Agentic systems must be able to predict the consequences of its own actions to effectively plan and operate in the physical world.
>>108849531>by 2025>by 2026>by 2027>...>by 2100
What will be obvious
>>108849602perpetual >>>/v/ irginity
>>108849531>am le cuntkeklol
>>108849531LeCunt is delusional, machine learning will never "understand" anything, it cannot ever have a real world model.He only happens to be a little less delusional than the average aitard which doesn't even acknowledge the unresolvable problems of machine learning (just add more data bro! just add more compute bro! it will fix itself!), but still delusional enough to deny that the whole foundations are flawed.
>>108850345I hear you but that doesn't solve real world problems. Right now we just have to keep iterating and doing more context engineering
>>108850370Machine learning can solve a lot of real world problems even without being perfect or having a world model, but not agents, those need a different approach.To me it's crazy that there are people right now that have their whole life connected to these unreliable LLM "agents", it's a disaster in the making.
>>108850345What are the unsolvable problem in machine learning?
>>108850687See Pearl’s causal hierarchy.Machine learning is all about making associations, but for agents the AI would need to understand interventions (which means understanding causation) and counterfactuals (which means that based on the cause and effect it should reason back from the possible outcomes), this is also what LeCunt is saying btw, but his solution is still more machine learning just in a more abstract "representation space" instead of token space.And I tell you machine learning can't do it, it can only solve those problems where associations is all you need because at its core it's close to statistics, it's same problem as "correlation does not imply causation".(note that what LLM people call "reasoning" is not actual reasoning but just a mimicry cause it's trained on data (text) of humans reasoning)
>>108850894Interesting, for me the real unsolved problem is continuous learning.>"correlation does not imply causation".I've been doing some experiments by teach various networks to play a simple competitive multiplayer game and I see this in action all the time. The "players" do these actions with out rhyme or reason and some times they work (kill an opponent), I feel like they randomly do this "behavior" without know why and just by pure luck it results in a positive out come. I really want to see some real competitive agency and long term planning and strategizing.
>>108851293I hope you're not messing up my Dota games...
>>108851307Don't worry they are for my own games.
>>108849531>Agentic systems mustpeople can't do this and yet we're doing fine
>>108849602that AI was a bubble
>>108849602That vague posting is... Well you know
>>108851081You will never take my precious clankers from me.
>>108850894an effort post? without obvious slop signs? i must be dreaming
>>108853644Why? This isn't red dit.
