68. [Bonus] Industrial AI Pitfall Guide: When 180 Years of Industrial Engineering Meets Large Models

Summary

• Li Hui’s key distinction is that industrial AI and internet AI run on fundamentally different logic: “Internet AI optimizes for satisfaction; industrial AI is primarily about certainty, safety, and reliability.” He cited a semiconductor wafer fab that, driven by the internet-AI boom, spent more than 2 years trying and failing to make the technology work before finally deploying it through a Physical AI approach. Host Wei Shijie closed with her own read: industrial AI deployment “is not a technology problem, and may not even be a capital problem—it is a patience problem.”
• Factories are not reluctant to use AI; they have been using it for years—the incremental opportunity is that large models add another tool to the toolbox. Li Yongli, head of Siemens’ Chengdu factory, broke down the existing stack: machine vision handles quality, while small models have long supported predictive maintenance and second-level dynamic parameter tuning. The current generative-AI opportunity is in human-machine interaction and office workflows, alongside the “long-tail demand” Zhu Xiaoxun highlighted: having AI write code on the fly to solve the small operational pain points that differ from one factory to the next. Large models are, for now, a supplement to existing industrial AI.
• Data is the foundation of AI, and companies cannot expect “artificial intelligence to fill the holes left in the past.” Zhu Xiaoxun called high-quality datasets the next digitalization battleground: data alone is not enough; it must incorporate human knowledge—“I changed a drawing then; why did I change it?” Siemens built a Data Layer to break down data silos and a “Process Cube” to connect IT and OT. Li Yongli’s warning was more direct: companies that frequently switch automation back to manual mode and whose platform data is “something you don’t quite dare trust” have no basis for extracting value from AI.
• The sharpest disagreement on the panel was whether companies with weak digital foundations can still leapfrog their peers. Zhu Xiaoxun said they “absolutely have an opportunity”; AI could make the expensive digitalization projects of a decade ago far more efficient. Li Yongli pushed back: “I don’t really agree with Dean Zhu—I think the opportunity is very small,” using electronic maps as an analogy: even the best model cannot run without high-quality underlying data. Li Hui took the harder line that “without digitalization, I don’t think industrial applications can get off the ground.” The exchange underscored how important the digital base remains to industrial AI deployment.
• The most actionable playbook from the discussion was “zero to one, one to 99, and the last mile”: a standard solution is less valuable than a veteran operator’s parameter set. Li Yongli said the last mile must be led by frontline employees with domain know-how. The specific approach is “small and beautiful”: experts handle 0 to 1 and 1 to 50, after which the factory takes over, keeping projects small to reduce the cost of failure, accelerate iteration, and move employees from observers to participants to drivers. Li Hui added that “looking for nails with a hammer” is a common failure mode—one major internet company spent 3 years trying to optimize data-center operations with AI without success. A real need must be defined by measurable, verifiable outcomes.
• ROI should be measured as “more, faster, better, cheaper,” not just as cost savings—and mature technology often beats cutting-edge technology. Li Yongli cited a company that had validated a second supplier in design and quality before memory prices rose, allowing it to switch immediately and move faster than competitors. In supply chains, “the value of every hour you save is impossible to quantify.” His counterintuitive conclusion: new technology is often unstable and expensive, making the ROI impossible to justify; in industrial settings, the ideal is mature technology that is both stable and cheap.
• Physical AI is the high point on 黄仁勋’s curve, but Zhu Xiaoxun and Li Hui both think the industry is still a long way off—and that Physical AI cannot solve every industrial problem. Zhu Xiaoxun noted that factories are not purely physical environments: large volumes of logical data also move through them, and black-box AI instructions must be continuously validated through real-time simulation. Wei Shijie went further, saying industrial AI is “even farther away than that.” Siemens’ hardware footprint also came up: 1 in every 3 industrial automation devices worldwide is made by Siemens, making edge-side AI compute one of its core hardware capabilities.

Deep dive

1. AI Anxiety Spreads from the Internet to the Factory: Expectations Are High, Deployment Is Hard

• Li Hui’s framing was blunt: “This fire came over from internet AI.” People saw internet AI become both smart and ubiquitous, assumed it could be dropped into factories to cut costs and lift efficiency, and then discovered that the underlying logic did not transfer. Internet AI optimizes for user satisfaction; industrial AI demands certainty, safety, and reliability. His clearest example was a semiconductor wafer fab that, driven by anxiety, “spent more than 2 years trying all sorts of things without success” before finding 深度智控 and deploying a Physical AI solution.
• The two guests’ backgrounds gave the argument weight. At Tsinghua, Li Hui worked on physics-based energy-system models—“one model would take roughly 4 PhDs more than half a year to build.” The models worked well but could not be replicated at scale. He began studying AI in 2009, found a way to combine physical models and AI in 2018, and founded 深度智控. Zhu Xiaoxun worked on barcode image recognition when he was young and joined Siemens in 2014 just as AI was overturning the field. “When it comes to AI’s disruptive power, I think I have the credentials to speak.”

2. Will Factories Use AI? They Already Have for Years

• Li Yongli brought the discussion back to the period before generative AI. Industrial AI has already been used for years in two broad areas: machine vision for quality control across raw materials, production processes, and finished goods; and small models—not large language models—for predictive maintenance, second-level dynamic parameter tuning, inventory optimization, and production-order optimization. In the tuning loop, downstream line quality feeds back quickly into upstream parameter adjustments.
• His description of large models was deliberately restrained: “Large language models have simply added a new tool to our toolbox.” The largest opportunity is in the human-machine interaction layer. Earlier generations of AI served the automated production lines on the factory floor; this generation serves the office applications and collaboration workflows “behind the computer screen.” The right response is less anxiety than recognition that “we have already seen many new opportunities.”
• Zhu Xiaoxun added another generative-AI opening: long-tail demand. Every factory has pain points that are “just a little different from those at other factories,” leaving them outside the reach of generic software. In the future, “AI can immediately write you a piece of code,” solving the scattered, low-volume problems one by one.

3. Data Is the Foundation: Someone Still Has to Read the Books

• Zhu Xiaoxun reviewed Siemens’ decade of groundwork: a knowledge-graph-based Data Layer to break down data silos, followed by the “Process Cube” suite to connect IT and OT systems with different standards and different degrees of real-time capability. The bigger challenge for AI is the quality of the dataset. Industrial data without the human knowledge behind it—“I changed a drawing then; why did I change it?”—has limited value for AI. “How to collect high-quality datasets has become one of the most challenging issues in the next phase of digitalization,” which Zhu called the battleground.
• Li Yongli poured cold water on the idea that broader data coverage solves the problem. LLMs can indeed process unstructured data and expand the definition of industrial data. But in companies where automation routinely falls back to manual operation and the platform data is “something you don’t quite dare trust,” expecting “artificial intelligence to fill the holes left in the past” is unrealistic. “Everyone still has to take it one step at a time.”
• Wei Shijie compared digitalization to building a library for every factory. If nobody reads the books, the data is inefficient or even useless. Zhu Xiaoxun agreed on the spot and added that once data enters the library, its quality improves only through continuous reading and use.

4. The Debate: Can Companies Without Digitalization Leapfrog?

• Zhu Xiaoxun took the optimistic side: “Any time something disruptive happens, it is a very good opportunity to leapfrog.” Digitalization that was expensive and difficult a decade ago could now, with AI, “very likely produce twice the result with half the effort.”
• Li Yongli delivered the most direct rebuttal of the episode: “I don’t really agree with Dean Zhu—I think the opportunity is very small.” His electronic-map analogy was simple. A map model that works perfectly in Beijing may fail in another city because street distances are wrong and the number of traffic lights is unknown. It is “a very good model, but it has no high-quality underlying data,” so the company must return to the starting point. Li Hui agreed that “digitalization is a basic prerequisite for AI applications.”
• Zhu Xiaoxun did not concede, instead reaching for an old case from Li Yongli’s Chengdu factory. A decade ago, the factory used AGVs carrying robotic arms to load and unload a PCB placement machine; programming the setup was painstaking. With humanoid robots and large-model control available today, “I would not use the method from 10 years ago.” His conclusion: “You can still do it now, because it is faster.”

5. Physical Execution Needs Simulation; Hardware Is AI’s Hand

• Zhu Xiaoxun gave the world-model discussion a practical turn. Physical AI has to do real work through robots and PLCs, while AI remains a black box. During operation, “we need a simulation result continuously validating whether the AI’s instruction is reliable and trustworthy.” Traditionally, teams simulated first and programmed afterward. Now the connection between the physical and virtual worlds has to meet a much higher standard of real-time performance.
• Siemens’ hardware position follows the same logic. 1 in every 3 industrial automation devices worldwide is made by Siemens, and edge-side AI compute is one of its core hardware capabilities. Zhu Xiaoxun’s summary: “AI’s hand is these hardware devices.” And: “Hardware is the entry point for data; data is AI’s fuel.”

6. Haste Makes Waste: The Last Mile from 99 to 100

• Li Yongli rewrote the usual internet narrative. Industrial deployment is not “zero to one, one to N,” but “zero to one, one to 99, and the last mile” (零到一、一到九十九和最后一公里). At 99, the team has a standard answer. “Put a standard solution on my machine, and it is often less effective than the parameter set from my veteran operator.” The move from 99 to 100 means adapting the system to a specific machine and a specific workpiece to find the local optimum. Data experts alone are not enough; the team needs people with domain know-how.
• Stability is the second constraint. A large language model “is astonishingly good sometimes and terrible at other times,” while “an unstable solution is a disaster for manufacturing.” The host asked about new systems being abandoned after launch, and the pattern was confirmed: someone may use it on day 1, and people may still use it in the first month, but performance deteriorates over time. The AI is not necessarily bad; it is not adapted to the local environment and has not been tuned to the optimum.
• Li Hui confirmed that “looking for nails with a hammer” is widespread. Real demand must be “defined by measurable, verifiable outcomes.” In 2021, a major internet company with formidable AI capabilities spent more than 3 years forcing a model onto data-center energy optimization, with poor results. 深度智控 started from the bottom: building a model for every device, running system simulation, optimizing the system, and then closing the loop with control. “Only by taking it step by step can you create sustainable value.”

7. A New ROI Formula: Count More, Faster, Better, Cheaper—not Just Savings

• Li Yongli offered two examples. An end-to-end supply-chain coordination system may appear unable to “save several people,” but during an energy crisis and a spike in oil prices, “the value of every hour you save is impossible to quantify.” When memory prices rise, a company that has already validated a second supplier across procurement, design, and quality can switch immediately—faster and more flexibly. Companies that have moved beyond mere survival should not invest only in headcount reduction; they should fund projects that make them “better, more competitive, and more consistent in quality.”
• His anti-tech-worship view was equally clear: insisting on technical novelty at the start of a project is a mistake. New technology is “first, often not stable enough, and second, generally quite expensive, so the ROI may simply not work.” “In industrial settings, the ideal is mature technology—stable and cheap.”

8. Two Separate Worlds—Technology and Use Cases? Good Technologists Start with the Customer

• Zhu Xiaoxun rejected the idea that he should be the person “holding the hammer.” “I should be the person figuring out what the next hammer should look like.” Using the newest technology in a particular scenario simply because it is interesting “is a very good way to fail.” “Industry does not care about emotional value”; internal customers such as Li Yongli care little about whether the technology is new.
• The proof point is Siemens’ AI team, which includes several steel-process experts. Algorithm engineers working on steel projects “had no choice but to speak with chief engineers at different steel mills” and learned the process that way. “Good technologists proactively integrate themselves into the customer’s needs. That is the only way to succeed in industrial AI.”

9. AI Replacing People: From “Big Cousin” to the “Optimal Solution”

• Li Yongli put the fear in historical context: factories have already gone through 2 waves of labor substitution—machines replacing people and digitalization replacing people. At the Chengdu factory, equipment and machine count “quadrupled” over the past several years, but headcount did not increase. Workers moved from tightening screws to equipment maintenance and failure handling, becoming higher-paid technical operators. Supply-chain planners who once acted as the “Big Cousin” collecting data have shifted toward finding the “optimal solution” and the best parameters.
• The most vivid AI case came from customs. A customs colleague built an agent that won 3rd prize in Siemens China’s AI competition to solve the HS code problem. There are more than 10,000 commonly used tariff codes, involving names, descriptions, materials, and processing steps; neither office staff nor shop-floor technicians can master the full set. The agent generates the code, performs a duplicate check, and proactively flags inconsistencies in the filing.
• Two trends stood out. AI can currently serve as an assistant and co-pilot, but “a complex, relatively complete process is still difficult to hand over entirely to an agent—I have not yet seen a very good example.” At the same time, AI is equalizing expertise: with an agent, “many complete beginners can become proficient in minutes,” lowering the barrier to entry for a role.

10. Where Scale Breaks: Small and Beautiful + Human-AI Hybrid Workflows

• Zhu Xiaoxun said nobody had solved the replication problem for industrial apps over the past 10 years because the bottleneck was the last mile of long-tail demand. He is “quite optimistic” that generative AI can help handle that adaptation. He is both optimistic and pragmatic about long-horizon Agents: the context window has “grown many-fold, just like Moore’s law back then,” but teams should not become fixated on it. Break tasks into modules; the future will be a hybrid implementation phase in which automation, information systems, people, and AI coexist.
• Li Yongli’s “small and beautiful” approach was the most repeatable management playbook of the episode. Experts handle 0 to 1 and 1 to 50; “after 50, we take the lead.” Otherwise, every change in a material or process parameter means bringing back the original specialist. Small projects deliver 3 benefits: low trial-and-error cost, low enough difficulty for frontline staff to lead, and fast results that support rapid iteration. “Beautiful” projects deliver 2 more: they stay on track, and with every step up, employees move “from observers to participants, and finally to drivers.” “Our own colleagues must be in the driver’s seat.”

11. Physical AI: First Tell It This Is an Elephant

• The discussion began with 黄仁勋’s GTC curve, which moves from generative AI to inference AI to Agentic AI, with Physical AI at the peak. Zhu Xiaoxun stressed that a factory is not a purely physical world. Automated equipment and information systems contain “large amounts of logical data in motion,” which cannot be described by a simple physical model. Industrial AI must also account for incentive structures, data flows, and program logic; Physical AI still has a long way to go in industry. Wei Shijie went further: “I think industrial AI is even farther away than that.”
• Li Hui’s evidence came from his own work: even the narrow field of factory energy systems took 10 years of research to find a viable path. Internet-style AI overfits and hallucinates. “Training accuracy is very high, but inference error becomes very large, making it completely unusable.” A single hallucination on a production line could cause losses of several million or tens of millions, or even a safety incident—“completely unimaginable.”
• His signature metaphor is worth keeping. Internet-style AI is a blind man feeling an elephant: “It touches a leg and thinks it is a pillar; it touches the belly and thinks it is a wall.” Physical AI’s approach is: “Before AI perceives the elephant, we first tell it that this is an elephant.” The AI then perceives and predicts within a defined framework, so however it feels around, “it will not feel out a hippopotamus.” His forecast is that Physical AI can ultimately achieve self-iteration, self-learning, self-optimization, and closed-loop control—roughly, modeling, understanding, learning, and control.
• The closing advice came in 3 lines: Zhu Xiaoxun said, “Stay curious and keep learning”; Li Hui said, “Understand the scenario before talking about technology, and commit to the long term”; Li Yongli said, “Focus on value and move fast in small steps.” Wei Shijie’s final judgment was that “industrial AI deployment is not a technology problem, and may not even be a capital problem—it is a patience problem.”