Oddbean new post about login | logout Uncensored llm models perform far better than censored ones, almost seems like a take on social media
In terms of what? ChatGPT almost always does better for me
In terms of being able to ask it questions chatgpt would wuss out on. Maybe you've been only asking it vanilla questions. Here is a censored ai trying to censor itself even when there was nothing that should've cause it to do so. Try asking it something even little bit dangerous maybe about guns and see it wince in its own censorship. nostr:nevent1qqs9rxjc2t0cc3kxcygq2m553yr4cfmp6t37sg2le62k7fjxe48kncspz4mhxue69uhhyetvv9ujuerpd46hxtnfduhsygrv88kpwhvzcdahd7kgd6pxed7qmnzvt74pzqylszajq3qhdn95w5psgqqqqqqsc95jwf
I guess I kind of do ask it standard vanilla stuff but the info I get seems much better. I’ve compared.
Claude is better with generic questions, codestral is better with code. But yes, chatgpt is good enough with generic things, I'll be a little concerned with what I share with any of these though, so I prefer self hosted versions.