Creator @iruletheworldmo countered that the output quality remains strong
you have to spoonfeed them and need to be extremely precise
otherwise they will usually make stupid decisions
after using GPT-5.5-xhigh for the past week for my research project I'm much less bullish on RSI
models are not opinionated and have 0 taste, like they just return training eigenslop
Negative users call GPT-5.5 unusable and nerfed for research due to lacking taste and worse instruction following, while positive users praise its implementation strengths after iterations and view timeline delays as minor.
the Pro models also aren't what they used to be
they think for 5 minutes and call it a day
after using GPT-5.5-xhigh for the past week for my research project I'm much less bullish on RSI
models are not opinionated and have 0 taste, like they just return training eigenslop
did they nerf GPT-5.5?
after using GPT-5.5-xhigh for the past week for my research project I'm much less bullish on RSI
models are not opinionated and have 0 taste, like they just return training eigenslop
let's wait and see how Mythos and GPT-6 fare
but I think that experience alone lengthens my timelines for RSI by a year, at least for machine learning / CS
math is probably a bit easier
after using GPT-5.5-xhigh for the past week for my research project I'm much less bullish on RSI
models are not opinionated and have 0 taste, like they just return training eigenslop
still good for implementing stuff
after using GPT-5.5-xhigh for the past week for my research project I'm much less bullish on RSI
models are not opinionated and have 0 taste, like they just return training eigenslop
All slop is uniquely decomposable into eigenslop.
after using GPT-5.5-xhigh for the past week for my research project I'm much less bullish on RSI
models are not opinionated and have 0 taste, like they just return training eigenslop
@scaling01 Nothing really surprising there.
after using GPT-5.5-xhigh for the past week for my research project I'm much less bullish on RSI
models are not opinionated and have 0 taste, like they just return training eigenslop
@scaling01 the output feels as good though tbh.
@scaling01 in cog sci/graph related work
it keeps ignoring my instructions for dynamic systems i ask for
and just adds hardcoded/deterministic values.
main reason i‘ve been using claude for such work nowadays.
@scaling01 5.4pro is significantly better than 5.5pro
@leothecurious it's not that bad
we just need to scale RL
taste isn't magic
@scaling01 It's really anoying too, because they are very good at lying to you so the stupid decisions pile on extremely fast if you aren't watching outputs like a hawk.
@scaling01 I think this is mainly an issue with OpenAIs design philosophy.
GPT is trained to be a tool. It's not supposed to be have good, independent ideas. It's supposed to do exactly what you tell it to.
We don't have to design them that way, and not all labs do.
@scaling01 damn
@rinsetsuyua @scaling01 Except with Claude you run out of credits in 10 minutes
@scaling01 Seems nerfed, instruction following looks worse, my agents ignoring instructions even on high
@scaling01 are they being told to think less to conserve compute?
@scaling01 "opinionated with taste" is gonna be the next buzzword benchmark mark my words
@scaling01 Don't think so
Haven't seen any difference
@scaling01 Ehh in my experience it can think for 3 minutes at least but for some prompts it still genuinely thinks for 30 plus minutes. The longest I’ve ever experienced was for 5.2 pro and it thought for 9 hours on three different prompts spending about 3 hours on each.
@scaling01 Yeah exactly! GPT 5.5 Pro used to take 20-30 min to think, now it's done in a few mins. They severely lowered the juice
Creator @iruletheworldmo countered that the output quality remains strong
you have to spoonfeed them and need to be extremely precise
otherwise they will usually make stupid decisions
after using GPT-5.5-xhigh for the past week for my research project I'm much less bullish on RSI
models are not opinionated and have 0 taste, like they just return training eigenslop