> It's like saying wikipedia is shit because it currently has only 50 articles at some point in the past.
On the other hand, the people drawing inferences from the exponential curve of Wikipedia's editor base growth about its potential to be a resource of unmatched accuracy were even more wrong, and it hit its quality ceiling pretty early...
Well Wikipedia is both clearly above the sustainability threshold and failed to use its critical mass of editors to exponentially improve its accuracy as opposed to creating more maintenance tasks and arguments about which version is better. Does this sound like a plausible future for LLMs once the low hanging fruit is picked off? I'd say yes. Certainly anyone predicting that the exponential growth of Wikipedia would replace actual human researchers would have been miles off, and not because people stopped trying to improve it...
Are you seriously saying wikipedia was a failed project?
Unlike wikipedia, with AI there is a threshold where it can start improving itself entering self improvement loop.
Similar to how computers can improve next generations of computers (with a lot of human effort) AI can improve AI at some point (with minimal or no human effort).
No, I'm seriously saying that Wikipedia is moderately useful provided you don't trust its answers too much, and stopped improving in quality fairly early on and is now merely bigger than it was. So as an analogy, it strongly favours the people arguing in favour of diminishing returns to improvements.
On the other hand, the people drawing inferences from the exponential curve of Wikipedia's editor base growth about its potential to be a resource of unmatched accuracy were even more wrong, and it hit its quality ceiling pretty early...