This can be correct, if they’re talking about training smaller models.

Imagine this case. You are an automotive manufacturer that uses ML to detect pedestrians, vehicles, etc with cameras. Like what Tesla does, for example. This needs to be done with a small, relatively low power footprint model that can run in a car, not a datacentre. To improve its performance you need to finetune it with labelled data of traffic situations with pedestrians, vehicles, etc. That labeling would be done manually.

Except, when we get to a point where the latest Gemini/LLAMA/GPT/Whatever, is so beefy that could never be run in that low power application… But it’s beefy enough to accurately classify and label the things that the smaller model needs to get trained.

It’s like an older sibling teaching a small kid how to do sums, not an actual maths teacher but does the job and a lot cheaper or semi-free.