ASIA
23 August 2026
The Myth of Artificial Intelligence
Artificial Intelligence is a subject of many myths.
Erik J. Larson argues that Artificial Intelligence (AI) is a subject of many myths, in his excellent book “The Myth of Artificial Intelligence: Why Computers Can’t Think the Way We Do, published by Harvard University Press in 2021”.
Larson is able to bring to his analysis a wealth of hands-on practical experience, based on his career as a computer scientist and tech entrepreneur since the turn of the century.
At the time, the computer was a new thing. The aspiration was to programme it to do things that human minds could do. In other words, the ambitious goal was to reproduce human cognition. This ambition still animates the field and is widely envisioned by the public.
But according to Larson, the current backstory is quite different. Today’s computer scientists are developing really powerful systems, but are not headed toward the original goal of programming to do things that human minds can do.
He sees both these scenarios as being more mythology than being related to the real work of computer scientists. But we live in a world where everyone has an opinion about AI, and there are armies of people who align themselves with one or the other scenario, often based on ideology or religious impulses.
Larson’s views on the bigger picture of AI have not fundamentally changed in the more than two decades he has worked in the field. And the more he actually works building AI systems in a hands-on capacity, the more he feels these highfalutin visions about AI are silly. To repeat myself, today's computer scientists are developing really powerful systems, but are not heading in a revolutionary direction, one way or another.
So, if you try to improve self-driving cars, you'll notice that these cars are stuck. We don't have level five autonomy, because we don't have AI systems that actually understand that there's a real world out there. That is the cyber world is not the real world.
A self-driving car may never be able to understand a crash or the need to merge into a single lane or a four-way stop where another driver might wave it on or a broken red light which is replaced by a traffic cop. All those cases fall outside of the training data that are given to those systems. And even if these issues can be dealt with, the problem is not solved because an unlimited number of scenarios can be imagined.
AGI was a term coined to capture what AI was originally supposed to be, which was this kind of general intelligence that we humans have – in contrast to the narrow AI which enables you to play chess or Go, or identify cancer in images. You can’t give a narrow AI system a new task, as you would have to redevelop the system from scratch with a new data set. The best we've been able to do is “wide AI” which is when you have a very large analysis of a small world. This is what chat BT does. AGI is a dream that has always sat about 10 years in the future! Larson notes that experts don’t even have scientific analysis and definition of what terms like general intelligence mean.
If AGI is not an end point for artificial intelligence, where are we heading? Larson argues that we are now at the end point of the machine learning paradigm which began really decades ago and went on steroids with the success of the Internet in the early 2000s. Larson thinks that we need a radical new innovation. At the moment, developers are now just making the large models larger, they are not coming up with big new ideas.
Larson is able to bring to his analysis a wealth of hands-on practical experience, based on his career as a computer scientist and tech entrepreneur since the turn of the century.
The birth of AI
The field of AI has a storied history going back to the 1950s, recounts Larson. The Dartmouth Summer Research Project on Artificial Intelligence, held from June 18 to August 17, 1956, is widely considered the founding event for AI as a formal academic discipline. Organised by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon, the historic workshop took place at Dartmouth College in Hanover, New Hampshire.At the time, the computer was a new thing. The aspiration was to programme it to do things that human minds could do. In other words, the ambitious goal was to reproduce human cognition. This ambition still animates the field and is widely envisioned by the public.
But according to Larson, the current backstory is quite different. Today’s computer scientists are developing really powerful systems, but are not headed toward the original goal of programming to do things that human minds can do.
Different AI camps
Larson writes about the different camps in the AI community. First, there are people who are really excited about AI, what it can do, how it can help us and make our lives easier. And then there's a group that's a little afraid of what AI can do, what it might do to us and what it might mean for humans (the “Terminator” model).He sees both these scenarios as being more mythology than being related to the real work of computer scientists. But we live in a world where everyone has an opinion about AI, and there are armies of people who align themselves with one or the other scenario, often based on ideology or religious impulses.
Larson’s views on the bigger picture of AI have not fundamentally changed in the more than two decades he has worked in the field. And the more he actually works building AI systems in a hands-on capacity, the more he feels these highfalutin visions about AI are silly. To repeat myself, today's computer scientists are developing really powerful systems, but are not heading in a revolutionary direction, one way or another.
The challenge of self-driving cars
The author believes that machines don't interact dynamically with their environment in the way biological intelligence does. Consider the case of large language models like Chat GPT which have an amazing ability to interact with us in a conversational capacity. But if you take that into the physical world, it thinks that the words on the internet are actually the world itself. So it doesn't think at all.So, if you try to improve self-driving cars, you'll notice that these cars are stuck. We don't have level five autonomy, because we don't have AI systems that actually understand that there's a real world out there. That is the cyber world is not the real world.
A self-driving car may never be able to understand a crash or the need to merge into a single lane or a four-way stop where another driver might wave it on or a broken red light which is replaced by a traffic cop. All those cases fall outside of the training data that are given to those systems. And even if these issues can be dealt with, the problem is not solved because an unlimited number of scenarios can be imagined.
Artificial General Intelligence
Artificial General Intelligence (AGI) is a theoretical type of AI that can match or beat human skills across any thinking or intellectual task. It was the dream of lan Turing, an English mathematician, computer scientist, logician, cryptanalyst, philosopher and theoretical biologist, who was a sort of AI pioneer.AGI was a term coined to capture what AI was originally supposed to be, which was this kind of general intelligence that we humans have – in contrast to the narrow AI which enables you to play chess or Go, or identify cancer in images. You can’t give a narrow AI system a new task, as you would have to redevelop the system from scratch with a new data set. The best we've been able to do is “wide AI” which is when you have a very large analysis of a small world. This is what chat BT does. AGI is a dream that has always sat about 10 years in the future! Larson notes that experts don’t even have scientific analysis and definition of what terms like general intelligence mean.
If AGI is not an end point for artificial intelligence, where are we heading? Larson argues that we are now at the end point of the machine learning paradigm which began really decades ago and went on steroids with the success of the Internet in the early 2000s. Larson thinks that we need a radical new innovation. At the moment, developers are now just making the large models larger, they are not coming up with big new ideas.