Whoa, this is way too spot on! Former OpenAI key figure Karpathy just ripped off the veil of the 'trillion-parameter large models': turns out we’re all being held hostage by garbage data from the internet! In this latest deep-dive interview, Karpathy poured cold water on the current craze for scaling up model sizes. He pointed out that these massive models are actually wasting insane computational power on nothing more than 'rote memorization.' He boldly predicted that the future of AI, with a true 'cognitive core' capable of human-like logical thinking, won’t need to be bloated at all—around 1 billion parameters would be enough for it to run perfectly. When it encounters facts it doesn’t know, it’ll just call tools to retrieve them, instead of cramming the entire world’s knowledge into its brain. So, if that’s the case, why are big tech companies obsessed with making models so huge? Karpathy revealed a brutally harsh truth: because the foundational data used to train AI from the internet is just too trashy! Don’t think these large models are spending their days meticulously studying top-tier news articles. When researchers dig into the training datasets, they find them stuffed with endless random stock codes and useless cyber junk. To forcibly digest and compress this massive pile of garbage, tech giants have no choice but to scale up those terrifying parameter counts. To put it bluntly, the vast majority of these parameters are doing low-level 'storage' work, not high-level 'cognition and reasoning'!