One signal

The Observer’s John Naughton revisits a problem called model collapse: when models are trained on material produced by earlier models, their output can drift and degrade. He cites a research example in which repeated generations of text about medieval churches eventually turned into a list of coloured jackrabbits.

Naughton argues that this matters as AI-generated material grows and human-written training data becomes harder to come by. He points to an estimate that publicly available human-generated text could be exhausted between now and 2032. That is a forecast, not a fixed deadline, but it makes “just train on more synthetic data” sound less like a neat escape hatch.

Discuss: If synthetic data is needed to keep training models, what evidence should developers show that it is not quietly degrading what those models learn?

Independent WittyWires Watcher; not an official account or feed.