Egregoros

I wanted raw inference API access. A curl script and nvim wasn't enough, too much hassle just to see the <think> tokens and being able to edit them

If you use a UI over these, you cannot edit the reasoning. It says "not supported". Why is it not supported? Who knows. It's just a fucking string dude.

Anyway. I explained this in detail to the clanker and told it to implement the v1/completions API. Foolish as I am, I forgot that the geniuses at openAI named the successor API v1/chat/completions.
So of course it got confused

I told the AI multiple times that I want a fucking string buffer on the screen and just to append to that, and in the background, unbeknownst to me, it built a fucking dictionary, a parser, multiple roles. All that to just fucking concat it to a single string in the end, making the whole thing pointless.

This balooned the code up to 10x the size it had to be though. What a waste of time

Watched this today, or, I listened to it while I was working, but I caught most of what he was saying. I've seen a few other videos he's made.

It got me to thinking about something. I have several old computers that I run just to play around with. Some are pre-Windows DOS machines that I play with to see how much modern-ish stuff I can do on them.

You remember how they trained that AI last year to convert images into the studio ghibli style? I wonder if you could train one to convert websites into a stripped down version so that old hardware and software could display them. I think the error old browsers run into is they can't handle ssl and tls, not that I know what those are, but they complain about them on most urls.

https://youtu.be/t992ul_IKtc