Interesting to see OpenAI learn that humans are strange. When an errant update a few months back made ChatGPT openly try to deify the user, they've now dialled that way back and SURPRISE! Now the people that liked the sycophant style are furious.
They're in such an interesting situation. They've been openly mocked for having so many terribly named models that you could select in the dropdown, so they solved it with GPT5 which decides what to use for you... but people also like control and those who learned how to use those terribly named models are freaking out.
I'm more than a little surprised they didn't directly train it on the strawberry question. It's a non-issue unless your job is to legitimately count letters in words, which would be... weird. But it also stands as this direct method for people to lose the plot. Maybe they just don't care about such attacks?
I haven't seen the GPT5 issues many are upset about. It's been quite a bit more to-the-point and I haven't caught it in a hallucination yet. I just had it grab me the emails of 10 people I'm supposed to reach out to today, and not only did it nail 9/10, but the one it missed, it flagged for me that it was taking a stab at it based on the company's general naming convention and suggested I dig further myself on that one.
So just by virtue of hallucinating less and maybe being able to question if it knows something for sure, it's a huge upgrade.
4o was touted as an incredible creative writing model and a lot of people who use it seem upset... but the tests I've run are actually really good. Maybe I'm just blinded by optimism, but I've had it make a few things and they've all been as good or better than 4o.
|