Don't ask AI about its new feature

I know that AI doesn’t know much about itself, but I can’t help asking an AI about its new feature. Today, I asked DeepSeek if it has a new memory feature, and it confidently hallucinated that it has a new DeepSeek++ Chrome plugin.

As it turned out, DeepSeek++ is actually a popular third party project. What’s hilarious is that humans never mix up a 3rd party feature as their own 1st party one.

This is quite interesting. Something must be very different between human memories and AI ones. Why do we never make such a mistake? This is probably because human memory has much wider and multi-dimensional context than AI’s. We remember when, where, with who, weather, what happened before/after, etc. about a memory.

Yes, AI’s memory is context-based. But the context is limited to its context size. Also, in its training phase, it’s given many different information one after next. For AIs, information’s importance is based on its frequency.

This must be one reason why AIs confidently make critical mistakes. DeepSeek consumed lots of information about its memory feature (or the lack of). When I casually asked, “Do you have a new memory feature?”, it generated a reply based on its knowledge. It was mostly correct. DeepSeek++ is a browser extension that enables memory and other features. But the answer was critically wrong in one important point. DeepSeek++ actually is not by DeepSeek team, but a 3rd party project.

Gemini also hallucinates about its features often. I speculate like this. As the Google product, Gemini is given its roadmap, alpha/beta features much more than other AIs. Those early information often times isn’t eventually correct. And corrupts Gemini’s knowledge.

This leads to an interesting conclusion: if you ask an AI’s new feature, ask other AIs. For example, Grok is good at gathering and validating latest information.