anyone should understand the dynamics of language, you’d hope it would be those trying to calibrate machines to process language.
I’ve often seen people on Lemmy confidently state that current “AI” thinks and learns exactly like humans and that LLMs work exactly like human brains, etc.
I believe OP is attempting to take on an army of straw men in the form of a poorly chosen meme template.
