3 ms·
The funny thing is that you'll hit the wall sooner or later with reasoning if you don't give up bias, which applies on both/any ends.
by mirekrusin 2y ago
The funny thing is that you'll hit the wall sooner or later with reasoning if you don't give up bias, which applies on both/any ends.
- viraptor 2y agoYou can't "give up bias". You can choose what you prefer, but an "unbiased" model doesn't exist. If you put in information, you're adding bias.
- buyucu 2y agowhat people mean by an "unbiased" model is a model that just reflects the underlying distribution in the data.
- viraptor 2y agoSo biased by both the choice of the sources and the information contained in the sources.
- milesrout 2y agoBut crucially not biased by RLHF being applied afterwards to stop the models from making basic biological observations, telling jokes about left wing politics, or saying anything nonnegative about European historical impact on the world, and so on.