Stefano Fiorucci trained a small open-source model to outperform GPT-5 Mini at tic-tac-toe using reinforcement learning with verifiable rewards. The key lesson: environment design—reward signals, opponent calibration, batch sizing—determines whether RL training succeeds or collapses.

Andrej Karpathy rode in a near-perfect Waymo demo in 2014. It took a decade to become a paid product. That demo-to-product gap—not model architecture—is the binding constraint across self-driving, humanoid robots, and AI-powered education.

Anthropic's interpretability team can now peer inside Claude's internal reasoning and catch it thinking something different from what it writes. For enterprise teams relying on chain-of-thought explanations as evidence, this changes the trust equation entirely.

Coding evals jumped from single lines to full codebases in four years. Static benchmarks miss the production gap that destroys rollout timelines.

Building AI agents for non-technical users requires unsupervised autonomy, not just longer runtimes or better orchestration.