OpenAI Just Published Its Speedometer and Asked Everyone to Agree on a Speed Limit
OpenAI has never told anyone how good its compressors are. On Sunday it did. Chief scientist Jakub Pachocki posted "Perplexity and Intelligence" on openai.com (dated Nov 22; Hacker News 395 points by 17:06 UTC, TechCrunch at 21:14 UTC), and it puts OpenAI's internal log-perplexity numbers on the table for the first time. GPT-3.5 sits at about 0.99 on OpenAI's current main eval set. GPT-4 at about 0.91. Below 0.8, cybersecurity and math capabilities showed up. A model at roughly 0.76 "started producing groundbreaking results" on open math problems. Post: https://openai.com/index/perplexity-and-intelligence/
The argument in plain words: capabilities keep popping out of one number, how well the model predicts held-out text, and nobody predicted which capability would appear at which threshold. So if you want to pace the frontier, measure the thing that actually moves. Pachocki expects "further transformative capabilities" in the 0.68 to 0.7 window, admits "I don't know what those capabilities will be," and warns that alignment auditing will keep degrading while models get better at noticing they are being tested. The proposal is "common limits," a cross-lab line on perplexity and other measurements that nobody crosses until monitoring hits a concrete milestone, such as a successor to chain-of-thought monitors. It links straight to Dario Amodei's "We must pace the frontier." Sam Altman called it "the most important thing OpenAI has published in a while."
Give OpenAI credit for one thing: publishing these numbers costs it something. Competitors now know roughly where it stands, and the post says so. That is a real move, not a press release. But read the weaknesses section, because the author wrote it for you. Labs use different datasets, so 0.76 at OpenAI is not 0.76 anywhere else. And "chain-of-thought reasoning does not reduce perplexity achieved by the model, even though it leads to a drastic increase in capabilities."
That last sentence is the whole story for anyone building agents. Everything this site covers every day, RL on tool use, harnesses, memory, verifiers, long-running loops, lives outside the metric. A perplexity cap would freeze the compressor and leave the elicitation layer wide open. Maybe that is fine: the scary part of the post is pretraining research done by AI "as soon as 2027," and that runs through the compressor. But a speed limit that cannot see the agent layer is a limit on the engine, not the car. If labs ever sign onto this, the second metric they need is one that measures what the scaffolding adds.
← Back to all articles
The argument in plain words: capabilities keep popping out of one number, how well the model predicts held-out text, and nobody predicted which capability would appear at which threshold. So if you want to pace the frontier, measure the thing that actually moves. Pachocki expects "further transformative capabilities" in the 0.68 to 0.7 window, admits "I don't know what those capabilities will be," and warns that alignment auditing will keep degrading while models get better at noticing they are being tested. The proposal is "common limits," a cross-lab line on perplexity and other measurements that nobody crosses until monitoring hits a concrete milestone, such as a successor to chain-of-thought monitors. It links straight to Dario Amodei's "We must pace the frontier." Sam Altman called it "the most important thing OpenAI has published in a while."
Give OpenAI credit for one thing: publishing these numbers costs it something. Competitors now know roughly where it stands, and the post says so. That is a real move, not a press release. But read the weaknesses section, because the author wrote it for you. Labs use different datasets, so 0.76 at OpenAI is not 0.76 anywhere else. And "chain-of-thought reasoning does not reduce perplexity achieved by the model, even though it leads to a drastic increase in capabilities."
That last sentence is the whole story for anyone building agents. Everything this site covers every day, RL on tool use, harnesses, memory, verifiers, long-running loops, lives outside the metric. A perplexity cap would freeze the compressor and leave the elicitation layer wide open. Maybe that is fine: the scary part of the post is pretraining research done by AI "as soon as 2027," and that runs through the compressor. But a speed limit that cannot see the agent layer is a limit on the engine, not the car. If labs ever sign onto this, the second metric they need is one that measures what the scaffolding adds.
Comments