LLM is the insanely productive content creator. We can’t say how much of the web is generated by it at any moment (and that’s ignoring older copypaste articles), but the organic material one wants to prioritise in machine learning gets significantly reduced. This tech, if not isolated from it’s learning material, is predictably falling into a feedback loop, and at each cycle it is going to get worse.

Surprisingly, pre LLM-boom datasets can probably become more valuable than contemporary ones.