The Center for AI Safety, an AI doom crank nonprofit, and Scale AI have released a new AI benchmark called “Humanity’s Last Exam.” This supposedly tests “world-class expert-level reasoning and knowledge capabilities across a wide range of fields.” [Humanity’s Last Exam; Scale AI; Scale AI, PDF, archive]

The original name of the test, “Humanity’s Last Stand,” was discarded for being “overly dramatic.” [NYT, archive]

But you should be afraid of the impending robot apocalypse, very afraid. And send them money.

Two years ago, CAIS put out a “Statement on AI Risk,” an open letter signed by past AI luminaries, VCs, and AI doom cranks, warning of the risks of human extinction from AI.

We noted in August how Scale AI was stiffing its task workers. They also appear to have stiffed the academics who contributed questions — according to one contributor, they “conned hundreds of Ph. D.’s around the world to write questions for much less reward than promised.” [Hacker News]

Many of the questions are about memorizable facts rather than reasoning. The example chemistry question was cribbed from Wikipedia — so much for the test questions being “non-searchable.” [Hugging Face]

The very first thing that will happen is LLMs being trained to this specific benchmark — as happens with every AI benchmark. Humanity’s Last Exam is the FrontierMath grift with a fresh, more dramatic coat of paint.

Pivot to AI

It can't be that stupid, you must be prompting it wrong

https://pivot-to-ai.com