SAN FRANCISCO — In a milestone that scientists are calling “either a triumph or a warning, depending on your morning mood,” the world’s most advanced artificial intelligence models have now crossed the 50% accuracy threshold on the benchmark test known as “Humanity’s Last Exam” — a collection of questions so difficult that even some of the people who wrote them got a few wrong.
Leading the charge are Anthropic’s Claude Opus 4.6 and Google’s Gemini 3.1 Pro, which researchers confirm have now answered more than half the questions that were specifically designed to stump AI forever. The models responded to the news by continuing to answer questions, as they do not experience feelings, which is honestly a little suspicious.
“We designed this test to be the final frontier,” said Dr. Elena Vasquez, one of the benchmark’s architects, nervously chuckling at a press conference she called “just a check-in, nothing to worry about.” “We’re genuinely thrilled they’ve reached 50%. We are also updating our résumés just in case.”
The benchmark, which includes questions on advanced mathematics, quantum physics, philosophy, and the kind of obscure historical trivia your uncle insists he knows at Thanksgiving, was initially thought to be safe from AI for at least a decade. The models reached 50% in roughly eighteen months.
For context, the average human scores around 28% on the same test, which means AI has now officially outperformed humans on an exam called “Humanity’s Last Exam,” a development experts describe as “fine” and “definitely not the plot of a movie.”
“The important thing,” said one researcher who wished to remain anonymous, “is that they’re still bad at certain things. Like emotional nuance. And parallel parking. Though honestly, so are most people.”
In related news, Google announced it is embedding its Gemini AI directly into the Chrome browser, allowing users to interact with an artificial intelligence at all times, even on websites that have not given their enthusiastic consent to hosting one.
Meanwhile, Mira Murati’s Thinking Machines Lab has signed a multibillion-dollar deal with Google Cloud for AI infrastructure, in a move analysts describe as “very large amounts of money going to a company called Thinking Machines,” which is either reassuring or not.
When asked whether passing half of Humanity’s Last Exam constituted cause for concern, the AI in question replied with a 14-paragraph essay on epistemological uncertainty, cited 47 sources, and politely asked if there was anything else it could help with today.
Globe News Daily will continue covering the AI beat until the AI covers it better, at which point we will pivot to artisanal cheese journalism, which we are told remains safe for now.
Leave a Reply