[bevekspldnw]

That’s a pretty disingenuous framing. Obviously the content of the communication- what you ask the chatbot - can totally be personally identifiable, and is stored.

You aren’t providing communications content privacy, you’re providing a meta data proxy which is not remotely the same.

[seplox]

Or in other words:

If you submit personal information in your Prompts, it may be reproduced in the Outputs, but no one can tell whether it was you personally submitting the Prompts or someone else.

[bevekspldnw]

This was the same premise when AOL released an “anonymized” set of search data and it was easy to find the original people.

https://en.wikipedia.org/wiki/AOL_search_log_release

[lxgr]

True, but it's also the best possible thing until they can run an LLM in your browser (or maybe we figure out how to do homomorphic encryption of neural networks).

What else should they call it? I suppose "anonymous" could work (i.e. they only know about you what you tell them).

[bevekspldnw]

They could literally run Mixtral or Llama on their own machines and have total control. This is basically a ChatGPT VPN.

[lxgr]

What would make their machines more trustworthy than OpenAI's or Anthropic's?

[dumbfounder]

Them saying they will not use our data or allow 3rd parties to use our data. This effort is weak sauce.

[bevekspldnw]

Correct, that would be full communications content protection. This is not that, this is a pretty half assed product with a misleading set of promises.

Either use one of the many open models that beat GPT-3.5 and you have a real, and better!, product.

I literally have no idea what the point of this is, or why not make the actual step of rolling out a real private LLM. It’s literally inferior to what Mixtral provides right now, under Apache license.