This, again, sparks the "is this general ai?" question, which often results in low quality, borderline-flaming content... My take:

the point of this paper isn't "here, we solved general intelligence". It's "look, multi modal token prediction is a sound iteration". Look at the scale of the model in comparison to, say, gpt-3: this is a PoC, they didn't bother scaling it, because we've already seen where scaling these mechanisms leads.

What I would love to know is what kind of architectures deepmind et al are playing with in-house. Token prediction is a promising avenue, but it's more of a language that an intelligent agent may operate in, opposed to the self-sufficient structure of the intelligent agent itself -- the symbolic system that implements algos like gato. If that symbolic system will be the result of a generator-function, that generator function won't be token prediction by trade. I mean, maybe somewhere in the deep depths of a multi modal model, intelligent structure may emerge, but that would be a very weird byproduct.

>but it's more of a language that an intelligent agent may operate in, opposed to the self-sufficient structure

yes, this kind of functional intelligence seems distinct from an actual living entity, which is the thing that uses subordinate functions to pursue goals and has some interior state, motivations and some sort of architecture. To reduce intelligence to tokens predicting more tokens is kind of like saying f(x), just solve for intelligence. When prediction itself is only partially what intelligent systems are about.

Agent is a very important word because it's accurate ("a means or instrument by which a guiding intelligence achieves a result") And it's the latter I think we ought to be after when talking about 'general ai'.

It’s possible that in serving the function of prediction, the model forms a complex internal representation akin even to goals, motivations, etc. It is true that DL architectures are not explicitly designed to do this, not yet anyway. But my point is that the task of prediction can give rise to such architectural patterns. According to Karl Friston’s Free Energy Principle, biological brains serve the purpose of predicting the value of different actions available.

I agree with that, I think it's even necessarily true in natural intelligence which after all emerged by some means spontaneously. But scientifically I think it is a big problem because it does not supply us with a theory or systematized knowledge of the mind or intelligence. I think you could even imagine say, why not just make a primordial soup simulation, insert some DNA, and crank the speed up, intelligence is just a byproduct somewhere in there in a physics simulation. don't even bother with such details as neural nets.

Scientifically this is unsatisfying but also if for some reason this turns out to be an engineering dead-end we have a big hole where a concrete theory of intelligence should be, with its components, mechanisms and so forth. And sadly I think this is still the weakest link in AI.

To me it seems a little bit like if you trained architects instead of having a theoretical basis for architecture, you just showed them every building in existence and sent them to work. It may very well work, but if it didn't you have a problem. And even if it did, you'd still want to have an understanding of why it works.

Sometimes there is just no reductive theory for emergent phenomena. For example, in the case of ant routing we have a good understanding of the behavior of individual ants, and we can observe the intelligent behavior of the colony as a whole. One could ask for a concrete theory of how the micro scale behavior of each ant leads to the macro scale behavior of the colony, but it doesn't exist. If we had enough working memory to hold all the ants in our minds, we'd see that both micro and macro scale behaviors are different aspects of the same system, there's no "missing link" in the chain of explanations, and the dichotomy is only due to our own cognitive limitations.

There are a lot of problems that do not have analytical solutions, like the N-body problem. With these problems all you can do is numerical simulation, which is pretty much what modern machine learning is.

this assumes there is a finite list of available actions. for example, a primate has to see that "sharpen a stick to use as a spear" is an option, and add it to the list

It explicitly says in the article that the reason they capped it at ~1B parameters is because that’s the current limit of what they can achieve and hit their latency requirements? It has nothing to do with not being interested in scaling it further, as far as I’m aware.

Imo, the current "the bigger the better" trend is limiting further progress. Intelligence finds the simplest model that explains all data, e.g. a small set of equations or rules, while the today's wannabe-AI models are trying to remember all the data in a fuzzy lookup table.

I think there’s two parts to this. The first is that there’s a minimum number of parameters required to do an arbitrary task. For any sufficiently complex task (image recognition, large language models, etc), it’s not clear how to find that lower bound. And the bound probably depends on the model chosen. But we do know that the more parameters you add, the more complex of a function you can learn.

On the other hand, for many reasons (energy, training/inference deployment complexity, latency, even some sense of model elegance I suppose), we don’t want to massively increase the number of parameters unnecessarily. But again, I don’t think we have great methods to estimate the “ideal” minimum number of parameters for a model to achieve its goals. And what we keep finding is that if you increase the number of parameters, and you increase the training corpus, the model gets more accurate, more impressive, and it’s not stopping.

So while I definitely agree that size for size’s sake is wasteful, I also don’t think we necessarily even know how to define “wasteful” for things like large language models right now.

Well, while this is very interesting research, the title of the paper is a bit misleading. They way I understand the paper, their network is specialized in multiple domains, but that doesn't make it a generalist. Clearly there is a lot of potential, but skimming through quickly, I didn't find much about out of domain data. Generally, I think nomenclature in the ai world is horrible. A mishmash of different academic disciplines and hyped keywords. Attention isn't really attention (more like association), generative adversarial networks aren't really adversarial (they are designed to work together in the end), ...

> because we've already seen where scaling these mechanisms leads.

In the case of GPT-3, scaling seemed to continuously improve results, they just kinda ran out of data. Are you implying this must be the same for this model? Or were you intending to say something different that I didn't see?

To be precise, scaling improved results logarithmically. Accuracy seems to plateau with scaling and never reach any magical limit of true general intelligence.

There's been no observed plateu with scaling. As far as we can tell further scaling will keep improving results.

This paper caused quite a big shift in the Metaculus predictions on when "AGI" will be achieved, https://www.metaculus.com/questions/3479/date-weakly-general... and https://www.metaculus.com/questions/5121/date-of-general-ai/.

What Gato shows is with the proper datasets and training we can achieve general purpose like results. It doesn't show any sort of learning behavior. So your personal home robot may be able to dice up an onion, but it is not going to sit down at your computer, look up its own schematics, and start learning about its existence.

It doesn't have the active working memory to do so.

And frankly that's exactly what we need. When you really think about it the whole concept of self awareness is ultimately pointless for solving problems, and is mostly counterproductive.

Diced onions > existential dread.

Frankly I'd sleep much better with a great but dumb AI onion cutter in my kitchen than with one that does who knows what during the night. There are already billions of humans that can surprise me sometimes in unpleasant ways. No need to add artificial beings to the list. That said, true strong AI would be a wonderful achievement (/understatement) but I'm happy with useful partial results, like a 787 or a bird sized drone compare to an actual artificial pterodactyl or pigeon.

Isn’t that simply a matter of computational resource? Learning is a limit of frequency and time taken between events.

When it came out and I was soaking up the sentiment on lesswrong, it was the first time I noticed a genuine shift towards “we are past the fire alarm stage.”

Although I do think that phrasing for the implication is already heavy handed in a negative bias.

I find it incredibly exciting; and it deserves more open discussion without provoking anxiety.

One important thing to note here is that this model was trained purely in a supervised fashion. It would be interesting to see a paper at a similar scale that's based on reinforcement learning. The reinforcement learning context (specifically the exploring part) gives a lot more opportunities to see the effects of positive/negative transfer. That approach would of course be much more expensive, though.

Discussed a lot five days ago: https://news.ycombinator.com/item?id=31355657

Transformers managed to confuse me about what interpolation and extrapolation really mean in high dimensional space

There's a neat argument against these models doing interpolation: the manifold of the data is so sparse that it's infinitesimally unlikely for a good predictor to be doing interpolation between existing points on the manifold.

https://arxiv.org/abs/2110.09485

Yes, nice, I know some of these words

Maybe what transformers actually transform are us..

Autobots, print out.

I made some concept maps of the first parts of the paper. It might help with clarifying some of it.

https://twitter.com/izzyz/status/1525099159925116928

Hmm 1.2B params... that's what, roughly 5 GB of VRAM? Surprisingly compact.

This was done so that the IRL robot manipulation tasks could be done fast enough. In the future, we may always need small models mixed with large models for some tasks (e.g., for slow long term planning and fast short term planning), though compute does have a tendency to improve exponentially...

Aha, ok

There's a breakthrough that I've been waiting for that I haven't heard anything about: when will an AI agent (probably a language model) discover something scientific that humans had not at the time it was trained. What if there was a math proof, physics interaction, ... that emerged from the model's approximation of our world?

Right now, the state of the art AlphaZero models can destroy humans at Go. But what if the machine learning models could teach us things about how Go works that humans have not yet discovered.

It very likely already did, specifically in Go. The problem is that humans would still be required to comprehend what they are seeing :-) letting agents develop strategies in an unsupervised manner has already yielded strategies we haven’t figured out ourselves. Other examples that come to mind are video compression (see twominutepapers) and proteine folding!

Think about it like this: if the domain of a problem we want AI to solve is so complex that we can barely formulate the question, how could we be confident that we can understand 100% of the answer we get? “Here, gpu, make sense of this 20-dimensional problem my brain can’t even approximately visualize!”

In both Go and Chess the strong engines available now have already shown us human strategy was off the mark somewhat. Said simply, humans put too much weight on margin of victory vs probability of victory. In these games the difference means humans favor maintaining a material advantage more than the engines, who are more likely to trade material for a positional advantage.

You are describing most successful machine learning models. Take AlphaFold, it has surely discovered relationships that govern protein folding better than any human has ever previously understood.

That's not really science though. Science requires developing a hypothesis. I've yet to see an AI do this. Not trying to raise the bar or anything, just saying let's not call it scientific.

Folding proteins is a scientific endeavour. You are raising the bar, thought I understand you wish you weren't—metaphysical questions about science, a requirement for the AI to explain any results in natural language of meatsuits, and referring to the first step of a grade school analysis of the _scientific method_ (forming a hypothesis) as a requirement for _science_, all raise the bar.

I don't feel it was I who raised the bar. The parent citing "scientific" did. Science is a human construct in and of itself. If it is going to do science, then it needs to do so as we have defined as "the scientific method".

But I would be satisfied if it developed a hypothesis in any language (including mathematics), not necessarily natural language.

Developing a hypothesis is not the same as expressing one, though.

There was a link posted here recently that talked about an AI which could identify race from xrays while trained experts could not.

Maybe it discovered some new science to pull that off?

Narrow deep learning ai is generally not suited for this. However automated theorem provers are a thing and have proven major conjectures/theorems that weren't solved by humans before. E.g. The four color problem IIRC. Although the best results are generally obtained with semi-automated theorems provers

But still, this is not cleverness, this just show that raw bruteforce + a few tricks can solve a few problems, by generating proofs of multiple terabytes(yes this is absurd scaling). The asymmetry between compute power and computer lack of intelligence is remarkable.

https://en.m.wikipedia.org/wiki/Automated_theorem_proving

Watching this video on it now https://www.youtube.com/watch?v=wSQJZHfAg18

I was a little underwhelmed to see that the model was trained to imitate other SOTA RL agents, still very impressive results

Can anyone help me understand the difference between this and saycan?

Why is this approach better?

They call me Gato I have metal joints Beat me up And earn 15 Silver Points

thanks for this

Chrono Trigger reference?

Beat him up, and you’ll earn 15 silver points!

Roko’s Basilisk indicates that we all ought to support this project as much as possible.

This is essentially the birth of AI.

The lack of fanfare on this achievement is baffling.

I disagree. It's not even clear from the paper exactly how much learning transfer is actually happening. I think it's fair not to be rolling out the red carpet and showering the authors with awards.

Lack of fanfare? Every techie news outlet is plastered with it, and I'd expect it to diffuse from there.

With "intelligence₂" we mean "the ability to produce a usable solution replacing the work of a professional in a specific task". It is a consolidated use with a pretty long history, and whether the task-generalist ANN offers a breakthrough for intelligence₂ has to be discussed.¹

With "intelligence₁" we remain referring to "the ability to reflect on world representations achieving (recursively) refined concepts and selectively producing founded conclusions". It seems that the research relevant to this submissions is still alien to the context of AGI.

A claim that "this is the Dawn of It" requires clarifications and demonstrations (meaning, explicitness in clarifying why).

--

¹Some argument that it goes in that direction is in the Introduction: «Historically, generic models that are better at leveraging computation have also tended to overtake more specialized domain-specific approaches (Sutton, 2019), eventually».

This result is unsurprising. "Give a model a bunch of unique datasets and it can do a bunch of unique things." There's nothing showing any sort of generalized learning or capability here.

What is the achievement? It seems that the author has shown that this path is fruitful, but transfer learning is no where near being solved.

Every other answer to your comment was a bit pessimistic. I’m intrigued, we need to see more!

You're hanging out in the wrong (or right) circles if that's your perception.

So.. Any circles?