LeCun has been the most aggressive critic of the transformer/LLM consensus for years, and this is his magnum opus "much of real-world sensor data is unpredictable, and generative approaches do not work well." This basically epitomizes his view that real intelligence cannot come from scaling text prediction alone whether he is right or not, im excited to see a another school of thought come to market that in some ways restores agency back to the physical world. Worth "keeping an eye on it" :)
AMI Labs
AMI LabsMar 10, 13:04
Advanced Machine Intelligence (AMI) is building a new breed of AI systems that understand the world, have persistent memory, can reason and plan, and are controllable and safe. We’ve raised a $1.03B (~€890M) round from global investors who believe in our vision of universally intelligent systems centered on world models. This round is co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions, along with other investors and angels across the world. We are a growing team of researchers and builders, operating in Paris, New York, Montreal and Singapore from day one. Read more: AMI - Real world. Real intelligence.
there is actually a really good reason why most ppl are inherently skeptical about the limits of LLM and nobody does a better job than judea pearl who explains it via "the Ladder of Causation" basically there are three rungs in this ladder of reasoning - rung 1 is "association" which is where you observe correlation in data. "people who carry lighters are more likely to get lung cancer" type of stuff. this is what the transformer does exceptionally well with infinite scale rung 2 is "intervention" which is to understand what happens when you actually DO something, and forms the basis of the "scientific method" that we all learned in elementary school. "if i MAKE someone carry a lighter, does it CAUSE cancer?" rung 3 is the "counterfactuals" which is where you reason about what WOULD have happened if you did xyz. its basically the highest form of reasoning, which is retrospective causal reasoning. this is the foundation of moral philosophy basically, there is no amount of compute that will help you get from rung 1 to rung 2 or 3 (which to me are more representative of the human experience of intelligence). this requires causal graphs, which means you need to have structured observations. this is fundamentally different than the transformer architecture which basically just curve fitting i think a lot of why there is real hesitation on this view of unlimited "all or nothing" spend on the current AI meta is that the gain is actually more finite than investors and scientists alike would think without returning to some first principles truth
77