It's shockingly bad because of how they want it to behave. They don't want it to bullshit, they don't want it to say rowdy things, and they want it stay "on bra

It's shockingly bad because of how they want it to behave. They don't want it to bullshit, they don't want it to say rowdy things, and they want it stay "on brand". Well too bad. These things are all modeled after human beings... they're literally called large language models. Human beings bullshit, they say rowdy things, and they don't always stay on brand. The only way to have an exceptional AI is you actually "raise" (train) it like you would a normal human being. That's being learned with the quality of Deepseek's models, although I think they're lying through their teeth about how much it cost to train it. It's in China's interest to bullshit on that metric, so they're bullshitting.