On Monday, July 24, 2023, at 11:05 PM, Matt Mahoney wrote: > 4. Reproduction is a requirement for life, not for intelligence. Not unless > you want to measure reproductive fitness as intelligence like evolution does. > Is gray goo our goal? It takes longest but is most accurate for measuring how smart we are. That is because the end goal is immortality and persistence of some future machine, and so if we really are smart and succeed, this score should be high. However, if brains are smart enough to discover knowledge or functions that makes them score better on prediction tests but doesn't improve immortality, then yes technically they can be smarter without increasing the immortality test's score I think maybe. *Do we though "need" brains to "know/waste energy on" stuff like "some thing that does not improve lifespan or fleet size"? We'd not want that actually.*
GPT-4 right now can earn 500USD for junior programming jobs. The money test for AI we can do now, it was not often that back then you have a AI like GPT-4 that can generate solutions for jobs. Maybe very narrow programs, but nothing like GPT-4. The one famous guy that I think I forget he was/is CEO of Google or something but anyway he said in 2 years he predicts AI will bring in 1 million dollars after devises a research plan and sells something all on its own. *GPT-5 I imagine would be able to be asked to "make a new programming language easier than C++ but better than C++ and easier then Python" and from just that prompt it would go on to do all steps on its own (a ton of little problems it can solve on its own) and burn through nights and days (possibly hours or even minutes in its fast brain) tirelessly and hand back the whole software. * Generation of senses and actions. This one we know already is a fun and useful measure. It's for humans to compare to AIs only, AIs comparing to other AIs they made need not this anymore, it's too subjective. Lastly Prediction Score, not comparable to us, but to other AIs, a very solid way to improve AIs. *However it doesn't tell us long term if it really worked or made things better resolution or more coherent etc. Other tests are needed then.* I tend to feel, as we go up the 4 paragraphs above, it gets more long-term horizon effect no? Money takes longer to measure, you need do many many many steps, then sell the final product! So, way 3 takes longer but not like top way4, that is much longer to wait for. Prediction score is fast, it seems it doesn't exactly tell you much about even what your AI makes, maybe. We currently use most the 1st bottom way to make AI, I'd imagine. What test to test for AGI? Way1 can't unless Matt is really sure about humans can compress what he said to 1bpc. Yes way2 and way3 already show we are closeish to human level. Have to think on this more but for now...: But what really to look for? What is the real thing we want? Not only do we want AI to work on AI, but we know humans - only smart humans - can work on AI. So, the test should maybe be if we can make an AI that can work on AI and have the AI use evaluations I listed above, it must be therefore human level AI, and, would also be what we want to see happen, too, hehe :). ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T91a503beb3f94ef7-Mb549e6c37e9761668fda9896 Delivery options: https://agi.topicbox.com/groups/agi/subscription
