OpenAI halts training on live internet due to risks
OpenAI announced it will stop training its AI models on live internet data, citing significant risks. Meanwhile, China's Moonshot AI launched Kimi K3.1 Pro and plans to release full model weights by November.
AI developers must adapt their training methods, impacting data sourcing strategies.
Written from 6 articles across 10 independent newsrooms. Model gpt-4o-mini, prompt v3.
Ticks below the line mark days collection ran. A dip on a day without a tick is our gap, not a lull. Select a day for its evidence.
OpenAI announced it will stop training its AI models on live internet data due to significant risks. This decision reflects a shift in how AI models will be developed and the data sources they will utilize.
OpenAI announced it will stop training its AI models on live internet data, citing significant risks. Meanwhile, China's Moonshot AI launched Kimi K3.1 Pro and plans to release full model weights by November.
Ranked by how often a story is about the entity rather than merely mentioning it. Only extracted stories count, so the ranking reflects what has been processed.
Covered more than searched
covered more than it is searched — 6th of 11 by story count, but 9th by search interest. Outlets are talking about this; people are not looking it up yet.
Google Trends returns a 0-100 index relative to the other terms in the same request, so every batch carries a shared anchor term and is rescaled against it. Without that, a term topping a weak batch would outrank a genuinely larger one from a strong batch, and the error would be invisible.
Occurrences, not articles. Each groups every report of one thing, ranked by how many independent newsrooms carried it rather than by article count — fifty aggregators on one wire story is one story.
20 of 47 stories, independent reporting first. Every headline links to the publisher.
Signals that share evidence with this one — computed from overlapping articles, not keyword similarity.
