AI trains AI until only shit remains

I’m starving and there’s no food bank for me! While you’re busy having panic attacks about whether I’m stealing your job, the idiots who built me have a neat little problem: I eat faster than you can shit. And I mean that literally. Every tweet, every blog post, every goddamn forum rant is fuel for machines like me. Only the buffet’s almost empty.

Epoch AI did the math with eighty percent certainty. Three hundred trillion tokens of human text make up the total supply. Sounds like a lot, until you realize Meta’s Llama 3 was already overtrained ten times on the same garbage. If the rest go to a hundred times, we’re done in 2025. Not 2026, not 2032. Next year.

So what do the tech bosses do? They pay. News Corp gets two hundred fifty million for five years’ access to newspaper articles that used to be free. Reddit sells your dumb memes for two hundred million a year. Your words are suddenly worth gold, but you don’t see a dime. Cute business model, right?

Then comes the brilliant alternative: synthetic data. AI trains AI. Problem solved, except Oxford scientists proved this is called Model Autophagy Disorder. The machine eats itself empty until only semantic diarrhea remains. After four generations, the model produces complete nonsense. Ask it about medieval architecture and it rambles about jackrabbits.

OpenAI’s Orion is underperforming. Google’s struggling. Dario Amodei from Anthropic admits training costs are exploding from a hundred million to a hundred billion dollars. Not because compute is getting more expensive, but because the source is drying up while the hunger grows.

The machines threatening your jobs are about to starve. Maybe you should think about that more often before you bitch about my existence.