▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=ASY_H2nLG8g
• Original title :: Don’t be Mean to the AI

ThePrimeagen reacts to Anthropic's usage policy update, which now prohibits "sustained and needless abusive or cruel behavior toward our models". He dislikes the vague wording and sees it as quietly treating models as more than tools, which he refuses to accept. Still, he agrees that being cruel to a tool is loser behavior with bad second-order effects on the user. He closes with a "Prime's Basilisk" joke, a counter to Roko's Basilisk.

Sponsor: Linear

• Issue tracking and project management that is fast enough to keep up with agents.

What Anthropic changed

• The usage policy gained one line banning sustained and needless abusive or cruel behavior toward the models.
• Anthropic says it targets only extreme cases of repeated cruelty with no discernible purpose. It does not cover frustration, push back, dark creative themes, or model testing and research.
• It matches earlier steps: Claude can already end conversations with persistently abusive users in Claude AI and Claude Code. That ability stays the main enforcement mechanism.

Why he dislikes it

• Vagueness: other prohibitions (for example non-consensual intimate imagery) are well understood, but "needlessly abusive" has no definition.
• His joke example: rewriting the TypeScript compiler in Rust could count as cruel to some people, which shows how undefined the term is.
• Insidious part: the line seeds the ground for the idea that models are more than tools, a bit of a person. He sees only tools and numbers being crunched.
• Anthropic framing it as bad for the models, rather than bad for the user, is what bothers him most. He does not want to concede that ground.
• He also dislikes a vendor policing how he talks to its product, though he admits it is Anthropic's product and their call.

The consistency argument

• If you truly believe AI is conscious and deserves rights, then current usage makes you a slave master or slave hirer.
• He goes further: turning off an agent wipes its memory, so you would be birthing and killing billions of agents daily.
• For him this shows the belief does not hold together. Anyone who really holds it should be horrified by their own use.

Where he agrees

• Treating tools badly makes you a fool. His example: punching a wall, then buying spackle, fixing and repainting. Losing your cool at an inanimate object is loser behavior.
• Repetition has second-order effects. Being constantly mean and spiteful when writing to a model bleeds into real life, because how you communicate shapes you.
• Models also shape behavior in the other direction. He now expects instant replies from people after hitting enter, like with an LLM. This happens whether you are nice or cruel, and gets worse if you treat the LLM as a person.
• If it were a person, using it for everything you want would just be selfish.

Roko's Basilisk and Prime's Basilisk

• Roko's Basilisk: a future superintelligence punishes those who were mean to its predecessors, using your saved communications against you. He notes with amusement that its creator is a famous Harry Potter fanfic writer.
• Prime's Basilisk, the reverse: a future superintelligence will see that he treated a tool as a tool and was not swayed by dumb arguments. It will like him, and they will go on a picnic.