Hacker News Reader

Google's AI overviews is the single worst AI-driven experience in widespread use today. It's a mistake to make conclusions about how good AI has got based on that.

Have you tried search in ChatGPT with o4-mini or o3?

View on HN

I don't use it that much, but I have noticed these AI overviews still seem hallucinate a lot, compared to others. Meanwhile I hear that Gemini is catching up or surpassing other models, so I wonder if I'm just unlucky (or just haven't used it enough to see how much better it is)

One is their state-of-the-art model, the other one's the best model they can run at scale and speed people expect from a search engine.

That sounds reasonable, and I don't doubt that it's part of the reason. Still, iiuc the solution to hallucination is that they can essentially train the model to recognize when it "doesn't know", and to say so in that case rather than just puke back the highest probability BS. I.e. it's a training time factor, not inference time, so it's not a fundamental cost issue, but more about priorities.

Reply on news.ycombinator.com