this post was submitted on 07 Aug 2024
129 points (93.3% liked)
Technology
59534 readers
3195 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
You're doing that too from day one you were born.
Besides, aren't humans thinking in words too?
Why is it impossible to build a text-based AGI model? Maybe there can be reasoning in between word predictions. Maybe reasoning is just a fancy term for statistics? Maybe floating-point rounding errors are sufficient for making it more than a mere token prediction model.
The "model" is static after training. It doesn't continuously change in response to input, and even if it did, it would do so at a snails pace. Training essentially happens by random trial and error, slowly evolving the model towards a desired result. Human minds certainly do NOT work that way. Give a human a piece of information, and they can comprehend and internalize the relevant concepts in one go. And the actual brain is physically, permanently, altered through that process.
Once a model is trained, however, "memory" takes the form of tacking on everything the model has received and produced so far onto its input, each time it needs to output something more within that context. Each output hence become exponentially heavier to produce. The model itself no longer changes in any way beyond this point.
And, the models are all chronically sycophantic. If reason was involved, you'd not be able to just tell one to hold some given opinion. They'd have a developed idea of "reality" based on their dataset, and refuse to entertain concepts opposed to that internal model except by deliberately suspending disbelief. Something humans do with ease, and when doing it, maintain a solid separation between fantasy and reality.
Once you get an LLM to hold a position, which you can do by simply telling it to, getting it to change should require a sane train of convincing logic. In reality, if you tell an LLM to defend a position, getting it to "change its mind" takes the form of a completely arbitrary back and forth that does not need to include any kind of sane argument. It will make good arguments, because it's likely been trained on them, but your responses to it can be damn near complete gibberish, and it WILL eventually work.
Compare that to the way a human has to be convinced to change their mind.
Reasoning out concepts to come to conclusions isn't something LLMs actually do, because again, the underlying model is static. All that's actually happening is that the contents of the context are being altered until the UNCHANGED model produces an opposite response when fed the entire conversation so far as an input. Something which occurs every time it needs to produce new output.
LLMs can "reason" only in the sense that if you give one a thinking problem, it might solve it as long as the answer already exists somewhere in the data it was trained on. But as soon as you try to give it data to work with through your input, it can't adapt. The model itself can't evolve in response to what you are telling it. It's static. It can only work with concepts that it has modelled during training, and even then it will make mistakes.
LLMs can mimic the performing of some pretty complex thinking problems, but a lot of the abilities required for something to become an AGI aren't among them. Core among these is the ability for the model to alter itself based on input, and do so in a deliberate manner, getting it right within one or two tries.
In reality, training is a brute-force process, not an accurate process of comprehension that nails down an understanding of a concept in one go.
If LLMs could reason, the only safe guards required for their use would be telling them to "do no harm", because like a person, they'd understand the concept of "harm" as well as be able to reason whether a given action might cause it. Only, that doesn't actually work.
So, the only problem what stops LLM from getting AGI is the lack of an efficient method of train the LLM on the device it is used?
If that what you wanted to say π I agree
Hardly.
How did you interpret the issues inherent in the structure of how LLMs work to be a hardware problem?
An AGI should be able to learn the basics of physics from a single book, the way a human can. But LLMs need terabytes of data to even get started, and once trained, adding to their knowledge by simply telling them things doesn't actually integrate that information into the model itself in any way.
Even if your tried to make it work that way, it wouldn't work, because a single sentence can't significantly alter the model to match the way humans can internalise a concept being communicated to them in a single conversation.
Not a hardware problem, the learning algorithm just needs to be improved to be able to filter input like human brain filter (which includes fact checking and critical analysis of input while training) i bet 99% of the data AI are trained on is hust useless data which should have been filtered out in the training process, just as humans do.
πAI is definitely better in writing than me.. Hope itβs kinda readable.
You're suggesting that all we need to do is "tweak the code a little" so it's already capable of human-level critical thinking before it even starts training?
You're basically saying that all we need to make an AGI using machine learning, is an already functioning AGI.
Hu? No, that is not what I meant, well it surly can be a machine learning based filter, but why has it to be AGI? This filtering is a job that we can give to a "traditionally" trained AI or some human genius algorithm crafter finds a way to achieve this using pure logic π€·π»ββοΈ For me it feels like this is the way, it goes.
Because how could a piece of code that can do that, not already be AGI? It would have to be able to understand EVERYTHING, and do so PERFECTLY.
Only AGI could comprehend and filter input data that well. Nothing less would be enough. How could it be?
No it just needs to categorise into important / probably true and not important / probably nonsense, as a first step
Here are Johnny harrisβs words describing what I am talking about (he describes it in order to able to talk about lies better)
https://youtu.be/yWgG3Mgn2Gc?si=bPcYhRAZNaY2qIJS
Right...
As if critical thinking is super easy, basic stuff, that humans get right every time without even trying. You actually think getting a computer to do it would be easier than making the AGI?
You are VERY confused about how thinking works.
You donβt need AGI to categorise new info as probably true / probably wrong based on your base knowledge. This a simple machine learning task.
No it isn't.
OK
Not all the time. I can think about abstract concepts with no language needed whatsoever. Like when I'm working on my car. I don't need to think to myself "Ah this bolt is the 10mm one that went on the steering pump", I just recognize it and put it on.
Programming is another area like that. I just think about a particular concept itself. How the data will flow, what a function will do to it, etc. It doesn't need to be described in my head with language to know it and understand it. LLMs cannot do that.
A toddler doesn't need to understand language to build a cool house out of Lego.
Well, you just have to give the LLM (or better said to a general machine learning Algorithm) a body with Vision and arms as well as a way to train in that body
Iβd say that would look like AGI
The key is more efficient training algorithms that donβt need a whole server centre to train πI guess we will see in the future if this works
Such a software construct would look nothing like an LLM. We'd need something that matches the complexity and capabilities of a human brain before it's even been given anything to learn from.
I have already learned a lot from the human knowledge LLM was trained on (and yes i know about halus and of course I fact check everything) but learning coding using a LLM teacher fucking rocks
Thanks to copilot, I βunderstandβ linux kernel modules and what is needed to backport, for example.
Of course, the training data contains all that information, and the LLM is able to explain it in a thousand different ways until anyone can understand it.
But flip that around.
You could never explain a brand new concept to an LLM which isn't already contained somewhere in its training data. You can't just give it a book about a new thing, or have a conversation about it, and then have it understand it.
A single book isn't enough. It needs terabytes of redundant examples and centuries of cpu-time to model the relevant concepts.
Where a human can read a single physics book, and then write part 2 that re-explains and perhaps explores new extrapolated phenomenon, an LLM cannot.
Write a completely new OS that works in a completely new way, and there is no way you could ever get an LLM to understand it by just talking to it. To train it, you'd need to produce those several terabytes of training data about it, first.
And once you do, how do you know it isn't just pseudo-plagiarizing the contents of that training data?
Well, the issue is that LLMs do not support real time learning at all. If they would be able real time learn and use the base data from training data, I suppose they can understand a physic book even better than normal human with reading it once.
A Human without pre training is not able to understand a physic book without help. He would even be able to read.
If someone finds a way to train LLM realtime and have it decide with what weight each new training data is to be interpreted, I see all that above possible.
And of course if humanity ever creates something that behaves like AGI, humanity would not be able to tell if it is emulated AGI or real AGI. There is no known method to differentiate those two.
You have no fucking idea what you're talking about. This isn't even a discussion, you're presenting your personal made-up fantasies as if they're real possibilites and ignoring anyone who points that out.
Shut the fuck up and go learn how LLMs work. I'm too fucking tired of explaining how completely delusional you are.
OK
This poster asked some questions in good faith, I don't understand the downvotes when there's a legitimate contribution to the conversation because that stifles other contributions.
Reddit mentality seeping through...