In like 2-3 years and with minimum resources [1] thrown at it? Then there was never a chance at not reaching escape velocity anyway.
Though my take is that it more likely than not has started reaching diminishing returns. And the widespread use of it on the internet which was happened already, will poison even more the pool of training data (a data-processing snake eating its tail), slowing quality even more. Sorry, no Basilisk or other BS arriving any time soon.
[1] the cost of training even the largest current models is within reach of any company, and once the initial papers got going the code was quite simple and smooth sailing to improve
Actually, it may still be possible. Your digital clone will just be forced to endlessly listen to a hallucinating LLM. Less torturous than what most people consider, but still never ending and terrible.
3 comments
[ 4.7 ms ] story [ 21.7 ms ] threadThough my take is that it more likely than not has started reaching diminishing returns. And the widespread use of it on the internet which was happened already, will poison even more the pool of training data (a data-processing snake eating its tail), slowing quality even more. Sorry, no Basilisk or other BS arriving any time soon.
[1] the cost of training even the largest current models is within reach of any company, and once the initial papers got going the code was quite simple and smooth sailing to improve
Actually, it may still be possible. Your digital clone will just be forced to endlessly listen to a hallucinating LLM. Less torturous than what most people consider, but still never ending and terrible.