We're optimizing language models for mediocrity.

LLMs are trained to find the most likely next word. But history's greatest ideas were, by definition, were improbable.

The future belongs to models that don't predict what most would say, but what one in a million might say to change everything.

3. Convenience is not necessarily contradictory to learning. Friction is not necessarily a pre-requisite. We want _quick_ and _lasting_ learning which is both _correct_ and _applicable_ to other contexts

2. Good teaching is, at its core, conveying the _intuition_ of the algorithm in a way that is most readily captured by the learner's existing mental models.

1. We need now, more than ever, to separate the task of learning from the task of proving that you know how to do something. They used to be one-in-the-same.

Fantastic paper from @majeedkazemi.bsky.social on how learning to code in the age of AI.

AI writing code negatively impacts learning, no surprise there. This paper looks at how we can fix that with embedded learning mechanisms. 3 thoughts from reading this →

"Do not wear too fine, nor yet too frowzy, a toga" - Seneca On the Philosopher's Mean (V)