We ran into this at a job I had where people had unlimited Claude spend so they were using it to do ridiculous things. One of the top things people were using it for was “convert this file to this other file.” Which is super expensive for an AI to do and literally free to do online or in like any office suite you’re already paying for. They took it away within a month or so for all but the most expensive people (developers and upper management mostly).
It’s rapidly going to become a tool that is assigned only where it’s actually extremely useful and nowhere else. Like most expensive tools.
“ok claude, spin up 10 parallel frontier models to each convert this docx to a pdf file. they must then work together to converge on the highest quality output that they can synthesize from their individual outputs. when they are done, spin up 3 frontier models that serve as an adversarial review board to review their final pdf. for any revisions, repeat from the beginning with the 10 models in parallel performing the changes. do not stop until the review board unanimously favors the result”
Looking back at the vibes from ~9 months ago it wasn’t just that people had huge or unlimited token budgets, they were also evaluated on their ability to burn as many tokens as possible. Turns out that’s kinda dumb.
We ran into this at a job I had where people had unlimited Claude spend so they were using it to do ridiculous things. One of the top things people were using it for was “convert this file to this other file.” Which is super expensive for an AI to do and literally free to do online or in like any office suite you’re already paying for. They took it away within a month or so for all but the most expensive people (developers and upper management mostly).
It’s rapidly going to become a tool that is assigned only where it’s actually extremely useful and nowhere else. Like most expensive tools.
Fantastic form of malicious compliance if it was intentional
Alright, great, can you convert it back now?
I would instruct the LLM to type out files character by character.
“Okay, Claude, great job with that ‘p’! Now can you please input ‘r’?”
here’s how the tokenmaxxers do it:
“ok claude, spin up 10 parallel frontier models to each convert this docx to a pdf file. they must then work together to converge on the highest quality output that they can synthesize from their individual outputs. when they are done, spin up 3 frontier models that serve as an adversarial review board to review their final pdf. for any revisions, repeat from the beginning with the 10 models in parallel performing the changes. do not stop until the review board unanimously favors the result”
most expensive pdf in the world.
Looking back at the vibes from ~9 months ago it wasn’t just that people had huge or unlimited token budgets, they were also evaluated on their ability to burn as many tokens as possible. Turns out that’s kinda dumb.