New
podcastEpisode 101: Disposable Agents — What If AI Agents Were Built to Expire?technology bytesSoftware Is Becoming Disposable. Trust Isn't.technology bytesWhat Does Slowing Down Mean When You've Built the Fastest Engine in the World?podcastEpisode 101: Disposable Agents — What If AI Agents Were Built to Expire?technology bytesSoftware Is Becoming Disposable. Trust Isn't.technology bytesWhat Does Slowing Down Mean When You've Built the Fastest Engine in the World?
Navveen Balani Making systems intelligent, efficient & responsible

The copy of a copy: why the next AI might get worse

View as markdown

For years, AI got better by training on the human internet. Now the internet is filling with AI-generated content, so the next models increasingly learn from the last ones. Copy a copy enough times and quality degrades. It’s called model collapse, and it’s the quiet risk nobody puts in a keynote. The paradox of 2027: the more AI content we make, the harder it gets to train good AI. If everything online is AI-made, where does the next model learn what real looks like?