What I find equally interesting is finding out how well an AI model knows an open source software library you’ve built.
E.g I’m quite pleased to see most frontier-level models are very competent setting up agent systems using the Langroid framework, in development since Apr 2023:
I find it astonishing that people didn't turn off these memory systems the moment they were introduced. Half the point of these things is that I control exactly what information is introduced in each query.
Sometimes Claude randomly references my geolocation to show off and it's incredibly creepy.
That’s actually one of the use cases I think Corpus could exist for.
You could turn native memory off in Claude/ChatGPT entirely and use Corpus as the only persistent memory. Save something there when you explicitly want it remembered, then let either AI pull it back later.
Basically: the AI doesn’t remember you. Corpus does.
Yep, geolocation is so creepy. And claude code will read your personal data, rc, .env, other projects files if you forgot to restrict it. So I have global settings for that.
I feel very sad. My AI knows essentially nothing about be. The only thing it said it knew about me was my location, which it got wrong. It completely missed the books, magazine articles, RFCs and web pages I've authored. I did ask it about a very old project I wrote as a test of the github integration and it threw substantial shade that I was using 2010 era idiomatic javascript on a project that was written in 2010 and suggested I should have used ECMAScript modules that were introduced in 2015.
I am constantly amazed by what AI gets right... but sadly... equally amazed by what I can't seem to reason well about.
Just to throw a little shade back at Claude... OpenCyc would never have made that mistake.
Are people logging in and volunteering their identity with every query?
Even if you’re not paranoid about tracking and retargeting, I find the experience is actually better when it knows less about the user.
Same is true of YouTube’s algorithm - you can get to a point with an account you can’t find anything. It’s the same dozen or so videos it seems like.
And then you log out and it’s a huge internet again
Lumo, Proton's AI
E.g I’m quite pleased to see most frontier-level models are very competent setting up agent systems using the Langroid framework, in development since Apr 2023:
https://github.com/langroid/langroid
Sometimes Claude randomly references my geolocation to show off and it's incredibly creepy.
You could turn native memory off in Claude/ChatGPT entirely and use Corpus as the only persistent memory. Save something there when you explicitly want it remembered, then let either AI pull it back later.
Basically: the AI doesn’t remember you. Corpus does.
I am constantly amazed by what AI gets right... but sadly... equally amazed by what I can't seem to reason well about.
Just to throw a little shade back at Claude... OpenCyc would never have made that mistake.