

Discover more from Klement on Investing
In the Terminator movies, the US Department of Defence started using Skynet as an AI system for defence purposes in 1994. The AI learns fast and by 1997, the AI had gone out of control and started a war with humans that lasted until 2029. That timeline was a bit premature, but with the jumps AI technology has made recently, some people warn that the arrival of Artificial General Intelligence (AGI) could be just years away and it could have disastrous consequences for humanity. Rupert Macey-Dare from Oxford University played around with a simple exponential growth model to calculate when AGI will arrive.
For those who don’t know what AGI is, it is a level of artificial intelligence that is equivalent to human intelligence. Currently, AI is always domain-specific. Some AI is good at playing chess, while other systems are good at reading text or manoeuvring across rocky terrain. But no AI can do all the things humans can do at least as good as they can.
When trying to predict when AI research will be able to generate the first AGI, we have two major unknowns. We do not know how close we are to achieving AGI at the moment, and we do not know how fast AI research will progress.
Macey-Dare thus makes simulations assuming that we are somewhere between 5% and 30% of the way toward AGI and that the capabilities of AI increase between 10% and 100% every two years.
Based on these assumptions, he estimates that AGI might arrive sometime between 2032 and 2048 with the most likely date 2041. That is just 18 years away. But in a worst-case scenario (or best-case scenario, depending on your point of view), AGI could be reached in 2028, just five years from now.
If true, that would make many jobs that currently exist redundant within the remaining careers of most readers. And if you are having children right now, they may graduate school into a world where their skills are about to be replaced by machines with human-like intelligence. Fun times.
But if you ask me, this exercise is not only fraught with extremely large uncertainty but with methodological error as well. The uncertainty is large because we have no idea if we are 5% or 30% or 0.05% on the way to developing AGI. Macey-Dare seems reasonably confident that we are somewhere between 5% and 30% on the way to developing AGI.
That just reminds me of the famous quiz about overconfidence. Humans are notoriously unable to forecast an accurate range and frequently predict that the range of outcomes is much narrower than it really is (if you don’t believe me, do the quiz in the link). Hence, the assumption that we are somewhere between 5% and 30% on the way to generating AGI is just that. An assumption and probably a poor one.
But the bigger problem I have with Macey-Dare’s simulations is that it assumes exponential growth. This is the mistake that economists and investors make all the time. They look at a growth process in the past and then decide that the observed growth rate will continue unabated into the future. In reality, developing technology does not follow an unlimited growth path but instead follows a logistic function where growth slows down over time. This error is what I have called the most common mistake in forecasting and why futuristic technologies like nuclear fusion have been promised to come to market for decades, yet never arrived.
I think the development of AGI will follow a similar path. We will make rapid advances at the beginning, but at some point, these advances will be harder and harder to achieve, and the goal of AGI will move further and further away. That is not to say that we won’t get there eventually, but in my view, predicting that we will get there by 2041 or even 2028 seems vastly overconfident.
When will Skynet arrive?
I took the overconfidence quiz. I got 8 out of 10. But even though I didn’t know the exact answer, I had a rough idea of what the result of each one could be. On question 9 and 4, I could even do a calculation knowing the circumference of the earth. The only two I had no idea about were 7 and 10 and I guessed a range that I thought was huge in order to get the answer right. For both, I wrote a range between 50 and a million, not even getting close to the answer. And the funny thing is that I consider myself a person with a great sense of measure. For example, on Sunday I predicted Sinner’s victory over Tsitsipas in an hour and 20 minutes and I was wrong by 5 minutes. But when it comes to something for which we have no proved reference points, then everything falls apart.
Actually any one specific technology will not continue exponentially (floppy disk or optical disk for example) but the ability to do the broader thing like store information can. There are many examples going back hundreds of years in technology.