> No, we haven't made GPT-4 dumber. Quite the opposite: we make each new version smarter than the previous one. > Current hypothesis: When you use it more heavily, you start noticing issues you didn't see before.
This latest model we trained on X% more data by integrating novel training data.
We now see user conversion funnels working better.
User feedback in closed betas indicate that 89.6% feel like answers are "generally correct most of the time".
Etc etc.
For example if their definition of "smart" means refusing to answer questions on delicate or specialized topics, then in some way they've made the model smarter but in ways they made it dumber.
It's obvious that this is damage control from OpenAI whenever their snake-oil product regularly starts malfunctioning and hallucinating. They have to bring up more excuses to plug the holes leaking from their black-box AI model.
The fact is, people are beginning to learn their limitations, guardrails, etc and are not blindly trusting whatever these LLMs are outputting as they are known to produce nonsense which they have to always check for.