The Arrival of Thinking Machines: A Prediction for the Future

The concept of an artificial intelligence that was capable of actually thinking for itself instead of just relying on input from its programmers used to be considered the realm of science fiction. In spite of the fact that this is the case, recent advances in the field of AI that were brought about by the release of chatbots like ChatGPT that use Large Language Models make it seem like thinking machines are more realistic than might have been the case otherwise.

However, when will these thinking machines actually into existence? First, we must define what a thinking machine actually is. Firstly, we have Artificial Narrow Intelligence, with virtual assistants like Siri and Alexa fitting neatly into this category with all things having been considered and taken into account. Secondly, we have something called Artificial General Intelligence, with the fictional AI named Jarvis from the Marvel Cinematic Universe being a great example of that.

Finally, there is the rather mythical concept of an Artificial Super Intelligence, which is essentially an AI that is capable of doing any type of cognitive task that humans can accomplish far more efficiently and effectively. Alan Turing was the first to propose a test that would try to determine how similar to humanity an AI has managed to become back in 1950.

With all of that having been said and now out of the way, it is important to note that this is just one of many benchmarks that experts can use to gauge how advanced AI has managed to become. It turns out that AI image classification has reached a 91.1% success rate, and English language understanding has reached 91.3% as well.

Visual reasoning lags behind somewhat with 84.3%, but this is still higher than the human benchmark of 80.78%. English language understanding among AI also appears to have advanced beyond the capabilities of the average human being, since most humans were only able to get to 89.8%.

Language capabilities are important because of the fact that this is the sort of thing that could potentially end up determining what an AI will truly be able to do for us. Natural language inference is yet another area where AI has already managed to move beyond human skills. The human benchmark for natural language inference currently sits at around 92.9, whereas AI has reached a 93.65% success rate.

As for question answering, AI was able to complete this task successfully in 90.94% of occasions. Humans only got the answers right 86% of the time, which just goes to show that AI is becoming far more powerful and intelligent than previous generations could have imagined.

One thing that bears mentioning is that the growth in these scores has begun to hit a ceiling. In 2022, only a 4% increase in question answering was acquired on average. This seems to indicate that narrow AI is plateauing, which makes it far less likely that AGI will get made anytime soon. It will be interesting to see where things go from here on out, since AI is clearly facing one of its biggest challenges in terms of taking the tech to the absolute limit and making it a machine that can think like a human.

 H/T: VisualCapitalist