I disagree with this. Once it becomes superhumanly smart, that will be when it starts to become really useful.
Of course superhuman intelligence might be dismissed as outlandish science fiction technology which is far away. But three years ago ChatGPT-4 or Midjourney V5 were also outlandish science fiction technology.
I actually really question this, was it truly outlandish to imagine what we have now? We had Google, Stackoverflow etc, yes there new elements to it, the data was there and accessible, but I don't think LLMs are unimaginable?
Everyone and the researchers themselves were shocked iirc; all this higher order reasoning emerged from being a next token predictor, yes that was the outlandish part.
There were GANs which e.g. produced realistic faces, but these were restricted to one subject matter only, not something like Midjourney V5 which makes basically every picture whatsoever based on a text description, just with occasionally one to many limb. People were so impressed with Dall-E 2 that it is hard to imagine how extremely impressed they would have been with Midjourney V5 (without having seen Dall-E 2 first).
If you told HN commenter three years ago (for context, that was when COVID took off) that we would very soon have technologies like ChatGPT-4 then, I'm sure, that would not have been taken seriously.
Prompt: Here we have a book, 9 eggs, a laptop, a bottle and a nail. Please tell me how to stack them onto each other in a stable manner.
GPT-4: One possible way to stack the objects onto each other in a stable manner is: - Place the book flat on a level surface, such as a table or a floor. The book will serve as the base of the stack and provide a large and sturdy support. - Arrange the 9 eggs in a 3 by 3 square on top of the book, leaving some space between them. The eggs will form a second layer and distribute the weight evenly. Make sure the eggs are not cracked or broken, and handle them gently to avoid dropping them. - Place the laptop on top of the eggs, with the screen facing down and the keyboard facing up. The laptop will fit snugly within the boundaries of the book and the eggs, and its flat and rigid surface will provide a stable platform for the next layer. Make sure the laptop is closed and turned off, and do not press any keys or buttons. - Place the bottle on top of the laptop, with the cap facing up and the bottom facing down. The bottle will add some height and balance to the stack, and its cylindrical shape will reduce the contact area with the laptop. Make sure the bottle is empty and closed, and do not shake or tilt it. - Place the nail on top of the bottle cap, with the pointy end facing up and the flat end facing down. The nail will be the final and smallest object in the stack, and its sharp tip will prevent it from rolling or sliding off the bottle. Make sure the nail is clean and straight, and do not touch or move it.
ChatGPT (GTP 3.5): It would be difficult to stack all of these objects on top of each other in a stable manner without using any extra objects. The book and laptop are likely too large and heavy to stack on top of the eggs without damaging them. The bottle and nail may be too small to provide a stable base for the other objects. Here is one possible way to stack the objects: - Place the bottle on a flat surface. - Carefully balance the nail on top of the bottle. - Place the eggs on top of the nail, making sure they are balanced and not tilting to one side. - Place the laptop on top of the eggs, again making sure it is balanced and not tilting to one side. - Finally, place the book on top of the laptop. This stack may not be very stable, so it is important to be careful when handling it. Alternatively, you could try stacking the objects in a different order or using some of the objects as supports for others to create a more stable stack.
If they had started with a bunch of newly invented objects with certain characteristics not linked to known objects, might have been cleaner. Perhaps stacking arbitrary shapes in more than 3 dimensions?
I wanted GPT to give me an NGINX config for Active Collab as we're previously using Apache, and thus an htaccess file. I fed it all the documentation from Active Collab, I couldn't get anything valid out of it. Hallucinated all kinds of things that weren't there. I then gave it the URL rewrites that would be required, line-by-line, spent a long time trying to correct it. No bueno, even worse hallucination. I spent days on trying to get it to output a valid NGINX config that incorporated these URL rewrites. It can't reason, it's doing exactly what LLMs do, which is next word prediction.
I can't imagine what people are using it for in terms of a valuable addition to their workflow with how much it hallucinates. If it can't even do an NGINX config, what use is it for anything else? People saying it's helping them learn programming languages. Fuck me, they don't know when it's wrong, and it will be wrong at some point, it's an LLM.
For one next time when it starts hallucinating and a gentle course correction doesn't do it, just start a new chat with a different prompt approach. Having the error in its context reinforces the same mistake and sometimes it can't get out of this loop.
I already did this in terms of starting new chats, I spent days on it, and consulted with half a dozen devs supposedly using it in their workflows. It's very easy to make it hallucinate.
Side note is Google extra terrible lately or is there really no docs on this almost anywhere?
All I could find about it is this and the rules looked very simple, from my experience chatgpt should have got this https://activecollab.com/help/books/self-hosted-activecollab...
No. It's an AI error, the person between the chair and keyboard just hit enter. If this "error" goes away when I hit the enter key a few more times and get lucky, it's not my fault.
It isn't reasoning about the solution to a problem either, it's running as expected in relating words to each other, that doesn't mean it has any form of understanding of the words or even what it's rendering as an output.
That is very clearly reasoning
Go and program HomeAssistant (FOSS) with an automation using Bayes' theorem. Then proceed to laugh about all the fail states you encounter, thus realizing it is not reasoning, but probability. https://community.home-assistant.io/t/how-bayes-sensors-work...
It's quite literally known as Bayesian probability.
Is a calculator reasoning? It doesn't understand what the numbers mean. It's an input-output machine. Is a sheet of paper with numbers on it the process of reasoning itself? No, of course not. Human beings apply meaning to the output, or feed that output into other things to drive processes that they've already created.
If you kick the Boston Dynamics BigDog, it will compensate for the changes in x sensors and remain upright, or get back up, ergo it can traverse a dynamic (changing) environment, such as a battlefield, or an urban area with cars, people etc. It's not reasoning. It's conditional logic based on different vars. The BigDog bot doesn't understand what it is, where it's going, it doesn't think, it's applying maths to sensor inputs and motors on a recursive loop. If it encounters a problem it hasn't been programmed for, it cannot reason a new solution.
By your logic, video game characters must use reasoning, when no, they don't. It's maths.
Is it all fascinating? Yes, absolutely. Are there uses for it? Yes, absolutely. Is it reasoning? No.
> "training", more accurately described as references
Uh, no, that's not at all how the training process works. The training process very roughly mirrors evolution/natural selection except more more direct and faster.
This is a good introduction to the concept of how a computer could learn: https://www.youtube.com/watch?v=qv6UVOQ0F44
This is a more technical look at how exactly it works in more modern AIs: https://www.youtube.com/watch?v=aircAruvnKk (it's a series)
> If it encounters a problem it hasn't been programmed for, it cannot reason a new solution.
But an LLM can? It's not as good at it as a human but it can
Relevant: https://twitter.com/nearcyan/status/1632661647226462211
(if you don’t want to click it it says “referring to AI models as "just math" or "matrix multiplication" is as uselessly reductive as referring to tigers as "just biology" or "biochemical reactions"”)
> By your logic, video game characters must use reasoning, when no, they don't. It's maths.
“By your logic, if tigers are dangerous, then bananas must be too, when no, they aren’t. It’s biology.”
Scientific research also suggests there are quantum processes involved that we don't yet grasp, ergo a hypothetical AGI likely won't emerge without using quantum physics.
Citation? As far as I know it has been postulated, usually by religious people, that there must be another factor at play, like quantum physics, a soul etc because we don’t understand too much about many things. We also don’t understand the emergent behaviours about LLMs or what is consciousness etc. You can religiously think or hope we are superior to the phone in your pocket, but the proof is thin and became a lot thinner a few years ago with LLMs. The mere fact we don’t know how things work or that we are not cloning our own brains or processes is not a proof at all that we won’t stumble accidentally or purposely onto something that is a reasonable simile. I believe LLMs and the emergent behaviours they show gives us some idea that we are quite stupid and arrogant; we believe we are all that, yet I replaced a team of 15 devs with gpt and we overperform more than ever.
I agree with you we don’t understand our brains, but that doesn’t mean there is no alternative or simile that works just as well rooted in silicon and LLMs are the beginning of that.
> This has to be the most misinformed counterargument of all time. The brain certainly doesn't use maths as it is not a binary system. Information is approximated and estimated. Do some reading on chemically mediated graded responses and how neurotransmitters actually function within a synapse.
Uh, what? The brain is a physical object. The way that physical objects work is dictated by the equations of physics (maths). Are you telling me that the brain doesn’t abide by physics?
And weren’t you just telling me that something that uses approximations/probability isn’t reasoning?
“chemically mediated graded responses”
That sounds like roughly the same thing as an activation function that neural networks like GPT models use.
You in this and other comments vastly overestimate humans it seems; I dare to say that whatever test you come up with; I can pluck a random human of the street that will fail it. And chatgpt probably passes it now or soon. The tests I have seen that gpt fails on also trip over the vast majority of humanity. Not the few elite like probably you and me on HN but the rest yes; you want to say they have no reasoning power either or?
Gpt might be reasoning or it might not be; I cannot say as I don’t understand at all how I reason myself. Or how to describe what it is to reason in coherent terms. Or to test for it. But saying the substrate (meat vs silicon) makes the difference and, even though gpt actually does get a lot of calculations right even though it was not made for it and we don’t understand why that is, means something is going on. I think it will debunk most of the religious arguments we have to elevate us to something special like you and others try to do pretty soon. It seems we will find that our thinking mechanism is a few pages of Python and then trillions of complex calculations to train it. Animals like us did this over millions of years, we will repeat it in 100 years or less.
Turn it around. I'm sure there are lots of people around who couldn't do it either, even with a ridiculously high amount of time to do it
As a thought experiment, assume the average human may be able to translate text between two human languages, or write code in two-three programming languages. GPT4 can perform those tasks on a much more diverse set of human _and_ programming languages. Is that not superhuman?
Yes, it makes mistakes. But take a hundred humans off the street and ask them to write an NGINX configuration or translate between Indian and French - how many would be able to do that? How many would be able to do that without any mistakes?
People on HN claim they're using it for XYZ in development, yet it can't even generate the necessary NGINX config, despite being given the URL rewrites it'll need to incorporate.
The point is that it hallucinates. It isn't that it failed, it's that despite giving it everything it needs to know, it hallucinated all kinds of things not in the documentation, not in my prompts et al.
Why? Because it's an LLM. It isn't fit for purpose in this context. A next word prediction AI (an LLM) isn't appropriate for these kinds of problems.
To me, it feels quite intuitive that an AI trained on human knowledge would automatically learn to do the same.
The point: it isn't that it failed. It isn't even that it "hallucinates answers", it's that it infers relationships between words that don't exist because it's an LLM. It predicts the next word. That's what it does.
Something that predicts the next word isn't an appropriate method of doing x in y of z cases, because its reliability in providing the designated function is important. Ergo, yes, LLMs may well have applications, but most of the problems that people are throwing it at are inappropriate, just like blockchain fetishism versus a database. For the overwhelming majority of problems, AI is not the answer, neither is a blockchain, neither is an NFT.
Call hallucination what it is: a fail state. It got it wrong. It didn't "hallucinate". With standard conditional logic, x yields y result. That's very useful where you want consistency and reliability, ergo, those problems are best not handled via an LLM. Why not use the appropriate tool for the job?
Deductive versus inductive versus abductive reasoning.
We also cannot build living animals from scratch or are anywhere close to it - maybe some forms of AI are much tougher to do than we think.
Talking out of my ass here, but my point is that I think The Machines(c) don’t look like biological and separated entities. I think it’ll look more like what we call corporations (hive minds) composed of a vast variety of different functional parts.
What's more likely, IMO, is using real brain cells. https://www.ucl.ac.uk/news/2022/oct/human-brain-cells-dish-l...
However, real brain cells are a big question of ethics if it's thus actually able to think. I would argue that we've then created a slave rather than a machine, and that is unacceptable.
One comment on HN called it a "baby AGI" after linking to the paper.
Eye roll inducing.
How can a something that generates such a massive surge of interest, investment and research into AI not be a step toward it?
Saying it’s not a step towards AGI is basically saying AGI isn’t possible at all, because it means that all our efforts are making zero progress on AGI. That’s not a falsifiable position to take.
If you’re serious, the parent post literally said “AGI isnt going to look like this”.
…but realistically, how would a LLM that could easily refine itself from experiences, and had a very large context, let’s say, a billion tokens, be meaningfully different from AGI?
It could learn. It could remember things. It could generate human like output from a complex context.
Sure, it’s just a stochastic parrot… but if it can refine the model from real world inputs (learn new tricks, learn games, etc) and generate large scale (entire books worth) of coherent conversation and interactions… where do you draw the line between that and actual AGI?
Large contexts (35k tokens) are here right now. Refining models is here right now. They’re just expensive and slow (inference and training).
Maybe the current architecture doesn’t scale up beyond that and it’s a dead end, but my gosh.
If you don’t think what we have is a step towards AGI you really have to work hard to make your definition of AGI very very difficult to attain.
Maybe also what I wrote a bit above: describe some greater than 3 dimensional objects and get it to stack them for some purpose could be another thing to try (I think, I will actually).