@risottobias @seldo Know your tools, maybe?
Because it also doesn’t matter how AI’s getting answers incredibly right. And not everyone wants AI to always get answers right. Especially when there are no right or wrong answers.
@pointlessone @risottobias@tech.lgbt @seldo
Isn't it exactly how it works? It doesn't "give answers" it just attempts to predict a criteria that you do not define explicitly. Then it "hallucinates" some totally random stuff (that takes the most computational resources) and matches it against that criteria, presents to you what statistically matched it the most — that's it!
@m0xee @risottobias @seldo That's true. No magic. But there's utility. You don't throw out your hammer because it can't drive screws. You accept the tool’s limitations and use it for what it’s good for.
@m0xee @risottobias @seldo Well, let me point out that “totally random” and “statistically matched” kinda clash.
I’d say there’s significant correlation between a question and the right answer to it. Even with all the flexibility of natural language there’s only so many ways a right answer can be phrased.
So if we take those statements as true we have a decent chance of getting a right answer from an LLM. We have to take into account that LLMs are trained on the internet. That is, most questions don’t have only right answers to them in the training data. Most of internet also is not questions and answers. So yeah, you’re getting a statistically plausible continuation of your prompt, not an answer to your question, and not necessary the right answer to your question.
Extracting knowledge from an LLM is not the main useful thing about them either. There are other functions that you can get from statistical language model. For instance, summarisation. Remove less significant tokens and repetitions and you have your input summary. Translation. If an LLM is trained on multiple languages tokens that have similar meaning would naturally settle close to each other in the token space because they’re used in similar contexts.
Hallucination can also be the desired function. An LLM with high temperature setting (cranked up randomness) can be used as a brainstorming tool. You have your idea, the prompt, and you want to get a bunch of related avenues to explore. They won’t be absolutely novel but there might be something that you personally have not thought of.
There’s no magic here. It can’t solve al your problems. But there are useful functions in there.
Sorry for switching to a different account — I don't have character limit on this one and I don't want to split my reply in 5-6 parts.
]
> Well, let me point out that “totally random” and “statistically matched” kinda clash.
No, not really, they don't. I might be pulling balls out of a box and the balls might be totally random, but I'm only interested in red ones so I leave those out and throw out the rest. If the box doesn't contain red ones (or I didn't find one in say hundred attempts) I might go with orange ones — or even purple if nothing even remotely red comes up — that's how they are statistically matched.
That's what I mean and I wasn't talking about LLMs only — just about how neural networks are applied in general. I think those image generators work the same way. Neural network doesn't know what a cat is, it was just trained on thousands of images of cats so it can match against that. Suppose it's given a request to generate a "floral cat" — it doesn't know what a flower is, it doesn't even know what a cat is, but it can match against those. How about this random image that I have generated: does it match against flower, does it match against cat? Fine — then we return it.
Exactly same as me pulling balls out of that box!
I mean sure there are quirks: I can pre-seed it by using fragments of images that are known to contain a cat to speed things up. But does it really change much? I think for simplicity we can still consider those totally random.
This leads us to the problems we can point out, first one's with the datasets our neural network has been trained on: suppose our metaphorical box doesn't contain any red balls at all and no matter what I do — I will never be able to pull out a red ball: orange ones, purple, maybe burgundy on a lucky day — I will never find red ones because no one put them there.
The second one stems from people putting too much trust into this "AI" thing — they ARE being told that it is indeed intelligent and they start believing it can actually com up with something we do not yet know. Suppose such a person has never seen a red ball and requests one, but it gives him/her purple — you tell them that it's wrong and they respond with something like: "But that is what AI told me, you're just too skeptical!"— but you aren't even skeptical, you just know precisely what red is.
Of course neither of these are problems with the tool itself — they are people problems, same as guns don't kill people — people kill people. A knife's a very useful tool, but it's also a dangerous one — no, I don't think we should ban knives on these grounds, but when I'm around a person who looks unstable and hostile towards me I sure would put all knives away. Same here, but there are not only people who use the tool for wrong things, there are people who convince other people and businesses to use the tool for all the wrong reasons presenting it as a golden bullet which it sure isn't.
And of course the inefficiency aspect — that one I fully agree with. People are spending megawatts to generate said floral cats and other useless stuff — for tasks that are easily and more efficiently solved with other tools.
Of course I might be wrong about some of the above or even all of it — please correct me if so.
@m0xee
@m0xEE@breloma.m0xee.net @risottobias @seldo @m0xee@librem.one I’m not entirely sure where you were going with the balls metaphor so I might be a bit off mark commenting on it but I’ll still do. :)
Neural nets are not random. If anything, they’re extremely deterministic. LLMs have special machinery to introduce some randomness in order to give more varied output. That is, you don’t randomly take ball out of a black box. It’s rather you put a ball into one side of a black box, and get a statistically correlated kind of a bad on the other side of that box. You can put in random balls and out put might look somewhat random in that case. Correlation might not be perfect, too. But a trained NN will encode some degree of correlation. Current generation of image generators use Diffusion approach. They start with some noice and tweak it to be closer a bit to an image that has desired features. It does it multiple times in small increments and eventually the noise is transformed into an image with high degrees of “floweriness” and “cainess”. But again, this is a totally deterministic process. Starting noice is there to increase variety. If you run the NN on the same noisy input you’ll get the same image.
The point about red ball, I think is correct. If LLM sees a new token it won’t know what to do with it. There’s no statistical relation for that token to any other token. So it can not produce that token. But even in more practical sense, if there was just a handful examples of some tokens in the training data there might be not enough to produce robust relationships to other tokens. Say, ChatGPT is quite good with widely used languages such as English, German, Spanish, etc. but its performance declines proportionally with the amount of training data. You will not get good performance in languages that poorly represented on the internet. Say, some dead languages or a language spoken by some Amazonian tribe of 200 people won’t give you much.
On knowledge. This is s ticking point of most modern AI scepticism. LLMs (and neural nets in general) do not know things. That is not how NNs work. NN that perfectly encodes factual knowledge is said to be overfit to its training data. However, that doesn’t mean there’s no correlation between input and output. Language encodes a decent amount of knowledge. There’s a significant chance token “nice” can be found somewhere after token “69”. This statistical correlation effectively encodes knowledge about a meme. LLM doesn’t have any understanding/knowledge what “nice”, “69”, or “meme” means but it still encodes it. Same goes for image generation NNs. They don’t know what “flowery” or “cat” means but they encode statistical relation between those words and patterns/features in images.
On novel discoveries. Current generation of LLMs can’t do it as far as I can tell. However, that is mostly because that is what we’ve coded them to do. Maybe a meta-LLM might notice patterns in relations between different groups of tokens and make that connect in some way? But that’s just speculation. What you’re describing is a regular grift. Same as homeopathy, essential oils. In my opinion it’s a different problem. It is absolutely a problem but it’s not the same as AI having no utility. Learn how stuff works and you’ll have better chance not to fall for it. We’ve known what homeopathy for decades and people still fall for it. However, unlike homeopathy AI system do have real utility. AI scepticism and AI grift scepticism should be two different things.
On efficiency. It’s true that huge LLMs require a lot of power. We’re unlikely to bring them down to the level of efficient sorting algorithms, for example. But work is being done to increase efficiency and it’s done at a pretty good pace. GPT-3.5 is 175 billion parameters. We have open source models of 13B parameters that are very close on some tasks. That is 10x improvement in just a couple of years. People are trying alternative architectures like collectives of specialised models that can only activate relevant models greatly reducing required power. It will become better. I mean, it will become more efficient. It might still require a lot of power but it will probably be better at what we want it to do. Maybe it will even learn to fact check its output if we really want that.
@m0xEE@breloma.m0xee.net @risottobias @seldo @m0xee@librem.one BTW, I’ve touched in some of these topics before.
About distinguishing AI the technology and AI the grift: https://status.pointless.one/@pointlessone/111807312813079938
About power consumption: https://status.pointless.one/@pointlessone/111778211684985340
@pointlessone @risottobias@tech.lgbt @seldo
I mean it doesn't know any answers that we don't already know, the trick is to expand very vague criteria it's given with what people are likely to expect — no magic there 🤷