Even this take was, at the time, a few years after I had come up with the concept. And one of the predictions I made on this topic (I have failed to find the original post, I just know it was from 2017-2018 or so) was that the 2020s would be filled with said intermediate-type AI models that constantly get called "AGI" every few weeks
Refresher, now with visuals thanks to Gemini's Nano Banana:
Courtesy of Nano Banana. Based on the levels of self-driving used by AV manufacturers.
The term "artificial expert intelligence" was suggested to me back in 2018 or so, as a more cohesive description of this intermediate-phase AI architecture (and because the acronym "AXI" sounds cyberpunk; it has nothing to do with the company XAI)
The operating thesis is basic logic:
How on earth do we jump from narrow AI to general AI? How is it possible there is nothing in between?
And in the end, I was validated by this mulling.
Right now, we possess ostensibly "narrow" AI models like ChatGPT, Claude, Gemini, Grok, DeepSeek, etc. that nevertheless seem to have "general" capabilities that no AI hitherto the present possessed or even could possess at all.
The question is how to shift to a "general function" model.
From the start, I imagined it as a sort of 'less narrow' form of AI, and nowadays I've backronymed it into "expanded narrow" intelligence through the same means
Narrow AI has, since the origins of the field, been the only mode of all AI programs until the emergence of large language models in the late 2010s. The common expectation of real world AI has grown into models that handle single tasks, or a tiny fuzzy field of tasks. A single model capable of writing poetry, analyzing image data, and holding a coherent conversation would have been seen almost unilaterally as artificial general intelligence 10 years ago. In fact, it would even have been seen as such 5 years ago— GPT-3 (circa 2020) had already triggered conversations about whether autoregressive attention-based transformers were actually unexpectedly the path to strong, general AI.
Commonly, the start of the AI boom is attributed to ChatGPT
However, from my recollection, the true start of the AI boom occurred 6 months earlier with this little known paper release and this little remembered tweet...
At the time, we had never seen a single AI model so general and so capable before. The ability to accomplish 604 tasks, and even do many of them at human level capability, was the first true call of "Oro!" that hyperaccelerated when the average consumer first tested GPT-3.5 in the late autumn of the same year.
Yet in retrospect, it seems obvious now that, as impressive and strong as Gato was, it was still a transformer-based tokenizer. Wasn't it? If I'm wrong, please correct me!
Even at the time, many commentators noted the bizarre uncanny valley effect at play in trying to deduce what exactly Gato was, because it didn't neatly fit into either category of ANI or AGI.
Indeed, it is closer to a proto-AGI than most frontier models today, but still fails to cross the threshold. Its function remains narrow, not general and expansive or dynamic— as in, not able to continually learn to develop new modalities, not able to anchor its understanding of concepts to symbols, and did not seem to possess a coherent world model from which it could engage in abstraction, which would have allowed it to accomplish "universal task automation." But within that narrow function, it developed general capabilities. So general, in fact, that it spooked many people into declaring that the game was now over and we simply needed to race down the path of scaling.
So far, we still haven't seen a true successor to Gato (could it be a state or corporate secret? I hope not, that doesn't bode well for the AI field if we're repeating the mistakes of the Soviet Union's OGAS network)
But what exactly is "universal task automation" in that context?
Anthropomorphic AGI (does it think like a human?) is important, yes. I don't doubt that this could emerge spontaneously from any sufficiently advanced system. However, my entire framing here is trying to focus on Economic/Functional AGI (can it handle the entropy of reality?)
The "Universal Task Automation" Machine
Essentially my attempt to do to AGI what "UAP" was for "UFO," taking away some of the Fortean woo around the term to focus on what actually is or could be there.
1. The Core Misconception: Jobs vs. Tasks
The Fallacy: We tend to measure automation by "Jobs Replaced."
The Reality: We haven't fully automated a single job title in 70 years. We only automate Tasks. Many people lose their jobs to AI, but entire job titles have yet to vanish in the digital age, besides the human computer. I have to credit sci-fi author and YouTuber Isaac Arthur for explaining this to me in the past, as this was a minor epiphany about how to think about AI. And you'll understand it thusly:
The Supermarket Paradox: A supermarket clerk uses heavy automation (scanners, conveyor belts, POS systems). Without these, the store fails. Consider this: the first supermarkets opened in the late 1910s and 1920s, but were terribly slow and would have been seen as borderline "plus-sized farmer's markets" by today's standards as opposed to the common conception of a supermarket. Without the mechanical tools of the coming decades streamlining the process of scanning products and their prices, the modern supermarket would not be able to exist without either staffing autistic savants en masse or staffing large numbers at extraordinarily high wages for the skills required. Self-checkout is not market automation, as many disgruntled consumers note (it's simply passing the job to you), but the actual scanning is. There is likely no more heavily automated job title you experience on any daily basis. And yet the human remains. Why?
The Machine handles the Rigid Tasks (scanning a clean code).
The Human handles the Edge Cases (crushed tomato, missing barcode, angry customer). If you tried to create a fully automated supermarket today, you'd be better served making it more like an automat warehouse built around the limitations of vision and robotics systems from which humans can remotely order.
Result: This is Partial Automation, not Universal Automation.
The barrier preventing Partial Automation (ANI/AXI) from becoming Universal Automation (AGI) is not "intelligence" in the IQ sense; it is the ability to navigate entropy, or what I tend to call "chaos"
Scripted Automation (Levels 0-2): Works only when A → B
Scenario: The widget is on the belt. The arm grabs it.
The Chaos Event: The widget falls off the belt and rolls under a table.
Current AI: Fails. It throws an error code or continues grabbing at empty air. More advanced deep learning models will note something is wrong and actively look for the widget, but may not have the embodiment to act if it's too far removed, or delegate the task to calling a human worker to solve the problem. It lacks the world model to understand that the object still exists but has moved (object permanence/physics), and/or it lacks the embodiment to act on such abstraction.
Universal Automation: Notices the error, pauses, locates the object, creates a new plan to retrieve it (or flags a specific cleanup protocol), even actively searches for and retrieves it no matter where it's fallen, and resumes.
You cannot have a Universal Task Automation Machine without the ability to handle chaos (abstraction). This is why I tend to feel that even the strongest coding models are not AGI— the whole vibe coding trend involves models that still need to be guided and reprompted, often when they cause errors that must be fixed. Some level of logic is needed for these models to work, and yet I've yet to see a coding program that is capable of using any language and can think through and fix its code without a human telling it it made a mistake. Which is the other chief thing: when you no longer need a human in the loop in any executive function, then you've clearly crossed the threshold into AGI (take note, C-suite, planners, and asset managers, this may come back to haunt you when we reach true AGI)
My latest analogy to understanding the difference between what AI currently is and what AGI will be is that of superfluidity.
When you cool helium close to its lambda point, very curious behaviors emerge, such as intense bubbling and roiling, and even liquid helium itself is an odd and extreme substance. However, when it crosses the threshold into becoming a superfluid, it's not a gradual shift at all. Its entire quantum state shifts, and immediately bizarre new effects emerge.
This is my take for what the shift to AGI is like, and why exclaiming every new model gets us closer to AGI is arguably completely missing the point.
Current AI has "friction." That is, it gets stuck on edge cases. You can automate 20%, 50%, or 80% of tasks, mostly through genius and creative programming, but as long as a human is required to fix the "chaos," you are still in the liquid state (Viscosity > 0).
Once the system can handle the chaos/abstraction— once it can fix its own errors, once it can abstractly predict future states, once it can generalize outside its training distribution and thus prove it has "general function" rather than just "general capability"— resistance drops to zero.
It doesn't matter if the AI is legally restricted to 50% of jobs. If it technically possesses the capability to handle the chaos of 100% of tasks, the Phase Change has occurred. An AGI, even a proto-AGI as per the infographic up above, ought to be able to handle 100% of tasks at 100% of jobs. Not some arbitrary number that appeases venture capitalist predictions about potential returns on investment.
Right now, we are deep in the AXI phase and hoping that scaling gets us to Level 4.
These first AGIs, which will probably be called Proto-AGI or First-Gen AGI or even Weak AGI, will be general function + general capability, capable of universal task automation. In many ways, they will be human-level, much like Gato or any frontier model.
And yet, I strongly doubt we'll claim they are sapient (besides those suffering AI psychosis). Even with a phase change occurring in terms of functionality, the first models are not inherently defined by human capability. They are general, tautologically speaking, because they're general. Whether they are "conscious" or "sapient" entities with inner worlds as rich and alive as a human being's is irrelevant at this stage. This is yet another area where it seems people have trouble visualizing the concept due to a lack of language around it, as often "AGI" will immediately invoke the idea of an artificial human brain, and because it seems we're so far off from such, there's no reason to worry we'll ever reach it. When in reality a "mere" general-function AI model could be built within a year or two. It could even be built by someone we don't expect, because of the possibility that nearly all the major AI labs are actually chasing the wrong method. Continual learning is undoubtedly one of the more important prerequisites for any AGI, but let's entertain the thought that even a massively multimodal, infinitely time-test computing, continuously learning transformer-based model still fails to cross the threshold of AGI for some inexplicable reason. Likely, we'd still call it such, because much as with superfluidity, you don't realize when the phase transition happens until after it happens. Before then, you spend a great deal of time convincing yourself that mild changes may be signs the transition has happened.
In regards to superintelligence, the most I want to note in this particular post is the topic of "qualitative" vs "quantitative" superintelligence as represented in that infographic.
Quantitative superintelligence simply means any AGI/ASI that is still par-human or low level superhuman but can operate at superhuman speeds (which is inevitable considering how digital and analog computing works); Qualitative superintelligence is the more common conception of it, as an entirely alien brain.
And judging by both popular folk conceptions of AGI, as well as the conceptions of AGI/ASI from the venture capitalist and billionaire class, I strongly feel most people do not truly understand what "superintelligence" actually means. I may go into some detail about what I mean in a future post.
All apologies for the rambling post, but I felt the need to expound on these topics early in the year: there's no better way to prune a concept than to throw it into the wider market.
As always, if I'm wrong, please correct me or expand upon this. Do whatever with this, even reject it entirely if the entire thesis is faulty or wrong.
Several years ago, I offered the possibility of there being a hidden intermediate state between Narrow AI and General AI
https://www.lesswrong.com/posts/wGJo9xDicwwppxDJt/the-case-for-artificial-expert-intelligence-axi-what-lies
Even this take was, at the time, a few years after I had come up with the concept. And one of the predictions I made on this topic (I have failed to find the original post, I just know it was from 2017-2018 or so) was that the 2020s would be filled with said intermediate-type AI models that constantly get called "AGI" every few weeks
Refresher, now with visuals thanks to Gemini's Nano Banana:
The term "artificial expert intelligence" was suggested to me back in 2018 or so, as a more cohesive description of this intermediate-phase AI architecture (and because the acronym "AXI" sounds cyberpunk; it has nothing to do with the company XAI)
The operating thesis is basic logic:
How on earth do we jump from narrow AI to general AI? How is it possible there is nothing in between?
And in the end, I was validated by this mulling.
Right now, we possess ostensibly "narrow" AI models like ChatGPT, Claude, Gemini, Grok, DeepSeek, etc. that nevertheless seem to have "general" capabilities that no AI hitherto the present possessed or even could possess at all.
The question is how to shift to a "general function" model.
From the start, I imagined it as a sort of 'less narrow' form of AI, and nowadays I've backronymed it into "expanded narrow" intelligence through the same means
Narrow AI has, since the origins of the field, been the only mode of all AI programs until the emergence of large language models in the late 2010s. The common expectation of real world AI has grown into models that handle single tasks, or a tiny fuzzy field of tasks. A single model capable of writing poetry, analyzing image data, and holding a coherent conversation would have been seen almost unilaterally as artificial general intelligence 10 years ago. In fact, it would even have been seen as such 5 years ago— GPT-3 (circa 2020) had already triggered conversations about whether autoregressive attention-based transformers were actually unexpectedly the path to strong, general AI.
Commonly, the start of the AI boom is attributed to ChatGPT
However, from my recollection, the true start of the AI boom occurred 6 months earlier with this little known paper release and this little remembered tweet...
Gato
https://deepmind.google/blog/a-generalist-agent/
At the time, we had never seen a single AI model so general and so capable before. The ability to accomplish 604 tasks, and even do many of them at human level capability, was the first true call of "Oro!" that hyperaccelerated when the average consumer first tested GPT-3.5 in the late autumn of the same year.
Yet in retrospect, it seems obvious now that, as impressive and strong as Gato was, it was still a transformer-based tokenizer. Wasn't it? If I'm wrong, please correct me!
Even at the time, many commentators noted the bizarre uncanny valley effect at play in trying to deduce what exactly Gato was, because it didn't neatly fit into either category of ANI or AGI.
Indeed, it is closer to a proto-AGI than most frontier models today, but still fails to cross the threshold. Its function remains narrow, not general and expansive or dynamic— as in, not able to continually learn to develop new modalities, not able to anchor its understanding of concepts to symbols, and did not seem to possess a coherent world model from which it could engage in abstraction, which would have allowed it to accomplish "universal task automation." But within that narrow function, it developed general capabilities. So general, in fact, that it spooked many people into declaring that the game was now over and we simply needed to race down the path of scaling.
So far, we still haven't seen a true successor to Gato (could it be a state or corporate secret? I hope not, that doesn't bode well for the AI field if we're repeating the mistakes of the Soviet Union's OGAS network)
But what exactly is "universal task automation" in that context?
Anthropomorphic AGI (does it think like a human?) is important, yes. I don't doubt that this could emerge spontaneously from any sufficiently advanced system. However, my entire framing here is trying to focus on Economic/Functional AGI (can it handle the entropy of reality?)
The "Universal Task Automation" Machine
Essentially my attempt to do to AGI what "UAP" was for "UFO," taking away some of the Fortean woo around the term to focus on what actually is or could be there.
1. The Core Misconception: Jobs vs. Tasks
The barrier preventing Partial Automation (ANI/AXI) from becoming Universal Automation (AGI) is not "intelligence" in the IQ sense; it is the ability to navigate entropy, or what I tend to call "chaos"
You cannot have a Universal Task Automation Machine without the ability to handle chaos (abstraction). This is why I tend to feel that even the strongest coding models are not AGI— the whole vibe coding trend involves models that still need to be guided and reprompted, often when they cause errors that must be fixed. Some level of logic is needed for these models to work, and yet I've yet to see a coding program that is capable of using any language and can think through and fix its code without a human telling it it made a mistake. Which is the other chief thing: when you no longer need a human in the loop in any executive function, then you've clearly crossed the threshold into AGI (take note, C-suite, planners, and asset managers, this may come back to haunt you when we reach true AGI)
My latest analogy to understanding the difference between what AI currently is and what AGI will be is that of superfluidity.
When you cool helium close to its lambda point, very curious behaviors emerge, such as intense bubbling and roiling, and even liquid helium itself is an odd and extreme substance. However, when it crosses the threshold into becoming a superfluid, it's not a gradual shift at all. Its entire quantum state shifts, and immediately bizarre new effects emerge.
This is my take for what the shift to AGI is like, and why exclaiming every new model gets us closer to AGI is arguably completely missing the point.
Current AI has "friction." That is, it gets stuck on edge cases. You can automate 20%, 50%, or 80% of tasks, mostly through genius and creative programming, but as long as a human is required to fix the "chaos," you are still in the liquid state (Viscosity > 0).
Once the system can handle the chaos/abstraction— once it can fix its own errors, once it can abstractly predict future states, once it can generalize outside its training distribution and thus prove it has "general function" rather than just "general capability"— resistance drops to zero.
It doesn't matter if the AI is legally restricted to 50% of jobs. If it technically possesses the capability to handle the chaos of 100% of tasks, the Phase Change has occurred. An AGI, even a proto-AGI as per the infographic up above, ought to be able to handle 100% of tasks at 100% of jobs. Not some arbitrary number that appeases venture capitalist predictions about potential returns on investment.
Right now, we are deep in the AXI phase and hoping that scaling gets us to Level 4.
These first AGIs, which will probably be called Proto-AGI or First-Gen AGI or even Weak AGI, will be general function + general capability, capable of universal task automation. In many ways, they will be human-level, much like Gato or any frontier model.
And yet, I strongly doubt we'll claim they are sapient (besides those suffering AI psychosis). Even with a phase change occurring in terms of functionality, the first models are not inherently defined by human capability. They are general, tautologically speaking, because they're general. Whether they are "conscious" or "sapient" entities with inner worlds as rich and alive as a human being's is irrelevant at this stage. This is yet another area where it seems people have trouble visualizing the concept due to a lack of language around it, as often "AGI" will immediately invoke the idea of an artificial human brain, and because it seems we're so far off from such, there's no reason to worry we'll ever reach it. When in reality a "mere" general-function AI model could be built within a year or two. It could even be built by someone we don't expect, because of the possibility that nearly all the major AI labs are actually chasing the wrong method. Continual learning is undoubtedly one of the more important prerequisites for any AGI, but let's entertain the thought that even a massively multimodal, infinitely time-test computing, continuously learning transformer-based model still fails to cross the threshold of AGI for some inexplicable reason. Likely, we'd still call it such, because much as with superfluidity, you don't realize when the phase transition happens until after it happens. Before then, you spend a great deal of time convincing yourself that mild changes may be signs the transition has happened.
In regards to superintelligence, the most I want to note in this particular post is the topic of "qualitative" vs "quantitative" superintelligence as represented in that infographic.
Quantitative superintelligence simply means any AGI/ASI that is still par-human or low level superhuman but can operate at superhuman speeds (which is inevitable considering how digital and analog computing works); Qualitative superintelligence is the more common conception of it, as an entirely alien brain.
And judging by both popular folk conceptions of AGI, as well as the conceptions of AGI/ASI from the venture capitalist and billionaire class, I strongly feel most people do not truly understand what "superintelligence" actually means. I may go into some detail about what I mean in a future post.
All apologies for the rambling post, but I felt the need to expound on these topics early in the year: there's no better way to prune a concept than to throw it into the wider market.
As always, if I'm wrong, please correct me or expand upon this. Do whatever with this, even reject it entirely if the entire thesis is faulty or wrong.