logoalt Hacker News

thelaxiankey • today at 5:05 PM • 0 replies • view on HN

There's a distinction between reasoning with access to a turing machine and reasoning without access to a turing machine. In one case, you must sometimes become a turing machine yourself and execute "algorithms", which is something LLMs can struggle with.

When I was learning combinatorics, I was aggressively discouraged from using induction to prove things. Not because it was bad, or because it didn't work, but because if I used induction, I wouldn't be practicing 'combinatorics', I'd be practicing induction. You can make similar observations about solving ODEs numerically, or using mathematica for all your integrals.

One way to think of this is that banning coding prevents the agent from outsourcing some of the skills it's bad at (discrete sequential reasoning) to a computer. We are grading precisely the thing it's outsourcing because we suspect improving that might make it better.