this post was submitted on 07 Aug 2024
129 points (93.3% liked)
Technology
59627 readers
2807 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Language models are literally incapable of reasoning beyond what is present in the dataset or the prompt. Try giving it a known riddle and change it so it becomes trivial, for example "With a boat, how can a man and a goat get across the river?", despite it being a one step solution, it'll still try to shove in the original answer and often enough not even solve it. Best part, if you then ask it to explain its reasoning (not tell it what it did wrong, that's new information you provide, ask it why it did what it did), it'll completely shit it self hallucinating more bullshit for the bullshit solution. There's no evidence at all they have any cognitive capacity.
I even managed to break it once through normal conversation, something happened in my life that was unique enough for the dataset and thus was incomprehensible to the AI. It just wasn't able to follow the events, no matter how many times I explained.
Maybe the grown up human LLM that keeps learning 24/7 and is evolved in thousands of years to make the learning part as efficient as possible is just a little bit better than those max 5year old baby LLM with brut force learning techniques?
The 5 year old baby LLM can't learn shit and lacks the ability to understand new information. You're assuming that we and LLMs "learn" in the same way. Our brains can reason and remember information, detect new patterns and build on them. An LLM is quite literally incapable of learning a brand new pattern, let alone reason and build on it. Until we have an AI that can accept new information without being tolled what is and isn't important to remember and how to work with that information, we're not even a single step closer to AGI. Just because LLMs are impressive, doesn't mean they posses any cognition. The only way AIs "learn" is by countless people constantly telling it what is and isn't important or even correct. The second you remove that part, it stops working and turns to shit real quick. More "training" time isn't going to solve the fact that without human input and human defined limits, it can't do a single thing. AI cannot learn form it self without human input either, there are countless studies that show how it degrades, and it degrades quickly, like literally just one generation down the line is absolute trash.
A human not trained by other humans also just dies…
Nope, people are quite resilient. As long as it's not a literal new born, the chance of survival isn't THAT low. Once you get past 4 years and up, a human can manage quite well.
Also dying because no one takes care of you and you fail to aquire food and dying of a stroke/seizure are 2 very different things.
This is because of semi hardcoded stuff using the mechanics of hormones that interact with the neurons in the brain, I would say. They are hardcoded by the instructions provided by the DNA, I believe.
About the learning differences between human and LLM, there I believe that a sub-“module" of the brain functions very similar to how the LLMs work with just a way better/efficient learning algorithm that is helped by the other modules in the brain like the part that can simulate 3D space and interpret other sensory data like feeling touch, vision, smell etc
Current LLM models are being used in static manner without ability to learn in real time, so of course it can not do anything it has not learned yet.
It is just a theory and it can not be proven wrong since the understanding of neurons is not advanced yet.
Well, or at least, I did not hear a good argument that proves that theory 100% wrong.
You can think of the brain as a set of modules, but sensors and the ability to adhere to a predefined grammar aren't what define AGI if you ask me. We're missing the most important module. AGI requires cognition, the ability to acquire knowledge and understanding. Such an ability would make larger language models completely redundant as it could just learn langue or even come up with one all on its own, like kids in isolation for example.
What I was trying to point out is that "neural networks" don't actually learn in the way we do, using the world "learn" is a bit misleading, because it implies cognition. A neural network in the computer science sense is just a bunch of random operations in sequence. In goes a number, out goes a number. We then collect a bunch of input output pairs, the dataset, and semi randomly adjust these operations until they happen to somewhat match this collection. The reasoning is done by the humans assembling the input output pairs. That step is implicitly skipped for the AI. It doesn't know why they belong together and it isn't allowed to reason about why, because the second it spits out something else, that is an error and this whole process breaks. That's why LLMs hallucinate with perfect confidence and why they'll never gain cognition, because the second you remove the human assembling the dataset, you're quite literally left with nothing but semi random numbers, and that's why they degrade so fast when learning from themselves.
This technology is very impressive and quite useful, and demonstrates how powerful of a tool language alone is, but it doesn't get us any closer to AGI.
Do you know if current LLM models use static neural networks (like where each node is connected with each of the next layers)or if they can rearrange their connections into other layers?
They are indeed. It has multiple steps, but pretty much. 3Blue1Brown has a nice visual explanation: https://www.youtube.com/watch?v=wjZofJX0v4M
So if you want it more like a brain, you would have to have nodes that are able to form the connections while learning letting each node decide in what "direction” it wants to grow it’s connection in some sort, rather than having fixed connections where you only adjust the correlation of the nodes. And you would need multiple transformer (and most likely some hard logic algorithms as well) for different inputs as well as a main "thinker” that decides through which transformer (or algorithm) a input has to go and if the output of that transformer needs to be feeded in again as no input.
In theory. Then comes the question of how exactly are you gonna teach/train it. I feel our current approach is too strict for proper intelligence to emerge, but what do I know. I honestly have no clue how such a model could be trained. I guess it would be similar to how people train actual braincells? Tho that field is very immature atm... The neat thing about the human brain is, that it's already preconfigured for self learning, tho it does come with its own bias on what to learn due to its unique needs and desires.
😁🥳 I’d say, you would then need something that takes the role of hormones in that system (like hardcoded reactions to events in and outside of the AI brain/body(so called emotions I would say)) that trigger the connections to grow, shrink, get their values adjusted etc.
At least that would be my approach.
Calling the reward system hormones, doesn't really change the fact that we have no clue where to even start. What is a good reward for general intelligence? Solving problems? That's our current approach, which has the issue of the AI not actually understanding the problems and just ending up remembering question answer pairs (patterns). We need to figure out what defines inteligence and "understanding" in an easily measurable way. Which is something people knew almost a hundred years ago when we came up with the idea of neural networks, and why I say we didn't get any closer to AGI with LLMs.