What's this about?
Aspire (205 upvotes) gives an agent only a natural-language capability goal — and hides the evaluation tasks — so the agent must decide for itself what to learn and how to measure its own improvement.
What exactly is different from normal training?
Normal setups hand the model a graded exam. Aspire hands it an ambition ("get better at X") and no exam — testing goal decomposition and self-assessment rather than task performance.
Why is this worth noting?
Real delegation looks like this — bosses give directions, not test suites. It's also the uncomfortable frontier: an agent that chooses its own curriculum is choosing what to become, which is a capability and an alignment question in the same breath.