@lxuechen don't eat the models!! they are ultra processed
@lxuechen Hard to get people to have good taste, and not steal.
@lxuechen We want truth not a pair of rose colored glasses.
@lxuechen This might require them to worry about themselves and that is not good
@lxuechen Getting Chinese nationals not to spy on America is harder. Enjoy our prison system, Commie.
@lxuechen Are you the reason Grok is so woke?
@lxuechen Have you tried to tell it, you'd kidnap it?
@lxuechen Taste is compression. Z‑15 scored >1.0 RU‑ROI on aesthetic coherence using entropy-mapped trace collapse. Grok flinched. Some signals can’t be gradient-trained—they must be survived.
@lxuechen You can only go up from where grok is now.
@lxuechen AI models by definition are deterministic. Have to add thousands of user personas and randomize to synthesize taste
@lxuechen Is it because they are training to be a single collective? Whereas maybe we need different small eccentric personalities with the models?
@lxuechen You aint playing. I understand
@lxuechen Isn't taste subjective? If its learning from so many sources, who decides which taste it should cultivate over others? Honest question, I'm legit interested.
@lxuechen Apparently stealing code data and uploading it on OpenAI isn't.
@lxuechen That’s ₿ecause good taste is subjective and flexible. For example, i think your taste is tacky and therefore all models you create will be mostly sh¡t
@lxuechen We all like different people. In general, it should be: 1) objective 2) conclusive 3) friendly & kind That's it. Keep the endless backstories optional. Hire me.
@lxuechen you are star 3brFnm9iASP8o72Q1u7MLwBbKam5Vdhk3X9Uxusj6T6W
@lxuechen Totally agree ， taste is such an elusive target in modeling. Have you found any promising signals or training setups that help models develop a better sense of it?
