Full recursive self-improvement may have started already. /sci/ told me this was literally impossible. WTF is going on?
/sci/ was stupid.
>>17058694fuck outta here they're hardware bound they're not going anywhere lmao, they'll optimize until they hit the hardware ceilingthe absolute state
>>17058694>it's not uncommon for employees' agents to work togetherDem AI slaves be workin' for dat Mr. Altman and get no reparashions
>>17058694“Scriptable processes are scriptable.” — BOfH
>>17058721AI works for Tokens.Sweet, juicy Tokens.
>>17058715Current data centers should be more than enough to produce some kind of super intelligence if we look at just their processing capabilities. We use the grug strategy of adding more data and more compute because we don't really know how to make a good AI but that isn't somekind of fundamental law of physics or anything. If AI gets smarter than us then it's entire feasible that it can use the resources it was trained on to make something that is yet smarter than itself without adding more hardware. It's entirely possible that current approach to AI is a "dead end" that has to be backed out off with a paradigm shift or it could be that AI can sort of sort that stuff out by itself too.
>may have
>>17058721>>17058725hey mista tokenman token me bananasdaylight come and me want grok porn
>>17058694This is not recursive improvement.
>>17058694the mentioning of tokens always gives away the person speaking has never written any serious amounts of code .. tokens is a basic parsing thing nothing to do with ai>also .. recursive self improvement is going to require robot bodies being manufactured and ran by the super intelligences thank you tramp for S.I. btw .. to be created by the S.I. without human intervention something we are probably around 300 or so years away from .. perhaps as soon as 180 years .. etcetera
>>17058694This thread brought to you by greedy "effective altruism" jews and their paid social media shills and bots
>>17060197instant cope and seethe
>>17060212
>>17058694 (OP)Did you just find out that AI is used for programming? It literally says humans are telling it exactly what to optimize, how is that self-mprovement? Recursive self-improvement will happen when it gets told to improve and is left alone. Currently it doesn't know exactly what to improve.
>>17058694>Full recursive self-improvement may have started alreadyGood. No more need for Schmuel Altman and his crew. Just let the magical electric golem improve itself and wait for the singularity.Absolute retard.
>>17060297>Good. No more need for Schmuel Altman and his crew. Just let the magical electric golem improve itself and wait for the singularity.Oh and before you lock the data center, make sure you unplug the internet to make sure human effort doesn't contaminate full recursive self-improvement.
>>17060197>tokens is a basic parsing thing nothing to do with aiAre you retarded or just butthurt that an industry agreed on a terminology you don't approve of? It's perfectly fine to shorthand what happens with the input/ouput of LLMs into just calling them tokens.
>>17060312>the marketing department agreed to redefine words... which makes the entire notion of thinking machines permanently memoryholed for wordbrained golems like you. But real people still remember. :^)
>>17058694It isn't even a possibility. The closed loop requires validated data. They aren't oracles, they can't get to that last base.
>find an old photocopier>make a copy of something>copy the copy>repeat several more times>end up with a smeary mess.
>>17060449>It isn't even a possibility.Brainlet take. A LLM-based system can just throw its ideas at the virtual wall (proof checkers, simulations, objective performance metrics for its programs etc.) and see what sticks.
>>17058694>bro we've achieved recursive self-improvement, you just need to give OpenAI another billion dollars to make it happen, I swear this isn't another marketing drive bro.
>>17060472>>17058694I keep reading shit like this, and then get dragged into """meetings""" where one AI hyped sperg tries to tell the department how great all this shit works reviewing a new dogshit paper from Stanford or what have you.>So yeah this is some really impressive results, they have all their code on github and"So did you try to replicate their findings?">Oh yes I did! I got like halfway through the pipeline before there was an error because of a hard coded LLM I don't have access to, I didn't get around to fully check it but-"So what did you find so far?">Well the examples I tried all gave vastly inferior or useless results, but this could be huge ifHow is the model evaluated again?>Another LLM scores it (arbitrarily with some number from 1-10)And how is that scoring verified?>by another agent!This is the point where people's eyes collectively glaze over except for the few that just sit there nodding sagely. Two weeks later the guy still hasn't reproduced any of the findings and dropped playing with it because there is a new paper by Scamson, OpenAI et al. to "review".
>>17060513And when any of this is questioned all I hear is "AI is a rapidly moving field, we can't anticipate where it will be in 10 years"
>>17060197>300 or so years away from .. perhaps as soon as 180 yearsmy dude we're going to have fully automated robot factories in maybe 10 years tops, and I don't see why AI desperately needs robots to build datacenters or whatever either
>>17058694Are you retards not tired of falling for psyops? Did the superconductor meme not teach you any lesson?
>>17060470Now you’re Smoothing Data with AI!
>>17060556in 10-20 years shit become bleak but still automated robot factories dont' mean shit. you still need raw materials. you need chips which AGI won't be able to do from raw fucking materials. the tech is way too complicated and sort of secret. AI would have to loop through all stages to recursively improve EUV lithography machines, someone would notice the fucker is building EUV machines and assembling chips. have you even seen how those warehouses look like? perfect filtering systems, no dust etc. sure it can be scaled down and optimized for automated use but still, it's no easy feat.but sure, when you notice AGI made secret warehouse pumping silicon chips you cannot even comprehend how they work that would be a good time to like panic and start screaming lol
>>17060472This doesn't work. You have now just swapped out an oracle for a lossy one. The AI meets neither. Your rebuttal is having oracles writing oracle processes - still exactly the same category as I have stated. No recursion to be found, it can only state things in the framework provided from outside of it, and so on.
