Not only is there infinite incentive to compete, but theres decreasing costs to. The only world in which AGI is winner take all is a world in which it is extremely controlled to the point at which the public cant query it.
The first-mover advantages of an AGI that can improve itself are theoretically unsurmountable.
But OpenAI doesn't have a path to AGI any more than anyone else. (It's increasingly clear LLMs alone don't make the cut.) And the market for LLMs, non-general AI, is very much not winner takes all. In this announcement, OpenAI is basically acknowledging that it's not getting to self-improving AGI.
This has some baked assumptions about cycle time and improvement per cycle and whether there's a ceiling.
To be precise, it assumes a low variability in cycle time and improvement per cycle. If everyone is subjected to the same limits, the first-mover advantage remains insurmountable. I’d also argue that whether there is a ceiling matters less than how high it is. If the first AGI won’t hit a ceiling for decades, it will have decades of fratricidal supremacy.
How steeply the diminishing returns curve off at.
The most advanced tools are (and will continue to be) at a higher level of the stack, combining the leading models for different purposes to achieve results that no single provider can match using only their own models.
I see no reason to think this won't hold post-AGI (if that happens). AGI doesn't mean capabilities are uniform.
I wonder, do you have a hypothesis as to what would be a measurement that would differentiate AGI vs Not-AGI?
So one fundamental difference is that AGI would not need some absurdly massive data dump to become intelligent. In fact you would prefer to feed it as minimal a series of the most primitive first principles as possible because it's certain that much of what we think is true is going to end up being not quite so -- the same as for humanity at any other given moment in time.
We could derive more basic principles, but this one is fundamental and already completely incompatible with our current direction. Right now we're trying to essentially train on the entire corpus of human writing. That is a defacto acknowledgement that the absolute endgame for current tech is simple mimicry, mistakes and all. It'd create a facsimile of impressive intelligence because no human would have a remotely comparable knowledge base, but it'd basically just be a glorified natural language search engine - frozen in time.
If you took the average human from birth and gave them only 'the most primitive first principles', the chance that they would have novel insights into medicine is doubtful.
I also disagree with your following statement:
> Right now we're trying to essentially train on the entire corpus of human writing. That is a defacto acknowledgement that the absolute endgame for current tech is simple mimicry
At worst it's complex mimicry! But I would also say that mimicry is part of intelligence in general and part of how humans discover. It's also easy to see that AI can learn things - you can teach an AI a novel language by feeding in a fairly small amount of words and grammar of example text into context.
I also disagree with this statement:
> One fundamental difference is that AGI would not need some absurdly massive data dump to become intelligent
I don't think how something became intelligent should affect whether it is intelligent or not. These are two different questions.
You didn't teach it, the model is still the same after you ran that. That is the same as a human following instructions without internalizing the knowledge, he forgets it afterward and didn't learn what he performed. If that was all humans did then there would be no point in school etc, but humans do so much more than that.
As long as LLM are like an Alzheimer's human they will never become a general intelligence. And following instructions is not learning at all, learning is building an internal model for those instructions that is more efficient and general than the instructions themselves, humans do that and that is how we manage to advance science and knowledge.
Then when OpenAI does another training run it can also internalise that knowledge into the weights.
This is much like humans - we have short term memory (where it doesn't get into the internal model) and then things get baked into long term memory during sleep. AI's have context-level memory, and then that learning gets baked into the model during additional training.
Although whether or not it changed the weights IMO is not a prerequisite for whether something can learn something or not. I think we should be able to evaluate if something can learn by looking at it as a black-box, and we could make a black-box which would meet this definition if you spoke to a LLM and limited it to it's max context length each day, and then ran an overnight training run to incorporate learned knowledge into weights.
> So one fundamental difference is that AGI would not need some absurdly massive data dump to become intelligent.
The first 22 years of life for a “western professional adult” is literally dedicated to a giant bootstrapping info dump
The zero training version not only ended up dramatically outperforming the 'expert' version, but reached higher levels of competence exponentially faster. And that should be entirely expected. There were obviously tremendous flaws in our understanding of the game, and training on those flaws resulted in software seemingly permanently handicapping itself.
Minimal expert training also has other benefits. The obvious one is that you don't require anywhere near the material and it also enables one to ensure you're on the right track. Seeing software 'invent' fundamental arithmetic is somewhat easier to verify and follow than it producing a hundred page proof advancing, in a novel way, some esoteric edge theory of mathematics. Presumably it would also require orders of magnitude less operational time to achieve such breakthroughs, especially given the reduction in preexisting state.
The moment after human birth the human agent starts a massive information gathering process - that no other system really expects much output from in a coherent way - for 5-10 years. Aka “data dump” some of that data is good, and some of it is bad. This in turn leads to biases, it leads to poor thinking models; everything that you described, is also applicable to every intelligent system - including humans. So again you presupposing that there’s some kind of perfect information benchmark that couldn’t exist.
When that system comes out of the birth canal it already has embedded in it millions of years of encoded expectations predictability systems and functional capabilities that are going to grow independent of what the environment does (but will be certainly shaped in its interactions by the environment).
So no matter what, you have a structured system of interaction that must be loaded with previously encoded data (experience, transfer learning etc) with and it doesn’t matter what type of intelligent system you’re talking about there are foundational assumptions at the physical interaction layer that encode all previous times steps of evolution.
Said an easier way: a lobster, because of the encoded DNA that created it, will never have the same capabilities as a human, because it is structured to process information completely differently and their actuators don’t have the same type and level of granularity as human actuators.
Now assume that you are a lobster compared to a theoretical AGI in sensor-effector combination. Most likely it would be structured entirely differently than you are as a biological thing - but the mere design itself carries with it an encoding of structural information of all previous systems that made it possible.
So by your definition you’re describing something that has never been seen in any system and includes a lot of assumptions about how alternative intelligent systems could work - which is fair because I asked your opinion.
The next time your in the wilds, it's quite amazing to consider that your ancestors - millennia past, would have looked at, more or less, these exact same wilds but with so much less knowledge. Yet nonetheless they would discover such knowledge - teaching themselves, and ourselves, to build rockets, put a man on the Moon, unlock the secrets of the atom, and so much more. All from zero.
---
What your example and elaboration focus on is the nature of intelligence, and the difficulty in replicating it. And I agree. This is precisely we want to avoid making the problem infinitely more difficult, costly, and time consuming by dumping endless amounts of knowledge in the equation.
I’m curious what epistemological grounding you are basing your claim on
EDIT: There can be levels of AGI. Google DeepMind have proposed a framework that would classify ChatGPT as "Emerging AGI".
I don't know if it is optimism or delusions of grandeur that drives people to make claims like AGI will be here in the next decade. No, we are not getting that.
And what do you think would happen to us humans if such AGI is achieved? People's ability to put food on the table is dependent on their labor exchanged for money. I can guarantee for a fact, that work will still be there but will it be equitable? Available to everyone? Absolutely not. Even UBI isn't going to cut it because even with UBI people still want to work as experiments have shown. But with that, there won't be a majority of work especially paper pushing mid level bs like managers on top of managers etc.
If we actually get AGI, you know what would be the smartest thing for such an advanced thing to do? It would probably kill itself because it would come to the conclusion that living is a sin and a futile effort. If you are that smart, nothing motivates you anymore. You will be just a depressed mass for all your life.
That's just how I feel.
There can be levels of AGI. Google DeepMind have proposed a framework that would classify ChatGPT as "Emerging AGI".
ChatGPT can solve problems that it was not explicitly trained to solve, across a vast number of problem domains.
https://arxiv.org/pdf/2311.02462
The paper is summarized here https://venturebeat.com/ai/here-is-how-far-we-are-to-achievi...
Think about it - the original definition of AGI was basically a machine that can do absolutely anything at a human level of intelligence or better.
That kind of technology wouldn't just appear instantly in a step change. There would be incremental progress. How do you describe the intermediate stages?
What about a machine that can do anything better than the 50th percentile of humans? That would be classified as "Competent AGI", but not "Expert AGI" or ASI.
> fancy search engine/auto completer
That's an extreme oversimplification. By the same reasoning, so is a person. They are just auto completing words when they speak. No that's not how deep learning systems work. It's not auto complete..
It's really not. The Space Shuttle isn't an emerging interstellar spacecraft, it's just a spacecraft. Throwing emerging in front of a qualifier to dilute it is just bullshit.
> By the same reasoning, so is a person. They are just auto completing words when they speak.
We have no evidence of this. There is a common trope across cultures and history of characterising human intelligence in terms of the era's cutting-edge technology. We did it with steam engines [1]. We did it with computers [2]. We're now doing it with large language models.
[1] http://metaphors.iath.virginia.edu/metaphors/24583
[2] https://www.frontiersin.org/journals/ecology-and-evolution/a...
The General Intelligence part of AGI refers to its ability to solve problems that it was not explicitly trained to solve, across many problem domains. We already have examples of the current systems doing exactly that - zero shot and few shot capabilities.
> We have no evidence of this.
That's my point. Humans are not "autocompleting words" when they speak.
No, it's bringing something out of scope into the definition. Gluten-free means free of gluten. Gluten-free bagel verus sliced bread is a refinement--both started out under the definition. Glutinous bread, on the other hand, is not gluten free. As a result, "almost gluten free" is bullshit.
> That's my point. Humans are not "autocompleting words" when they speak
Humans are not. LLMs are. It turns out that's incredibly powerful! But it's also limiting in a way that's fundamentally important to the definition of AGI.
LLMs bring us closer to AGI in the way the inventions of writing, computers and the internet probably have. Calling LLMs "emerging AGI" pretends we are on a path to AGI in a way we have zero evidence for.
Bad analogy. That's a binary classification. AGI systems can have degrees of performance and capability.
> Humans are not. LLMs are.
My point is that if you oversimplify LLMs to "word autocompletion" then you can make the same argument for humans. It's such an oversimplification of the transformer / deep learning architecture that it becomes meaningless.
The "g" in AGI requires the AI be able to perform "the full spectrum of cognitively demanding tasks with proficiency comparable to, or surpassing, that of humans" [1]. Full and not full are binary.
> if you oversimplify LLMs to "word autocompletion" then you can make the same argument for humans
No, you can't, unless you're pre-supposing that LLMs work like human minds. Calling LLMs "emerging AGI" pre-supposes that LLMs are the path to AGI. We simply have no evidence for that, no matter how much OpenAI and Google would like to pretend it's true.
[1] https://en.wikipedia.org/wiki/Artificial_general_intelligenc...
> No, you can't, unless you're pre-supposing that LLMs work like human minds.
You are missing the point. If you reduce LLMs to "word autocompletion" then you completely ignore the the attention mechanism and conceptual internal representations. These systems have deep learning models with hundreds of layers and trillions of weights. If you completely ignore all of that, then by the same reasoning (completely ignoring the complexity of the human brain) we can just say that people are auto-completing words when they speak.
Sure, Google wants to redefine AGI so it looks like things that aren’t AGI can be branded as such. That definition is, correctly in my opinion, being called out as bullshit.
> obviously there will be stages in between
We don’t know what the stages are. Folks in the 80s were similarly selling their expert systems as a stage to AGI. “Emerging AGI” is a bullshit term.
> If you reduce LLMs to "word autocompletion" then you completely ignore the the attention mechanism and conceptual internal representations. These systems have deep learning models with hundreds of layers and trillions of weights
Fair enough, granted.
It is not a redefinition. It's a classification for AGI systems. It's a refinement.
Other researchers are also trying to classify AGI systems. It's not just Google. Also, there is no universally agreed definition of AGI.
> We don’t know what the stages are. Folks in the 80s were similarly selling their expert systems as a stage to AGI. “Emerging AGI” is a bullshit term.
Generalization is a formal concept in machine learning. There can be degrees of generalized learning performance. This is actually measurable. We can compare the performance of different systems.
The g in AGI is General. I don't what world you think Generality isn't a spectrum, but it's sure as hell isn't this one.
"A framework for classifying AGI by performance and autonomy was proposed in 2023 by Google DeepMind researchers. They define five performance levels of AGI: emerging, competent, expert, virtuoso, and superhuman"
In the second paragraph:
"Some researchers argue that state‑of‑the‑art large language models already exhibit early signs of AGI‑level capability, while others maintain that genuine AGI has not yet been achieved."
The entire article makes it clear that the definitions and classifications are still being debated and refined by researchers.
Edit: because if "AGI" doesn't mean that... then what means that and only that!?
"Agentic AI" means that.
Well, to some people, anyway. And even then, people are already arguing about what counts as agency.
That's the trouble with new tech, we have to invent words for new stuff that was previously fiction.
I wonder, did people argue if "horseless carriages" were really carriages? And "aeroplane" how many argued that "plane" didn't suit either the Latin or Greek etymology for various reasons?
We never did rename "atoms" after we split them…
And then there's plain drift: Traditional UK Christmas food is the "mince pie", named for the filling, mincemeat. They're usually vegetarian and sometimes even vegan.
It's kind of a simple enough concept... it's really just something that functions on par with how we do. If you've built that, you've built AGI. If you haven't built that, you've built a very capable system, but not AGI.
"Can", but not "must". The difference between an LLM being harnessed to be a customer service agent, or a code review agent, or a garden planning agent, can be as little as the prompt.
And in any case, the point was that the concept of "completely autonomous agentic intelligence capable of operating on long-term planning horizons" is better described by "agentic AI" than by "AGI".
> It's kind of a simple enough concept... it's really just something that functions on par with how we do.
"On par with us" is binary thinking — humans aren't at the same level as each other.
The problem we have with LLMs is the "I"*, not the "G". The problem we have with AlphaGo and AlphaFold is the "G", not the ultimate performance (which is super-human, an interesting situation given AlphaFold is a mix of Transformer and Diffusion models).
For many domains, getting a degree (or passing some equivalent professional exam) is just the first step, and we have a long way to go from there to being trusted to act competently, let alone independently. Someone who started a 3-year degree just before ChatGPT was released, will now be doing their final exams, and quite a lot of LLMs operate like they have just about scraped through degrees in almost everything — making them wildly superhuman with the G.
The G-ness of an LLM only looks bad when compared to all of humanity collectively; they are wildly more general in their capabilities than any single one of us — there are very few humans who can even name as many languages as ChatGPT speaks, let alone speak them.
* they need too many examples, only some of that can be made up for by the speed difference that lets machines read approximately everything
Stepping back for a moment - do we actually want something that has agency?
Name me a human that also doesn't need direction or guidance to do a task, at least one they haven't done before
Literally everything that's been invented.
To be fair, there is a section of the population whose useful intelligence can roughly be summed up as that or worse.
- Processing visual data and classifying objects within their field of vision.
- Processing auditory data, identifying audio sources and filtering out noise.
- Maintaining an on-going and continuous stream of thoughts and emotions.
- Forming and maintaining complex memories on long-term and short-term scales.
- Engaging in self-directed experimentation or play, or forming independent wants/hopes/desires.
I could sit here all day and list the forms of intelligence that humans and other intelligent animals display which have no obvious analogue in an AI product. It's true that individual AI products can do some of these things, sometimes better than humans could ever, but there is no integrated AGI product that has all these capabilities. Let's give ourselves a bit of credit and not ignore or flippantly dismiss our many intelligent capabilities as "useless."
No, I’m using useful problem solving as my benchmark. There are useless forms of intelligence. And that’s fine. But some people have no useful intelligence and show no evidence of the useless kind. They don’t hit any of the bullets you list, there just isn’t that curiosity and drive and—I suspect—capacity to comprehend.
I don’t think it’s intrinsic. I’ve seen pets show more curiosity than some folk. But due to nature and nurture, they just aren’t intelligent to any material stretch.
The turing test was succesfull. Pre chatGPT, I would not have believed, that will happen so soon.
LLMs ain't AGI, sure. But they might be an essential part and the missing parts maybe already found, just not put together.
And work there will be always plenty. Distributing ressources might require new ways, though.
The very people whose theories about language are now being experimentally verified by LLMs, like Chomsky, have also been discrediting the Turing test as pseudoscientific nonsense since early 1990s.
It's one of those things like the Kardashev scale, or Level 5 autonomous driving, that's extremely easy to define and sounds very cool and scientific, but actually turns out to have no practical impact on anything whatsoever.
Bots, that are now allmost indistinguishable from humans, won't have a practical impact? I am sceptical. And not just because of scammers.
In particular we redefined the test to make it passable. In Turing's original concept the competent investigator and participants were all actively expected to collude against the machine. The entire point is that even with collusion, the machine would be able to pass. Instead modern takes have paired incompetent investigators alongside participants colluding with the machine, probably in an effort to be part 'of something historic'.
In "both" (probably more, referencing the two most high profile - Eugene and the large LLMs) successes, the interrogators consistently asked pointless questions that had no meaningful chance of providing compelling information - 'How's your day? Do you like psychology? etc' and the participants not only made no effort to make their humanity clear, but often were actively adversarial obviously intentionally answering illogically, inappropriately, or 'computery' to such simple questions. And the tests are typically time constrained by woefully poor typing skills (this the new normal in the smartphone gen?) to the point that you tend to get anywhere from 1-5 interactions of a few words each.
The problem with any metric for something is that it often ends up being gamed to be beaten, and this is a perfect example of that.
And I did not looked into it (I also don'think the test has too much relevance), but fooling the average person sounds plausible by now.
Now sounding plausible is what LLMs are optimized for and not being plausible, still, I would not have thought we get so far so quick 10 years ago. So I am very hesistant about the future.
The two concepts have historically been inexorably linked in sci-fi, which will likely make the first AGI harder to recognize as AGI if it lacks consciousness, but I'd argue that simple "unconscious AGI" would be the superior technology for current and foreseeable needs. Unconscious AGI can be employed purely as a tool for massive collective human wealth generation; conscious AGI couldn't be used that way without opening a massive ethical can of worms, and on top of that its existence would represent an inherent existential threat.
Conscious AGI could one day be worthwhile as something we give birth to for its own sake, as a spiritual child of humanity that we send off to colonize distant or environmentally hostile planets in our stead, but isn't something I think we'd be prepared to deal with properly in a pre-post-scarcity society.
It isn't inconceivable that current generative AI capabilities might eventually evolve to such a level that they meet a practical bar to be considered unconscious AGI, even if they aren't there yet. For all the flak this tech catches, it's easy to forget that capabilities which we currently consider mundane were science fiction only 2.5 years ago (as far as most of the population was concerned). Maybe SOTA LLMs fit some reasonable definition of "emerging AGI", or maybe they don't, but we've already shifted the goalposts in one direction given how quickly the Turing test became obsolete.
Personally, I think current genAI is probably a fair distance further from meeting a useful definition of AGI than those with a vested interest in it would admit, but also much closer than those with pessimistic views of the consequences of true AGI tech want to believe.
It isn't close at all.
It is not hard to imagine a "cooking robot" as a black box that — given the appropriate ingredients — would cook any dish for you. Press a button, say what you want, and out it comes.
Internally, the machine would need to perform lots of tasks that we usually associate with intelligence, from managing ingredients and planning cooking steps, to fine-grained perception and manipulation of the food as it is cooking. But it would not be conscious in any real way. Order comes in, dish comes out.
Would we use "intelligent" to describe such a machine? Or "magic"?
A machine could be super intelligent at solving real world practical tasks, better than any human, without being conscious.
We don't have a proper definition of consciousness. Consciousness is infinitely more mysterious than measurable intelligence.
I don't think there has ever been a time in history when work has been equitable and available to everyone.
Of course, that isn't to say that AI can't make it worse then it is now.
"AGI" was already a goalpost move from "AI" which has been gobbled up by the marketing machine.
This is current research. The classification of AGI systems is currently being debated by AI researchers.
It's a classification system for AGI, not a redefinition. It's a refinement.
Also there is no universally accepted definition of AGI in the first place.
Here is a mainstream opinion about why AGI is already here. Written by one of the authors the most widely read AI textbook: Artificial Intelligence: A Modern Approach https://en.wikipedia.org/wiki/Artificial_Intelligence:_A_Mod...
Can ChatGPT drive a car? No, we have specialized models for driving vs generating text vs image vs video etc etc. Maybe ChatGPT could pass a high school chemistry test but it certainly couldn't complete the lab exercises. What we've built is a really cool "Algorithm for indexing generalized data", so you can train that Driving model very similarly to how you train the Text model without needing to understand the underlying data that well.
The author asserts that because ChatGPT can generate text about so many topics that it's general, but it's really only doing 1 thing and that's not very general.
I think we need to separate the thinking part of intelligence from tool usage. Not everyone can use every tool at a high level of expertise.
but so do I!
Of course they can. We already have computer controlled car systems, the reason LLMs aren't used to drive them is because AI systems that specialize in text are a poor choice for driving - specialized driving models will always outperform them for a variety of technical reasons.
That was my whole point. Maybe in theory an LLM could learn to drive a car, but they can't today because they don't physically have access to cars they could try to drive just like a person who can't learn to use a tool because they're physically limited from using it.
Likewise for "intelligent", and even "artificial".
So no, ChatGPT can't drive a car*. But it knows more about car repairs, defensive driving, global road features (geoguesser), road signs in every language, and how to design safe roads, than I'm ever likely to.
* It can also run python scripts with machine vision stuff, but sadly that's still not sufficient to drive a car… well, to drive one safety, anyway.
How about we have ChatGPT start with a simple task like reliably generating JSON schema when asked to.
Hint: it will fail.
This doesn’t imply that it’s ideal for driving cars, but to say that it’s not capable of driving general intelligence is incorrect in my view.
Same model trained on audio, video, images, text - not separate specialized components stitched together.
Last time I checked, in an Anthropic paper, they asked the model to count something. They examined the logits and a graph showing how it arrived at the answer. Then they asked the model to explain its reasoning, and it gave a completely different explanation, because that was the most statistically probable response to the question. Does that seem like AGI to you?
There’s a good reason why schools spend so much time training that skill!
It is easy to see why, since the LLM doesn't communicate what it thinks it communicates what it thinks a human would communicate. A human would explain their inner process, and then go through that inner process. An LLM would explain a humans inner process, and then generate a response using a totally different process.
So while its true that humans doesn't have perfect introspection, the fact that we have introspection about our own thoughts at all is extremely impressive. An LLM has no part that analyzes its own thoughts the way humans do, meaning it has no clue how it thinks.
I have no idea how you would even build introspection into an AI, like how are we able to analyze our own thoughts? What is even a thought? What would this introspection part of an LLM do, what would it look like, would it identify thoughts and talk about them the way we do? That would be so cool, but that is not even on the horizon, I doubt we will ever see that in our lifetime, it would need some massive insight changing the AI landscape at its core to get there.
But, once you have that introspection I think AGI will happen almost instantly. Currently we use dumb math to train the model, that introspection will let the model train itself in an intelligent way, just like humans do. I also think it will never fully replace humans without introspection, intelligent introspection seems like a fundamental part to general intelligence and learning from chaos.
It does have some weasel words around value-aligned and safety-conscious which they can always argue but this could get interesting because they've basically agreed not to compete. A fairly insane thing to do in retrospect.
"Instead of our current complex non-competing structure—which made sense when it looked like there might be one dominant AGI effort but doesn’t in a world of many great AGI companies—we are moving to a normal competing structure where ..." is all it takes
That's always been pretty overtly the winner-take-all AGI scenario.