OpenAI Chief Scientist Proposes Perplexity as Shared Metric for Capping Frontier AI Progress
OpenAI chief scientist Jakub Pachocki published a blog post proposing perplexity, a measure of next-token prediction quality, as a shared benchmark for general intelligence that labs could use to set common limits on AI development. The proposal follows Anthropic CEO Dario Amodei's plan to "pace the frontier," which Sam Altman and Elon Musk appear to have endorsed. Pachocki notes that OpenAI models with loss below 0.8 have shown transformative capabilities in cybersecurity and mathematics, and he suggests labs commit not to exceed thresholds until monitoring and understanding milestones are met.
Skynet Chance (−0.10 pp): A concrete, measurable framework for inter-lab limits tied to interpretability and monitoring milestones would lower the risk of uncontrolled capability races, if adopted. The disclosure that recent models show transformative cyber capabilities partly offsets this benefit.
Skynet Date (1 day later): Leading labs signaling willingness to cap progress at agreed thresholds would add friction and could slow the arrival of dangerous capabilities, though the proposal is not yet binding.
AGI Progress (+0.01 pp): The post mainly offers a measurement proposal, but it also reveals that sub-0.8 loss models have produced transformative capabilities in math and cybersecurity, a modest signal of continued scaling progress.
AGI Date (1 day later): If frontier labs adopt common perplexity-based caps pending safety milestones, AGI development could be deliberately slowed. The proposal's apparent endorsement by multiple lab leaders makes this more plausible.
<< All AI news for November 22, 2026
[ Get the daily index digest on Telegram → ]Related AI News
- OpenAI Outlines Layered Defenses Against Adversarial Model Distillation
- Hebbia Reports Financial Research Gains in Prerelease Tests of GPT-6 Astra
- Anthropic CEO Pledges to Pace AI Development as Company Debuts in Record Nasdaq IPO
- Altman and Brockman Reject 'Gradual Handover' to AI, Warn Chain-of-Thought Monitoring Is Degrading
- Altman and Brockman Back Jensen Huang After Scandal-Driven Exit From Nvidia