Search Results
7/12/2025, 2:12:25 AM
>>105875735
K2 also really, really likes dodging any requests it doesn't like (slurs, NSFW etc), similarly to nonnet, but unlike nonnet it won't listen to the instruction suppressing its own commentary. It's not like GPT 4.5 which writes anything if you JB it. K2 is pretty inhibited if under-jailbroken, it has a really persistent bias that is pozzing it to hell and back. If it likes the request though, it will be closer to v3 in creativity/plot incoherence.
>>105875802
>deepseek is pretty deterministic at 0.7 or lower
The model actually receives temp 0.2. In any case, K2 is the second model in existence after DS that is incoherent at temperatures normal for any other model. And it's really really similar in creative writing to DS. Hence the suspicion.
K2 also really, really likes dodging any requests it doesn't like (slurs, NSFW etc), similarly to nonnet, but unlike nonnet it won't listen to the instruction suppressing its own commentary. It's not like GPT 4.5 which writes anything if you JB it. K2 is pretty inhibited if under-jailbroken, it has a really persistent bias that is pozzing it to hell and back. If it likes the request though, it will be closer to v3 in creativity/plot incoherence.
>>105875802
>deepseek is pretty deterministic at 0.7 or lower
The model actually receives temp 0.2. In any case, K2 is the second model in existence after DS that is incoherent at temperatures normal for any other model. And it's really really similar in creative writing to DS. Hence the suspicion.
Page 1