OpenAI Opens the Kimono on Pretraining, Eyes Shared Safety Limits
Sam Altman says OpenAI will pull back the curtain on its pretraining progress, hoping rivals will agree on "common limits." Chief Scientist Jakub Pachocki wants perplexity and cross-entropy loss adopted as a common yardstick between model developers. It's a bid to turn AI safety into shared science rather than a solo race.
- Original source: breakingthenews.net.
- The source published it on Nov 22, 2026 at 5:53 PM.
- No second outlet has reported it yet. Treat it as a single-source claim.
OpenAI is opening up about what it builds before the models ship. CEO Sam Altman confirmed on Sunday that the company decided to offer what he called a lot of transparency into its pretraining progress, writing on X that he hopes agreements can be reached on common limits. He added that such a move would be very wise and help ensure safety and control.
The technical pitch came from Chief Scientist Jakub Pachocki in a blog post. He argued for establishing perplexity, or equivalently cross-entropy loss, as a metric that model developers can use to compare notes and inform shared safety standards. In his framing, more transparency into pretraining progress could be a step toward building a shared scientific basis for future pacing efforts.
The announcement builds on a theme Altman has been hammering for weeks. He has called on AI companies to put safety first and keep people in control as the technology advances, warning that the race dynamic in the fog of war is a bad thing. He has also described recent months as a newly challenging period, with more of the same ahead for a long time.
Neither Altman nor Pachocki said what specific data OpenAI would publish or when, and no other developer has signed onto the perplexity benchmark yet. The posts frame transparency as groundwork for collective pacing rather than a finished pact.
TopicsOpenAI · Sam Altman · Jakub Pachocki · X
Source: breakingthenews.net
Written by AI from verified reporting.
More stories
- Tech & AI cryptobriefing.com 2 sources 1 viewUS intel: China's spy agency ran a two-year hunt for Nvidia chips
- Tech & AI cryptobriefing.com 4 viewsCiti Lifts Big Tech AI Buildout Forecast by $500B, Now at $2.8T
- Tech & AI cbsnews.com Caught by @kylebrennan 1 viewMicrosoft's Smith: AI Set Off a Patching Race, and More Outages May Follow
- Tech & AI cbsnews.com Caught by @kylebrennan 1 viewBessent Says Nvidia Still Hasn't Answered China Chip Questions

