News

The most important charts in AI are wrong,' Ramez Naam argues

What's the deal? Technologist and author Ramez Naam published an analysis on November 21, 2026, arguing that the benchmarks underpinning predictions of imminent superintelligence are "badly misleading." He says new data on how AI is actually used inside OpenAI and Google contradicts the exponential charts from METR, the AI Futures Project, and the AI 2027 scenario.

The background: For years, a chart from the evaluation organisation METR has been one of the clearest ways to show AI capability growing exponentially. It measures how long a task an AI can complete, using how long the same task would take a human, at 50% or 80% success rates.

The counter-evidence: Naam says real-world work is far messier, with far shorter timeframes than the benchmarks suggest. At 80% success, OpenAI's internal task horizon grew much more slowly from June to October than either the AI 2027 scenario or extrapolations of the METR chart predicted.

Why it matters? By next year, according to those forecasts, AIs should be able to autonomously complete tasks that would take humans years. Naam argues the gap between forecasts and reality is widening every month, undercutting the case for a near-term singularity.

A measured critique: Naam is not assigning blame. "We were all building models in the absence of real-world data," he writes, adding that it is now "time to shift from being AI theorists to being AI empiricists."

The signal: The piece reframes a debate that has shaped billions in AI investment and policy around superintelligence timelines. Naam also calls for an AI transparency agenda, starting with Anthropic — a push to settle the question with data rather than extrapolation.

Read more: rameznaam.com

Image credit: Generated with Gemini

More top stories