←back to thread

Google's new AI mode is good, actually

(simonwillison.net)
129 points xnx | 1 comments | 07 Sep 25 14:43 UTC | HN request time: 0.202s | source
Show context
xnx ◴[07 Sep 25 15:58 UTC] No.45159332[source]▶
Interestingly, per the recent Google antitrust ruling documents, AI mode is extra fast because of a special FastSearch index: https://x.com/Marie_Haynes/status/1963031598829314161
replies(1): >>45159523 #
cj ◴[07 Sep 25 16:17 UTC] No.45159523[source]▶
Gemini in general is extremely fast, compared to ChatGPT 5 Thinking.

It also seems to excel at things ChatGPT 5 Thinking isn't good at. Simple things like "Here's a screenshot of text, please transcribe it" - ChatGPT 5 Thinking will spend 2 minutes and still get the results wrong, while Gemini Pro will spend 20-30 seconds and transcribe everything perfectly.

Obviously that's just 1 use case, but as someone who previously used ChatGPT exclusively, I'm increasingly impressed by Gemini the more I use it. Mainly due to the much faster thinking times that seem to provide equal or better results than GTP 5 Thinking.

replies(3): >>45162490 #>>45165112 #>>45171443 #
GlitchInstitute ◴[07 Sep 25 21:49 UTC] No.45162490[source]▶
Gemini is very fast because it runs on TPUsV7 mostly
replies(1): >>45165376 #
1. sigmoid10 ◴[08 Sep 25 07:03 UTC] No.45165376[source]▶
It is definitely because it's a smaller model. TPUv7 has ~10% lower flops at FP8 and 33% lower memory bandwidth than Nvidia Blackwell cards. Add CUDA to the comparison and they'll probably be even worse at real world utilization. Grok is already running on Blackwell cards and although there's little info on GPT5, I doubt they are behind.