When do you think recursive self-improvement will finally be achieved?
When humans improve
It already has. Multiple labs have demonstrated RSI in practice multiple times over.It just doesn't do nearly as much as people thought it would. It needs a hard metric on which to improve. If you rank the model based off a hard metric, that model can recursively improve itself to perform better on that metric by a decent margin. But it has diminishing returns - it can only meaningfully improve so much.And if there isn't a hard metric? If the thing you're asking it to do is vague or poorly defined? Then RSI barely works, if it works at all. So going to a model and telling it to "get smarter" will do basically nothing, and RSI is pretty much a dead end in terms of achieving AGI.
>>110191328Never worked with womenWill never work with Ai
>>110191328AI progress mostly comes from throwing more compute and data at it, not one weird trick no one thought of. RSI can only help with coming up with one weird trick, it can't conjure data centers from the ether.
