normative
LLMs validated via structural outcomes not text
LLM-based agents are more realistic than rule-based agents but much harder to validate, sitting awkwardly between high-validity empirical methods and parsimonious formal models; the way to use them responsibly is to ignore the text they produce and instead study robust structural outcomes that emerge from their behavior, trusting results in proportion to that robustness.
normativepending
Speaker
Petter TörnbergEvidence Quote
“if we were just looking at the kind of toxicity of their conversation, I would feel less confident that these are something that are actually playing out in real world. But given the robustness of these emergent patterns, I have more confidence.”
Created: 6/13/2026, 12:22:44 AM
Your notes
Loading notes...