qustrolabe

I wish this was true, my trial runs out

You as delusional as you claim them to be if you're so quick to disapprove consciousness so easy

test driven development is good until your agent gets "all tests pass" psychosis

this was an early GPT issue but now when they train tokenizers there's special regex that makes so that bunch of spaces become single string for BPE to "learn" tokens from

finished Antimemetic Division and ooh it was good

then went asking Claude to explain few confusing parts I missed, at first it completely hallucinated nonsense, asked it to keep doing web searches but best responses came from dropping entire pdf and asking questions piecing plot back together

for optimists:
imagine a future genie that can make most wishes come true and start writing down every good wish now to be prepared, maybe some wishes will become completable earlier than others, order them by complexity and importance

for pessimist:
preparing for every fail scenario where that doesn't happen

solar panels, homelab, downloading and storing entire internet knowledge while it's still up and open, some other post-apocalypse preparation stuff

Using image model while reading a book to confirm whether I properly understood confusing scene described

do zoxide next

should I?

They caught the Rust disease of never releasing anything into stable

Gmail. Most emails now have Unsubscribe button and it silently sends unsubscribe email from you.

They now even have separate tab listing all services that sent you annoying stuff recently with unsubscribe button near each of them

maybe they're migrating for new model or something because that thing is dead, even flash barely work

They're good for integrating p5.js or canvas into Astro page without meddling with passing IDs around when you want to have multiple of them on same page

the council has spoken, turn on the A/C

Nobody, source article more about password that made their way into the code.

But even checking GitHub with their provided examples most of these are comments and placeholders. Though few actually use passwords with those characters 😭, but then why passwords in the source code in the first place

what in the cognitohazard

github.com/openai/grok

waste of time and money

hashtag not all meatbags

Pluribus-like AI that can just answer that at any time and you can even chat with other users though it

lame

Maybe via Termux SSH via Tailscale into docker container running agent at home? Both safe isolated as it is and you supposedly could also attach into tmux session.

Or maybe some better interface on top of all that

Unironically this is true and Gemini right on that one. Even before LLMs became big widespread thing internet was in such awful state that I had to add "reddit" at the end of each google query just to prevent SEO optimized slop articles from appearing in first results. Now we have AI with web search

And the silly things is: if I had much more tokens to spare I no doubt would also build agents that build agents that build agents all over the place, that sounds way too cool and funny to miss on

I've built few small things, some published some not, I really love and use them but they do need a lot more polishing to get into somewhat presentable state.

"If AI so good why don't you show us something good it made" kind of discussion mostly fails because most AI generation produces stuff that only relevant to one person, and this applies all forms (images/video/music/vibecoding). It's rarely something worth sharing outside, especially to skeptics.

Mainly the fact that their CEO is active on AI twitter and I see his posts from time to time. And other big respected accounts recommend it there. Feels more "real" than some random site from search results. And their bot on twitter has been catching slop articles a lot