It’s not just OpenAI. I think AI winter is coming much sooner than expected. Throwing more compute at the current solution is not yielding better results it seems. Reasoning will not be solved by a bigger LLM
We don't even know how existing LLMs work, not really, yet we're done?
All signs are pointing to there still being no upper limit on results. Not yet.
Just because it's not on the leaderboard yet, doesn't mean we've heard the bell.
There are thousands of people around the world trying to reverse engineer what is going on in the billions or trillions of parameters in an LLM.
It's a field called "Mechanistic Interpretability." The people who do the work jokingly call it "cursed" because it is so difficult and they have made so little progress so far.
Literally nobody can predict before they are released what capabilities new models will have in them.
And then, months after a model is released, people discover new abilities in it, such as decent chess playing.
They are black boxes.
Also an artefact of how evals have been done on a pass fail basis. So that an LLM that gets 90% of a question right is just as much a failure as one that gets 0% of the question.
So that skills appear to emerge suddenly and surprisingly only due to the flawed way that we are forced to study them. Consider the training regime, and partial success towards a goal, and emergence is far less prevalent. There was a paper on that recently, I'll see if I can find.
Those same academics admit themselves that they're surprised at how well LLMs do considering how simple(?) rudimentary(?) the logic underneath is.
I don't quite understand what you're saying. That these academics were being lazy by not properly investigating/publishing their findings? That doesn't seem right.
All of that is imitation, nothing near thought.
We know how to build it. We don't understand how it's producing the output it does based off what we give it
We do not understand language, grammar, music, only partly emotions, or especially sentience and consciousness. Further, we don't understand how the disparate systems are integrated together.
The comparison itself is pretty telling; that the brain and AI work so similarly in specific ways
We know exactly how llms work (relatively simple maths), and to a large extent even why they work (backpropagation updates weights to more closely approximate the desired function). There are open questions relating to LLMs of course - we don't understand what the space of potential LLM-like things looks like and how the features in that space relate to subjective performance (although note that transformers were designed based on a theory that they would perform better, not just randomly generated or inspired by the muse). We also don't know to what extent the output of LLMs can be approximated by simpler symbolic systems, or how to extract such systems from LLMs when they do exist. Those are really interesting questions, but they're not questions about 'how LLMs work'.
I dislike the 'LLMs are magic' framing that seems to be taking over the world. Nobody thinks that Taylor expansion is magical, but LLMs are doing the same sort of thing - approximating a function through a bunch of weights on a bunch of simpler functions. Just because the function we're approximating (intelligent output) is not known in advance (but can be sampled), and multi-dimensional does not fundamentally change how mysterious the process is.
But we actually can't! We can build a program that can build a program that is the LLM, which is not the same! I'd argue that you're right insofar as training is concerned. We understand training very well. But the actual model, how it operates, what it actually knows, we don't know how to build that, we don't know what weights to put where.
Malbolge is an esoteric programming language designed to be impossible to use. The first program written in it wasn't written by a human, it was written by another program.
But since then, with that working program to learn from, people have figured out how to write programs in malbolge: https://lutter.cc/malbolge/tutorial/cat.html
Cloning animals or even humans did not automatically make us understand how brains work. In fact, these were quite unrelated endeavors.
> I dislike the 'LLMs are magic' framing that seems to be taking over the world
Don't take that out on me. That's not what I'm saying. I'm saying there is a lack of determinism (mathematically provable, per se) in our current understanding of all AI (LLM included). There are many attempts to solve this problem. I've sat in on seminars about it myself. So far, we're not there yet
I agree. It's not copying that I'm saying is understanding, it's modelling.
> I'm saying there is a lack of determinism (mathematically provable, per se)
What do you mean by a lack of determinism in this case?
The path ahead to AGI is deceitful, there are no gradients leading towards it. It's based on exploration and discovery. It works in populations for reasons of diversity - evolution is that way. And evolution is a slow process, not a singularity kind of event.
I am an apostate. Technology does not have any innate vector for linear development.
The choice to amplify dystopian social trends remains a wrongful choice.
So-called AI is bad for humanity and the planet, in finite terms at this moment in time. It's essence is wrongful.
Turning bad ideas up to 11 does not make them good ideas.
As far as images and video are concerned, we're done. Frame generation is almost perfect, and we don't really need any more training data. Now it's time to build product and enhance how the models work.
LLMs, though? That field appears to have hit a wall for now.
AI for media is going to be a rocket ship. AI for knowledge and text and reasoning will take longer. People will recognize this soon.
Is there actually enough data to train an AI on? Photos seem to be a success, like you say, but everything else?
I’ve heard many people make claims about where AI content will be most useful. A persistent theme is an AI made VR world, customized video games, and endless AI generated TikTok videos.
I genuinely question if the training data exists for these use cases. Photos are cheap and easy, but quality annotated 3D models? Short form videos? What about long form video? Are we really there (assuming inference was cheap)?
These are being solved as we speak. I'm working on this problem directly and the level of control and consistency achievable is incredible. Video is just a special case of images.
Take a look at the ComfyUI space and the authors of plugins and papers.
> quality annotated 3D models
The research is progressing at a fast pace. We can get good surface topologies, textures, and there are teams working on everything from rigging to animation.
Named-Entity Recognition (NER) and Text Classification will allow you to figure out what kind of text you're looking at and extract structured data.
LLMs are not good at this because they're not specialized for it, but you can build a specialized NER model to extract custom entities from unstructured data today.
That said, I don't really think this is some yet untapped potential of AI so much as an area of ML that just hasn't been applied enough.
ETA: also, in general I think AI is going in the direction of basically just having an LLM route tasks to more specialized ML models (for corporate tasks at least). That's what Google's Vertex AI agents sort of do (and I am guessing the GPT 4 agents as well).
"AGI" LLM's can handle all the nuance, quite simply, they don't need specialist infrastructure or specialist programming, way cheaper, upfront cost. Way easier to scale.
Individual employees can ask it for specific things that make their lives easier and it'll give it to them/or do it. No need to ask your manager, motivate for funding and hire an engineer/purchase new software/equipment.
I can't see how AGI from that perspective _isnt_ just an LLM routing tasks to more specialized NLP models to be honest.
Unless you're proposing that a bigger LLM (training data, neural network, etc..) will develop the emergent capability to accomplish this without the need for specialized agents.
In which case, I can't see how that would perform more efficiently than an LLM routing requests appropriately amongst agents, as it necessarily requires processing much more data.
But even if somehow it did perform more efficiently while needing much more data, I don't think the no-agent AGI approach will cover all use cases appropriately.
It might be an easier drop in solution, but if I need it to behave a specific way in a specific context, I don't see how an AGI is going to consistently do that more accurately than fine tuning a model for the specific use case and having an LLM routing to it.
There is also this other process at work that is shaped by quarterly earnings reports in public markets. Maybe if we started over with a mirror economy that we rebuild from the ground up, most of these things could be automated.
I think there are many inefficiencies in the system that Homo economicus wouldn't be able to deal with even if Homo economicus actually existed.
The definition of it isn't clear but from what I gather it's basically an aggregate of emergent capabilities that work together to produce a singularity.
Maybe with enough resources it's possible but I highly doubt it'll be economically feasible given how much has gone into it so far and how far we really are away from something like that with current models.
In an age of such hopelessness about the future, this looks a lot like an emotional crutch wrapped in the veil of rationality - just the thing an anxious materialist needs to make sense of the world.
Like many cults and religions it mistakes the plausible for possible and possible for probable.
The problem with religious beliefs like these is that they don't just disappear with evidence or sufficient reasoning.
I don't think that particular bubble is bursting anytime soon.
It's useful but basically every method of quality control requires a human.
I've found that components of general intelligence specialized beyond human capability are much more useful than a model that can mimic a human.
I think an LLM is just trying to do too much at once, all of the individual NLP algorithms most of them are made of are very useful to us, but an LLM is just not specialized enough to be any more useful than a human without specialization.
Which isn't to say they're _useless_, but obviously not as useful as a specialist (in special contexts, denoted by whatever kind of specialist they are)
ETA: as an aside, I'd like to contextualize my presumption that AGI is about AI singularity with the fact that Sam Altman casually stated that he doesn't care if it takes $50 billion to reach AGI.
In the real world, with 50 billion dollars, you can do something much more useful than trying to build a product that's basically contradictory by definition.
An AGI is (presumably) a general intelligence model but it's implicitly touted as being extremely useful for specialized tasks (because, humans can specialize), but once you specialize, I would argue your general intelligence tends to weaken. (For example I wouldn't expect a Harvard PhD to be 100% up to date with modern slang terms, but I'd be shocked if I went to a local bar and met someone who didn't know what rizz means).
This is basically just trying to squeeze two opposite ends of a spectrum together, which sounds kind of like a singularity to me.
I get that. I guess my point is this already seems to exist. We could combine AI with machinery to replace almost everything humans can do already, someone just has to build for that solution (e.g. train some models).
AGI just sounds like a sort of automation of that process. And I don't think a bigger LLM will accomplish that task. I think more developers will.
Which I wager would be cheaper and arguably more fortuitous to the human race than $50 billion thrown into one pot
But yeah developers are needed, a bigger LLM won't fix everything.
The money's a funny one. Global GDP is about $85,000 bn/yr so if someone can spend $50bn on getting AGI and taking it over it's a bargain. But if you spend $50bn and just get a loss making chatbot then less so.
Also, I still think you can probably build something (or rather, many, many somethings) with existing tooling to accomplish exactly that.
> A bigger LLM won't fix everything
I'm not sure if there's a camp that says it probably won't fix anything, but I'm in that camp if it exists.
If you think about how humans actually work, I think a basic, non AGI LLM routing information to different agents/models is closer to how most humans behave (when productivity is their goal).
E.g. a person's behavior is driven almost entirely by the current context they are in most of the time.
It's not that our minds become overexcited by loads of previous information and we magically are able to do other specialized tasks, we decide based on context what specialty in our toolset best fits the scenario.
> The money's a funny one. Global GDP is about $85,000 bn/yr so if someone can spend $50bn on getting AGI and taking it over it's a bargain. But if you spend $50bn and just get a loss making chatbot then less so.
If that's true then the same could be said of just dumping $50 billion into grants/research/funding for education around AI so that developers worldwide have an easier time developing AI enabled technologies and services.
At least with that plan, there is extremely little risk of creating nothing more than a chatbot (and extremely low risk of tech companies monopolizing labor the same way they try and monopolize everything else; I don't have much faith that if a few companies automate all or most labor that they'll redistribute wealth)
I've half seriously considered the possibility a large portion of the hype has been manufactured in an attempt to shock stagnating economies back to life, post-COVID, post low interest rates.
"AI winter !!!!????"
It's been a pattern in AI research since the 70s. Sure, the current boom is unprecedented, but that doesn't mean there won't be a relative bust. AI winter doesn't mean chatgpt will disappear. It just means research funding may get significantly scaled back of the hundred billion dollar investments of today don't generate trillion dollar returns
He was expressing doubt that there is an AI winter near. He's more of the:
>Geoffrey Hinton, dubbed the 'Godfather of AI,' warns technology will be smarter than humans in five years"
school of thought. https://www.dailymail.co.uk/sciencetech/article-12610845/geo...
They dont want lose to Apple,Google and Meta