The search for AGI

In the summer of 1974, when I was eight years old, my family moved from Guelph, Ontario to Kemptville, a small town just south of Ottawa. As we settled in to our spacious rural home, my parents did something my brothers and I had hoped for but weren’t expecting. They went to the local appliance store[1] and bought our first television set, a small black and white model with rabbit ears that could pick up both local and national broadcasts.

Needless to say—my parents’ views on the benefits of TV were still pretty conservative—our TV time was strictly regulated. With our father, we watched Hockey Night in Canada during the NHL playoffs, and my mother loved Sunday night family shows like Little House on the Prairie. Weeknight viewing was also limited, but one show that sticks in my mind is Project Blue Book, which aired Thursday evenings for a couple of seasons in 1978 and 1979. It was based on a real 1950s US Air Force program with the same name and featured two officers charged with the mission to investigate claims of UFO sightings and alien encounters across the United States. Invariably, they would find a logical, natural explanation for each phenomenon they examined.

What I find interesting about Project Blue Book is its ability to navigate a fine line. Unlike science fiction shows such as Star Trek, it did not take for granted the existence of extraterrestrial life, but it also did not absolutely deny such existence either. However, it set a very high bar. Using all the scientific and logical tools at their disposal to explain each event they encountered, the protagonists also always kept an open mind to the possibility that sometime, somewhere, a real alien encounter might actually have taken place.

What’s all the fuss about?
Much like the mid-20th century stories of UFOs and alien abductions, there’s growing hype and polarization in the conversations around AI. Many, including Sam Altman, CEO of OpenAI and Darius Amodei, CEO of Anthropic, claim that if artificial general intelligence (AGI) is not already here, it’s imminent, and all they need is more data and more GPUs for their generative AI systems to make the breakthrough. Altman explicitly states that his company’s goal is to create an AGI system that will benefit humanity. At the other extreme are a few die-hard skeptics who insist that AGI will never happen. And then there is a growing community of technologists who believe for various reasons that generative AI is the wrong path to achieving AGI and we need a completely new approach. Among these are Yann LeCun, chief AI scientist at Meta and Mira Murati, former CTO of OpenAI and now CEO of her own startup, Thinking Machines Lab, which recently landed $2 billion in seed funding.

I think Project Blue Book can be a useful guide to help us evaluate the technology and the arguments for achieving AGI. It seems to me that AGI has a lot in common with UFOs and alien encounters—in both cases I accept that, theoretically, it might occur at some point but, on the other hand, I think that many, if not all, of the claims made thus far can be explained in other ways. I believe we also need to be very careful of the definitions and terminology we use and avoid as much as possible the anthropomorphism of the technology we have.

Let’s take a closer look at whether or not generative AI is a viable path toward AGI.

Defining a duck
Having been educated in mathematics, I believe it’s critical to have clear, accurate and broadly accepted definitions of the terms you are going to use, before you try to prove anything. AI and AGI, I fear, lack that clarity and accuracy of definition. I would trace the origin of the term AI to Alan Turing and his work in the 1950s and 1960s. As much as I believe in Turing’s genius (he did lay the foundations of modern computer science), I have reluctantly come to the conclusion that his eponymous Turing Test, also known as the Imitation Game, is not a sufficient demonstration of the achievement of real artificial intelligence or AGI. I tend to agree with philosopher Eric Schwitzgebel’s argument that mimicry is not a sufficient condition for exhibiting the real thing.

The Turing Test examines whether a computer system can mimic human conversation, and although I was initially skeptical, I now think that most of the current generative AI systems have achieved this milestone. A recent controlled experiment at the University of California ran hundreds of trials of the Turing Test, pitting several generative AI systems against humans in blind conversations with expert interrogators. The computers achieved a pass rate of about 73 per cent. Does this mean they’re in any meaningful way, actually intelligent or human? When I worked at IBM in the 1980s and 1990s, my colleagues often invoked the saying “if it walks like a duck, and quacks like a duck it’s probably a duck”—sort of the technological expression of Occam’s Razor[2]. But there’s more to duck-ness than waddling and quacking, and there’s much more to human intelligence than speaking in complete sentences. So, let’s try to look deeper than the easy superficial answers.

For starters, we can define, or confine, the term AI to what we have today—a set of technical tools that emulate specific human behaviours to accomplish particular tasks. Beyond this, AGI is a broader notion of technological sentience—a machine that replicates all the attributes of human intelligence and sentience. To get there, we will need a better definition of human intelligence, or more than that, what it means to be human and therefore distinct from other species as well as our potential technological rivals.

… Therefore I am
In the transition from the late Renaissance to the Enlightenment, René Descartes and David Hume were among the first post-classical philosophers to consider, in a secular context, the meaning and definition of humanity—and their approaches could not have been more different.

Hume’s premise was that it is our memories, or more particularly our experiences, that uniquely define us as human beings. He argued that every person from birth accumulates a unique set of experiences throughout their life, and subsequently suggested that each individual can therefore be defined as that unique “bundle” of experiences. In other words, what happens to us, and how we organize and reflect on those happenings, define who we are. There may well be merit to that.

Descartes, of course, penned the famous phrase cogito ergo sum—I think, therefore I am. My ideas, my thoughts, my dreams—the things that percolate up in my brain both consciously and subconsciously—those are what define me as a human being. Certainly there’s also merit to the notion of original, creative thought and intentionality as the criteria for human intelligence.

I feel, therefore I am; or I think, therefore I am. I won’t suggest which of Hume or Descartes is right or wrong. Maybe the truth is in the middle, or more likely, both are aspects of a larger truth about human intelligence[3]. There’s interplay between our inner and outer selves—our experiences affect our thoughts, and vice versa: our thoughts affect how we interpret and classify our experiences. Hume’s approach lacks the sense of agency that Descartes brings, while Descartes lacks Hume’s sense of context.

To define the human condition we need both, as well as the notion of intelligence as a measure of one’s ability to complete cognitive tasks and store and organize large quantities of information.

Is this a clear and rigorous definition of humanity and human intelligence? Maybe not, but hopefully it’s a step in the right direction—and enough to draw some tentative conclusions. The protagonists of Project Blue Book probably didn’t have a clear idea of what alien intelligence would look like were they to find it, but they were able to use rational arguments to explain alien-looking phenomena. Hume and Descartes can help us on a similar quest to explain what appears to be AI or AGI, and maybe point us in a different direction to eventually find it—or not.

Road to nowhere?
Generative AI is the outward and most recent manifestation of a technology stack based on machine learning and large language models (LLMs). It took only three years to capture the public and corporate imagination, possibly more so than any previous technology has. In August 2025, OpenAI released ChatGPT version 5. CEO Sam Altman claimed that this was a “Ph.D. level personal assistant” and reiterated his company’s mission to bring about AGI in short order. Impressive as ChatGPT-5 is, it didn’t take long for it to be plagued with the same issues of hallucinations and inaccuracies that also bedevil all the other generative AI platforms. Of course, humans are biased, humans err and humans hallucinate as well, so that by itself doesn’t disqualify generative AI from being an AGI candidate.

But we humans learn from our mistakes, using our experience and our ability to think as a guide. Can generative AI do the same? I don’t think so, and the reason goes back to the fundamentals of how it works.

A common trick in the past year or two has been to ask ChatGPT or other systems to tell you how many times the letter ‘r’ occurs in the word strawberry. Quite often, the system would confidently announce that the answer was two, or four, or some other number but not three[4]. If the user tried to correct it, ChatGPT would either turn contentious and insist that it was right, or sometimes meekly admit that it was wrong and that the correct answer was, indeed, three. If you shut down your session, and came back later to try again, exactly the same scenario would play out with no better odds of getting the answer right.

Why can’t generative AI count, or do better the next time? Well, because it doesn’t actually count, nor does it really remember. Bear in mind that LLMs were initially designed only to translate text between languages, by breaking down input text into tokens—words or phrases—which are compared to similar tokens in the target language. Similarly, every prompt generative AI receives is broken down into tokens that are used to search its database and produce a plausible response. Brilliant and complex as the search algorithms are, they still simply look for likely patterns and then assemble sentences. A prompt token of “how many” suggests that an integer response is expected, but any integer will do. The system does not enumerate anything[5]. In a ChatGPT session with multiple prompts and responses, it is able to keep track of context so that follow up questions can build on previous questions and answers, producing a lifelike conversation. Now, close that session and start a new one, and you have a clean slate. Memory and context are not preserved across sessions.

But, you might say, I’ve given a pretty simplistic example. ChatGPT can do much more than look at letters. In fact, I recently used it to plan out a driving itinerary through northern Spain and southern France. It did a remarkably good job, when I prompted it with just the total number of days along with the start and end points of the journey[6]. It mapped out all the stopovers in the right order, recommended good hotels and restaurants, and even suggested some excellent wineries en route as well as sightseeing and other activities. How should I explain this seemingly intelligent behaviour, or generative AI’s success with other even more complex tasks?

Project Blue Book would say, and I would agree, simply that the pattern-matching algorithms are pretty good when you give generative AI a good prompt. There’s plenty of data available on the internet about historical cities in Europe, points of interest and driving times. Matching this to tokens in a clear, well-defined prompt with multiple refinements is not too difficult. In this sense, generative AI makes a pretty good federated search tool, as long as you can put up with the occasional and random glitches that occur when the pattern matching goes off track. In the case of my trip, I did cross-reference ChatGPT’s suggestions against other sources just to be safe—I wouldn’t want to find myself driving to a nonexistent town or visiting an imaginary landmark. Kind of a human retrieval augmented generation, if you will, but it worked well and much faster than doing the original research on my own.

However, there was no original thought, no long-term memory and no thinking ‘outside of the box.’ Generative AI is nothing more than a closed system of sophisticated rules acting on a large but finite set of data. Descartes would look at it and conclude that there is no agency, no ability to think or reason intelligently. Hume would look at it and also conclude that there is no lived experience, no bundle of memories that form a unique living being.

So, generative AI is still not intelligent. As it gets bigger it will complete more complex tasks, and outperform humans in some areas, but this is the way of every technological advancement. Underlying it all is no spark of human creativity, just a clever way of combining and matching existing data to input tokens. If there is a road to AGI, this is at best a wrong turn or more likely a dead end.

Where next?
Of course, as the theologians often point out, absence of evidence is not evidence of absence. Just like Project Blue Book did for UFOs, eliminating one candidate for AGI does not mean that AGI will never exist. Maybe we just need to keep looking, somewhere else.

Yann LeCun, in a recent profile in Newsweek, posed an interesting question. How is it that the average human teenager can learn how to drive a car in, probably, about 20 hours but after decades of experimentation and millions of hours of training data and road tests with AI, we’re still far from achieving self-driving cars? He suggests that it’s not about the quantity of data but how the data is perceived and processed. Mira Murati, on the other hand, is more interested in the interaction between humans and AI systems. After breaking with OpenAI over its monolithic approach to building large-scale generative AI, she launched Thinking Machines Lab with the aim of enabling human-AI collaboration. Instead of building fully autonomous AI systems, the company plans to launch smaller, more personalized, task-oriented AI that is also multi-modal—communicating and collaborating with people not just via text prompts but also visually and orally.

While we wait for an elaboration of their approaches to appear in my next post, I will close with words of wisdom from Piet Hein—the Dutch/Danish mathematician, physicist and poet who, as always, points his finger exactly at the heart of the problem with AI.

The soul may be a mere pretense,
the mind makes very little sense.
So let us value the appeal
of that which we can taste and feel.[7]

We’ll have a long time to enjoy the pleasures of our human existence before the machines even come close to doing the same.

[1] The “Bad Boy” chain owned and operated by future Toronto mayor Mel Lastman.

[2] The mediaeval philosopher William of Occam famously said “do not multiply entities beyond necessity,” by which he meant, the simplest explanation for any phenomenon is often the correct one. Another modern expression of this is “if you hear hoofbeats behind you, think horses, not zebras.”

[3] An assignment in my first-year philosophy class in university was to reconcile Hume’s and Descartes’ approaches.

[4] The strawberry meme became so famous that OpenAI declared the problem fixed in ChatGPT v5. It didn’t take long for users to switch to the word raspberry, with the same results.

[5] In fact, don’t ever use generative AI as a calculator. It doesn’t do mathematics; it simply reads and interprets text. For example, I’ve noticed lately that Google’s Gemini search assistant constantly misinterprets 10²⁴ as 1,024.

[6] Of course, when I asked for an additional stopover, it was added to the end of the trip instead of being inserted into the appropriate geographic location between existing cities on my route. But this was easily fixed with an additional refinement to the prompt.

[7] Hein, Piet: A Toast. Grooks, MIT Press, 1966. Tranlation by Jens Arup.