4 ms·
Plenty of people working on the Manhattan Project were totally committed to building the bomb, from General Groves on down. But it doesn't seem like a valid com
by nradov 3y ago
Plenty of people working on the Manhattan Project were totally committed to building the bomb, from General Groves on down. But it doesn't seem like a valid comparison. When the Manhattan Project started, there was already a fairly clear understanding of nuclear physics and it was mainly an engineering problem. Whereas we still have no fundamental theory of cognition and are clueless about how to design a true artificial general intelligence. Some researchers believe they can get there by just scaling up LLMs but there is no real scientific evidence to support this prediction.
- circuit10 3y agoIf scaling up LLMs (or more like neural networks in general) has worked so far, isn't that a good indication that it could continue to work? Do we have any reason to believe it will stop working before human-level intelligence? It could, but it also could not so we should be cautious
- logicchains 3y ago>clueless about how to design a true artificial general intelligence. Some researchers believe they can get there by just scaling up LLMs but there is no real scientific evidence to support this prediction. What do you mean by "true artificial general intelligence"? Is there anything that a 100 IQ human could do that GPT4 with a body and more memory couldn't do? The only thing preventing GPT from having a body is the engineering problem of getting LLMs to process live images and audio fast enough for realtime.
- RandomLensman 3y agoWhy would you want a machine that is as poor and inconsistent at reasoning as humans when building machine intelligence?
- chongli 3y agoChatGPT does not have any capability to control a body. It’s an LLM, not a control loop. There is far more to human knowledge than what we write down and what can be read by machines. You can read all the books in the world about tennis but that won’t teach you how to play the sport at a competent level. You need to practice a lot! Playing a sport involves the development of deep muscle memory and strength + fitness. Muscle memory is tacit knowledge [1]. It can’t be conveyed through words, it must be developed through practice and correction from an expert instructor. There are large categories of human knowledge that fall under the tacit umbrella, and LLMs will never master any of them. [1] https://en.wikipedia.org/wiki/Tacit_knowledge https://en.wikipedia.org/wiki/Tacit_knowledge
- nkohari 3y ago> Is there anything that a 100 IQ human could do that GPT4 with a body and more memory couldn't do? The "more memory" part is carrying a lot of weight here, and it represents a very difficult problem. (Specifically, finding a way to identify and access contextually-relevant data in real time.) Definitions of AGI vary, but almost everyone agrees that GPT4 is not AGI. Can transformers continue to be scaled to the point where they can power an AGI? Sure, maybe. But like the person you responded to said, there's currently not enough evidence to draw that conclusion.