itworks with the Humans"
mimisubmit ademo
consciouslymaking a little joke when they say it?
Claudeis not a person. Right?
reallydon't like it, but i don't know why yet
mimi
mimiclaude
program does this automatically when you use it to commit code. and i think github when you use copilot?
mimi—
, ;
, proper capitalization) also make me wary nowseem rightat first blush, but have subtle mistakes the kind a person never makes
needto write something
chee
mimifeliennehermans
makemistakes
mimi* have*to be assistants, they don't have to talk in prose
NataliaTalmina came in to talk about ai and conversation she mentioned that there's linguistics research that people adapt to each other's mental states when they are talking to each other
mimi
chee****from people outside of https://claude.ai and vscode
partof why the models are SO compute expensive and therefore money expensive. The output of pre-training is a base-model.
fine-tunethe model by feeding it loads of conversations in the style that we're aiming for (sycophantic simpering) and retrain some layers of the model towards the target output.
instructmodels rather than
basemodels.
{
"user_message": "Generate a tool that does X",
"model_response": " { @patchwork: more javascript json stuff goes here, blah blah blah you get the idea}",
...
}
And then it would, hopefully, be really good at generating the javascript for patchwork tools.
mimi* LLMs don't have to generate human-like language AT ALL.*It is a design / training choice that LLMs generally 'have' personalities. We could train an LLM to generate Patchwork tools or JavaScript and they wouldn't have any capabilities to generate human-like language like most of them do now. Our input could
alsobe entirely different / richer than a text description of the thing we want out.
exactlywhat i want
We've had a go at anotherway , but there are so many more to try :))
great. do we wanna talk about the stuff re: a new punk? punk computing? like, about craft vs art that's veneer and music with guttural screams and mistakes left in and... etc?