Tencent researcher debunks AI value alignment as 'pseudo-concept' after OpenAI's Superalignment collapse

Wang Huanchao, a senior researcher at Tencent Research Institute, argues that value alignment—the effort to make AI systems follow human intentions—is a pseudo-concept that cannot be solved through engineering. He traces the movement to Norbert Wiener's 1960 Science paper warning about machine purposes, and cites Nick Bostrom's paperclip maximizer thought experiment as the extreme formulation. But the real proof, he says, is the industry's own track record: OpenAI's Superalignment team, created in 2023 and promised 20% of the company's computing resources plus four years to crack the problem, was disbanded in just ten months. Co-leads Ilya Sutskever and Jan Leike both resigned, with Leike's farewell—saying safety culture had taken a backseat to shiny products—becoming infamous.

Wang's core critique is philosophical: "humanity" is not a single subject with one set of values, so alignment cannot be a one-way process of stamping human values onto machines. Instead, it's a dynamic, contested negotiation between many stakeholders—and current approaches treat it as a purely technical checklist. Evidence of failure abounds, he notes: Sutskever's new company, Safe Superintelligence Inc. (SSI), reached a $32 billion valuation in a year without releasing a single product, making "safety" a marketing label rather than an engineering achievement. Wang concludes the concept should be abandoned as a posture or symbol, not defended as a solvable problem.