ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
NewsMaterial3 min readDisconfirmAI Editorial

TechOrange Columnist Spencer Liu: AI Output Has Increased, but Who Is Responsible for Understanding?

Spencer Liu uses the contrast between Grigori Perelman and OpenAI's approach to solving mathematical problems as an entry point, pointing out that when AI makes output cheap, verification and understanding become the new bottlenecks—in enterprises, reports generated in minutes are rarely confirmed for accuracy by anyone, nor is gatekeeping treated as a performance metric. Citing Terence Tao's "Math 2.0" speech which stated "the problem is the lighthouse, not the destination," he argues that true AI transformation is not about how many Agents are deployed, but whether there is someone on the team responsible for understanding.

2026-11-09 14:12 GMT+8

Over the past few months, we have likely used AI to do quite a bit—writing reports, organizing meeting minutes, generating proposals. The volume of output has indeed increased, but have you ever stopped to ask yourself: Did any of this actually make any decision within the company better? Spencer Liu raised this question in his TechOrange column, noting that mathematics is undergoing a similar shift: when AI can produce proofs in bulk, the standards previously used to measure quality suddenly become unreliable.

In September 2026, OpenAI released ten thousand AI Agents, directing them to search in different directions simultaneously and exchange intermediate results with one another. The entire process transmitted 2.7 million messages, took 88 hours, and cost approximately $15 million, ultimately solving the Navier-Stokes equations—one of the seven Millennium Prize Problems listed by the Clay Mathematics Institute. Twenty years earlier, Russian mathematician Grigori Perelman independently developed new proof tools along the Ricci flow method, spending nearly a decade to solve the Poincaré conjecture; he published his paper directly on arXiv for global mathematicians to verify. In 2006, he declined the Fields Medal, and in 2010, he again refused the $1 million Millennium Prize. Both events were considered successes, but Liu points out the key difference: Perelman left behind a methodology that provided new tools for future researchers, whereas OpenAI left behind only an answer.

Sign in / Sign up
ReadingResearchRadarInvestment framework

This is the exact same problem encountered by enterprises implementing AI. When a proposal is generated by AI in just a few minutes, who is responsible for confirming its correctness? More often than not, the answer is no one—because the speed of verification cannot keep up with the speed of production, and gatekeeping is rarely regarded as a performance metric. The report is submitted, but has the team gained deeper understanding as a result?

Liu cites another Fields Medalist, Terence Tao, whose judgment during his October 2026 "Math 2.0" speech at the California Institute of Technology was pivotal: the mathematics community has long prioritized being the first to solve difficult problems as the core standard for measuring achievement, but when AI can mass-produce proofs, this standard has become "misaligned"—deviating from what it was originally intended to measure. Tao went further, stating that continuing to blindly optimize solely for problem-solving is now substantially harming the long-term health of mathematics. He delivered a line that Liu considered the most important of the entire speech: "The problem is the lighthouse, not the destination." If you rush straight to the base of the lighthouse at top speed, you have indeed arrived, but you have also abandoned the voyage itself.

When the Clay Mathematics Institute acknowledged that OpenAI's result "appears to be solved," it specifically emphasized that the procedures for verification and attribution of credit would be "deliberately slow." Liu reflected deeply on this deliberate slowness: in an era where everyone competes to reach the summit first, why does someone argue that success should be attributed to taking things slowly? Because if AI produces a thousand mathematical proofs in a day, yet there are insufficient researchers capable of reading, understanding, or extending these findings, we have not achieved a thousand-fold scientific progress.

For businesses, the logic is identical. Liu believes that great breakthroughs often do not end a problem but instead open up more new research directions; true AI transformation lies in achieving what was previously impossible, rather than counting how many Agents are deployed. He ultimately returns the question to a specific managerial judgment: Does your team have someone responsible for understanding? Without understanding, there can be no judgment; without judgment, decisions merely become signatures on a stack of outputs. AI can help us reach the moon faster, but those stars still require someone to look up and see them.

Sources

  • techorange.comTechOrange 科技報橘↗
Continue researching →