Pony.ai CTO: Engineers Must Shift from Teaching AI to Enabling Self-Improvement
At LEAP East 2026, Pony.ai CTO Tiancheng Lou argued that AI progress is redefining engineers' role from step-by-step instruction to creating training environments for self-improving AI systems.
2026
2016
What Happened
In the past, engineers taught autonomous systems what to do step by step. Today, according to Pony.ai CTO Tiancheng Lou, their role is to build training environments, define objectives, and establish evaluation frameworks that allow AI to find better solutions on its own. Looking ahead, engineers will increasingly serve as assistants to AI, helping it learn from human experience and accelerate its own evolution.
“The goal has to shift from driving like a human to driving well.”
Lou traced this shift to AlphaGo Zero, which learned entirely through self-play rather than human data, surpassing earlier versions. The lesson: human knowledge provides a valuable starting point but can also limit progress. In autonomous driving, a system trained only on human data may inherit poor habits, so imitation alone is insufficient. Reinforcement learning in simulation allows AI to explore better strategies.
- Create training environments that simulate real-world driving scenarios
- Define clear objectives and evaluation criteria for AI performance
- Set up feedback loops for AI to iterate and improve on its own
- Support AI self-improvement as the system identifies weaknesses and proposes solutions
World models are central to this process, enabling AI to practice complex long-tail scenarios at scale without real-world risks. Reducing the "sim-to-real gap" is a key technical challenge. As world models improve, AI can take the lead in the self-improvement cycle, identifying weaknesses, determining needed data, and closing the loop between real-world operations and simulation.
Lou highlighted a "double standard" where society tolerates human error but views a single AI mistake as failure. Autonomous systems are seen as infrastructure, expected to deliver consistent performance. The answer is not to lower expectations but to develop AI that is substantially safer than human drivers, validated with evidence and built through gradual public trust and regulatory oversight.
Beyond safety, commercial viability matters. Robotaxi success depends on cost structure, fleet scale, operational efficiency, and passenger experience. Regulation plays a key role: a clear framework creates a path for evaluation and confidence. As deployments grow, more data enables better performance, and greater confidence allows further expansion.
Why this matters
As AI moves into the physical world through autonomous driving, the way engineers work is fundamentally changing. Instead of programming every action, they now build virtual environments where AI teaches itself through trial and error. This approach, championed by Pony.ai, aims to produce AI drivers that are safer than humans, but requires engineers to master new skills in simulation, reinforcement learning, and world modeling.
Terms in This Story
- foundation models
- Large AI models trained on broad data that can be adapted to various tasks.
- reinforcement learning
- A training method where an AI learns by trial and error, receiving rewards for desired actions.
- world models
- AI models that simulate physical environments for safe training.
- agent loop
- A process where an AI generates, tests, evaluates, and iterates on its own outputs.
Summarised from the linked release; details can be imperfect — always verify against the original source.