How to measure the pace of progress
Debate over slowing AI development intensified after a former OpenAI and Anthropic researcher said companies were “playing with our lives.” Anthropic’s IPO filing also said the technology could pose “catastrophic or existential risks to society.”
Anthropic CEO Dario Amodei then published a plan to regulate the pace of advanced model development. OpenAI CEO Sam Altman and SpaceX CEO Elon Musk appear to have backed it. Amodei’s proposals include shared safety standards and limits on the pace of unchecked AI progress.
To limit that pace, companies need to agree on how to measure it. Pachocki believes perplexity could serve that purpose; it is unrelated to the company of the same name. Altman called Pachocki’s publication one of OpenAI’s most important recent works.
What perplexity shows
Benchmarks that assess AI capabilities on practical tasks quickly become outdated as the technology advances. Perplexity, by contrast, remains informative, Pachocki says: it shows how well a model predicts the next token on average. The lower the score, the more accurate the predictions.
The publication includes a chart showing the decline in perplexity across OpenAI models starting with GPT-3.5. Pachocki notes that GPT-4, with a score of about 0.91, was already using chain-of-thought reasoning effectively. That prompted researchers to deliberately identify the capability and build modern reasoning models. When the score fell below 0.8, transformative capabilities emerged in areas such as cybersecurity and mathematics.
Pachocki proposes making perplexity a shared measure of general intelligence that companies could use to set joint limits. They would have to agree not to cross the chosen threshold until the ability to understand and monitor AI behavior reaches a specific level.
Dataist's take: The proposal ties limits on model development to a measurable indicator, but does not itself define a perplexity threshold or criteria for determining whether control systems are ready. Companies still need to agree on those parameters.
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X