Sonnet (specifically sonnet 5) is absolutely retarded as fuck now. I looked into it. What they have done is they heavily penalized it during fine-tuning and rhlf to not make overconfident statements and not trust its own weights (they did this because midwits complained about "hallucinations" too much). So now instead it sits there on google for 10 minutes of its "thinking" time and falls back to sentences it reads from "credible sources" and refers to the sources as absolute reality. If you don't have search on or are using the api without search, then it just tries to route around giving opinions that don't cite other people directly.
Will gemini end up reigning supreme in the long run? Relying on searching the web for every prompt is not sustainable this shit should be loaded into the pretrains and I can see goog being the only company with the capability
Idk I trust the weights of a strong model more than the human generated slop it’s citing. And to the extent that I don’t it’s because they had human “experts” rhlf the shit out of it to confirm various human generated hallucinations and current historical biases