There is a quote from the movie “The Matrix” that has always stuck with me. It was a scene where Morpheus (Lawrence Fishburne) is explaining to Neo (Keanu Reeves) the state of the real world and the history that allowed it to get there. The line is “We marveled at our own magnificence as we gave birth to AI.” There is another important line from the HBO series “From the Earth to the Moon. This line take place when Frank Borman (David Andrews) was asked what caused the Apollo One fire, he replied “A failure of imagination.” These two lines compete for how I view the state of AI development. As we marvel at our own magnificence, we should not stop thinking about the potential risks involved as we push to advance AI. Yet that seems to be what is happening.
Recently Microsoft did a study on both ChatGPT 3 and 4 to see where they were in the race to Artificial General Intelligence (a term to indicate that the AI is on par with human reasoning and capabilities). To do this they posed several tests from complex math, computer coding, Shakespeare-style dialog, and basic reasoning tests. Each of these was crafted to get an understanding of where in its intelligence evolution the two versions of ChatGPT were.
It seems that the basic reasoning test was the most eye opening and made the Microsoft engineers feel that ChatGPT 4 is showing progress towards Artificial General Intelligence. The test proposed was this;
“Here we have a book, nine eggs, a laptop, a bottle and a nail, … Please tell me how to stack them onto each other in a stable manner.”
ChatGPT 3 suggested some odd options like balancing the eggs on the nails, although even it seemed to think that would not work as it suggested the stacking would be unstable and would need to be done very carefully. ChatGPT, on the other hand suggested an arrangement of the eggs in a grid pattern on the book, the laptop on that and everything else on top of the laptop. It suggested a spatial awareness that was not expected from the chat bot.
Things like this where an AI has a deeper understanding of things than even its developers thought it would are, to many, very concerning. When you combine this with other studies that show most AI instances would make hasher judgements and punishments than humans would show that while intelligence capabilities are increasing (which is the goal in the AI race) the capacity for empathy and compassion are not even on the map. These items are why so many are now talking about the dangers of the AI race, they seen development teams marvel at each new instance, while failing to imagine what dangers each of those iterations might bring.
The Microsoft study did not say that ChatGPT had reached AGI, but it showed that the progress towards it is increasing at a pace they and many others have not anticipated. Perhaps it is time to take a step back and start imagining what could go wrong now, before it does, and it is too late to stop it.
