3 ms·
I’ve seen degradation in the app and via the API, so if I had to bet, they’ve probably kneecapped the model so that it works passably everywhere they’ve been ma
by bugglebeetle 3y ago
I’ve seen degradation in the app and via the API, so if I had to bet, they’ve probably kneecapped the model so that it works passably everywhere they’ve been made it available vs. works well in one place or another.
- TeMPOraL 3y agoYes. I think 'sirsinsalot is likely right in suggesting[0] that they could be trying "to hoard the capability to out compete any competitor, of any kind, commercially or politically and hide the true extent of your capability to avoid scrutiny and legislation", and that they're currently "dialing back the public expectations", possibly while "deploying the capability in a novel way to exploit it as the largest lever" they can. That view is consistent with GPT-4 getting dumber on both OpenAI proper and Azure OpenAI - even as the companies and corporations using the latter are paying through the nose for the privilege. Alternative take is that they're doing it to slow the development of the whole field down, per all the AI safety letters and manifestos that they've been signing and circulating - but that would be at best a stop-gap before OSS models catch up, and it's more than likely that OpenAI and/or Microsoft would succumb to the temptation of doing what 'sirsinsalot suggested anyway. -- [0] - https://news.ycombinator.com/item?id=36135425 https://news.ycombinator.com/item?id=36135425
- cma 3y agoIf it got faster at the same time it could just be bait and switch with a quantized/sparsified replacement.