Matthew Berman ran Jev over 193,355 video hooks for $0
A big classification run with a headline price of nothing, and no breakdown of how the price got there.
Matthew Berman says he had Jev analyze 193,355 video hooks and that the run cost him $0. The task was a three way labelling job over each hook, covering what's in it for me, why do i care, and what do i see first. He reports that the best combination of those three breaks out 33.1% of the time, against 9.1% for the worst.
i had jev analyze 193,355 hooks for $0
That spread is the interesting number, not the price. On a corpus of 193,355 hooks, a roughly 3.6x gap between the best and worst labelled combination is the kind of result that is worth having even if you do not make "how to" videos, because it is a concrete example of Jev being used as a bulk labeller on a corpus large enough that per item cost usually decides whether the project happens at all.
The $0 is unexplained
Berman does not say how the run came to cost nothing. There is no mention of which tier he was on, whether credits were involved, how many tokens went through, or how long the batch took. The post is a result and a price, with nothing between them.
So treat the price as his, not yours. If you are sizing a classification batch anywhere near 193,355 items, this gives you no unit economics to work from, and nothing here says the same run would be free for someone else.
The other thing the post does not give you is an accuracy check. There is no held out set, no human labelled sample, and no statement of how often Jev picked the right hook category. The 33.1% and 9.1% figures describe how often hooks in each bucket broke out, which is a claim about the content, not a measurement of the classifier that sorted it.
Berman says he put the findings into a free PDF, linked from the post.