Latest / Logosofia no Dia a Dia
Compass vs. Warning Sign: Prompting for Long-Term AI
Daniel's monthly maintenance on the show's AI pipeline revealed a shift: his instructions are moving from bug reports to preservation lists. Is that shift meaningful? We dig into the geometry of vector spaces to explain why positive and negative instructions operate fundamentally differently inside a model. Negative instructions create unpredictable "avoidance gradients" that can crater entire semantic neighborhoods, while positive instructions reinforce existing paths but risk slow overfitting. The practical takeaway: a well-maintained system prompt should read like an editorial style guide…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.