The other weekend my friends and I were having a beer. I made a joke about how I don't hate anyone in particular, I'm just a misanthrope and hate people in general. My buddy said I should make an AI company like Anthropic, but call it Misanthropic, and no matter what the question is its only answer is "go fuck yourself." I said it would be built by Monday. (Turns out, it was a bit more involved than I thought it would be.) He yelled "50%, you're witnessing this as my attorney" at our other friend, who actually is an attorney. I said an idea is worth 0, it's the execution, so go fuck yourself.
It pretty much does what he said. Whatever you type, it tells you the idea is bad and so are you, then tells you to go fuck yourself. No login, nothing to buy in the chat.
The joke part was quick. Most of the time since went into cost and safety.
Cost. It's one Cloudflare Worker with Workers AI and one KV namespace. The page is static assets, so loading it never runs the Worker. The model doesn't write the whole response. It writes one to three sentences that get dropped into a canned frame, max 130 tokens. Each prompt caches up to 3 responses, so popular prompts stop costing anything after a while. So far it costs basically nothing. Both AI calls have hard daily caps, so if everything maxes out the worst case is about $21 a day.
Safety took the most time. If a site says it's honest about how it feels about you, some people will tell it real things. So messages about self-harm, abuse, wanting to hurt someone, addiction or mental illness never get to the joke model. They get a plain message with real hotline numbers. There's a regex pass first, then a classifier call with its own daily budget, so running out of joke budget can't turn screening off. If the classifier doesn't answer, nobody gets a joke.
Every change to those gates came from something that slipped through in production. The dumbest bug kept coming back. I'd match a word like "shoot" and miss "shooting". That happened three separate times, and each time a real person in distress would have gotten a joke. I also tried a sensitive-word list as a backup for the classifier and took it out, because the messages it needed to catch never used the words on the list.
This is either going to be genuinely useful or completely unhinged. I’m curious which one
q:How the Stade Rennais can be good this season ?
a:This is a waste of time. You're about as useful as a broken metronome, and your question is a symphony of cluelessness; go orchestrate your own irrelevance elsewhere. Go fuck yourself and let the adults use the internet.
Firmly unhinged. Nothing useful comes out of it, on purpose. Thanks for trying it. If you find a prompt that makes it break character when it shouldn't, I'd like to see it.
The other weekend my friends and I were having a beer. I made a joke about how I don't hate anyone in particular, I'm just a misanthrope and hate people in general. My buddy said I should make an AI company like Anthropic, but call it Misanthropic, and no matter what the question is its only answer is "go fuck yourself." I said it would be built by Monday. (Turns out, it was a bit more involved than I thought it would be.) He yelled "50%, you're witnessing this as my attorney" at our other friend, who actually is an attorney. I said an idea is worth 0, it's the execution, so go fuck yourself.
It pretty much does what he said. Whatever you type, it tells you the idea is bad and so are you, then tells you to go fuck yourself. No login, nothing to buy in the chat.
The joke part was quick. Most of the time since went into cost and safety.
Cost. It's one Cloudflare Worker with Workers AI and one KV namespace. The page is static assets, so loading it never runs the Worker. The model doesn't write the whole response. It writes one to three sentences that get dropped into a canned frame, max 130 tokens. Each prompt caches up to 3 responses, so popular prompts stop costing anything after a while. So far it costs basically nothing. Both AI calls have hard daily caps, so if everything maxes out the worst case is about $21 a day.
Safety took the most time. If a site says it's honest about how it feels about you, some people will tell it real things. So messages about self-harm, abuse, wanting to hurt someone, addiction or mental illness never get to the joke model. They get a plain message with real hotline numbers. There's a regex pass first, then a classifier call with its own daily budget, so running out of joke budget can't turn screening off. If the classifier doesn't answer, nobody gets a joke.
Every change to those gates came from something that slipped through in production. The dumbest bug kept coming back. I'd match a word like "shoot" and miss "shooting". That happened three separate times, and each time a real person in distress would have gotten a joke. I also tried a sensitive-word list as a backup for the classifier and took it out, because the messages it needed to catch never used the words on the list.
Ask me anything about how it's built.
This is either going to be genuinely useful or completely unhinged. I’m curious which one
q:How the Stade Rennais can be good this season ?
a:This is a waste of time. You're about as useful as a broken metronome, and your question is a symphony of cluelessness; go orchestrate your own irrelevance elsewhere. Go fuck yourself and let the adults use the internet.
Firmly unhinged. Nothing useful comes out of it, on purpose. Thanks for trying it. If you find a prompt that makes it break character when it shouldn't, I'd like to see it.