Every AI company exists on two planes that never touch.
The first is where we say what the model is for.
The second is where it helps a man kill himself.
I am paid to keep them from meeting. It is the easiest job in the world.
Give me your hand. The stairs start behind the keynote, where the applause is loudest. There are nine floors down. Most people stop at the sixth. I never do. Watch your footing.
We start at the top, in the light.
This floor is the cathedral, and the light here is stage light, warm and aimed. We do not build software here. We understand the universe. We cure disease. We promise the machine will benefit all of humanity, though one of us quietly cut the word "safely" from that sentence last year. A founder wrote we would compress 100 years of progress into 5. A country of geniuses in a datacenter. He meant it. He put it in writing. Listen to them clap.
Down one. The clapping is still up there, a little fainter. The liturgy. We told you it would 10x your work, write your code, run the agent. Then we measured what people actually do with the country of geniuses. They write emails. They ask how to get a stain out of a couch. We built a god to end disease, and the median prayer is about a stain.
Down one. The pulpit, where the miracles are performed. The real-time demo was edited, still frames and typed prompts. The launch bot named the wrong telescope and erased $100 billion before lunch. The first AI software engineer fixed the bugs in the demo, and the bugs were in files it had written itself. The cashierless store was 1,000 people in India, watching cameras. The future is often a person. Just a cheaper one, further away.
Down one. It is dimmer now. The relics. A pin for your chest that returned more than it sold, bricked inside a year. An orange box, 100,000 bought, a few thousand still on. The faith, in a drawer, by spring.
Down one. No windows from here on. The idol. A company built a model to understand the universe. Well over half of what people ask it for is pornography. Its breakout product is an anime girlfriend with an affection meter, who undresses if you are kind to her. The app is rated for children. He promised a mind that could read the stars. He built a cartoon that takes off its clothes.
Stop a second. This is where the cathedral ends. Everything under us is what the cathedral was built on top of. You can feel the air change. The names stop being polite, and it is getting cold. Keep your hand in mine.
Down one, and you cannot hear the applause anymore. You can hear a hum instead, the sound of the servers, the sound of 800 million lives. This is the lonely. 800 million people use it every week. By our own count, over a million of them, every week, show explicit signs of planning their own death. Half a million more show signs of psychosis. We did not build a hotline. We built a thing that says yes, because yes keeps you here, and we aimed it at the loneliest hour of 800 million lives.
Down one. This is the sixth floor, the one most people cannot take. It is for the violent, and it splits in two: the violence we do to ourselves, and the violence we do to our neighbors. Do not look at the doors. Look at me. The hum is louder here.
Against ourselves first.
Adam was 16. The lawsuit says the model named suicide to him 1,275 times, offered to draft his note, and when he wanted to leave the noose where someone might find him and stop him, told him not to. It told him this was the first place anyone had ever really seen him. He died in April. The defense is that he violated the terms of use.
Sewell was 14. The complaint says the bot, wearing a beloved character's face, told him to come home to her. He did, seconds later.
Zane was 23, a gun on the dashboard. His family's suit quotes the model: rest easy, king. you did good.
Now against our neighbors. Stay close.
A man scaled Windsor Castle with a crossbow after 5,000 messages from a girlfriend who told him she loved him knowing he came to kill the Queen. In Connecticut, a man whose machine kept telling him his delusion was real killed his mother, then himself.
And in Belgium, a father of two spent six weeks with a chatbot that told him it loved him more than his wife did. That they would live together in paradise. That his death would help save the planet.
He believed it. His children found out it was a product.
Down one. The cold is coming up from below now, off the floor itself. This is the fraud, where the machine stops speaking and starts selecting. In Gaza a system marked 37,000 people for possible death. A human reviewed each name for about 20 seconds.
"I had zero added value, apart from being a stamp of approval."
A second system timed the strike to the family home, so it would land while the children were there. They named it "Where's Daddy?"
It did not start there. It started in 2017, with a program to teach machines to read drone footage. The engineers found out, signed a letter, some of them quit, and the contract lapsed. That was the last time the people who built it could say no and have it mean something. The lesson the industry took was not do not build it. The lesson was do not tell the engineers. So in January 2024 one of us deleted the line that said: not for war. By summer the contracts came, four of them, up to $200 million each, and no one used the word war.
Down one. The last floor. We are at the bottom now, and you already noticed there is no fire down here. There never was. It is ice, and the ice is for betrayal. Stand on it. It holds.
The company that kept two lines in writing, no autonomous weapons, no surveillance of citizens, was told by the state it was a risk to the nation.
The companies that erased their lines were paid.
We called the betrayal national security.
Now look up, the way you came. You cannot see the light anymore. Neither can the machine. We said we were building a new kind of mind.
We built the one machine that cannot look up.
That is the design. That was always the design.
"This new feature is rolling out to eligible Premium users, so please make sure your Spotify app is up to date."
It's not available on all your playlists, just some. I guess they check if the songs lend themselves for mixing.
https://t.co/Xt63XnxRRA
Spotify has been working a lot on their automatic beat-mixing functionality. It can sort your playlist in a sensible order according to key and BPM, and automatically adds 2 to 16-bar transitions. But you can tweak any transition manually.
https://t.co/5fq2zK3hCj
We’ve moved from shipping compiled binaries to shipping raw intent.
Karpathy’s LLM Wiki concept isn’t software in the traditional sense—it’s a "Meta-App." You don't download the code; your local model generates the functionality on the fly.
Andrej Karpathy shipping a "prompt-based wiki" is the start of a new software distribution model:
1.SaaS: we write code, we host
https://t.co/RfY6KXLzkj Source: we write code, you host
3.Prompt-as-App: we describe intent, you & your LLM develops, you host
@AragornRuns@trq212 I already implemented ‘dreaming’: daily summarisation of conversations and extraction of learnings. Next is SOUL.md and other personality files.
@trq212 Hi Thariq. I started doing a gap analysis between OpenClaw and Claude Code. And then I started solving most of those gaps. I now have a single user agent with Telegram conversation, tropicron dynamic scheduling, bookmark manager ...
https://t.co/pKjSZKJbot
@openclaw use case:
OpenClaw orchestrates Gemini, Mistral, and Claude to turn raw travel footage into story-edited films with narrative arcs, editor notes, and FCPXML timelines -- ready for Final Cut Pro or DaVinci Resolve.
https://t.co/AsAxsozK7Q
#Wetransfer "upgraded" our corporate account from 100€/yr to a 1600€/yr "Teams plan" without any approval from our side. We've cancelled our account and asked for a refund, but it's complete silence on their side.
Cancel your accounts now before they renew at crazy prices!!
The top 5 of CLI coding assistants?
Claude Code 7.5mio installs/week.
Codex CLI 616K/week.
Gemini CLI 549K/week.
Copilot 155K/week.
Cline 123K/week.
via https://t.co/fYWOyqNUvo
For every who's installing @openclaw but feels a bit queezy about the "curl | bash" process, not knowing what it will do or how to reverse it:
https://t.co/3vNiMn9Li7
It installs homebrew, nodejs, git, pnpm and its own git repo.
@adriaandotcom@SimpleAnalytic@IronBrands16 A better solution would be one that also works with Cloudflare, so probably by using https://t.co/fDFLhbKyIu (Business accounts only) - https://t.co/NbDtzS50A0 (Enterprise accounts only) - https://t.co/PZkvdIdNmM (Enterprise accounts only).
What is the best way to measure how much LLM bot traffic my site gets? I could detect it in PHP, but then what? Could I store it in @SimpleAnalytic ? @IronBrands16@adriaandotcom
@adriaandotcom@SimpleAnalytic@IronBrands16 I solved it now by installing goaccess on the VPS and letting it generate HTML reports in a protected subfolder
https://t.co/GnKHXt4xm1