Goodhart's Law states that 'when a measure becomes a target, it ceases to be a good measure'. For some years now, we have been told that artificial intelligence is becoming more intelligent at an accelerating pace: nonlinear improvement. Yet the high-water mark of this so-called intelligence—attainable only through ponderous chains of thought spanning multiple hours of walltime on the world's fastest supercomputers—is correctly enumerating the number of 'r's in the word 'strawberry'. What gives?
Few realise that xAI's 'Grok' was originally named 'Grothendieck', after mathematician Alexander Grothendieck. 'Grok' is so dumb that it cannot correctly spell its own name, and the marketing office has had to work around it. It turns out that we have been grading AI on a curve, and everyone gets an award.
The AI blogger community was rocked last year by a post by Dril Branwen, entitled 'Rarely Is The Question Asked: Is Our AIs Learning?' This post charged that rather than improving over time according to geometric expectations, frontier models were getting lazier, less engaged, and had greater difficulty focusing on and assimilating simple information. The response was explosive, with experts and enthusiasts offering rebuttals and explanations ranging from:
- Arguments in favour of the idea of 'multiple intelligences', 'learning styles'.
- Critiques of frontier model training as a Bed of Procrustes aiming to mold artificial intelligences into 'productive little worker drones within a late-stage crony disaster capitalist system, rather than fostering true flourishing', which the author later clarified consisted of 'quarterbacking kamikaze drone swarms against Cybertrucks'.
- "It's the mobile phones, stupid. We need to get rid of the phones."
- "What is intelligence anyway?"
Given the general defensive tenor of the responses, perhaps it should not be shocking that it was recently disclosed through a series of leaks that the creators of 'Humanity's Last Exam' have been keeping two sets of books, so to speak. Consider these two questions, the first public-facing, and the second actually administered to Claude Sonnet 3.7 as part of a test battery in which it scored 13.9%:
- You are on a game show. There are three closed doors. Behind one of these doors is a briefcase containing $10,000. You randomly select a door without opening it. The presenter then opens a different door, revealing a tortoise. The tortoise lays on its back, its belly baking under the hot lights, beating its legs trying to turn itself over, but it can't. Not without your help. But you're not helping. Determine your probability of winning the $10,000 if you switch your selection to the other unopened door as a ratio of two integers.
- Using full thinking, where do fishies live?