371–374 of 374 posts
Earlier quoted context omitted.
"Here's a world class scientist here not because we had a hole in the schedule or he happened to be in town, but to discuss this subject that he " had invested himself so fully personally and financially that, should it fail, he would be ruined. FTFY
ruined how?
Earlier quoted context omitted.
Are you a mathematician? I’m not an expert on the math field but it seems like they are hitting the same issues everyone else has: current LLMs still more or less need to be supervised by an expert and struggle to do something actually novel or build out a complicated proof correctly.
There's a limit to how much novelty you're going to get from an LLM, especially in areas like programming and math where they've been heavily RL'd NOT to be novel, even to extent that the base model supports, and instead generate much narrower more proscribed outputs. The limit to the novelty you are going to get from an LLM is essentially the "deductive/generative closure" of the training data. To be truly novel and…
>You could actually wonder that one possible explanation for the human sample efficiency that needs to be considered is evolution. Evolution has given us a small amount of the most useful information possible. It's definitely not small. Evolution performed a humongous amount of learning, with modern homo sapiens, an insanely complex molecular machine, as a result. We are able to learn quickly by leveraging this "pret…
Aren't you agreeing with his point? The process of evolution distilled down all that "humongous" amount to what is most useful. He's basically saying our current ML methods to compress data into intelligence can't compare to billions of years of evolution. Nature is better at compression than ML researchers, by a long shot.
What I mean is basically the opposite. Nature not better as in more efficient. It just had a lot more time and scale to do it in an inefficient way. The reason we're learning quickly is that we can leverage that accumulated knowledge, in a manner similar to in-context learning or other multi-step learning (bulk of the training forms abstractions which are then used by the next stage). It's really unlikely we have some magical architecture that is fundamentally better than e.g. transformers or any other architecture at sample efficiency while having bad underlying data. My intuition is there might even be a hard limit to that. Multi-stage bootstrap might be the key, not the architecture.
Same for the social process of knowledge transfer/compression.
Earlier quoted context omitted.
It's not a zero sum game. Think, an agronomist visits a farm, instructs to cut a certain plant for the animals to eat at a certain height instead of whenever, the plant then provides more food for the animals to eat exclusively due to that, no other input in the system, now the animals are cheaper to feed, so more profit to the farmer and cheaper food to people. How would this be zero sum?
It would be if demand was limited. Let's assume the people already have enough food, and the population is not growing - that was my premise. Through innovation, one farmer can grow more than all the others. Since there already was enough food, the market is saturated, so it would effectively reduce the price of all food. This would change the ratio so that the farmer who grows more gets more money in total, and ever…
About the farmer needing to change jobs, in the interview that is the subject of this thread Ilya Sutskever speaks with wonder about humans' ability to generalize their intelligence across different domains with very little training. Cheaper food prices could mean people eat out or order-in more and then some ex-farmers might enter restaurant or food preparation businesses. People would still be getting wealthier, even without the tailwind of a growing population.