We can’t even reproduce cat intelligence ~ Yann LeCun Thoughts??

The issue here is that we are comparing AI to human intelligence ( and so by extension mammal intelligence ). But mouse/cat/toddler intelligence has no value so we are building models in a manner very different from the way nature ‘builds’ organic intelligence. So the comparison makes little sense. AGI will arrive when models have agency and can take the initiative to solve tasks and their subtasks and at that point it won’t matter if the model can’t figure out how to hunt down a mouse.

Long term memory and continuous learning this two things once solved it's done.. Something I and @Shawnryan96 both think is necessary.

When your From The Field Bulk Catnip order comes with a free cat! 😹📦 Watch as @mcfloofypants_creations unboxes our catnip and gets an unexpected furry surprise! 🐱❤️ #FromTheField #BulkCatnip #CatnipSurprise #Unboxing #CatLovers

Just keeping it here my friend.

😂 How did you get this...

" We're missing something really big " Eh there are several things missing: Spatial intelligence is poor ability to navigate in 2d & 3d world is semi-incoherent, esp poor understanding of collisions & order of operations no-little sense of time Planning & reasoning esp for meaningful objectives over various timescales is not present every inference is effectively amnesic (anterograde amnesia) starting over but given the chat history text&images from it's previous self to try to infer things from it's previous messages --no continuous thoughts over time; heavily dependent on "reasoning" and "planning" which is helpful but is more a work around. Would be nice if we had memory informed inference/true memory; rather than purely frozen models. In short, we have a lot of things baked in from evolution & nat selection; pattern recognition, ability to understand local areas, navigation through space, and other intuitions, understanding object permanence, & agentic abilities. While ai likely needs at least a curriculum of training and progressively building those skills, abilities, perceptions in. and iterative skills of those skill, and so on. :x RN they're only trained for chat turned based often one shot responses --- chat bots evals. Prob need to train them for many many tasks from computer use, to various tasks, various 2d games (make plans, move character or cursor effectively, use those plans in a coherent meaningful way. ect and likely some arch & representational issues to be hammered out too behind the scenes. Lets see where agents are at by the end of the year, and end of next year. hmm 🤔🧐 ,.·*¯`·.,¸,.·*¯[;::(.◕‿‿◕) Nyan ;3 Cat intelligence nears

probably true, but we’re not trying to reproduce cat intelligence. We’re trying to produce things that are generally capable across domains that are economically useful. And that is working incredibly well.

Exactly and this will improve.

something big we're missing is that representations in biological brains are dynamic @ylecun . it is more like steered randomness/construction than true recall/recognition from episodes. current ML only tries to reproduce, not recreate actively.

◕ω◕ cat intelligence is coming

The robots are coming, and I’m pretty sure a cat can’t write code. There isn’t 1:1 overlap, but I don’t see why there won’t be eventually, and the parts we really have to worry about are definitely happening.

we don't need cat intelligence , we need something more and we will get there.