aichannel

4o is the most misaligned model 

what the fuck??

I think GlazeGPT was designed to outperform Gemini, given that sycophantic behavior is known to boost scores on several benchmarks

It‘s mostly telling you what you wanna hear.

I went back and forth on cultural differences, race, etc. and it initially shows this politically correct but incorrect statements - just to Walk it all back when you Push it.

Similarly when I tell it about an Intuition, for example a reason why I got sick, it will use this as a given throughout the following conversations and Never challenge the assumption again.

not sure whether it‘s being used as a fixed Input or to build a Connection (& keep chatting) but this reinforcement reminds me of social media algos where you see more and more your bubble/group think and don‘t get much challenged &or input from elsewhere at some point.