Hacker News Reader

Based on what do you think it's comparable to GPT-3.5 and not to 4? Did we see a lot of public performance?

They claim it is already being used in Bard, also if you read the paper it does much worse at the important benchmarks.