Is Yann LeCun’s Vision on Autonomous Machine Intelligence a Game Changer For The AI Community?
On June 27th 2022, Yann LeCun, one of the godfathers of artificial intelligence and Head of AI at Meta released his vision on how to build autonomous AI systems. Here is the link to the paper.
First of all, I really suggest you to read this paper. As mentioned in the prologue, the text is written with as little jargon as possible. It uses as little mathematical prior knowledge as possible to appeal to readers with various backgrounds. It’s essentially a vision of what might direct the research efforts at Meta and elsewhere in the industry.
When you start reading the paper, quite quickly, you realize that this vision is very ambitious and futuristic. After all, Yann is describing an autonomous and polyvalent AI system.
Just to put things into perspective, according to a data science maturity chart like the one below, companies’ average maturity level is still around level 2-ish as they are still developing intuitions about their data. Furthermore, only a few companies can leverage machine learning to create predictive analytics (level 3) as only about 15% of AI systems are successful (Gartner, 2020), and even fewer can optimize planning using prescriptive analytics (level 4). In his paper, Yann proposes a vision to tackle level 5, the holy grail of data-driven decision-making.
After reading this paper, here are my biggest takeaways:
Boy, what a complex system to develop!
First, making a system fully autonomous and polyvalent takes a LOT of work. To predict any type of outcome, Yann proposes to generate a full world model. Not only is this a challenging task to predict the world’s results, but this is just the beginning as the world’s model becomes the input towards predicting all kinds of other, more specific outcomes.
The ability to reason is critical
Yann’s vision comprises a “system 1” component that is short-term and does not involve complex reasoning and a “system 2” component that requires reasoning and planning. Although the means are different to get to a similar destination, Yoshua Bengio, also recognized as a godfather of AI made the same parallel in 2019.
It is very centralized around machine learning
I am surprised to see very few mentions of operational research or heuristics techniques, as they could reduce the overall complexity and make some building blocks more data efficient. An iterative way to get to his vision would probably involve these more straightforward techniques.
In conclusion
In conclusion, getting full autonomy OR polyvalence is very challenging. Targeting both is near impossible… right now anyway. If you want to get benefits out of AI, you are better off building a solution that supports your team in THEIR operations.
What do you think? What is missing to develop such an ambitious vision? Please let me know in the comment section!
