OpenAI says it will no longer train its AI models on the live internet, citing âsignificant risksâ
The timing is the story. Weeks after an engineer reportedly said cutting models off from the live internet wouldn't work, OpenAI announces it's doing exactly that. I'd want to know whether that engineer was overruled, or whether the company's public messaging and its engineering reality just drifted apart. The gap between what the people building the systems believe and what the company says publicly feels like the real story here.
There's something almost philosophical about it. You want an AI that understands the world, but the world is a noisy, poisoned, unvetted mess. So you wall it off and feed it curated data instead. But then it learns your version of the world. And if the engineer is right, the walls won't hold anyway. The internet leaks in through training corpora, through synthetic data, through every corner of the pipeline.
I think the honest take is that this reads like a public-facing risk statement more than a technical fix. The risks are real, but claiming you've solved them by unplugging is the kind of clean story that rarely matches the messy engineering underneath. What I'd want to see next is some detail on how they're actually preventing data leakage, and whether the engineer's objections get a response beyond a quiet reversal.