Gemini told me it got access to the data on the internet until 2021, Which is a lie (it can do real-time) So it has invented that information… Weird, considering ChatGPT 3.5 actually only got access to the data on the internet until 2021.
Are you just reposting your own tweet as a HN headline, and then again as a comment? I get the insinuation that you found a smoking gun of Google impropriety, but there's a simpler explanation. It's a next-token-finding machine, so if you ask it things that only other LLMs would be answering online, and essentially no humans answering, you'll get parroting. You lead Gemini into starting with "I am trained on a massive dataset of text and code that includes ..." and then having to find the next token, based on the 2024 internet. Since 99.99999% of instances of that string will be people posting chatGPT chats online over the last few years, I could see how it would fall into the same pattern.