AGI Is Very, Very Far Away

Behind this confusion lies a commercial appetite. Everyone, related or not, who sees the size of the pie is after somehow profiting from this business. AGI is also a topic that those with the biggest slice of the pie keep constantly on the agenda. It gets reheated and served to us at the smallest development.

We know very little about intelligence

There's a lot we don't know. We don't know the nature of intelligence. We don't know its relationship with memory. We have no idea about consciousness. We're not even sure the whole story happens in the brain, since our knowledge about the human brain and how it works is still next to nothing. Despite such advanced technology, it was only last year that we managed, for the first time, to map the complete brain of a mammal (a mouse). Just the map. In an equation with this many unknowns, making bold claims about intelligence is dangerous.

Let's start with definitions

Let's take mathematics. One camp considers mathematics a branch of science. Another camp, which I belong to, argues that mathematics is merely a language. But when it comes to certainty, almost everyone agrees that mathematics is the most powerful weapon we have. We owe this certainty essentially to definitions. In mathematics, everything starts with definitions. When you can't find your way out of a problem, as Ali Nesin puts it, you need to "go back to the definition." So before I start supporting my claim that AGI is not on the horizon, let me give the definitions:

"a highly autonomous system that outperforms humans at most economically valuable work." In other words, a highly autonomous system that performs better than humans at the most economically valuable work. This is OpenAI's definition of AGI.

Although this definition is widely accepted, it is both incomplete and flawed. It's incomplete because it ignores certain tasks that have no economic value but are very simple for humans. The best example of this is François Chollet's ARC-AGI project. Chollet and his team designed puzzles made of various shapes using only a grid of variable size (think of it like graph paper) and a color palette. The goal is to solve the logical relationship between the shapes (I discussed this project in detail in an article published in HBR Türkiye; those interested can take a look).

Moreover, performing complex tasks that are hard for humans cannot be a criterion on its own either. One of the best examples of this is the test that DeepMind co-founder Mustafa Suleyman proposed as a replacement for the Turing test. As Mustafa Suleyman put it, starting with $100,000 in capital and turning it into $1 million cannot be an AGI test. Just as losing the $100,000 wouldn't make you stupid, turning it into $1 million doesn't mean you're intelligent either. I would first expect the system to grasp what $100,000 even means, that is, the concept of money. That would be a tremendous step. Douglas Lenat's legendary Cyc project aimed at exactly this: giving artificial intelligence common sense. Even though Ramanathan V. Guha, one of the project's creators, admitted that the idea of creating a system with common sense had failed, the project is still ongoing. A glimmer of hope...

A system that doesn't know why it does what it does cannot be intelligent.

Now, one could say that OpenAI's not-yet-released o3 system achieved a historic success on ARC. True, since the congratulations came straight from the source: Chollet himself. But it's worth being cautious. First, there are dozens of puzzles that o3 couldn't solve and that are quite simple for humans. Moreover, Chollet has signaled that a brand-new puzzle set will be announced very soon, on which he expects human success to be around 95% while systems like o3 drop to around 30%. Second, when GPT-3 was released, also by OpenAI, statements to the effect of "congratulations, AGI is here" were in vogue. I don't want to burst anyone's bubble, but the supposedly super-intelligent GPT-3's score on ARC puzzles was 0%! It couldn't correctly solve a single one of the puzzles that an average person solves 80% correctly. I'll leave the interpretation to you.

It is also flawed because many systems already perform better than humans. We've started to see more of such systems, particularly in the field of medicine. Performance measured on physical, or even cognitive, tasks can only be one component of intelligence. Nothing more.

Now let's also look at François Chollet's definition of AGI: "AGI is a system that can efficiently acquire new skills outside of its training data. More formally, the intelligence of a system is a measure of its skill-acquisition efficiency over a scope of tasks, with respect to priors, experience, and generalization difficulty." In other words, AGI is a system that can efficiently acquire new skills outside the data it was trained on. More formally, a system's intelligence is a measure of its skill-acquisition efficiency over a scope of tasks, with respect to priors, experience, and generalization difficulty.

AGI is not yet part of our lives

Let's state it clearly: AGI is not yet part of our lives. Whether it will be is unknown. It being realized is our greatest hope, but it's even doubtful whether we're on the right path toward it. By saying it's coming any minute now, or no, 3 years left, sorry, I meant 5, actually let's think of it as 10 years: with every failed prediction, we're only fooling ourselves by generously "buying time." The whole world is holding its breath waiting for "what will OpenAI do now," but as far as I'm concerned, the birthplace of AGI will not be companies like these but the scientific community. Just like François Chollet's prediction that the 85% threshold on ARC puzzles (while also satisfying all the other rules) will be surpassed outside the AI industry...