@cobaltdigital33@marcelmarais@ThePrimeagen The issue is that it’ll optimise for anti-sycophantic behaviour then. But in a way it’s good because the backlash might get you to think harder. The kind of cross-questioning you get when you work in a Software company.
Hmm. I still think this is a rosy retrospection. For example, this is similar to how we remembered games from 2012 as ultra realistic, but when we play them again in 2026 we find them to have pretty bad graphics. What must have happened was some of you found the jump from 3.5 to 4o so high (along with couple of regressions in GPT 5) that it caused a distorted view of improvements post that.
Personality is subjective so not much comment on that. Also, I disagree with the explaining every little detail part as, I find myself explaining every little detail to current models way lesser than anything previously.
I would be more interested in a study where we bring back 4o and A/B test preference of 4o vs other models. Since lot of the claims are subjective this is the only way to definitely prove that 4o was more preferred to anything we have currently.
@ChShersh I’d say this is one of the biggest hurdles to cross for hard tech to improve as fast as software development . It’s a much needed breakthrough.
@theo@fogellxy Except you have 100s of engineers, managers, directors, and VPs working on fixing a Sev1 issue.
With this PTSD paging sound waking you up - https://t.co/LiXkeXXEdE
@stark4833@gdb I don’t really get what’s with this 4o movement. Now, I do agree they should probably bring it back to enlighten us on how much things have progressed ever since.
@ThePrimeagen I wonder what it would have looked like if deepseek had anthropic’s budget? Would they have still been hyper obsessed with efficiency or would be have AGI?