Hacker News Reader

Comment by qntmfred | Hacker News Reader

I told ChatGPT my mom's name the day my account got persistent memory, last april. I also told it to figure out my dad's name. Once a month or so I would ask it my mom and dad's name. By november it had figured out my dad's name.

https://x.com/kenwarner/status/1859637611595214888

View on HN

Unfortunately I feel like the fact that your dad's name is the same as yours somewhat diminishes that accomplishment.

I think that is the accomplishment. It progressed from not being able to give an answer because it did not have the direct knowledge to being able to make a guess based on a pattern of naming of others in the family and a clue.

I asked Gemini and Gemini thinks even knowing Ken's uncle is a junior Ken's father is more likely to be named "John" or "James".

If Gemini is correct ChatGPT is dumb and simply got lucky.

Could absolutely be that. Or it is so smart that it realizes that the author believes they have given enough information and that it should not have to land on a low chance guess. So that pattern is the only one that make sense in that case.

Maybe unlikely that is that smart though

What the author provided is not necessarily the same as what the software forwarded to the model, especially if some sort of "recall" feature is being used.

our middle names are James.

Fair concern, but:

1. This is via Gemini, and I had been using Gemini for only a few days prior to this experiment. I assure you I never mentioned anyone's name to it before this.

2. This was via API. Whether Gemini or OpenAI, they do not train/retain this kind of memory if accessed via API. 99% of my interaction with LLMs is via APIs. I don't think I've ever used the ChatGPT interface.

Is the tweet saying that you also told it your name and then it guessed that your Dad's name was the same as yours?

correct

One important note is that chatgpt has a memory you cannot see, besides chat history and besides memory. You cannot purge or manage this memory. I don't yet know how long it lasts. I don't know if it's some form of cached recent interaction or is a hidden permanent memory.

This is not true. How do you come to this conclusion?

By specifically testing it. I even made an extra account to get a clean state. You can check its memory interface and find nothing, delete all chats. It will still remember it. If you delete that and start a new thread, it may even mention the fact then say it forgot it at the user's request.

You can't tell me "that's not true". If my account's memory is empty and I've deleted all chats and it still remembers things there is some hidden form of memory. It may be intentional. It may not. But it's retaining information in a way that users can't manage.

Sorry, but I refuse to believe you until you provide proof. What exactly did it remember? I think you are misreading the hallucinations here.

If this was true, there might even be laws here in Europe that they are breaking.

> there might even be laws here in Europe that they are breaking

You're telling me an american technology corporation might have violated european laws? i can't imagine such a thing happening...

No - but a random hacker news commenter wouldn't be the only one noticing this.

I am not sure how I'd provide proof. But I'd encourage you to test it. It's always possible it's a bug. You can check with something like telling it that your real name is something very identifiable/not a typical name and working from there.

No the name of the user is part of the system prompt. How would you think this works? You can get ChatGPT to tell you all it knows about you, which is more than memories but never anything out of old conversations

Not the name of the user, a separate name. I'm aware of this.

That is also not true, it can access old conversations, this is a known feature. I have been able to have it access back to the beginning of my using the site.

Wrong approach in this day and age. Data is big business. Snowden revs already over a decade old. Today: https://news.ycombinator.com/item?id=44062586

Oh, and soft deletion is a common pattern. Prove a tech company is not hoarding data—is the useful hypothesis for the last decade.

AFAIK it will have f e access to your account, browser info and f e location information. Just from that it can figure stuff out. Some guy tested that when they asked to locate a photo.

No I'm talking about specific information not related to that. You're right that it has access to that sort of rough information.

Reply on news.ycombinator.com