Post

This is a really nice example of how LLMs learn shortcuts — heuristic strategies that can get the right answer for data within its training distribution. But as this example shows, these shortcuts can generalize poorly to data outside this distribution (e.g. here adding larger numbers).

Colin@colin-fraser.net

here's an example of a test I did over two years ago using gpt-4o. I am asking not only for it to compute the sum but to return the answer in words. 4o was not a so-called "thinking" model so there is no intermediate step between the request and the response. It just blurts out the sums.

3/10/26, 6:11 pm0 reposts2 likes

Aatish Bhatia

@aatishb.com

I think the idea of shortcuts provides some helpful intuition beyond 'stochastic parrots'. The latter suggests that LLMs are merely recombining their training data. Shorcuts imply they are learning heuristic algorithms that generate such data (and these heuristics can fail in unexpected ways).

Aatish Bhatia

@aatishb.com

I wrote about how LLMs learn shortcuts here: aatishb.com/blog/2026/ai... and this is an interesting technical paper that talks about how such shortcuts make it hard for LLMs to generalize arxiv.org/abs/2210.10749

aatishb.comHow a Simple Puzzle Reveals the Jagged Edge of AI’s AbilitiesAn interactive explanation of how AI can fail in surprising ways.