OpenAI proposes industry-wide safety caps keyed to pretraining perplexity
gdb · x · 2026-11-23
OpenAI co-founder Greg Brockman highlighted researcher merettm's essay "Perplexity and Intelligence," framing it as a transparency step toward industry-wide shared safety bars.
Key points:
- OpenAI publishes pretraining progress transparency, proposing common limits based on perplexity and other measurements that developers commit not to cross.
- Progress beyond those limits would be gated on concrete milestones in understanding and monitoring AI behavior, such as a successor to CoT monitors.
- Central claim: AI will help advance alignment and monitoring, but likely not fast enough to permit full pretraining acceleration — safety risks lagging model intelligence.
It's the first public proposal to anchor safety thresholds to a quantifiable pretraining metric like perplexity.
More from Companies & People
- A better yardstick for 10x engineers: how much of their work gets reused — vboykis · 2026-11-23
- OpenAI and Anthropic are hiring life sciences roles, with Anthropic posting 5x more — dr_alphalyrae · 2026-11-23
- Snorkel Founder Talk: AI Data Shifts From High-Volume Tasks to Expert Work — ajratner · 2026-11-23
- David Sacks: Anthropic pitching recursive self-improvement to IPO investors, and he's worried — DavidSacks · 2026-11-23
- Developer: Apple rejecting developer accounts and new Apple IDs with no explanation given — FuSheng_0306 · 2026-11-23
- Sarvam praised as India's most responsive AI team; in-house speech tool rivals $2B-valued Wispr Flow — cneuralnetwork · 2026-11-23