Like a Financial Times the title says: “AI in finance is like “going from typewriters to word processors”‘” (June 16, 2024). But, I think, not much further, despite all the excitement (see “Ray Kurzweil explains how AI will transform the physical world“, The Economist, June 17, 2024). At least doubts are justified about the “generative” form of AI. (IBM defines generative AI as referring to “deep learning models capable of generating high-quality text, images, and other content based on the data they were trained on.” “)
The conversational and grammatical capabilities of an AI bot like ChatGPT are impressive. This robot writes better and seems to be a better conversationalist than what must be a significant proportion of human beings. I’m told that he (or she, except the thing has no gender and I’m using the neutral “he” anyway) performs object identification and classification tasks effectively and that it does simple coding. It’s a very sophisticated program. But it depends crucially on its enormous database, in which it makes millions of comparisons with raw electronic force. I had the opportunity to verify that his analytical and artistic abilities are limited.
Sometimes they are surprisingly limited. Very recently I spent a few hours with the latest version of DALL-E (the artistic side of ChatGPT) trying to get it to correctly understand the following query:
Generate an image of a strong individual (a woman) walking in the opposite direction of a crowd led by a king.
He just couldn’t understand. I had to elaborate, rephrase and re-explain several times, like in this modified instruction:
Generate an image of a strong, individualistic individual (a woman) walking in the opposite direction of an indescribable crowd led by a king. The woman is in the foreground and walks proudly from west to east. The crowd led by the king stands at the back and walks from east to west. They are going in opposite directions. The camera is to the south.
(By “near background” I meant “near background”. Nobody is perfect.)
DALL-E was able to repeat my instructions when I tested it, but it couldn’t see the glaring errors in its visual representations, as if it didn’t understand. He created many images where the woman on one side, and the king and his supporters on the other, walked in the same direction. The first image below provides an intriguing example of this fundamental misunderstanding. When the robot finally drew a picture where the woman and the king entered opposite directions (reproduced as second image below), the king’s supporters had disappeared! A child learning to draw recognizes his mistakes better when explained to him.
I said of DALL-E “as if it couldn’t understand”, and that’s the problem: the machine, in fact a piece of code and a large database, simply does not to understand. What it does is impressive compared to what computer programs have been able to do so far, but it’s not about thinking or understanding – intelligence as we know it. This is a very advanced calculation. But ChatGPT doesn’t know that it’s thinking, which means it doesn’t think and can’t understand. It simply repeats the patterns it finds in its database. It looks like analogical thinking but without thinking. Thinking involves analogies, but analogies do not imply thinking. It is therefore not surprising that DALL-E did not suspect the possible individualist interpretation of my instruction, which I did not explain: a sovereign individual refused to follow the crowd loyal to the king. A computer program is not an individual and does not understand what it means to be one. As the featured image of this article suggests (also drawn by DALL-E after much research and reproduced below), AI cannot, and I suspect never will, understand Descartes’ idea . cogito ergo sum (I think so I am). And it’s not because it can’t find Latin in its databases.
Nowhere in its database could DALL-E find a robot with a cactus on its head. The other Dali, Salvator, could easily have imagined it.
Of course, no one can predict the future or how AI will develop. Caution and humility are required. Advances in computing will likely produce what we consider miracles today. But from what we know about thinking and understanding, we can safely deduce that electronic devices, however useful they are, will probably never be available. clever. What is missing in “artificial intelligence” is intelligence.
*****************************
Erroneous illustration of DALL-E from a simple request from P. Lemieux
Another bad interpretation by DALL-E of a simple request
Self-portrait of DALL-E, under the influence of your humble blogger
