Sqrt-1

ai generated image of doakes (from the tv show dexter) crying

this is how it feels not being able to say "thank you" to claude because it would only waste your credits

what the duck is claude wetness

permission to cite this the next time you make a dunk post?

I asked my locally running qwen thinking vision to OCR a PDF and it did everything right, but near the end hallucinated a request I never made, and proceeded to hallucinate a response to that hallucinated request.

guess this is what you get for 6GB VRAM

git push showing a drastic reduce in the lines of code

most satisfying refactor

And yet movies still get made.

Dude what even is this image

I may be being hyperbolic but that line made my day better. Best advertisement I have till yet.

And if it can't, now you know what *not* to do.
Which can be as important.

unrelated but I found this really cool person called bsky.app/profile/astrra.space online I think you'll like her

YouTube video by Kane Pixels Backrooms | Official Teaser HD | A24

www.youtube.com/watch?v=BjRn... wooooo

*quantizes you*

I had always been optimistic of AI and the future but what I had not predicted was that we would get the technological progress we have in AI today _before_ we got the societal progress needed for it to be exclusively helpful.

crack the skull of the human and keep them strapped in a shoulder-width room with thousands of pins above to perform active neurosurgery on their weights while the human reads the context they are allowed to see from a printer directly connected to a shredder

I can't solve a Rubik's cube but I can tell if someone is good at solving them. I can't make movies but I can tell if Nolan is better than Michael Bay. I can't write code but I know I would rather just use my brain than use Apple Maps. also this has terrifying implications for scalable oversight

7 sentence horror

I heard a good analogy somewhere, which went "debating whether LLMs can think or not is like debating if Submarines swim."

using claude code is too time consuming. so I asked claude code to write code that would allow claude code to use claude code for me automatically (with --dangerously-skip-permissions enabled) just 20 minutes before my session resets (if I have 20% credits left) by scraping my todolist

yeah we haven't made a "installing lift kits on land cruiser" benchmark yet. in 2027 we will have benchmarks for how well AI makes benchmarks. 2028 benchmark benchmark benchmarks

the "AOCification of Zohran Mamdani" is a phrase that's going to live in my brain rent free

you guys probably already knew this, but you can just ask claude code to sort your media or downloads folder.

Devastating when this happens.

I use "prodography" in my notes

this is currently the most convincing argument to my brain against AGI.

I once had a dream where I found a robot that could do so many things for me, eventually it became a full time job just trying to find things for it to do.
The fact that this is now reality is both terrifying and exciting.

Show & Tell July 2025: AI Therapist, AI Clone, Wargames for Peace Trauma-dumping at an AI, mind-uploading myself, and the only winning move is not to play

Do you know any pre-existing techniques or thingies for AI therapy? This is one I read about.

blog.ncase.me/show-and-tel...

By their very definitions, AI cannot be empathetic. But I agree with you on the rest.

benchmark benchmark benchmarks

benchmark benchmarks

Philosophy degree! Wohoo