Gemini lying about having data until 2021 (opens in new tab)

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.

This is probably a result of Gemini having ingested plenty of ChatGPT conversations that include ChatGPT telling the user that it only has access to information up to 2021.

This is not a new problem, either. Once, I had a model (I think it was Llama 2 7B) tell me that its name was Bard. Yeah, right...

Thank you for copying the twitter post text, although I cannot even parse what it is supposed to mean.

I would suggest that it would be polite that when someone decides to post some word vomit from twitter, they should also include a translation in the comments.

It's Google's LLM that powered Bard (Since renamed to Gemini)

Whoops, forgot to message you to ask if you knew before posting.