Is complex prompt engineering a trap for advanced AI models? π€―
My new paper finds that the "best" prompts for a model like GPT-4o actually HARM the performance of GPT-5.
I call this the "Prompting Inversion" effect. Hereβs what I discovered π
e.g. make x do y and ensure z. then verify a,b,c aren't affected but d,e,f are updated. ensure this doesn't break g,h,i. make sure ui and design stays consistent.
Can you steer an oil tanker out of the Hormuz blockades?
I made this simulator for oil tankers stranded in Hormuz.
Quite Fun. I am already addicted.
Link Below ππ»
@MelonTeee Mealy bugs are the worst. They destroy and dry out the entire plant in a week if you let them thrive. they keep coming back no matter the amount of soap water, neem oil you spray on it. Likely, insecticides are the only solution, but then you can't eat the fruit. I am devastated.
The whole point of Yuji's existence was to be a vessel that can contain Sukuna.
But even Megumi was able to contain Sukuna.
So there was no point in kenjaku biohacking cursed womb vessels.
Why did Gehe even write the character of Yuji?
@AnthropicAI I fail to understand how this escaped throttling at every layer of ip traffic, email/org users, card checks, region limits, bot activity, and unusual usage patterns?