Humanity’s Last Exam

This month Meta Superintelligence Labs announced Muse Spark — native multimodal reasoning, visual chain of thought, a “Contemplating” mode that runs several reasoning agents in parallel. It scored 58% on a benchmark that researchers, without much irony, named Humanity’s Last Exam.

Meanwhile the lab’s own superintelligence chief said their next model, still in training, has caught up with the current frontier. The pace is not slowing. Everyone in the field can feel it.

In the same weeks, a group of AI researchers published a proposal to delay superintelligence until 2040 — deliberately, by agreement — to give governments and societies time to deal with loss of control, concentrated power, and job disruption.

Read that again. The people closest to the machine are asking for more time. Not because it will fail, but because it might work.

There is a very old question in the book of Job that no benchmark has ever touched: where shall wisdom be found? Job asks it after listing everything human ingenuity can reach — miners cutting shafts through rock, finding sapphires in the dark, turning rivers aside. We are extremely good at reaching things.

And then he says: but not that. Not the one thing. You can build a system that answers every question we know how to write down, and it will still have nothing to say about what happens to you.

Because there is a last exam, and it is not a benchmark. It is not multiple choice, it is not graded on a curve, and there is no model you can put in your seat.

The standard is God’s law — the commandments you already half-remember — and the marking is done by someone who was present for every answer you have ever given. This is the appointment underneath all the others, and no amount of compute moves it by a day.

Here is the part that no amount of intelligence would have predicted. God’s solution was not to lower the standard, and it was not to grade you kindly. It was to send someone to take the exam and the sentence in your place — a substitute who had never once failed it, punished as though he had failed all of it, so that guilty people could be counted righteous.

That is not a system upgrade. It is a swap. And unlike every benchmark ever published, it has already been passed on someone else’s behalf.