I cannot express how annoyed I am a researcher could use such a shitty definition.
It only makes sense to say "most" if you have a clear idea of what constitutes the majority. "Most people are male" yeah, fine..50% + epsilon of humans are males. That's more or less decidable (maybe a little vague because of intersex folks). I believe it's false because there are slightly more females but it's obviously measurable.
Now, most cognitive labor...what does that mean? Is it most of the time? Most of the tasks? Most of the value? Most of the job descriptions?
If I am a developer, and the majority of my code is written by AI, but I'm still in the driver's seat, is that most of my cognitive labor? Probably not. Ok, what if my company fires 60% of its developers, does that mean most development cognitive labor is automated? Well, it's most of the expense, and most of the butt in chair time, and it's most of the individual jobs, but it's not most of the job descriptions.
Of course, there's no way that all these researchers making pronouncements are giving consistent answers to what they mean by "most". They're probably not using his phrasing either.
Edit: The four options I threw out above: time, tasks, value, job descriptions are each interesting in their own way. My point is not that they're bad questions to be asking, it's that they're all separate questions that matter in different ways.
You can make them separate questions rhetorically but it doesn't mean they need to followed up on as such. It's pretty simple...
Most of the time? Well it includes the word most, so yes.
Most of the tasks? Well it includes the word most, so yes.
Such is a common way of writing. Think of it as a kind of compression. Researchers consider the rhetoric more than you want to give them credit for with your knee jerk "I don't personally understand so these researchers are idiots" ad hominem.
Models do contain a mathematical happy path to answer most questions that have been asked and answered when the model was trained. The issue is not whether those answers exist but finding them. That's what the bulk of the bleeding edge of model work is focused on atm.
One of my reports just sent me a giant design doc that Claude enthusiastically generated packed with plausible looking technical detail. Unfortunately the problem it's trying to solve is completely misguided and we shouldn't be doing it at all. So I'd say as answer to the question posed by this title: a while.
Yeah. I find a lot of the gripes people have a bout LLMs, these days, could also be said about employees.
Like, "the agent went and did a bunch of unrelated changes for the task I asked it to do!"... Have you ever worked with other engineers? This happens all the time
That's because you're still in the loop to point out it is trying to solve a misguided problem. Once it is LLMs all the way up - or at least far enough up so that the layers above it don't have the required technical knowledge to deduce whether the machines are following the correct track - they'll be solving problems 'till the cows come home no matter whether they're worth solving.
Yeah which is very fair but I still see arguable assumptions, the most crucial ones imho are that humans are exceptionally good at deciding what is worth doing and that AI will not get better on taking autonomous decisions.
Look, you might be the most knowledgeable person on this planet but we cannot verify. Your “report” might be brilliant and you are just a dinosaur. I mean, it’s impossible to tell.
Also, deciding on whether something is a good idea for the business is not a thing LLMs are currently trained for. We are busy automating development which will in turn accelerate all other automation. Deciding whether some line of thinking is a good idea does not strike me as particularly outside the range of capabilities we are witnessing them exhibiting today.
The current HN submission title ("AGI timelines shift with whichever lab is dominant") is very bad. It is neither the title of the article nor is it the thrust of the content.
The title of the article is "How long until AI automates all cognitive labor?"
The main point of the article is summarized by its intro: "Recently, though, I noticed that many great researchers have now published two or more precise forecasts, all using similar definitions of AGI, and all providing confidence intervals. So I was able to visualize how their forecasts changed over time."
The closest the article comes to saying the HN submitted title is:
> And every single person who updated their timelines from January 2026 to April 2026 has moved their timeline to say AGI is coming sooner, myself included.
> So I think the data supports the impression I got from Daniel, Eli, and the AI Futures team. One way I could characterize it is: in the ChatGPT era, people updated towards AI coming sooner. Then in the xAI, Meta, and Gemini era, people updated towards it coming later. Then in the Anthropic era, people updated towards AI coming sooner. Take from that what you will.
Original title took one framing from the back half of the post (3 update cycles that can loosely be called the "ChatGPT era, then xAI/Meta/Gemini era, then Anthropic era"), but definitely not the point here. Thanks for flagging
There was a ChatGPT era, and now an Anthropic era (less so though than the initial boom was 'ChatGPT dominated'), but there never was an xAI, Meta, and Gemini era.
> The overlapping AGI definition I use here is "Most purely cognitive labor is automatable at better quality, speed, and cost than humans".
That's a poor definition. Nowhere have I seen cheapness as being a requirement to count as AGI. If we have something that can do everything people can do and more, but it costs a lot means it's not AGI?
Author here, I drew on this from AI 2027. Yes, a very-expensive AGI, e.g. $1 million / day to simulate a smart human, would be a huge deal. But it would have meaningfully different effects than a cheap one.
Here's one definition AI 2027 used [1]: "Superhuman coder (SC): An AI system for which the company could run with 5% of their compute budget 30x as many agents as they have human research engineers..."
I've got no problem with your concept, and even think it's useful. I just don't think that concept and AGI are the same thing. Economically useful has no relation to what has been called AGI before.
I take it as a sign of how close it is (or how close people think it is). When AGI was SFnal magic, merely having it at all is a fascinating concept. Now that (people think) it's on the horizon, there are more practical concerns, like the fact that running these things might cost a substantial amount of money.
If it's merely human equivalent but somehow costs a lot more than actual humans, then it's actually pretty marginal until the cost comes down. There are a lot of humans.
So you could technically have AGI without entering a true AGI era. "95% as good as an average Harvard graduate across the board, but it costs $5 million/year to run" is impressive and scientifically interesting, but not economically transformative.
But if it costs $50,000/year to run, then everything changes really fast. And not necessarily in a good way.
Plenty of C-level executives have salaries around that number. Replacing them with AGI would be cost-effective. Cost is contingent, shouldn't be part of the AGI definition.
You could replace them with a potato and get similar economic outcomes for their companies.
They get this much money not because their work is worth this much. It's just how the system is set up.
AGI couldn't be CEO BECASUE it can't receive millions of dollars in compensation the same way potato can't. Getting this much money is what the CEO does. Apart from that they do very little when summed up.
Well, let's look at someone like Einstein. Just for argument's sake let's say he has a flat salary demand of $5 million dollars. It's not cost effective to hire Einstein to write your CRUD apps in this situation. That doesn't mean there isn't somewhere that he would have a value of $5 million.
No insult to the Harvard grads I know. But the median grad isn't Einstein, and they won't magically earn back $5 million/year. They're not that special, on average.
Now, if you have an AGI that can reliably and repeatedly do Einstein-level science, then I'd argue that we're starting to talk about ASI, aka, "superintelligence." Which would be providing something that humans can't consistently produce at any cost. So cost becomes much less relevant.
But if the best you can do is replace an ordinary smart human for $5 million/year, you have to compete with ordinary smart humans. Who are abundant and who very rarely cost more than $500,000/year, if you're willing to shop around and gamble a bit.
It's also a very lame definition. Intelligence - and humans - are more than just labor.
(You'll forgive me for conflating humanity and intelligence - we are homo spaiens, after all. Thinking man.)
I'm not _confused_ why these "AI" "Labs" are using that definition though. It's extremely clear they're trying to eliminate the need for the non-owner class. They're not selling LLMs (some companies are, but not these companies). These companies are selling the idea of labor without laborers to people who hate and fear laborers - and their utter dependence on them - more than anything else in their lives.
Really looking forward to the scam collapsing. Crypto wasn't very satisfying to me because too many of the victims were just idiots. This time, it's class warfare.
If it's not cheaper, it won't be automated. So saying that it's gonna get massively automated already includes assumption that it's gonna be cheaper. No harm in mentioning it again.
> The overlapping AGI definition I use here is "Most purely cognitive labor is automatable at better quality, speed, and cost than humans". For some of these researchers, saying they use this definitions is a bit of a stretch, but I included everyone who I judged as close enough to be informative.
Seems "AGI" is on the same level as "art" or "love" in that everyone knows what we're talking about but no one can nail down unanimously what it is.
I have no idea why this "AGI is not even well defined" meme gained so much traction recently.
AGI is something that can do everything better than humans. Write a novel, seduce someone, prove a theorem, fix a pipe, whatever. And it's clear right now we don't have it.
Your definition is closer to ASI than AGI. And that's the explanation for your first sentence: it's not well defined because you ask 10 people and get 12 different definitions. And it gets even worse if you ask experts in the field :)
Then you have the process of drifting definitions (or, more colloquially moving the goalposts). Hassabis has said this himself: his definition of AGI has shifted. And we know that's true, because we have his definition from 2010 when he started DeepMind. His definition then was much much "simpler", and there are arguments to be made that we already have that. But, alas, he's changed the definition. As did most of us. Seeing the progress will do that to you.
Even going by your definition, even adjusting it for "General" instead of "Super", it's still not clear. What's better? Is a poem written by a nobel laureate better than one written by a lit student? Probably. Is one written by a nobel laureate better than another written by another nobel laureate? Maybe? Is the one scribbled on a card by your 5yo for your birthday better? It most certainly is better for you. And so on...
We're not dealing with easy to define things here. Hell, I could make arguments that every word in Artificial General Intelligence is so hard to define or ambiguous that you'd never reach a consensus between a group of people. There are good arguments to be made in ever each direction. That makes it by definition not well defined. It's all ... relative :)
I don’t think the definition is that clear cut at all. A human can remember the smell of something and invoke it right then and there even if only for half a second. Are we expecting an AGI to do the same?
Then again, transformers seem super-human in some ways already. Who do you know who can more or less recite and make associations from (even if not always intelligently) hundreds of billions of text fragments? Transformers already are better at math than your average human.
My bet is we’ll land in a weird place in between where these systems clearly have some superhuman intelligent capabilities but still are far from “do everything better than humans”.
Which humans? "Humans" are not fungible objects, no matter what the gray-wool-suit set says. The LLMs are already replacing human workers on the bottom of the food chain. Are they perfect? No. Are the humans they are replacing perfect? No. At that point it becomes about tradeoffs.
If AGI is "better at every human at everything" that is ASI, which is a different breed of cat.
You make a good point that not having it is easy to spot. But what precisely would flip the switch?
Seducing someone for example, how often would that have to work? On all people? Maybe that was just thrown out as an example but it points to how subjective these goal posts are.
It's been a big problem for a while. The big Metaculus question about AGI has depends on the game "Montezuma's revenge" (!), and there have been many debates about this going back to at least 2020: https://www.metaculus.com/questions/3479/date-weakly-general...
It's not just that, it can also learn without having to be retrained. Which goes back to the issue, the real issue is people like Scam Altman can claim AGI is near, but then later say "well my view of AGI was that it is x, y and z" if not pressed to define what they think AGI is in that exact moment they're commenting on AGI, they can just later redefine it.
AGI is just "artificial" (a program) version of general intelligence (the general purpose intelligence humans have).
Nothing in AGI implies "surpass humans in every cognitive task".
Not even "match in every cognitive task" is really required. There are humans that by definition have "general intelligence" that still don't match other humans "in every cognitive task", just in some.
Why should AGI need to match ALL humans in EVERY congitive task then? An AGI just needs to be as good as an average (or even slightly below average) human, in human-like cognition.
AGI is simple: the model does not need to be endlessly trained, I can hand it a PDF about a brand new programming language, and the next person to talk to the same model should get an answer at the same speed and knowledge as if it were trained. We are clearly nowhere near this, we're in a state where we can 100% fake this, but nobody has shown this to be the case yet. I think its certainly possible, but I am also convinced that it will require rethinking how we do LLMs today.
So find the group that cares about the collection of capabilities you care to talk about. Regardless of whatever line is drawn for AGI, it's obvious that should some tech advances come to pass, we'll all care about the threshold of many jobs going away. Does that mean AGI? The people who care about jobs won't quibble, they care about the jobs.
If the issue you care about is jobs going away then I think you'll find a growing movement with a common base of beliefs.
So it's not a human intelligence. The transformer works very differently. We're trying to emulate human intelligence on a very different architecture.
Although, for the most part, what we actually seem to care about is that the job gets done. It's just that all the training data we have is "guy shaped" (linear), not transformer shaped. We haven't actually figured out how to train a transformer yet.
I really liked Dario's metaphor that in the 80's, we could have said someday we'll have "supercomputers", which can do all the calculations we did except WAY faster. When, in reality, the AI's just get smarter over time, even if the frontier is jagged. AGI is just vibes only for "smart enough, consistently enough".
AGI means no needing to retrain the model, it should be able to learn on the fly. That's the true meat of AGI. Any CEO or exec saying any remark about AGI should be forced to define what their definition of AGI is in that moment, or be completely shunned by the industry, since it seems they can just reframe what they meant by AGI later if they don't define it in that moment.
Most jobs do not require congnitive labor. They require following rules, policies and best practices. Independent thinking might even be discouraged. Much of accounting/finance, law, medicine, and business administration fall in this area.
Research and problem-solving in these fields may still need cognitive work, but the day-to-day practice of jobs does not. AI will take all of that work soon.
On the point of the advent of artificial general intelligence it is worth considering the expected reduction in human intelligence which comes with the increased offloading of cognitive activities to thinking machines. My daughters both remarked on how many of their friends seem to use 'ChatGPT' for just about everything no matter how trivial. Just like unused muscles tend to waste away the same is true for unused cognitive circuits: use 'm or loose 'm. Those ChatGPT-ing girls are doing their part in advancing the advent of AGI by strengthening the botware while weakening the wetware.
Top researchers say AI will automate all cognitive labor by 2045 at the latest. 5 years ago I would've thought you're joking. Today, I think you're being serious, but I still think you're wrong.
It's a specific subtype of Gell-Mann Amnesia effect. They "know" AI cant do their own jobs, but it seems pretty good at summarizing what others do without understanding the nuance. It seems to really apply to the AI Lab CEOs who appear "shocked" everyone isn't simply replaced with LLMs by now so the timelines get kicked.
i like this and i think it's adjacent to gell-man effect rather than a subtype of it. Any CEO claiming AGI is here will never say, "AGI is here because it can automate what I do today." They are saying AGI is here because they have some loose understanding of what other humans seem to be doing and think the LLMs can also do this. In a way it also makes me think that these CEOs are kind of operating like LLMs - they sound confident, they don't have the full nuanced picture (its impossible to have a nuanced understanding of everything), they are not doing the actual labor that they want to replace.
I linked elsewhere in a comment, Metaculus has AGI forecasts.
You can also now use AI forecasters like FutureSearch [1] (disclaimer: I work there), which are competitive with the best humans / teams of humans. And since you aren't depending on a human crowd, you can ask any variation of AGI questions with any definition, even ask conditional questions.
Amodei still predicts 2028, the same year when we'll have full self driving and Mars settlements.
So far all he has is this little code stealing application that could be replaced by git clone and sed for stripping the license.
The times before the Internet when Scientology people had to go into the streets to recruit people were nice. I wish we could put him and his ilk on some Claudology remote island, cut all Internet cables and enjoy the world without dorks and criminals that have been given a megaphone.
Until the context window gets superceded with some groundbreaking new architecture, not ever.
Even if LLMs become incredibly, undeniably brilliant 1000000 IQ, they cannot keep track of what's going across long horizons. Imagine a supergenius, but in Memento.
No amount of MD scribbling or embeddings will remove that limitation, but it may obfuscate it further and make it seem like progress is being made.
At the end of the day, being fully autonomous means that something can keep track of context, goals, complex and shifting relationships, over LONG time horizons without drift. If you need to be there to prompt, it is not truly automating. Until the continuity becomes real instead of simulated, context no longer has to compact, and weights update on-demand, you will always need a prompt wrangler leading the effort. And prompt wrangling is cognitive labor.
The article misses an important clarification for a general audience: current LLM architecture is not AGI by most scientists working on intelligence and cognition, even if its impact is already extraordinary and in many tasks exceeds human performance. AGI implies a broader set of traits.
"Most purely cognitive labor is automatable"
I cannot express how annoyed I am a researcher could use such a shitty definition.
It only makes sense to say "most" if you have a clear idea of what constitutes the majority. "Most people are male" yeah, fine..50% + epsilon of humans are males. That's more or less decidable (maybe a little vague because of intersex folks). I believe it's false because there are slightly more females but it's obviously measurable.
Now, most cognitive labor...what does that mean? Is it most of the time? Most of the tasks? Most of the value? Most of the job descriptions?
If I am a developer, and the majority of my code is written by AI, but I'm still in the driver's seat, is that most of my cognitive labor? Probably not. Ok, what if my company fires 60% of its developers, does that mean most development cognitive labor is automated? Well, it's most of the expense, and most of the butt in chair time, and it's most of the individual jobs, but it's not most of the job descriptions.
Of course, there's no way that all these researchers making pronouncements are giving consistent answers to what they mean by "most". They're probably not using his phrasing either.
Edit: The four options I threw out above: time, tasks, value, job descriptions are each interesting in their own way. My point is not that they're bad questions to be asking, it's that they're all separate questions that matter in different ways.
Some big names in AI made predictions by pulling random dates based on vibes, the author collected this and called this data.
You can make them separate questions rhetorically but it doesn't mean they need to followed up on as such. It's pretty simple...
Most of the time? Well it includes the word most, so yes.
Most of the tasks? Well it includes the word most, so yes.
Such is a common way of writing. Think of it as a kind of compression. Researchers consider the rhetoric more than you want to give them credit for with your knee jerk "I don't personally understand so these researchers are idiots" ad hominem.
Models do contain a mathematical happy path to answer most questions that have been asked and answered when the model was trained. The issue is not whether those answers exist but finding them. That's what the bulk of the bleeding edge of model work is focused on atm.
What's a definition of AGI you would use, for either time, tasks, value, or job descriptions?
No one has to provide a definition to argue that your definition is inadequate.
One of my reports just sent me a giant design doc that Claude enthusiastically generated packed with plausible looking technical detail. Unfortunately the problem it's trying to solve is completely misguided and we shouldn't be doing it at all. So I'd say as answer to the question posed by this title: a while.
>Unfortunately the problem it's trying to solve is completely misguided and we shouldn't be doing it at all
Isn't that like 90% of project plans by pre-LLM managers too?
Yeah. I find a lot of the gripes people have a bout LLMs, these days, could also be said about employees.
Like, "the agent went and did a bunch of unrelated changes for the task I asked it to do!"... Have you ever worked with other engineers? This happens all the time
yeah, those damned managers.
Yeah but these models are the worst they’ll ever be! They’ll get better! You need to give it better context! Did you try multiple agents?
Sarcasm aside, I am just so tired.
You have to loop its weighted tool harness tighter until it gets iterated.
That's because you're still in the loop to point out it is trying to solve a misguided problem. Once it is LLMs all the way up - or at least far enough up so that the layers above it don't have the required technical knowledge to deduce whether the machines are following the correct track - they'll be solving problems 'till the cows come home no matter whether they're worth solving.
Your point of view is flawed. You should take the ratio of good / bad AI plans activity and compare it with the same ratio by human work.
My point is a lot of cognitive labor goes into deciding what is even worth doing, and an LLM will happily run with anything you give it.
Yeah which is very fair but I still see arguable assumptions, the most crucial ones imho are that humans are exceptionally good at deciding what is worth doing and that AI will not get better on taking autonomous decisions.
We should trust your assessment.. why?
Look, you might be the most knowledgeable person on this planet but we cannot verify. Your “report” might be brilliant and you are just a dinosaur. I mean, it’s impossible to tell.
Also, deciding on whether something is a good idea for the business is not a thing LLMs are currently trained for. We are busy automating development which will in turn accelerate all other automation. Deciding whether some line of thinking is a good idea does not strike me as particularly outside the range of capabilities we are witnessing them exhibiting today.
The current HN submission title ("AGI timelines shift with whichever lab is dominant") is very bad. It is neither the title of the article nor is it the thrust of the content.
The title of the article is "How long until AI automates all cognitive labor?"
The main point of the article is summarized by its intro: "Recently, though, I noticed that many great researchers have now published two or more precise forecasts, all using similar definitions of AGI, and all providing confidence intervals. So I was able to visualize how their forecasts changed over time."
The closest the article comes to saying the HN submitted title is:
> And every single person who updated their timelines from January 2026 to April 2026 has moved their timeline to say AGI is coming sooner, myself included.
> So I think the data supports the impression I got from Daniel, Eli, and the AI Futures team. One way I could characterize it is: in the ChatGPT era, people updated towards AI coming sooner. Then in the xAI, Meta, and Gemini era, people updated towards it coming later. Then in the Anthropic era, people updated towards AI coming sooner. Take from that what you will.
You're right, just updated.
Original title took one framing from the back half of the post (3 update cycles that can loosely be called the "ChatGPT era, then xAI/Meta/Gemini era, then Anthropic era"), but definitely not the point here. Thanks for flagging
Nice!
There was a ChatGPT era, and now an Anthropic era (less so though than the initial boom was 'ChatGPT dominated'), but there never was an xAI, Meta, and Gemini era.
Author here, I agree, I'd be happy if admins want to change the title of this submission to the title of the piece.
> The overlapping AGI definition I use here is "Most purely cognitive labor is automatable at better quality, speed, and cost than humans".
That's a poor definition. Nowhere have I seen cheapness as being a requirement to count as AGI. If we have something that can do everything people can do and more, but it costs a lot means it's not AGI?
Author here, I drew on this from AI 2027. Yes, a very-expensive AGI, e.g. $1 million / day to simulate a smart human, would be a huge deal. But it would have meaningfully different effects than a cheap one.
Here's one definition AI 2027 used [1]: "Superhuman coder (SC): An AI system for which the company could run with 5% of their compute budget 30x as many agents as they have human research engineers..."
[1] https://ai-2027.com/research/timelines-forecast
I've got no problem with your concept, and even think it's useful. I just don't think that concept and AGI are the same thing. Economically useful has no relation to what has been called AGI before.
I take it as a sign of how close it is (or how close people think it is). When AGI was SFnal magic, merely having it at all is a fascinating concept. Now that (people think) it's on the horizon, there are more practical concerns, like the fact that running these things might cost a substantial amount of money.
If it's merely human equivalent but somehow costs a lot more than actual humans, then it's actually pretty marginal until the cost comes down. There are a lot of humans.
So you could technically have AGI without entering a true AGI era. "95% as good as an average Harvard graduate across the board, but it costs $5 million/year to run" is impressive and scientifically interesting, but not economically transformative.
But if it costs $50,000/year to run, then everything changes really fast. And not necessarily in a good way.
Plenty of C-level executives have salaries around that number. Replacing them with AGI would be cost-effective. Cost is contingent, shouldn't be part of the AGI definition.
You could replace them with a potato and get similar economic outcomes for their companies.
They get this much money not because their work is worth this much. It's just how the system is set up.
AGI couldn't be CEO BECASUE it can't receive millions of dollars in compensation the same way potato can't. Getting this much money is what the CEO does. Apart from that they do very little when summed up.
Well, let's look at someone like Einstein. Just for argument's sake let's say he has a flat salary demand of $5 million dollars. It's not cost effective to hire Einstein to write your CRUD apps in this situation. That doesn't mean there isn't somewhere that he would have a value of $5 million.
No insult to the Harvard grads I know. But the median grad isn't Einstein, and they won't magically earn back $5 million/year. They're not that special, on average.
Now, if you have an AGI that can reliably and repeatedly do Einstein-level science, then I'd argue that we're starting to talk about ASI, aka, "superintelligence." Which would be providing something that humans can't consistently produce at any cost. So cost becomes much less relevant.
But if the best you can do is replace an ordinary smart human for $5 million/year, you have to compete with ordinary smart humans. Who are abundant and who very rarely cost more than $500,000/year, if you're willing to shop around and gamble a bit.
It's also a very lame definition. Intelligence - and humans - are more than just labor.
(You'll forgive me for conflating humanity and intelligence - we are homo spaiens, after all. Thinking man.)
I'm not _confused_ why these "AI" "Labs" are using that definition though. It's extremely clear they're trying to eliminate the need for the non-owner class. They're not selling LLMs (some companies are, but not these companies). These companies are selling the idea of labor without laborers to people who hate and fear laborers - and their utter dependence on them - more than anything else in their lives.
Really looking forward to the scam collapsing. Crypto wasn't very satisfying to me because too many of the victims were just idiots. This time, it's class warfare.
If it's not cheaper, it won't be automated. So saying that it's gonna get massively automated already includes assumption that it's gonna be cheaper. No harm in mentioning it again.
> The overlapping AGI definition I use here is "Most purely cognitive labor is automatable at better quality, speed, and cost than humans". For some of these researchers, saying they use this definitions is a bit of a stretch, but I included everyone who I judged as close enough to be informative.
Seems "AGI" is on the same level as "art" or "love" in that everyone knows what we're talking about but no one can nail down unanimously what it is.
I have no idea why this "AGI is not even well defined" meme gained so much traction recently.
AGI is something that can do everything better than humans. Write a novel, seduce someone, prove a theorem, fix a pipe, whatever. And it's clear right now we don't have it.
Your definition is closer to ASI than AGI. And that's the explanation for your first sentence: it's not well defined because you ask 10 people and get 12 different definitions. And it gets even worse if you ask experts in the field :)
Then you have the process of drifting definitions (or, more colloquially moving the goalposts). Hassabis has said this himself: his definition of AGI has shifted. And we know that's true, because we have his definition from 2010 when he started DeepMind. His definition then was much much "simpler", and there are arguments to be made that we already have that. But, alas, he's changed the definition. As did most of us. Seeing the progress will do that to you.
Even going by your definition, even adjusting it for "General" instead of "Super", it's still not clear. What's better? Is a poem written by a nobel laureate better than one written by a lit student? Probably. Is one written by a nobel laureate better than another written by another nobel laureate? Maybe? Is the one scribbled on a card by your 5yo for your birthday better? It most certainly is better for you. And so on...
We're not dealing with easy to define things here. Hell, I could make arguments that every word in Artificial General Intelligence is so hard to define or ambiguous that you'd never reach a consensus between a group of people. There are good arguments to be made in ever each direction. That makes it by definition not well defined. It's all ... relative :)
I don’t think the definition is that clear cut at all. A human can remember the smell of something and invoke it right then and there even if only for half a second. Are we expecting an AGI to do the same?
Then again, transformers seem super-human in some ways already. Who do you know who can more or less recite and make associations from (even if not always intelligently) hundreds of billions of text fragments? Transformers already are better at math than your average human.
My bet is we’ll land in a weird place in between where these systems clearly have some superhuman intelligent capabilities but still are far from “do everything better than humans”.
Is seducing someone a cognitive task? In a way I guess it is but often there are a lot of meatspace factors at play as well.
Maybe a bit off topic but your comment made me wonder.
I think generally we don’t have a good definition of what intelligence is.
Which humans? "Humans" are not fungible objects, no matter what the gray-wool-suit set says. The LLMs are already replacing human workers on the bottom of the food chain. Are they perfect? No. Are the humans they are replacing perfect? No. At that point it becomes about tradeoffs.
If AGI is "better at every human at everything" that is ASI, which is a different breed of cat.
You make a good point that not having it is easy to spot. But what precisely would flip the switch?
Seducing someone for example, how often would that have to work? On all people? Maybe that was just thrown out as an example but it points to how subjective these goal posts are.
This problem was inherent even in the original turing test. Is it AI if it fools just one person? Or does it have to fool everybody?
It has to fool the average person. Fooling someone with low IQ/bad perception is not really a feat.
It's been a big problem for a while. The big Metaculus question about AGI has depends on the game "Montezuma's revenge" (!), and there have been many debates about this going back to at least 2020: https://www.metaculus.com/questions/3479/date-weakly-general...
>AGI is something that can do everything better than humans. Write a novel, seduce someone, prove a theorem, fix a pipe, whatever
That was never the concept (which predates LLMs).
AGI was something that can think like a human.
Not necessarily better, and not necessarily do everything any human can do.
Well, the 2nd one requires a human form (I think? Or at least video), and the 3rd one requires robotics.
By the 3rd example we won't have AGI until we have plumber-level robotics, and by the 2nd example we won't have AGI until the plumber is really hot.
It's not just that, it can also learn without having to be retrained. Which goes back to the issue, the real issue is people like Scam Altman can claim AGI is near, but then later say "well my view of AGI was that it is x, y and z" if not pressed to define what they think AGI is in that exact moment they're commenting on AGI, they can just later redefine it.
Given your definition what's the difference between AGI and superintelligence?
AGI should at least match, not surpass humans in every cognitive task.
AGI is just "artificial" (a program) version of general intelligence (the general purpose intelligence humans have).
Nothing in AGI implies "surpass humans in every cognitive task".
Not even "match in every cognitive task" is really required. There are humans that by definition have "general intelligence" that still don't match other humans "in every cognitive task", just in some.
Why should AGI need to match ALL humans in EVERY congitive task then? An AGI just needs to be as good as an average (or even slightly below average) human, in human-like cognition.
I guess AGI is the breaking point and superintelligence is everything above?
AGI is simple: the model does not need to be endlessly trained, I can hand it a PDF about a brand new programming language, and the next person to talk to the same model should get an answer at the same speed and knowledge as if it were trained. We are clearly nowhere near this, we're in a state where we can 100% fake this, but nobody has shown this to be the case yet. I think its certainly possible, but I am also convinced that it will require rethinking how we do LLMs today.
So find the group that cares about the collection of capabilities you care to talk about. Regardless of whatever line is drawn for AGI, it's obvious that should some tech advances come to pass, we'll all care about the threshold of many jobs going away. Does that mean AGI? The people who care about jobs won't quibble, they care about the jobs.
If the issue you care about is jobs going away then I think you'll find a growing movement with a common base of beliefs.
So it's not a human intelligence. The transformer works very differently. We're trying to emulate human intelligence on a very different architecture.
Although, for the most part, what we actually seem to care about is that the job gets done. It's just that all the training data we have is "guy shaped" (linear), not transformer shaped. We haven't actually figured out how to train a transformer yet.
I really liked Dario's metaphor that in the 80's, we could have said someday we'll have "supercomputers", which can do all the calculations we did except WAY faster. When, in reality, the AI's just get smarter over time, even if the frontier is jagged. AGI is just vibes only for "smart enough, consistently enough".
AGI means no needing to retrain the model, it should be able to learn on the fly. That's the true meat of AGI. Any CEO or exec saying any remark about AGI should be forced to define what their definition of AGI is in that moment, or be completely shunned by the industry, since it seems they can just reframe what they meant by AGI later if they don't define it in that moment.
Most jobs do not require congnitive labor. They require following rules, policies and best practices. Independent thinking might even be discouraged. Much of accounting/finance, law, medicine, and business administration fall in this area.
Research and problem-solving in these fields may still need cognitive work, but the day-to-day practice of jobs does not. AI will take all of that work soon.
On the point of the advent of artificial general intelligence it is worth considering the expected reduction in human intelligence which comes with the increased offloading of cognitive activities to thinking machines. My daughters both remarked on how many of their friends seem to use 'ChatGPT' for just about everything no matter how trivial. Just like unused muscles tend to waste away the same is true for unused cognitive circuits: use 'm or loose 'm. Those ChatGPT-ing girls are doing their part in advancing the advent of AGI by strengthening the botware while weakening the wetware.
Top researchers say AI will automate all cognitive labor by 2045 at the latest. 5 years ago I would've thought you're joking. Today, I think you're being serious, but I still think you're wrong.
It's a specific subtype of Gell-Mann Amnesia effect. They "know" AI cant do their own jobs, but it seems pretty good at summarizing what others do without understanding the nuance. It seems to really apply to the AI Lab CEOs who appear "shocked" everyone isn't simply replaced with LLMs by now so the timelines get kicked.
i like this and i think it's adjacent to gell-man effect rather than a subtype of it. Any CEO claiming AGI is here will never say, "AGI is here because it can automate what I do today." They are saying AGI is here because they have some loose understanding of what other humans seem to be doing and think the LLMs can also do this. In a way it also makes me think that these CEOs are kind of operating like LLMs - they sound confident, they don't have the full nuanced picture (its impossible to have a nuanced understanding of everything), they are not doing the actual labor that they want to replace.
Spot on. That is what I was trying to get at.
Reverse imposter syndrome.
Any chance there is a prediction market for this that we can use, since research has shown they tend to be more accurate than experts?
I linked elsewhere in a comment, Metaculus has AGI forecasts.
You can also now use AI forecasters like FutureSearch [1] (disclaimer: I work there), which are competitive with the best humans / teams of humans. And since you aren't depending on a human crowd, you can ask any variation of AGI questions with any definition, even ask conditional questions.
[1] https://futuresearch.ai/app
Zero chance, the outcome is far too vague to set a betting line.
That is an absolutely beautiful infographic and should become the standard for time series change!
Thank you! Tok me a few hours, without Claude Code I don't think I would have even attempted this.
somewhat relevant longread:
https://paoloanzn.github.io/2026/04/26/agi-will-always-be-on...
Thanks for sharing this, it made for a very fascinating read
Amodei still predicts 2028, the same year when we'll have full self driving and Mars settlements.
So far all he has is this little code stealing application that could be replaced by git clone and sed for stripping the license.
The times before the Internet when Scientology people had to go into the streets to recruit people were nice. I wish we could put him and his ilk on some Claudology remote island, cut all Internet cables and enjoy the world without dorks and criminals that have been given a megaphone.
The timelines you’re quoting are too short but “ai is just git clone and sed” is not a good faith argument in 2026.
Until the context window gets superceded with some groundbreaking new architecture, not ever.
Even if LLMs become incredibly, undeniably brilliant 1000000 IQ, they cannot keep track of what's going across long horizons. Imagine a supergenius, but in Memento.
No amount of MD scribbling or embeddings will remove that limitation, but it may obfuscate it further and make it seem like progress is being made.
At the end of the day, being fully autonomous means that something can keep track of context, goals, complex and shifting relationships, over LONG time horizons without drift. If you need to be there to prompt, it is not truly automating. Until the continuity becomes real instead of simulated, context no longer has to compact, and weights update on-demand, you will always need a prompt wrangler leading the effort. And prompt wrangling is cognitive labor.
I still can't believe we're going to get AGI before 2050. What's coming is amazing and frightening.
I predict that it already happened in 99 and we are all living in the Matrix Out AI optimist me if you dare
I'm curious to learn whether people think the anthropic era will last
it will literally never do that.
At this rate never. The performance of AI has been somewhere between bad and terrible, like a new but well meaning intern.
The article misses an important clarification for a general audience: current LLM architecture is not AGI by most scientists working on intelligence and cognition, even if its impact is already extraordinary and in many tasks exceeds human performance. AGI implies a broader set of traits.
how long until i stop seeing this nonsense shoveled at me from every direction
[flagged]