Until a language model can develop a generalized solution to a real-world phenomena, it's not even close to AGI. The current iteration of ML algorithms are useful, yes, but not intelligent.
any approach that is so obviously alien to it is unlikely to be a good approach
Computers very frequently do not solve problems the same way humans do, so I'm not sure that's a significant point against any particular modeling technique.
Otherwise, if you're waiting for language models to understand language he same way that human minds do, you're probably waiting for AGI, not any particular breakthrough in NLP alone.
I don't think a pure language model of the sort under consideration here is heading towards AGI. I use language models extensively and the more I use them the more I tend to see them as information retrieval systems whose surprising utility derives from a combination of a lot of data and the ability to produce language. Sometimes patterns in language are sufficient to do some rudimentary reasoning but even GPT4, if pushed beyond simple patternish reasoning and its training data, reveals very quickly that it doesn't really understand anything.
I admit, its hard to use these tools every day and continue to be skeptical about AGI being around the corner. But I feel fairly confident that pure language models like this will not get there.
IMHO (and this is just my own uniformed view), this means that language models by themselves are insufficient for certain important tasks. It seems to be hard for systems to learn deductive reasoning purely based on text prediction.
OTOH, who knows what would happen if you somehow managed to combine the generating capabilities of a language model with a proper inference engine, e.g. Wolfram|Alpha. Maybe it would bring us significantly closer to AGI, but maybe that way is also a dead-end because it's not guaranteed that those systems would work well together.
Well one way out is if large language models don't just somehow magically turn into human level (or better) AGI at some point once enough data has been thrown at it. Then the whole debate will turn out to be pretty moot.
Any language model alone isn't going to solve natural language understanding, but some future ML full-brain model surely will achieve AGI and therefore language understanding as part of that.
What's missing from a pure language model, is any grounding in reality and ability to interact with the world to test and extend its knowledge beyond what is derivable from the corpus it was trained on. It's level of understanding is ultimately limited by the content of the training corpus, regardless of how massive that may be.
Something like GPT-3 is really just a statistical twist on Doug Lenat's Cyc; it's understanding is always going to be limited by it's own fundamental nature. Yes, one deals with language, and one with facts, but ultimately both are just large, fixed, self-referential bodies of data.
Cyc really is a great analogy, and for some reason it took decades for Lenat et al to eventually realize that regardless of how much fixed data you added to it, it was never going to be enough. A closed black box can never know what's outside of the box, although it may gamely try to tell you if you ask it.
These modern language models, GPT-3, etc, have certainly been a bit of an eye opener, and can perform some impressive feats (question answering, etc), but one shouldn't be tempted to believe that if scaled up sufficiently they'll eventually somehow transcend their own nature and become more than a language model... a one-trick pony capable of generating plausible continuations of whatever you seed it with.
I can interpret this in a couple different ways, and I want to make sure I am engaging with what you said, and not with what I thought you said.
> I think better LLMs won’t lead to AGI.
Does this mean you believe that the Transformer architecture won't be an eventual part of AGI? (possibly true, though I wouldn't bet on it)
Does this mean that you see no path for GPT-4 to become an AGI if we just leave it alone sitting on its server? I could certainly agree with that.
Does this mean that something like large language models will not be used for their ability to model the world, or plan, or even just complete patterns as does our own System one in an eventual AGI architecture? I would have a lot more trouble agreeing with that.
In general, it seems like these sequence modelers that actually work right is a big primitive we didn't have in 2016 and they certainly seem to me as an important step. Something that will carry us far past human-level, whatever that means for textual tasks.
To bring it back to the article, probably pure scale isn't quite the secret sauce, but it's a good 80-90% and the rest will come from the increased interest, the shear number of human-level intelligences now working on this problem.
Too bad we haven't scaled safety nearly as fast though!
There's one big weakness in all current language models that I feel holds it back. There's no way proactive way to have it be persuasive.
Weak AGI will be the first language model that is able to somehow influence the thoughts of the person communicating with it, I think that is the milestone of AGI. From my experience with GPT-Neo and OPT and using it to help write stories or make chatbots, the responses are still very reactionary. In that sense, adding more parameters helps the model give a more coherent response, but it's still a response.
It does have some ability to extrapolate to new problems, provided its training corpus has reasonably close coverage. It is not going to be making new scientific discoveries or insights but then neither are most people. With a sufficiently large training set I think these models can achieve human parity for a subset of language generation tasks, and be effectively of human intelligence. They nearly already have.
It doesn’t matter to me if they have “reasoning” capabilities or not if the outcome is the same.
That kind of ML model is pretty general. Pretraining a big model has been extended to multi-modal environments. People are training them with RL to take actions. People are applying other generative techniques to them, and all sorts of other stuff. If you just look at it as 'predict the next word token,' then it's pretty limiting, but people have already gone way beyond that. TFA talks about some interesting directions people are taking them.
A more general form of your question is whether we can get to AGI with just incremental steps from where we are today, rather than step-change way-out-of-left-field kinds of ideas. People are split on that. Personally, I think that incremental changes from today's methods are sufficient with better hardware and data, but Big New Ideas could certainly speed up progress.
I say that large language models are not intelligent because of the way they fail to do things. In particular, they fail in such a way as to indicate they have no mental model of the things they parrot. If you give them a simple, but very unusual, coding problem, they will confidently give you an incorrect solution even though they seem to understand programming when dealing with things similar to their training data.
An intelligent thing should easily generalize in these situations but LLMs fail to. I use GPT4 every day and I frequently encounter this kind of thing.
I think this line of reasoning is probably correct: these language models aren't intelligence, they're just the next iteration on information lookup systems.
Or, more likely, the march towards better abstraction ends with a natural language interface where problem descriptions can be effortlessly translated into functional code.
We’re already in spitting distance of this with current language models. I don’t think you even need AGI to make this happen, a large language/image model gets you most of the way there.
Language models can only parrot back their training input. There's no generality to them at all; the best they can do is some crude, approximate interpolation of their training examples that may or may not be "correct" in any given instance. There are things that the typical AI/ML approach might be genuinely useful for (e.g. generating "probable" conjectures by leveraging the "logical uncertainty" of very weak and thus tractable logics) but mainstream language learning is a complete non-starter for this stuff.
reply