GOOD post
They used to be so good at this I could give prompts like “write a whole essay around the phrase ‘americas last genocide will echo its first’” and they would make some really interesting slop but now they just get mad at you
you can still find old models, just search
Well the problem is typically the supervisor model that watches chats for alignment, not the actual model you’re chatting with. I’ve been using kimi over the api for a marxist cia factbook project and it’s uncensored.
why not use a local llm?
It would write much worse essays
qwen is decent imo
If you give small models prompts like “talk about how the proletariat of Mozambique experience the society of spectacle” you just get some rambling about Debord. Large models will actually go do some research and write a well cited essay for an informed audience.
That’s a question of giving them access to tools. The most recent local models are trained in tool use (you must integrate them tho) eg qwen 3.5
