Wang Haozhi doesn't give many interviews. When he does, he speaks in measured paragraphs, pausing before technical terms as if weighing their real-world consequences. That habit—caution before certainty—defines his work. As one of the earliest researchers to push for a structured AI safety protocol in China, Wang has spent the better part of a decade arguing that alignment isn't a Western problem to be solved at conferences. It's a local engineering discipline, built case by case。model by model.I first came across his name in a 2019 workshop paper that almost nobody cited. The paper described a method for auditing reinforcement learning agents in simulated grid environments—boring stuff on the surface. But buried in section four was a proposal: treat reward hacking as a security incident, not a training quirk. That framing was unusual at the time. Most labs treated specification gaming as an annoyance. Wang treated it as a breach. He later told a colleague that the distinction mattered because security teams understand incident reports;they don't understand loss curves. That kind of cross-disciplinary translation shows up again and again in his work.His more recent contributions focus on evaluation infrastructure. Wang Haozhi leads a small team that builds adversarial test suites for large language models—not to break them for headlines, but to create reproducible failure logs that regulators and engineers can actually use. A public report from his group in early 2024 documented how minor prompt perturbations caused a customer-service model to leak internal policy details across three separate deployment configurations. The finding wasn't glamorous. It was useful. Several firms quietly patched their systems after reading it. Wang's insistence on publishing negative results。even when they embarrass industry partners, has earned him a reputation as difficult but reliable. "He won't tell you what you want to hear," one collaborator said. "He'll tell you what your logs already know but you're ignoring."There's a stubbornness to Wang Haozhi that I find increasingly rare in AI research. He still writes code. He still reviews pull requests. He refuses to let his team use the phrase "state-of-the-art" in internal documents unless they specify the benchmark, the seed。and the evaluation harness. During a panel discussion last year, someone asked him whether China was behind on safety research. He shrugged and said the question missed the point. "Safety isn't a race," he replied. "It's a maintenance schedule. You don't win maintenance. You just don't skip it." The audience laughed. He wasn't joking.

What makes Wang's approach worth watching isn't a single breakthrough. It's the accumulation of unglamorous artifacts—test suites, incident templates, audit checklists—that make safety work legible to people outside the lab. He has argued that the biggest risk isn't a sudden superintelligence;it's a slow erosion of accountability as models get buried inside products. His response is deliberately small-scale: one evaluation at a time, one documented failure at a time. Whether that pace matches the industry's deployment speed is an open question. Wang Haozhi seems fine with open questions. He'd rather have a fragile process that works than a perfect framework that doesn't ship.