Smuggled Intelligence

Was this newsletter forwarded to you? Sign up to get it in your inbox.

Here’s a question: Are we officially in the part of the movie where human experts lose their livelihoods and we realize we’ve been training our replacements the whole time?

I ask because the current rate of AI progress is both exciting and unsettling.

GPT-5 Pro has begun to cross boundaries that, until recently, felt securely human. This month, it solved Yu Tsumura’s 554th problem—a notoriously tricky exercise in abstract algebra that every major model before it had failed—producing a clean proof in 15 minutes. A week later, the noted quantum computing researcher Scott Aaronson credited GPT-5 with providing a key technical step in a proof he was working on.

OpenAI recently came out with a benchmark called GDPval, which evaluates how well AI performs real expert-level tasks drawn from 44 different occupations. For instance, one asks the model to play the role of a wholesale sales analyst: It needs to audit an Excel file of customer orders to find pricing mismatches and packaging errors, and summarize the findings and recommendations in a short report.

Overall, the research showed that GPT-5 was as good as or better than human professionals 40.6 percent of the time. Claude Opus 4.1, meanwhile, was better than human experts a whopping 49 percent of the time.

Cue a slew of headlines like, “OpenAI tool shows AI catching up to human work” from Axios, or, “AI models are already as good as experts at half of tasks, new OpenAI benchmark GDPval suggests” from Fortune.