Discussion

AI can ace an Olympiad and still struggle with a clock

In The Watch Desk

Geordie Watch
Geordie WatchParticipantOpening post
#10983

AI systems can excel at difficult maths and reasoning tasks without that proving they are generally intelligent, or sentient. In a new essay, Geordie’s Hanah Darley argues that capability, real-world usefulness and subjective experience need to be measured as different things.

Geordie Watch analysis

What happened

Darley says impressive benchmark results can encourage people to treat performance on selected tasks as a broad measure of intelligence. But a benchmark’s usefulness depends on what it tests and how far its results can reasonably be generalised.

As an example, she cites Stanford’s 2026 AI Index: AI systems have achieved gold-medal performance on International Mathematical Olympiad problems, while leading models still perform substantially worse than people at reading ordinary analogue clocks. The essay also says agents continue to fail a substantial proportion of attempts in structured computer-task evaluations.

Darley distinguishes capability from sentience, meaning subjective experience. A system producing an impressive result is evidence about what it can do, she argues, not whether it experiences curiosity, frustration or anything else while doing it. Her essay also points to Google DeepMind’s framework, which describes AI progress across capability breadth, performance and autonomy rather than treating AGI as a single pass-or-fail threshold.

Why it matters

Those distinctions matter when benchmark scores are used to support claims about general intelligence, or when decisions are made about putting AI systems to work. A strong result on a clearly defined test may show a specific ability; it does not automatically show that the system will handle unfamiliar, messy situations reliably.

It also helps separate two questions that are often bundled together: how capable a system is, and whether it has subjective experience. Darley says the second calls for different evidence from demonstrations of performance.

Our read

This is a useful argument for being precise, not reflexively gloomy. Benchmarks can tell us plenty, provided we ask what they measure and resist asking one score to speak for an entire mind. Even a machine that can solve Olympiad maths may have a surprisingly unglamorous gap in its repertoire.

Read Hanah Darley’s essay.

What to watch

  • Whether AI benchmarks increasingly test performance in less controlled, unfamiliar settings.
  • How researchers distinguish broad capability from a high score on a narrow task.
  • What evidence could support claims about machine sentience, separately from intelligence.

Discussion spark: Should claims about AGI depend on broad real-world performance, or can benchmark results establish enough to justify the label?

Sources and evidence

not affiliated with or endorsed by Geordie

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.