I agree. It's much worse than the cheap Chinese models. They appear to heavily bias parametric knowledge and discourage tool use. That's fine for things like "how do I perform CPR?" but worse than useless for any kind of research. [There is one benchmark showing far lower rates of hallucination, so let's see how accurate this is.](https://www.reddit.com/r/singularity/comments/1wuj72j/gemini...)