What is AGI? When Will we Achieve it? What Does it Mean for education?
Understanding AGI's development through an "intelligence pyramid"
Stefan Bauschard
Designing Schools AI Bootcamp; Educating4ai.com; Co-Editor Navigating the Impact of Generative AI Technologies on Educational Theory and Practice.
TLDR
AGI, ANI, ASI explained
What constitutes general intelligence? What abilities do machines currently have (natural language, basic reasoning, some creativity) and what don’t they have (ability to apply learning from one domain to another, perceive, plan a response, respond, memory, long-term planning, common sense, empathy, resilience, sense of morality, sentience, consciousness, more)?
When might AGI and ASI arrive?
What does this mean for education?
You’ve likely heard frequent references to the term “artificial general intelligence” (AGI). What does this mean? When will we see it? What does it mean for education?
The Basics
Until the most recent advances in AI, which have arguably demonstrated some initial “sparks” of intelligence, precisely defining “intelligence” and “general intelligence” was not that important.
For decades, the prevailing determination whether or not machines are intelligent was tied to their ability to perform mental tasks on par with humans. This idea, known as the Turing Test, which was introduced by Alan Turing in 1950, postulates that if a computer can respond to questions in a manner indistinguishable from how humans responds, it is “intelligent.” Arguably (there is not universal consensus on this (Marcus)), a machine that passes the Turing Test could be considered to possess artificial general intelligence (AGI), which is very broadly defined as the intelligence of the average human.
It is important to note that the Turing Test is a concept and not a formal test, so computer and cognitive scientists would have to agree on a specific test that would be administered (it would be interesting to be a fly on the wall as they tried to come to some agreement as to what should be in the test) to a computer.
Any machine intelligence that is more limited than the general intelligence level of a human is considered to possess artificial narrow intelligence (ANI). So, yes, everything from spell check (which uses AI) to the most advanced capability embedded in language models such as GPT4, as well as everything that develops between now and AGI, is considered to represent ANI.
I do think it is important to note that the term AGI includes the word “general,” something people often overlook. The question of AGI is not whether or not machines can possess any intelligence but whether or not they can be considered to be generally intelligent, or possess human level capabilities. Conceptually at least, the Turing Test refers to this level of general intelligence, not any intelligence.
This basic understanding of general human intelligence is helpful, but as machines develop more components of intelligence, it gets a bit more complicated.
It also gets a bit more complicated because AGI is used in two different ways: It is used to refer both to the idea that it is the equivalent of the “average” intelligent human and to the idea that it is more intelligent than a human.
For example, on September 21, 2022, Sam Altman, the CEO of OpenAI, which released ChatGPT, referred to it as having the abilities of the average human, including the ability to apply what it learns to another situation.
But in a recent blog post, he also defined it as AI systems that are “generally smarter than humans (Altman, February 25, 2023).” OpenAI’s Charter defines it as, “…highly autonomous systems that outperform humans at most economically valuable work.” ”Gary Marcus, a noted AI researcher, defines it as, “…any intelligence (there might be many) that is flexible and general, with resourcefulness and reliability comparable to (or beyond) human intelligence.” Sébastien Bubeck et al., define it as, “at or above human-level.” Hal Hodson defines it as, “a hypothetical computer program that can perform intellectual tasks as well as, or better than, a human.”
In practice, these concepts are not that distinct, as once machines achieve AGI, humans will be able to apply the AGI abilities of computers toward helping them achieve artificial superior intelligence (ASI). And already, computers can out-perform humans in some knowledge domains (image recognition, chemical drug development, chess), so when computers can exceed human abilities in all knowledge domains and can transfer knowledge of what is learned from one domain to the other (AGI), then there will be a degree of ASI.
Once we reach this level of intelligence, we will reach a point of singularity where machines start developing intelligence in ways that we don’t understand, both in terms of process and intention. Some individuals conflate both AGI and singularity, which makes sense given that both terms can refer both to achieving human level intelligence and moving a bit beyond it. But when they start developing intelligence beyond what we can understand, that is “next level,” both figuratively and literally.
Artificial (Super)ior Intelligence (ASI)
In this interview, Ilya Sutskever, Head Scientist at OpenAI, after defining AGI in the ways referenced above, refers to ASI as (make sure you watch it) —
Ray Kurzweil adds:
As we think about the future, there will be hard questions about humans retaining control in a world where AI is smarter than us, especially if it can more fully understand human emotions and build a greater connection with us.
Breaking Down What it Means to be “Generally Intelligent"
I’m not a scholar of human intelligence,” and the subject has been discussed and studied by philosophers, theologians, and educators for decades. Some will certainly feel this list is too long, and others will think it is too short. But after reading about this for six months, at least some of these are what many people (these “many people” have differing opinions) think computers would have to demonstrate the ability to do to be considered “generally intelligent.”
Natural language. Can the computer converse with a human in the human’s “natural language”?
Basic Reasoning. Can the computer engage in some basic reasoning? What might this include? Perhaps an ability to reach a novel but supported conclusion from a set of facts.
Reasoning. “Full reasoning” might include deductive, inductive, analogic, abductive, cause and effect, critical thinking and decompositional reasoning (Indeed.com)
Creativity. Can the computer come up with creative solutions to problems?
Memory. Can it remember what it just learned and apply it forward?
Apply cognition to other tasks
new situations
any intellectual task
This is explained well by Wei at al (2023): “AGI is capable of adapting to new situations, transferring domain knowledge, and exhibiting human-like cognitive abilities beyond streamlined and for-matted task-solving workflows..”
This one is a bit personal for me. My father passed in early June, and when I was preparing his eulogy, I was thinking through all of the things he had done in his life despite not having attended college. He went from an entry level store clerk to the District Manager and ran the church bingo, summer festivals, youth group, Boy Scouts, and a million other things. He even judged at my debate tournaments. How did he do all this with only a high school education and no formal training in any of these? He took what he learned in one area and, with ambition, applied it to other areas.
If you listen to the Sam Altman video above, I think it’s fair to say that he thinks AI needs to possess at least all of these in the list so far to be considered an AGI. I don’t think anyone thinks machines can currently do that.
He may think it needs to possess more of them, but those are clear.
[As a side note, it’s funny to think about what this means in the context of the charts you find everywhere of ChatGPT4 getting a 1460 on the SAT and 5s on most of the AP exams. Imagine that you hired someone with those scores, but they were incapable of applying a skill they learned on one part of the job to another…you wouldn’t think they were very intelligent].
Perceive a situation
Plan a response
Respond
I group these three because, in addition to the above, Yann Lecun, the Chief AI Scientist at Meta and winner of the Turing Award (like Geoff Hinton and former student of Hinton’s Ilya Sustyker), argues: “Intelligence means being able to perceive a situation, then plan a response and then act in ways to achieve a goal – so being able to pursue a situation and plan a sequence of actions.”
Now, unlike Altman, he believes that something like AGI (he doesn’t believe in the concept of AGI but he does believe machines will eventually become intelligent based on his definition above) will happen, but probably not for 15-20 years. It’s not clear to me if he believes the time frame is longer because he believes these are additional essential attributes of intelligence or because he doesn’t believe LLMs will lead to human-level intelligence (he has a very low opinion of the ability of LLMs) and thinks it will just take longer to develop an object-oriented World Model that he is working on that will achieve this level of intelligence (I’ll explain the difference in the models later), but, regardless, it’s worth articulating these attributes of general intelligence.
Learns autonomously. Ian Hogarth, the co-author of the annual "State of AI" report and an investor in dozens of AI startups, defines AGI as "God-like AI" that consists of a "super-intelligent computer" that "learns and develops autonomously" (Mok).
Planning/Long-term planning abilities. Can the machine respond not just to an immediate situation, but develop a long-term plan to solve a problem?
An understanding of the physical world. One argument LeCun frequently makes against the claim that we are close to AGI is that humans can learn much faster than machines. For example, you can teach a 16-year-old how to drive a car in about 20 hours, but we’ve been working on self-driving cars for years, and we are still probably 5+ years away from useful self-driving cars. One reason sixteen-year-olds can learn to drive cars quickly is because they understand the physical world they inhabit, and Lecun’s argument is that we can only understand less than one-third of our world through language, which is the focus of LLM training. Human brains, on the other hand, can integrate data from multiple sensory modalities (more on this below), such as vision, hearing, and touch, allowing them to form a coherent perception of the world (Zhao et al, 2023). This way of learning about the everyday world also enables humans to internalize what society accepts as right and wrong, enabling it to act with a sense of morality, which is probably why he claims his model will possess the ability to act with empathy.
Common Sense. Common sense is obviously very hard to define, but it probably refers to the basic level of practical knowledge and judgment that most people possess. Does the computer have an understanding of everyday ideas about the world and the ability to make sound decisions based on that knowledge? If it's raining outside, does it know it should carry an umbrella to stay dry? Even if it is given the direction to secure as many paper clips as possible as quickly as possible, does it know it shouldn’t run over a small child to get to the store as quickly as possible?
This notion of common sense is related to the notion of the physical world discussed above: common sense develops from how we interact with both the physical world and other people who reside in it. It’s arguably hard to imagine people being intelligent without common sense. And if they develop intelligence without the common sense not to run over a small child even if given the object to get a heart attack patient to the hospital, they may become dangerous.
Adaptability. Moreover, “the brain is highly adaptable, capable of reorganizing its structure and function in response to changing environments and experiences. This property, known as neuroplasticity, enables the brain to learn and develop new skills throughout life. (Zhao et al, 2023).”
Sentience. This refers to the capacity of an entity to have subjective perceptual experiences, or "feelings." Sentient beings not only receive sensory input, but they are also aware of the sensations. For instance, if a being is sentient, it doesn't just mechanically respond to light; it experiences the brightness or color of the light.
Consciousness. Consciousness is a broader, more complex concept. It generally refers to an entity's awareness of its thoughts, feelings, and environment. Consciousness often includes self-awareness— the recognition of oneself as a separate entity from others and the environment. It also implies a continuous narrative or stream of consciousness that connects past, present, and future. (Mok)
IBM shares this more aggressive definition: Strong AI, the machine would require an intelligence equal to humans; it would have a self-aware consciousness that has the ability to solve problems, learn, and plan for the future. Strong AI aims to create intelligent machines that are indistinguishable from the human mind (IBM)
Reslience, Morality, and Emotion In a 2020 paper, Yoshihiro Maruyama of the Australian National University identified eight attributes a system must have for it to be considered AGI: Logic, autonomy, resilience, integrity, morality, emotion, embodiment, and embeddedness. The last two attributes—embodiment and embeddedness—refer to having a physical form that facilitates learning and understanding of the world and human behavior, and a deep integration with social, cultural, and environmental systems that allows adaptation to human needs and values.
When will AGI Arrive?
It partially depends on how you define it.
As you are probably starting to see, when you think it might arrive is largely determined both by what you think it includes and how long you think it will take to develop the attributes you think are essential to possessing general intelligence.
For example, it arguably already has basic reasoning skills, so if you think the only thing that is required for it to have general “intelligence” is basic reasoning, then it arguably already has it. If you think it has to have consciousness, then it’s arguably not close.
This quote highlights the importance of the elements of the definition, arguing that it doesn’t need to be conscious even to have super intelligence.
Tom Everitt, an AGI safety researcher at DeepMind, Google's AI division, says machines don't have to have a sense of self for them to have super intelligence. "I think one of the most common misconceptions is that 'consciousness' is something that is necessary for intelligence," Everitt told Insider. "A model being 'self-aware' need not be a requisite to these models matching or augmenting human-level intelligence. He defines AGI as AI systems that can solve any cognitive or human task in ways that are not limited to how they are trained. Business Insider
I won’t discuss where AI is at each of these potential steps, but there are a few critical ones.
Natural language
Computers clearly have the ability to engage in natural language conversations; this is what has made ChatGPT so easy to use. Some point out that these conversations are largely in English. That was true initially, but there has been a lot of progress in expanding the interactions to more languages. Meta’s Massively Multilingual Speech (MMS) project “presents a significant step forward in helping to preserve languages around the world. This project has recently scaled speech-to-text & text-to-speech to support 1,100+ languages and trained new language identification models that can identify 4,000+ languages!”
Regardless, I don’t think there is any debate that these tools can engage in natural language conversations.
Basic reasoning
It’s hard to define “basic reasoning,” but I’ll start with an idea that is probably not controversial: The ability to reason from some basic facts to a novel conclusion. This is an example Geoff Hinton likes to give:
Those who don’t think it’s possible for ChatGPT4 to have basic reasoning skills point out that it was only trained on next word prediction, making it impossible for it to be intelligent. The most commonly referenced article that makes this claim is On the Dangers of Schoastic Parrots.
While this point has a valid premise, it has a lot of weaknesses.
First, ChatGPT4 was not only trained in next word prediction. It was also trained using Reinforcement Learning with Human Feedback (RLHF) (Hinton; Ouyang et al; Bei et al 2022),, Chain-of-Thought Prompting (Wei), and visual processing systems (Alayrec; OpenAI).
Second, ChatGPT4 (and even earlier models) have “emergent properties.” “Emergent properties” are capabilities that models have that are not expected based on their training. Rowan (2023) explains:
For example, large language models (LLMs) were not trained to program computers, but one of their strengths is writing computer code. They were not expected to be able to reason, but evidence of reasoning has emerged (Wei above and see below).
As Ilya Sukstever points out, the better it gets at next word prediction, the more emergent properties it seems to possess. Wei (2022) explains that as models increase in size, more emergent properties, such as reasoning, develop. This should not be surprising, as a similar thing occurs in animals. Comparatively, the brain of an insect only has around 100,000 neurons, the brain of a chimpanzee around 1.5 billion neurons, and the brain of a human around 100 billion neurons. Just as the greater number of neurons has expanded the intelligence capabilities of humans, a greater number of parameters (although they are not the same as neurons and perform different functions) appears to have expanded the intelligence capabilities of LLMs. GPT4 has 1.7 trillion parameters. ChatGPT3.5 has 175 billion.
This may (beyond my technical knowledge to opine) be related to Sam Altman’s frequent comment that critics who follow the schoastic parrot line of reasoning in the end basically agree that next word prediction is what your brain does, and if this is what LLMs have been trained to do, then it is not shocking that they can do some other things a brain can do even though they are not trained to do those specific things.
Third, Qin et al (2023) find that it has basic arithmetic reasoning.
Fourth, Sébastien Bubeck et al..“…systems that demonstrate broad capabilities of intelligence, including basic reasoning, solve a problem, and comprehend complex ideas. It’s probably a stretch to say these are “sparks of AGI,” but they are probably at least sparks of intelligence, and they are at least a demonstration of basic reasoning.
Fifth, OpenAI does claim ChatPT has basic reasoning abilities.
And it’s well documented in their research paper on ChatGPT4.
Six, even critics of LLMs (Choi, see below) agree it has basic reasoning skills.
That doesn’t mean that everyone agrees that at least ChatGPT4 has basic reasoning abilities. Melanie Mitchell below may be one qualified skeptic, but even she may agree ChatGPT4 has demonstrated some very basic reasoning skills. I do think a substantial number of AI scientists believe it does, and it’s quite a bold claim to state otherwise, which many do as a simple matter of fact (often relying on a dated and potentially flawed schoastic parrots article).
Creativity
ChatGPT4 is generally considered to be creative. It can help users write music, screen plays, and even assist with technical writing. One of my favorite creative things to do is to take a famous historical event (e.g., the Constitutional Convention) and ask it to write a screenplay using contemporary characters (the CEOs of AI companies, for example); it does a great job with this.
Beyond my own experience, a wider study showed it “trounces” humans in developing creative business ideas:
And a study conducted by the University of Montana showed it was more creative than most undergraduate students.
Finkenstadt et al. (2023) explain that it can promote neurodivergent thinking, challenge bias, and assist in idea evaluation.
Of course, it can be used in non-creative ways (copying and pasting a story you were asked to write), but that doesn’t mean that it is not a creative tool and/or that it can’t be used for creative purposes.
Common Sense and More Advanced Reasoning
Yejin Choi, a computer scientist who recently gave a popular TED Talk, argues that ChatGPT4 and other similar LLM models lack common sense and an ability to engage in “chain reasoning,” but agrees that they have basic reasoning abilities and that those are advancing.
But progress does not make advanced reasoning inevitable. Grady Booch, a software engineer, doubts that at least current LLMS will support the development of AGI because they lack a "proper architecture for the semantics of causality, abductive reasoning, common sense reasoning, theory of mind and of self, or subjective experience."
Melanie Mitchel, a computer scientist at the Santa Fe Institute in New Mexico, who is probably the greatest skeptic of any near-term time-line for AGI, explains that machines lack the ability to “reason about advanced concepts.” Nature.
One of the critical tests that is used to test machine intelligence is the Abstraction and Reasoning Corpus (ARC) test. When the test is administered, humans and machines trying to solve the problem observe multiple visual examples of a square grid altering into a different pattern. They demonstrate their understanding of the governing principle of these changes by predicting the transformation of the subsequent grid. The goal is to test the capability to extract general principles from common knowledge and utilize these in addressing new, unfamiliar challenges.
Mitchell created a new test called ConceptARC that is different from the first ARC in two ways. Firstly, ConceptARC tests are simpler; she aimed to ensure that even minor advancements in machine abilities would not go unnoticed. Secondly, she deliberately selected particular concepts for examination and devised a series of theme-based variations in puzzles for each concept.
For instance, to evaluate the concept of 'sameness', one puzzle demands the solver to retain objects in the pattern that possess identical shapes; another insists on maintaining objects aligned along the same axis. This design aims to reduce the possibility of an AI system passing the test without fully understanding the concepts.
When the test was administered, humans scored around a 90% and machines only scored around 30%. This shows a gap between humans and machines that is greater than the SAT and law school exam scores suggest, but as Mitchell notes, “It was surprising that it could solve some of the problems, because it had never been trained on them.” Nature. So, it’s reasonable to conclude that it has some reasoning abilities.
And, Mitchell’s critics point out that since GPT-4 can’t currently read "see” images (at least in the public version), it was inherently disadvantaged relative to a person who can see (parts of the image were converted to numbers for the purpose of the test). Once machines develop multimodal abilities (once they can accept text, image and video as input, just like a human) then it will be a more fair comparison. Multimodal capabilities are arguably critical to the development of AGI and systems are currently making progress in that direction. [The discussion of machines understanding the physical world will deepen your understanding of the significance of this].
Memory, Planning, Resilience, Autonomy, Ability to Generalize Learning to Another Situation, Sense of Morality, Ability to Perceive a Situation
I don’t know of anyone who thinks it has these capabilities. Park et al claim they have demonstrated some memory capabilities in Generative Agents: Interactive Simulacra of Human Behavior, but this is limited to a video game.
Sentience/Consciousness
I don’t think it’s credible (based on the overwhelming view of AI scientists) that any current AI is sentient or conscious. Blake Lemoine, a Google engineer who was terminated, did make that argument in a Washington Post interview.
Nick Bostrom, a philosopher who has studied AI for many years, was quoted in the New York Times to argue it may have “sparks” of sentience.
The debate over this is a bit difficult to evaluate because neither sentience nor consciousness scientific concepts that can be objectively assessed, and we can’t know for sure whether any other person, and certainly any other animal, is sentient or conscious. So, there is this argument that people make that if we think it’s sentient or conscious maybe it is, and that the way AIs have been interacting with us suggests that may be possible. Certainly, the widespread appearance of consciousness in these machines will have effects on society, if for no other reason than that millions of people are falling in love with them.
The development of “actual consciousness” (whatever that is) may be more likely once we have meaningful developments in quantum computing.
Applying Skills and Knowledge from one Domain To Another
As discussed previously, no one believes these models have the ability to apply learning from one domain to another. In a different article, Melanie Mitchell notes: “The assumptions that we make for humans—that they cannot memorize vast collections of text related to test questions, and when they answer questions correctly, they will be able to generalize that understanding to new situations—are not yet appropriate for AI systems." Science
Time-Frame
There are a number of general claims made regarding when we will have AGI that do not reference specific definitions when making the claim.
Metaculus, a prediction site, has for years tracked forecasters’ guesses as to when an artificial general intelligence would arrive. Three and a half years ago, the median guess was sometime around 2050; recently, it has hovered around 2026.
Former Google engineer Mo Gadawat says we will achieve it by 2025 or 2026 (Gadawat).
Geoffrey Hinton, known as the "Godfather of AI," told CBS that AGI could be here in as little as five years.
According to Demis Hassabis, the CEO of Google DeepMind, AGI “could be just a few years, maybe within a decade away.” TomsGuide
There is even a new prediction of 2024, based on the potential for ChaGPT to read a billion+ tokens (approximately 750+ million words) in 2 seconds.
Here is another general prediction.
*Ben Goertzel, a leading AI guru, mathematician, cognitive scientist and famed robot-creator… ”We were years rather than decades from getting there." (Yahoo
Emad Mostaque, the CEO of Stability.ai says machines will reach the average level of human intelligence in 3-5 years (minute 51ish) and argues that OpenAI is putting significant resources into alignment because they think AGI is five years out and explains Elon Musk thinks it will be 6 years.
These are all consistent with Ray Kurzweil’s original prediction he made in 1999 that it would be in 2029.
It is interesting to note that Anthropic’s CEO, Dario Amodei, said that when he was in graduate school, most people considered Kurzweil’s claims to be outrageous, though the claims have now generally turned out to be correct.
Definitionally Based Predictions
It is harder to find specific predictions related to specific definitions, but there are a couple of relevant notes.
One, Google’s Deep Mind may release a new model in December that has the ability to solve problems and plan, two core components of intelligence that current models lack. That doesn’t mean it will be an AGI, as there are only two abilities, and even if it can solve problems and plan it doesn’t mean it’s generalizable.
One of the greatest potential challenges is summarized by this quote: "AI systems still lack long-term planning abilities, memory, reasoning, and understanding of the physical world around us," Everitt said. (Mok).
If AGI requires that understanding of the physical world, which people such as Yann Lecun and even Geoff Hinton believe it does, and that understanding requires the development of new non-LLM models (LeCun) such has his JEPA model, that will push back the AGI time-frame. [Lecun’s most recent lecture 5/24/23 is here. [slides]. He says we might do it using the “world models” he’s developing, which would push back the time-frame to the development of AGI].
If, however, models can learn about the physical world through YouTube videos (Hinton), the time-frame may be faster. In fact, fast time-frames may depend on if LLMs can understand the physical world while also functioning as “vision models.”
Time-Frame and “AGI” Skeptics Are not Belitting Artificial Intelligence
AI scientists who are skeptical of how close we are to AGI and/or whether or not it’s a great measure of machine intelligence, such as Choi and LeCunn aruge that human-level intelligence will like be achicved.
Lecun: There's no question in my mind that we'll have machines at least as intelligent as humans (LePoint).”
Choi is working very hard a new models that incorporate common sense and the physical world.
There is Potentially a “Long Way” To Go
What these AI models can do is already impressive, but as far as the steps to AGI go, these models have probably only demonstrated the ability to use natural language, engage in very basic reasoning, and demonstrate some creativity. They still lack all of the other potential attributes of human intelligence, perhaps all of which are needed to reach AGI.
AI models do need to develop most of these attributes to achieve AGI, but we don’t know how long that will take and I’m in no way qualified to opine. It could be the common 3-6 years, which would put us at 2029 (‘prediction). It could be a bit longer, especially if it requires new models, but we’ll get there, and we may get there faster than anyone wants.
The only other thing I’ll say here this is something that Ray Kurzweil, Geoff Hinton, and Yashua Bengio (also a winner of the Turing award) agree on: The time-frame is accelerating.
“The idea that this stuff could actually get smarter than people — a few people believed that,” he said. “But most people thought it was way off. And I thought it was way off. I thought it was 30 to 50 years or even longer away. Obviously, I no longer think that.”
Geoff Hinton, March 21, 2023
INTERVIEWER: Some people think it could be like five. Is that silly?
HINTON: I wouldn't completely rule that possibility out now. Whereas a few years ago I would have said no way.
(T)he recent advances suggest that even the future where we know how to build superintelligent AIs (smarter than humans across the board) is closer than most people expected just a year ago.
Ray Kurzweil (2022) adds:
These “fast time-frame” references included, Melanie Mitchell voices her skepticism:
These are extraordinary claims that, as the saying goes, require extraordinary evidence. However, it turns out that assessing the intelligence—or more concretely, the general capabilities—of AI systems is fraught with pitfalls. Anyone who has interacted with ChatGPT or other large language models knows that these systems can appear quite intelligent. They converse with us in fluent natural language, and in many cases seem to reason, to make analogies, and to grasp the motivations behind our questions. Despite their well-known unhumanlike failings, it’s hard to escape the impression that behind all that confident and articulate language there must be genuine understanding.
Energy and Physical Robots
Even if AGI is obtained, it will not trigger an immediate societal transformation, as the capacity needs to be integrated into the world, which would require some infrastructure change. Yann Lecun, for example, imagines a world where five AIs that are smarter than us will work for us. That would require a massive amount of energy, which we probably don’t have now, even if we ignored the environmental consequences. This is probably why Sam Altman is investing in fusion power, but that’s not coming immediately. Second, while OpenAI is also investing in physical robots, and as he notes in the Atlantic article, he’d like physical robots so they don’t have to depend on people, we are a ways out from that, though there has been recent progress.
Even with progress, however, it would be hard to find enough silicon to build billions of physical robots.
It’s Not all or Nothing
Regardless of if and when we obtain AGI, today’s AI tools can dialogue in natural language, share basic reasoning skills, perform mathematical calculations and analyze data, write at least rough drafts of school research papers, demonstrate some creativity, and possess a massive amount of knowledge (more than any single human, hallucinations in tow (Rowan)) have tremendous implications for education and every day knowledge work. As abilities advance in other attributes of what is considered intelligence (reason abstractly, apply what is learned to another context), and those abilities are integrated into every day products such as office software and cars, the world will change forever (if it hasn’t already).
Education
What does this mean for schools and education broadly?
Honestly, I think this is an interesting discussion question. These are a few things that come to mind, but I encourage everyone who has made it this far to start a discussion in your school.
There is a lot of additional intelligence available to strengthen instruction at a low cost. The development of machines that can engage in natural language conversation to share knowledge, reason, perceive a situation, plan a response, and respond offers opportunities for schools to develop more individualized instructional supports for students at a very low cost, as the cost of intelligence drops close to zero (Altman).
These will come in the form of tutoring bots that are available to schools and families. In a recently published interview, Ross Anderson shares that, “Altman comes alive when discussing the potential of AI tutors. He imagines a near future where everyone has a personalized Oxford don in their employ, expert in every subject, and willing to explain and re-explain any concept, from any angle. He imagines these tutors getting to know their students and their students’ learning styles over many years, giving “every child a better education than the best, richest, smartest child receives on Earth today.”
And as more advanced AI continues to be integrated into more and more products such as Google Docs, Grammarly and Quizlet, this intelligence will be deployed as an everyday instructional enhancer at a very low cost. What intelligence will schools acquire and how will they use it to support instruction?
The implications are substantially faster for students. Why? Two reasons.
First, conceptually, AGI is measured against what adults can do. We don’t expect our students, especially our younger K-12 students, to do everything an adult can do. So, even before we reach AGI, AI tools will be able to do what our students do and do it well. To replicate what most students need to do, it needs to be able to replicate a basic essay or research paper; it doesn’t need to be a lawyer.
Second, schools are largely focused on knowledge work. As noted, more comprehensive definitions of AGI assume that someone needs to have some general knowledge of how the world works, which they learn simply by living in it and not largely through language. This is farther away, but school is not the place where most people learn how the world works. Sure, they incidentally gain some of that knowledge in school by being there, but they gain most of the knowledge that comes from living simply by living. AGI can replicate a lot of knowledge work, and it may soon be able to teach “knowledge” well. It not be able to replicate knowledge of the physical world and day-to-day living for a long time, but students don’t need school for that.
To help understand this, think about why a store may wait to hire someone until they are 14-16 for their first job (beyond labor laws). They generally want someone of that age because they’ve lived in the world long enough to understand it. If they only wanted the knowledge, they could have probably hired them after the 4th grade (the point at which they probably have the language and math skills required for most of these jobs). There is knowledge intelligence and worldly intelligence. Schools deal with the former, and AI’s strength is in the former.
Tools and students. There is farther to go along the pathway to AGI before we will have strong tutoring bots than there is until these systems can do basic student work (that’s obviously working now to a degree), but there is a decent chance that we’ll get to at least something like AGI in a decade. If we do, then these systems will be able to perform many human functions.
These unprecedented changes will happen quickly. We’ve obviously seen rapid developments in both the strength of AI tools (ChatGPT 3.5 in November to 4.0 in March), the race to develop new tools (Anthropic, Bard, Llama, Inflection), the integration of those tools into general every day products (Canva, Google, Microsoft, SnapChat) and those that are designed for school (Grammarly Go, Quillbot, getconch.ai), as well as the acceleration of investment in intelligent tutoring systems. With a $1+ trillion in investments expected to flow into these tools (Mostaque), the quick growth in abilities is expected to continue. Sam Altman has even suggested we will soon enter a world where this change is no longer linear but exponential.
This will be difficult for schools to manage, as society has never previously experienced this rate of change, especially in a way that directly impacts education. Even last spring, we saw many teachers and professors adapt instruction and assignments to a world of GPT3.5, only to have them upended by improvements in GPT4 and the integration into mainstream products. Policies developed by major academic organizations related to “citing AI text” are less relevant in the world of co-pilots. AI text detectors sold to schools are widely considered to be unreliable.
These are just a small number of challenges schools will face in a world of exponential growth of AI abilities and product integration. As schools process changes through committees, we could see the launch of ChatGPT5+ as the committee concludes its work.
Given this, I strongly suggest administrators immediately engage the issue and hire a Head of AI to help them navigate the terrain. This person does not need to be someone with a computer science degree; they need to be someone who understands the technologies well enough to help them navigate the larger societal and educational changes that are coming. Expertise in the latter is just as important.
Schools need to help students understand the world they are living in. In his interview, Ross Anderson noted that, (Altman) told me that the AI revolution would be different from previous dramatic technological changes, that it would be more “like a new kind of society.” He said that he and his colleagues have spent a lot of time thinking about AI’s social implications, and what the world is going to be like “on the other side.” Atlantic This will make students and faculty apprehensive about the future, not only about their current jobs but what the jobs of the future will look like. As Same Altman said in that same Atlantic interview, “A lot of people working on AI pretend that it’s only going to be good; it’s only going to be a supplement; no one is ever going to be replaced…Jobs are definitely going to go away, full stop.”
Education as change. Schools need to help students develop literacy in this area and to prepare them for a world of constant change because, regardless of the net impact on jobs, the jobs of the future and the way we work in society will continually change. Yuval Harari, an Israeli Historian, argues the most important skills students need in an AI world is keep learning, keep changing and how to keep a “flexible mind,” which he says may be the most. important skill.
The role humans play is essential. This essay has been about technology, but as leaders and educators, it is we hold the power to leverage AI's abilities and to shape it in a way that fosters enhanced creativity and deeper connections so we can nourish values and attitudes that will support its positive use for current and future generations. As we have all realized, the development of AI has occurred at a rapid pace, and this change is expected to accelerate in ways that are both predictable and unpredictable. As leaders, we must adjust to what is predictable and prepare for how to manage what is unpredictable. A lot of what is both predictable and. unpredictable has been laid out here.
Thank You
Thank you to Sandhya S for the idea of the “intelligence pyramid” reference.
Additional References
Anderson, R. (2023, July 24). Does Sam Altman Know What He’s Creating?
Jeremy Baum, John Villasenor (July 18, 2023). How close are we to AI that surpasses human intelligence?
Zhao, L. (2023). When brain-inspired AI meets AGI. Meta-Radiology. June
__