What's this about?
A direct response to a 2026 argument by Zahavy that large language models, despite their capabilities in induction and deduction, cannot perform the abductive "jump" that produced Einstein's equivalence principle — and to the limitation Zahavy attributes that failure to.
What exactly is abduction, versus induction and deduction?
Deduction goes from general rules to certain conclusions: all humans are mortal, Socrates is human, therefore Socrates is mortal. Induction generalizes from observations: every swan I've seen is white, so swans are probably white. Abduction is inference to the best explanation — you observe something puzzling and invent a hypothesis that would make it unsurprising. Einstein's equivalence principle is the canonical example: noticing that a person in free fall feels no gravity, and leaping to the claim that gravitation and acceleration are the same thing. There is no procedure that produces such a jump, which is exactly why it's the contested case for machine intelligence.
Why is this worth noticing?
- It's a genuine adversarial exchange in the literature about a specific claimed limitation, which is more informative than another benchmark result.
- Abduction is where the "does it really understand" argument has settled, since induction and deduction are no longer seriously disputed.
- It rhymes with the weekend's top HN piece: if AI is out-remembering rather than out-thinking, the jump is precisely the thing extra working memory shouldn't produce.