Stanford student Ken builds Sup AI, an ensemble system that outperforms individual models by 7.4% on Humanity's Last Exam through confidence-weighted synthesis.

Sup AI combines multiple AI models in parallel and weights their outputs based on confidence levels measured by entropy in token probability distributions

On Humanity's Last Exam benchmark, Sup AI achieved 52.15% accuracy compared to the best individual model's 44.74%, with statistically significant improvement

Ken (20-year-old Stanford CS student) built the system with his father Scott, an AI Research Scientist at TRI with a UCLA PhD, who contributes research insights

The key insight is that different AI models make unique, uncorrelated errors, so synthesizing their outputs while filtering low-confidence (high-entropy) predictions reduces hallucinations