>AI trained on content from the internet which is increasingly AI generatedlol, lmao even
>>16999288How is that different that what universities do?>Uni teacher, trained by content taught at university
>>16999724people can correct for degeneration."AI" can't./thread
>>16999726No they cant, its well known that education quality is lowering each year, for years on end.
>>16999288Why is that AlphaGo can play chess against itself and learn but AI can't train against itself with art and learn?
Not a problem with the right importance weighting.
>humans are educated by books which are increasingly human-authoredlol, lmao even
They need to split the AI into groups so it can award noble prizes and direction and agreement.
>>16999288https://www.youtube.com/watch?v=QADKN3hantI
>>16999739As a result of destructive leftist policies, DESPITE people's natural tendency to improve
>>16999757she literally says that it basically FUCKING DOESNT. At least watch what you link
>>16999834who put those policies into place? animals?
>>16999871Umm... have you ever heard of these people called Jews?!
It's about "post-training" now - "RL environments" and expert human data. Yields impressive results, but as Sutskever said on that Lex replacement's podcast, something important is missing."It".
>>16999288The answer structures are trained by human beings working for companies which are proxies for the tech giants.
>>16999288Yes—this is how you go from AAI to AGI to ASI and finally to AZI and beyond.
>>16999739So what? That doesn't mean that humans CANNOT correct for degeneration. How do you think we advanced this far? Retard.
>>16999835Well maybe xhe should've made a less retarded click bait title
>>16999288Researchers figured out they can generate synthetic data and make AI better over time.This happened few years ago, you slept on the news
>>16999871Politically motivated sociopaths. They are the parasites to the host, not the main body.
>>17000017>youtuber>less clickbait titleheh
>>16999288The future of AI is in cultivating smaller and better data sets for more specialized content and feedback. You don’t try to solve a particle physics problem with ChatGPT or whatever, you have a first pass fine tuning model trained on broader physics stuff and a second pass trained on particle physics stuff and potentially a tertiary model trained on niche material immediately related to the problem you’re investigating
>>16999288Cybernetics from the start ought to get looked at as some discipline that were concerned on that context with these concerns the concerns about and to mechanism and control and the integuments to nature and also the concerns about harmonic analysis to data sets and causal inference and that were exactly the role that cybernetics were ersatz as that might seem ian banks only wrote one cybernetician on the first culture book consider phlebas which serves as some sort of deus ex machina role to that book but that cyberniticians themselves were sorts to analog computers like from those dune books during the butlerian jihads and computing being done by hand on some of those dune books because of the war at the ai and dune really does take some sort to hard stance against artificial intelligence and the danger about artificial intellgence those were some house corino intrigues kevin j saunderson brian herbert the atreides anyway dune there were that story about artificial technology taking over at some point and that’s the deep lore dune the earth gets like blown up and that’s one perspective on ai and artificial intelligence like just the butlerian jihad like get to the butlerian jihad mattresses during times to war with the ai but that’s an extreme frank herbert brian herbert vision and cybernetics does off an other vision and one that was not seperate from what frank hebrert happening during the green mind sci novel and the cybernetic idea were that analog computer that were to get completed or something like that
Half of the improvements in AI model quality are because the training data is being lorded over by, collectively, thousands of people if not more. Continuous improvements in the available datasets, which is something of a linear process. About 2.5 years ago training datasets were just an absolute mess because of the "oops we can't find any more real text and this new simulated data thing is only recent". They're still messy, but much improved.It's a lot of the reason today's smaller models (20-30B class) are roughly equivalent to the 70B models of 2-3 years ago. We have things like Gemma4 26B with 128 experts and only 8 active, and more than half of the active parameter per tokens is just the routing layer. Think about the math of that. It's effectively a collection of 1.6B models with the output quality of yesteryear's large models.The other half of it of course is improvements in the software structure, new methods, better number crunching to preserve quality, etc.There are some SOTA techniques and additions coming down the pipe that will make todays models look like toys by comparison. 2030 will be interesting times. A phone with 16GB of RAM could be outputting comparable results as today's big big corporate stuff much the same way today's 20-30B classes can be compared to shit like gpt4.Also congrats /sci/ 17 million posts of mostly pure fucking garbage.
>>16999724Is not different. In times of rampant corruption academia is run by psychos and parasites so it produces nothing of value. Eventually the system and its rotten institutions collapse.
>>17000134But GPT is literally the best model for solving physics problems.
