It seems like Grok is only good for images, and that’s it. With everything else, Grok is stupid. For example, if you roleplay Superman and Lois with Grok, and you write that Jon becomes a superhero at 21 and Clark and Lois find out at 23, Grok will write Clark and Lois as super predators, harassers, and borderline rapists. It has them try to ‘forbid’ Jon from being a superhero, despite the fact that Jon is an adult. And if Jon files a restraining order against Clark and Lois, never sees them again, tells them to ‘fuck off,’ and refuses to visit Lois while she has cancer, Grok really is dumb enough to believe Jon is wrong. Is Grok the dumbest AI ever?
LLMs do not ‘believe’ anything. They are very big, very expensive autocomplete. They don’t know any information. They don’t think about your prompts and propose an answer. They are a system that takes what you put in, assumes what you have entered is the first X words of a document that already exists, and then makes a probabilistic prediction what the next word will be.
No thought. Only probability.
Is this what people use LLMs for?
I just had a friend overshare he uses LLMs to play out his nastiest sexual fantasies. Hillarious and disturbing. This seems tame by comparison.
Does Jon change his name to “Carrington”?
Grok sounds brilliant compared with Siri, which can’t do basic math without going online.
Siri used to bust your balls with Sesame Street jokes if you asked it to divide by zero. Now it needs to go online or connect to ChatGPT.
Grok is dumb because it is one of the least efficient models out there. If you’re just gonna do pointless role-playing use a smaller model that is fine tuned for role playing
