LLM, AGI, and why we may need a world model

This is worth a read: https://www.wsj.com/tech/ai/yann-lecun-ai-meta-0058b13c. In case you don’t have access or time here are a few notes:

• Ok, it’s not at all finance related but everything is AI now and LeCun’s view seemed interesting in a contrarian way, so why not write a quick post
• The idea is that human thinking involves modeling the world around us, i.e., things, animals, people, the relationship between those entities, etc., in our minds. We combine those with data to arrive at inferences and predictions
• E.g., if I push the cup over the edge of the table, based on the model I have of cups, tables, space, and gravity, I can make a prediction as to what’s going to happen
• I can formulate a similar conclusion by looking at enough sentences with the words ‘cup,’ ‘table,’ ‘edge’ and ‘push’ in it, but that formulation doesn’t quite seem to be rooted in the same kind of understanding. It also doesn’t lend itself to the same kinds of generalizations
• LLMs’ logic resembles the second approach a lot more than the first. They predict words based on other words.
• Because of this seeming fundamental difference between human thought and LLMs, a good number of people who don’t live in this space (including myself) have doubts that pure LLMs can get to human-level AGI
• Of course, anyone would be right to take this with a large grain of salt because those are not AI experts
• This is why it’s nice to see a bona fide AI guru who’s been at this for a long time agree with that intuition, presumably based on an incomparably deeper knowledge of the subject matter. This view may have existed within the AI community for a long time, I just never saw it bubble up in the popular press
• That is not to say AGI isn’t possible, but according to LeCun, it won’t happen until we somehow bring a ‘world model’ as he calls it into those systems. Therefore, he says we’re not nearly as close as a lot of folks seem to think

The wsj piece is here. It includes some of the Meta internal politics surrounding LeCun’s exit

This is a youtube version of the same argument