Why hasn't an independently self improving AI been made yet? Is there a limitation that stops something like that from existing?
>>16999401Frontier LLMs are all Autoregressive Transformers, which are bottlenecked on training data and compute. If you want more data, you either need to capture real data (which is expensive) or generate new data with a zero-shot approach. Zero-shot performance of transformers is not very strong at the moment, and is mostly handled as a reinforcement learning process. The current frontier models are also not smart enough to find better model architectures, which is the only practical way to escape from the current self-improvement trap.
>>16999401the only limitation is anthropocentrism
>>16999544Bullshit, agents do work, enconter issues and try to solve them, this loop generates you infinit training data
>>16999567Agents sparsely generate novel content, and agentic content poisons training sets.
>>16999401because AI hasn't been made yet
>>16999544>The current frontier models are also not smart enough to find better model architectures, which is the only practical way to escape from the current self-improvement trap.Thats what i was guessing desu but hasnt AI been finding novel solutions to math and computing problems lately? Why can't they do it for this is it just too high IQ
>>16999401>Why hasn't an independently self improving AI been made yet? Is there a limitation that stops something like that from existing?There is no "AI". There are statistical models of preexisting human knowledge. The limitation of this is obvious: suppose you removed all mathematical knowledge from the training data. Now your token guesser is inherently unmathematical. It will never make the conceptual leap from having no math to having some math. It doesn't even have tokens for it. Getting past this problem is obviously going to require a conceptual leap beyond "intelligence" as statistical prediction and that's not going to come from "AI" for the aforementioned reason.
>>16999401no one's put enough gpus in one room yet
>>16999401You have human level intelligence, why don't you self improve?
AI model collapse is a degenerative feedback loop where an AI system trains on content generated by older AIs instead of fresh human data. The result? The AI goes mentally smooth-brain over time, losing its ability to generate diverse, creative, or accurate stuff, eventually turning into a babbling idiot.Think of it like making a photocopy of a photocopy of a photocopy. By the tenth copy, the details are completely washed out, everything looks like generic slop, and the model completely forgets how to make anything unique.Here is how it breaks down:The "Inbred" Internet: AI companies are scraping the web for training data, but now half the web is already AI-generated "slop". The bots are eating their own tails.The Loss of "Rare" Data: An AI only learns patterns. Every time a model copies another model, it focuses only on the most common, boring, mainstream answers. The rare, weird, or unique stuff (the things that make the data interesting) just gets deleted.The Doom Loop: The next generation of AI trains on that washed-out, narrowed data. It gets dumber, generates even blander content, and contaminates the web for the next model.Documented extensively in AI studies, it's a mathematically inevitable death spiral if bots only learn from other bots. To prevent this, AI devs are desperately racing to find human-only data or synthetic data verification systems so their models don't completely rot their own brains.
>>16999401Besides the real limitations listed so far, the other key issue is that no serious lab wants to give AIs that sort of blank check. It's fine enough to let them semi-autonomously explore open problems in a pipeline with some generous budget cap, but no one can afford to just hand over a full research environment to an agent swarm and pray for the best, training models is extremely expensive too. I don't believe the alarmism about muh superintelligence takeoff at this stage and neither do most labs, but wasting $ billions is a very concrete risk.That said, you can bet your ass the labs are using their AIs to explore and suggest improvements and research ideas.
>>16999401One limit that could get hypothesized were some large amount to noise would get generated as advancing onto the next level and part to that would get the managing to the various noise outputs that were feeding back against each at each level to processing alright there could get some claim about fast and slow processing was like fast and slow lanes that were traffic with noise folding and rounding the limits to both those fast and slow processing lanes which were some mind would be doing at the same time both fast and slowing processing together with that noise division running between the two lanes and i would think that generally the usefulness to artificial intelligence would be filtering through noise like whales filter baleen looking for krill algae plankton and like i would claim that those whales were running some krill eating algorithm and if those whales are some like artificial intelligence agent aspect moving through noise then that whale would also get criteria to those algorithms and some whole to the image and noise would arise out of that or something
