Anonymous
8/21/2025, 5:12:42 PM
No.106336831
>>106336600
Won't apply to chinese smuggled cards, you already see plenty on ebay and it's often shipping from China.
So what are anons opinions of the new 3.1?
I have a feeling that it suffers from a similar problem that new Sonnet/Opus 4 suffer from, excessive focus on code, to some detriment to writing.
I did some loli ERP with it earlier, 30 turns.
I haven't gotten any refusals, but it was hesitant to initiate , made me wonder if filtered dataset or just overfit on synth slop from OAI and Gemini API. Earlier DS V3 and R1 had some Gemini data in it most likely and this seem to have a lot more of it.
The writing wasn't bad, but it took longer to initiate or do some things, although eventually got the hint, in earlier turns it really felt a bit avoidant of some things, almost positivity biased, but this did go away in future turns.
Overall I've found R1 and previous DS3 more engaging and with a lot more fun creative replies.
I wouldn't say the replies here were bad.
As far as instruction following goes, it ignored some stuff in the system prompt, but I did an OOC about maintaining certain details of the chara's personality and it did maintain it for all future turns perfectly, so that's a bit weird, often it's the other way around, sys prompt obeyed, inline instructions ignored after some turns. Here it maintained my suggestion for many turns (all), but did worse with the system prompt. It also referenced stuff many turns back fairly flawlessly. Overall though, I think I prefer older R1 and V3's writing more, this feels and tastes too much like Gemini to me.
It's not clear they actually censored anything on purpose, but overfitting on gemini or OAI synth slop is obviously going to make things more cucked.
I guess this release was aimed at avoiding the need to serve 2 models on their API and for people that don't want to download a reasoner and a chat model (size), agent for Coding, regression for writing like Sonnnet 4 vs 3.5 relative to Sonnet 3.5+
Won't apply to chinese smuggled cards, you already see plenty on ebay and it's often shipping from China.
So what are anons opinions of the new 3.1?
I have a feeling that it suffers from a similar problem that new Sonnet/Opus 4 suffer from, excessive focus on code, to some detriment to writing.
I did some loli ERP with it earlier, 30 turns.
I haven't gotten any refusals, but it was hesitant to initiate , made me wonder if filtered dataset or just overfit on synth slop from OAI and Gemini API. Earlier DS V3 and R1 had some Gemini data in it most likely and this seem to have a lot more of it.
The writing wasn't bad, but it took longer to initiate or do some things, although eventually got the hint, in earlier turns it really felt a bit avoidant of some things, almost positivity biased, but this did go away in future turns.
Overall I've found R1 and previous DS3 more engaging and with a lot more fun creative replies.
I wouldn't say the replies here were bad.
As far as instruction following goes, it ignored some stuff in the system prompt, but I did an OOC about maintaining certain details of the chara's personality and it did maintain it for all future turns perfectly, so that's a bit weird, often it's the other way around, sys prompt obeyed, inline instructions ignored after some turns. Here it maintained my suggestion for many turns (all), but did worse with the system prompt. It also referenced stuff many turns back fairly flawlessly. Overall though, I think I prefer older R1 and V3's writing more, this feels and tastes too much like Gemini to me.
It's not clear they actually censored anything on purpose, but overfitting on gemini or OAI synth slop is obviously going to make things more cucked.
I guess this release was aimed at avoiding the need to serve 2 models on their API and for people that don't want to download a reasoner and a chat model (size), agent for Coding, regression for writing like Sonnnet 4 vs 3.5 relative to Sonnet 3.5+