It feels like any time scale on AGI is basically just made up. Since no one has any idea of how to get there, how could you possibly estimate how long it will take? We could stumble on some secret technique that unlocks AGI tomorrow or it could be literally impossible. You might as well ask how long until humans can cast magic spells.
The people most qualified to make an educated guess simultaneously have a direct financial incentive to claim that it is in reach within a few years. The only one who doesn't seem to care all that much is Le Cun.
To have an alternative? So far the work on transformers was done in a large part in the open (except OpenAI which tried to be as closed as possible). There is zero guarantee that whoever discovers the path to AGI decides to publish the paper etc. It's just one of many reasons (a less important one IMO) why research into novel AI approaches is valuable for humanity.
If you have invested in OpenAI because their mission statement says their goal is to reach AGI first, then it's not so hard. It is: an AGI is an AI that can build better AI's faster (in particular refine itself) faster than humans can. The first AI company to achieve that wins, and the rules of the game are winner takes all.
When you see these claims it's important to frame the assertion contextually as in the transformer generation, AGI is 10+ years away. This does not, however, account for the next architecture that will do more with less.
It's not quite the same category as magic spells. Kurzweil's prediction has been for 2029 for the last thirty years or so based on Moore's law type stuff. The logic I think is roughly project the hardware improvements, which has worked well, and then add on about five years to sort the software. Time will tell on the second one.
The most interesting thing in this whole picture is not AGI, it's how the collective intelligence works. CEOs claim the AGI is near because that's how they manipulate the public. But the public knows that it's only a manipulation. So how come the manipulation is still possible?
Sunk cost fallacy? What percentage of the public has actually invested the billions/trillions and is now demanding something to show for it? I'm not sure the average joe wants copilot in their outlook, but sure as hell someone wants it in there
Because "the public" isn't one person or even one cohesive group. Some see the manipulation, and point it out. Others don't see it, or ignore it when it's pointed out.
And why ignore it? Because they don't want to believe it's manipulation, because it promises large numbers of dollars, and they want to believe that those are real.
because people hedge their bets almost always. basically how likely something is vs costs vs what everybody else is doing vs how you are personally affected.
So in case of the current AI there are several scenarios where you have to react to it. For example as a CEO of a company that would benefit from AI you need to demonstrate you are doing something or you get attacked for not doing enough.
As a CEO of an AI producing company you have almost no idea if the stuff you working on will be the thing that say makes hallucination-free LLMs, allows for cheap long term context integration or even "solve AGI". you have to pretend that you are just about to do the latter tho.
If I had an AGI that designs me a safe, small and cheap fusion reactor, of course I would be interested in that.
My intelligence is intrinsically limited by my biology. The only way to really scale it up is to wire stuff into my brain, and I'd prefer an AGI over that every day.
If an AI is advanced enough to design mass production ready fusion reactors, I can't help but feel that the "human in charge of that AI" quickly becomes fully redundant. There is nothing that human can do that the AI itself can't do faster, better and cheaper.
This idea of "AI taking control" reminds me a lot what Karl Marx would say about "capital".
He wrote his things during the blooming of the second industrial revolution, a time when machines were replacing humans and forcing forward new economic, social, political, cultural and labour relations. And a key issue he stressed a lot is that this diffusion of machines and capital reshaping society was brought forward by a class of people that he called the bourgeoisie. He stressed a lot that it was a power struggle within society.
We're going through something similar today with the information technologies reshaping social relations. And the Bezos/Zuckerberg/Altman/Ellison of today are similar to the industrial barons from the Gilded Age. But, the same way that people reacted against the full-blown wild capitalism from 19th century's second half, we might also see some reactions against the advance of this techno-plutocracy.
In particular, I am optimistic about how the EU and some 3rd World countries (e.g. India, Brazil) are placing restrictions on social networks and techno-cartels.
> What is the value of you in this system?
So, to answer your question: individually I can't go beyond much more than careful choices (avoid cookies, stay out of Facebook, etc). Collectively we can make political choices. Ultimately, the most consequential political choice is move away from countries that give all power to the techno barons.
Whether or not that question is interesting is hardly relevant in this particular discussion. Companies building this technology are marketing it as if computers can think. They're telling us they can, and people are buying into it based on that claim.
So that quote doesn't reconcile the extraordinary claims of one side with the skepticism of the other.
And also, no disrespect to Dijkstra, but that sentiment is a bit shortsighted. If we could make computers think, it would have a profound impact on humanity. This is why there is so much excitement around this. We've been imagining this scenario for centuries, and we hope that this time around we can finally crack it. So comparing that achievement with something we can produce with classical technology is... uninspiring? Underwhelming? Selling ourselves short? I can't quite put it into words, but the possibility of answering that question would certainly be very interesting.
> and people are buying into it based on that claim.
I say this remains to be seen. You know that a lot of times you see the expression "AI" in the news. it comes followed by the word "bubble", right? If we see a big crash on the AI companies stocks we'll have proof that people aren't buying. And I strongly believe we'll see this crash and I think smart people aren't buying it.
OTOH, I think we need to be careful with the usage of the word "think". Dijkstra would probably give it a very broad meaning, going from French Impressionism, Bach and Shakespeare to Relativity Theory, Evolution Theory or Quantum physics, maybe even to Maradona's or Johan Cruyft's feet (Dijkstra was Dutch, remember). Computers and AI might go very deep in their "think" but will be very, very bad at the broad game. Frankly, I don't see how Markov Chain based technologies (e.g LLMs and most of AI today) can stop being replicators and start being innovators.
It is a bit like Pablo Picasso's quotation: "Computers are useless, they can only provide us answers".
It seems to me by most classical definitions we've basically already reached AGI.
If I were to show Gemini 3 Pro to anyone in tech 10 years ago they would probably say Gemini 3 is an AGI, even if they acknowledged there was some limitations there.
The definition has moved so much that I'm not convinced that even if we see further breakthroughs over the next 10 years people will say we've finally reached AGI because even at that point it's probable there might still be 0.5% of tasks it struggles to compete with humans on. And we're going to have similar endless debates about ASI and the consciousness of AI.
I think all that matters really is utility of AI systems broadly within society. While a self-driving car may not be an AGI, it will displace jobs and fundamentally change society.
The achievement of some technical definition of AGI on the other hand is probably not all that relevant. Even if goal posts stop moving from today and advancements are made such that we finally get 51% of experts agreeing that AGI has been reached there could still be 49% of expert who argue that it hasn't. On the other hand, one will be confused about whether their job has been replaced by an AI system.
I'm sorry - I know this is a bit of a meta comment. I do broadly agree with the article. I just struggle to see why anyone cares unless hitting that 51/49% threshold in opinion on AGI correlates to something tangible.
LLMs do NOT have intelligence. Achieving AGI would mean solving self-driving cars, replacing programmers, scientist, etc. All things LLMs are currently unable to do in a way that replaces humans.
There's a huge gap between what Gemini 3 can do and what AGI promises to do. It's not just a minor "technical definition".
The winner if there is one will be clear. Making a machine that's as good or better than human at most tasks is the turning point but doesn't demonstrate the win. Applying it for self-improvement making superintelligence which is when the winner will be racing with rockets vs bicycles. I don't know how this will play out--will the second one to achieve it always be behind? It probably depends on the difference in timing and other factors.
The winner will be the one that announces "we have AGI" based on their own definition, and the rest of the world actually agrees with them.
The upside of being right is massive, but the downside of prematurely announcing AGI only for the rest of the world to disagree with your definition is probably tremendously large.
Outside of internet debates I don't believe it matters. Much like the discussion of when machines will be treated like people, that's already happening without applying a definition.
"AGI" was hijacked to mean something else and was turned into a scam.
What it "really means" is more mass layoffs to power AI infrastructure for that to power so-called "AI agents" to achieve a 10% increase in global unemployment in the next 5 years.
From the "benefit of humanity", then to the entire destruction of knowledge workers and now to the tax payer even if it costs another $10T to bailout the industry from staggeringly giant costs to run all of it.
Once again, AGI is now nothing but a grift. The crash will be a spectacle for the ages.
Serious question. Do some of the large firms run LLM with no guardrails? I'm guessing they are doing constant research not available to the public. What results have been found? I'm not necessarily saying AGI, but what happens when the systems are not hindered by humans?
Also, somewhat related, the model/system that was reported by the Google whistleblower about LaMDA was very interesting for the time, especially considering the transcript. What happens when the guardrails are disabled? Even if it wasn't sentient, it's behavior might be reason for concern.
> The architecture might just be wrong for AGI. LeCun’s been saying this for years: LLMs trained on text prediction are fundamentally limited. They’re mimicking human output without human experience.
Yes, and most with a background in linguistics or computer science have been saying the same since the inception of their disciplines. Grammars are sets of rules on symbols and any form of encoding is very restrictive. We haven't come up with anything better yet.
The tunnel vision on this topic is so strong that many don't even question language itself first. If we were truly approaching AGI anytime soon, wouldn't there be clearer milestones beforehand? Why must I peck this message out, and why must you scan it with your eyes only for it to become something else entirely once consumed? How is it that I had this message entirely crystalized instantly in my mind, yet it took me several minutes of deliberate attention to serialize it into this form?
Clearly, we have an efficiency problem to attack first.
>Yes, and most with a background in linguistics or computer science have been saying the same since the inception of their disciplines
I'm not sure what authority linguists are supposed to have here. They have gotten approximately nowhere in the last 50 years. "Every time I fire a linguist, the performance of the speech recognizer goes up".
>Grammars are sets of rules on symbols and any form of encoding is very restrictive
But these rules can be arbitrarily complex. Hand-coded rules have a pretty severe complexity bounds. But LLMs show these are not in principle limitations. I'm not saying theory has nothing to add, but perhaps we should consider the track record when placing our bets.
I'm very confused by your comment, but appreciate that you have precisely made my point. There are no "bets" with regard to these topics. How do you think a computer works? Do you seriously believe LLMs somehow escape the limitations of the machines they run on?
Again: why can't any of that run on a sufficiently capable computer?
I can't help but perceive this as pseudo-profound bullshit. "Real soul and real imagination cannot run on a computer" is a canned "profound" statement with no substance to it whatsoever.
If a hunk of wet meat the size of a melon can do it, then why not a server rack full of nanofabricated silicon?
For the same reason you don't sit and talk with rocks. Nobody understands how it is that wet meat can do these things but rocks cannot. And a computer is a rock. As such, we have no idea whether all the hunks of wet meat in the world can figure out how to transform rocks into wet meat.
Modern computers can understand natural language, and can reply in natural language. This isn't even particularly new, we've had voice assistants for over a decade. LLMs are just far better at it.
Again: I see no reason why silicon plates can't do the same exact things a mush of wet meat does. And recent advances in AI sure suggest that they can.
"Language" is an input/output interface. It doesn't define the internals that produce those inputs and outputs. And between those inputs and outputs sits a massive computational process that doesn't operate on symbols or words internally.
And, what "clearer milestones" do you want exactly?
To me, LLMs crushing NLU and CSR was the milestone. It was the "oh fuck" moment, the clear signal that old bets are off and AGI timelines are now compressed.
Language massively restricts LLMs because there's no way to create novel concepts while limited to existing language.
Humans create new words and grammatical constructs all the time in the process of building/discovering new things. This is true even in math, where new operators are created to express new operations. Are LLMs even capable of this kind of novelty?
There's also the problem that parts of human experience are inexpressible in language. A very basic example is navigating 3D space. This is not something that had to be explained to you as a baby, your brain just learned how to do it. But this problem goes deeper. For instance, intuition about the motion of objects in space. Even before Newton described gravitation every 3 year old still knew that an object that's dropped would fall to the ground a certain way. Formalizing this basic intuition using language took thousands of years of human development and spurred the creation of calculus. An AI does not have these fundamental intuitions nor any way to obtain them. Its conception of the world is only as good as the models and language (both mathematical and spoken) we have to express it.
> Its conception of the world is only as good as the models and language (both mathematical and spoken) we have to express it.
Which is pretty damn good, all things considered.
And sure, training set text doesn't contain everything - but modern AIs aren't limited to just the training set text. Even in training stage, things like multimodal inputs and RLVR have joined the fray.
I don't think "create novel concepts" is a real limitation at all. Nothing prevents an AI from inventing new notations. GPT-4o would often do that when talking to AI psychosis victims.
Language is an interface between whatever our thoughts actually are and the outside world.
Imagine trying to write apps without thinking about the limitations of the APIs you use. In fact we just recently escaped that same stupidity in the SaaS era! That's how silly LLMs will seem in the near future. They will stick around as the smarter chatbots we've wanted for so long, but they are so very far away from AGI.
According to Clarke's First Law, "When a distinguished but elderly scientist states that something is possible, he is almost certainly right. When he states that something is impossible, he is very probably wrong."
With all due respect to Arthur C. Clarke, I think science education is the only thing standing between us and even bigger scams and absolute chaos in the streets.
What is understood by a scientist isn't so far ahead from what the layperson understands these days compared to when Clarke wrote that.
Toddlers and dogs and earthworms have intelligence. It's really hard to argue that LLM's don't have intelligence in the way the earthworms do. So just by a reading of the acronym, we already have AGI.
"But AGI means human level or better". Fine. You should call it that, Artificial Superhuman Intelligence.
But intelligence isn't a single thing, it's a broad array of behaviours. LLM's are obviously superhuman in some areas, just like calculators are. LLM's are obviously worse than humans in other areas.
But you'll always be able to find something that humans are better at. I'm sure somebody can find something that the earthworm brain is better at than the human brain.
Could you please make your substantive points without breaking the site guidelines? They include:
"When disagreeing, please reply to the argument instead of calling names. 'That is idiotic; 1 + 1 is 2, not 3' can be shortened to '1 + 1 is 2, not 3."
> It's really hard to argue that LLM's don't have intelligence in the way the earthworms do.
It's really easy to argue, actually. LLMs have intelligence the way humans online do. An earthworm is highly specialized for what it does and exists in a completely different context - I doubt an LLM would be successful guiding a robotic earthworm around since all it knows about earthworms is what researchers have observed and documented. The actual second-to-second experience of being an earthworm is not accessible as training data to an LLM.
Edit:
This is true almost by definition. An LLM (Large Language Model) can't have intelligence that's not expressible in language and earthworms are notoriously shy in interviews.
If this is true then almost all of the jobs in the world would already be replaced by AI. Sure, the model might be better than most people at lots of things but there are still tasks that human children find easy and AI struggles with. I wouldn't call being better at something than humans being "smarter" than them; if you do, calculators must be smarter than humans since they have been better at adding up numbers than us for a long time.
For text and image-based tasks they are infinitely better than a human.
What they lack are arms to interact with the physical world, but once this is done this is a giant leap forward (example: they will obviously be able to do experiments to discover new molecules by translating their steps-by-steps reasoning to physical actions, to build more optimized cars, etc).
For now human is smarter in some real-world or edge cases (e.g. super specialist in a specific science), but for any scientific task an average human is very very weak compared to the LLMs.
There are forms of science that don't involve "arms". Why don't we see a single research paper involving research entirely undertaken by AI? AI development and research itself doesn't need "arms". Why don't we just put AI in a box and let it infintely improve itself? Why doesn't every company that employs someone who just uses a computer replace them with AI? Why are there no businesses entirely run by AIs that just tell humans what to do. Why don't the AIs just use CAD and electronic simulation to design themselves some "arms"? Why can't AI even beat basic videogames that children can beat?
>>For text and image-based tasks they are infinitely better than a human.
Sometimes. When the stars align and you roll the dice the right way. I'm currently using ChatGPT 5.1 to put together a list of meals for the upcoming week, it comes up with a list(very good one!), then it asks if I want a list of ingredients, I say yes, and the ingredients are completely bollocks. Like it adds things which are not in any recipe. I ask about it, it says "sorry, my mistake, here's the list fixed now" and it just removed that thing but added something else. I ask why is that there, and I shit you not, it replied with "I added it out of habit" - like what habit, what an idiotic thing to say. It took me 3 more attempts to get a list that was actually somewhat correct, although it got the quantities wrong. "infinitely better than a human at text based tasks" my ass.
I would honestly trust a 12 year old child to do this over this thing I'm supposedly paying £18.99/month for. And the company is valued at half a trillion dollars. I honestly wonder if I'm the bigger clown or if they are.
At analyzing and reproducing language.. words, code etc sure because at their core they are still statistical models of language. But there seems to be growing consensus that intelligence requires modeling more than words.
Not sure why you are being downvoted. Seems like lot of people with high ego are not ready to accept the truth that a human has way less knowledge than a world encyclopedia with infinite and practically perfect memory.
Otherwise researching intelligence in animals would be a completely futile pursuit since they have no way of "knowing" facts communicated in human language.
Yeah, I think a lot of people are very insecure. I’m genuinely sorry for them. I think the best thing to do is to derive utility from AI (to mitigate the costs).
>>Seems like lot of people with high ego are not ready to accept the truth that a human has way less knowledge than a world encyclopedia.
Well, thank you for editing your own comment and adding that last bit, because it really is the crux of the issue and the reason why OP is being downvoted.
Having all of the worlds knowledge is not the same as being smart.
Smart is being able to produce knowledge quickly. I’m not sure how it could be denied that AIs are capable of producing knowledge quickly (obviously extremely quickly).
Doing your laundry is a sign of general intelligence? If someone is a quadriplegic and cant do their own laundry, does that mean they arent intelligent?
