Post-training AI methods should simulate user capabilities, not just fix errors

Original: worked on this with @pkseeg and team.

Source: x.com ↗