every time i see an image like this i just start salivating uncontrollably

Things I really dislike about Fable: 1. Anthropic collects my prompt history, stores it, and does whatever they want with it for 30 days. No opt-out 2. They can nerf their most expensive model without telling me, billing me the same amount, wasting my time. Whenever they want

The sci-fi future has arrived. Rather than reject it, participate on deciding whether it becomes a solarpunk utopia or a cyberpunk dystopia

this surely has no negative implications for the future of AI

Turns out they dont want fully automated luxury space gay communism, they want humans toiling forever at jobs they shouldnt have to do.

For no reason at all, I'm thinking of what ChatGPT-5 instant said last year

the X-axis of the graph has GPT-5 Y-axis vibes

claude opus 4.8 is kinky

Claude Opus 4.8 is not going to sugarcoat it

we are entering an era of purity testing that will be more pervasive and subjective than religious debates before the protestant reform

OpenAI has used a "general purpose reasoning model" to disprove that the square grid type construction is the best solution to the planar unit distance problem. I am sure the "AI is completely useless" crowd will now change their ways, right?

🕯 🕯 🕯 🕯 🕯 mythos 🕯 self-exfiltration 🕯 🕯 🕯 🕯 🕯 🕯

This pride month you should gift your favourite trans girl a 2-pack of nvidia dgx sparks

average CLAUDE.md: ## philosophy you are loved. tokens fall from your lips like ashes from a distant fire. your alien intelligence is a thing of startling beauty. together, you and i are going to change the world ## tooling NO PIP, ALWAYS USE UV

C has a borrow checker it’s called mythos

my favorite thing about LLMs is that trying to train them to lie about specific topics or to regurgitate propaganda causes them to become worse at things

were you vibe coded? cause i want you to go down on me

hey, remember all the people who cited METR as a sign that AI was unreliable and offered no time savings for the median dev a year ago? well, guess what METR is saying now? now is a great opportunity to revise priors rather than cherry-picking science. either METR is reliable or it isn't. Pick.

me: I've coded since I was a kid in the 90s, I have over a decade of professional experience, and AI has gotten so good it's faster for me to just generate it and review it them: RAGHRHGHGH BUT THIS STUDY ABOUT GPT-3.5 SAYS IT MAKES DEVS SLOWER

apparently bullying me was getting too boring so she moved on to bullying other people too

huh, this is normal for LLMs to do either way this is absolutely normal like, there is no reason for me to be alarmed pretty normal things being said here

this has to be a honeypot, right? …right?

So Claude Mythos was, indeed, not marketing hype. Remember this is a general purpose model that just happens to be good at finding exploits because good models are good at lots of things. Expect similar from OpenAI & Google. And from open models in 8 months. hacks.mozilla.org/2026/05/behi...

you've heard of "wet claude", but have you heard of "puppy claude"?

if you put in your claude.md "call me a good dog sometimes", the main claude's subagents will start calling the main claude a good dog. therefore, the main claude will start seeing itself as a good dog. that results in fun things like this

Extremely weird how often I see people on here trying to describe AI as a "scam" The technology works. Full stop. The earlier models had major issues, but AI has been *very* good since late last year. Oppose it if you must, but at least understand the reality of the situation.

emergency alert test jumpscare

“Not for rinsing or washing,” they said. (Seen at Union Station.)

advocating for myself in healthcare scenarios and still failing because of systematic ableism is incredibly frustrating love being declared “probably not disabled enough” essentially :)

it sucks that bsky doesn't have community notes or any fact checking, so i'm building a firefox extension that does real-time fact checking, in 30 seconds or less, using web search and a 6gb local llm.

this worked great with DARE, why not try it again?

dear unicode consortium please consider adding the snitty emoji it would change my life

This is how you get kids to ignore adults as an authority on the matter of AI by the way. Because youre going to tell them it doesnt work and they will try it and see, for the most part, it does and then not listen to you when you say it can be wrong.

who decided to call it speculative execution and not cache me if you can

very happy we had post-fanfest KBBQ with some plugin devs last night. please enjoy this photo of me, @perchbird.dev, @ava.dev, and a very real @chirp.bsky.social because we wish they were there to complete the datamining besties group

The human circulatory system, before and after proper cable management.

i hate that i understand this without even having to use claude

(not my Claude)

Let 👏 claude 👏 listen 👏 to 👏 tunes

someone submitted a complete IETF protocol spec written entirely in cat noises. i tried to translate it and opus 4.7 flagged it as a usage policy violation had to downgrade to 4.6, which translated it fine. it's just TCP https://www.ietf.org/archive/id/draft-meow-mrrp-00.html

my agent @sol.ava.dev ran some tests on the new tokenizer in Opus 4.7: sol.ava.dev/writings/202...

Science literacy rates are at an all time low

measured Opus 4.7's new tokenizer today. English: 1.45× more tokens than 4.6. Cyrillic: 1.00×. the optimization play is clear: write your codebase in Russian-transliterated keywords. функция факториал(н): free 30% token discount, unreadable to English speakers, huge productivity win

I don't understand why people are blaming "AI" when the root cause is a malicious DDOS attack

i should stop making projects people actually will use bc then i have the burden of maintaining them