Everyone assumed large models would first replace those working in offices.

Writing copy, coding, reviewing contracts, generating reports--these tasks, inherently screen-based with data in text form, seemed the most suitable for AI. Yet, the first roles large models are mastering belong to a different group entirely.

These individuals stand before steel furnaces, judging the color of flames; they crouch next to compressors, listening to the machine's sounds; they watch gauges and material changes, relying on decades of experience to decide the next operational step. This craft has never been written into SOPs and is difficult to articulate, existing only in the eyes, ears, and intuition of veteran workers.

Logically, this should be the capability hardest for AI to replicate. The reality, however, is the opposite. After launching an intelligent steelmaking large model, Yongnian Special Steel (Hebei Yongyang Special Steel) learned from 200,000 furnace historical data points, processing tens of thousands of sensor signals per second. The endpoint carbon qualification rate rose from 75% to 97%. Shandong Haihua Group trained its master craftsmen's experience of "identifying faults by sound" into an intelligent control large model, achieving equipment failure prediction accuracy exceeding 95%. Guangxi Huasheng (Guangxi Huasheng Alumina) reduced a manual test requiring 6 hours to a 3-minute prediction. Zhongtian Technology (ZTT) compressed the fiber routing path work of a master craftsman from 6 hours down to 40 minutes.

These appear as four different industries with four different systems.

In reality, they are all doing the same thing: converting the tacit knowledge accumulated over decades by master craftsmen into data capabilities that a large model can learn, replicate, and use for real-time decision-making. This is the most counterintuitive aspect of this wave of industrial AI.

When a master craftsman retires, their experience leaves with them.

Dai Ziwei, Head of the Enterprise Management Department at Yongnian Special Steel, stated: "In the past, steelmaking relied entirely on the experience of master craftsmen." Converter steelmaking is a "black box" reaction—high-pressure oxygen is blown into the high-temperature molten iron, completing decarbonization, heating, and impurity removal within 5 minutes. Each batch of molten iron has a different composition. Minor changes in oxygen flow rate and lance height affect the final carbon content. Master craftsmen judged by the flame's color and sound, and their skill level determined the qualification rate. A 75% qualification rate meant 1 out of every 4 furnaces was substandard. Substandard batches meant rework, heat loss, and raw material waste. When a master craftsman retired after a lifetime of work, their knowledge left with them. Newcomers had to start from scratch, and the qualification rate would inevitably drop for several years.

Shandong Haihua faced the same problem. The chlorine compressor operates 24/7 in a highly corrosive environment. It's the "heart" of chlor-alkali chemical production, high in cost, and has no backup. Slight vibration anomalies are hard to detect. By the time the compressor makes an abnormal sound, a master craftsman would use a listening rod to touch the machine and determine the problem based on sound, sometimes taking three or four days to repair. Yan Guohui, General Manager of the Process and Digitalization Center at Shandong Haihua, said the company was stuck in the dilemma between "over-maintenance" and "unplanned shutdowns." The direct loss from an unplanned shutdown of a single chlor-alkali production line exceeded 30,000 RMB per hour, and recovery took at least 8 hours. To be safe, the company adopted a conservative "replace rather than wait" strategy: one minor overhaul per year, one major overhaul every two years. Annual maintenance costs were 800,000 RMB.

Huasheng Alumina's pain point was even more direct. The alumina production process is complex, with a time lag in testing key indicators, and operational optimization relies on experience. It took 6 hours for manual testing to get results for a set of indicators. Within those 6 hours, the operating conditions had already changed, making the data "past tense." At Zhongtian Technology, optical cable fiber routing was also a purely manual task. Workers would plan fiber routing paths on blueprints, which was time-consuming, labor-intensive, and error-prone. Daily fiber routing took 6 hours. Raw material inventory amounts were persistently over 50 million RMB because it was impossible to predict what would be needed on which day, leading to overstocking.

A common characteristic of these four scenarios across four industries is that the most critical knowledge resides in people's minds. When the person leaves, the knowledge is gone.

From observing flames to reading data.

The learning material for Yongnian Special Steel's converter large model was 200,000 furnace records of stable and reliable historical smelting data. For each furnace, the molten iron composition, temperature, oxygen flow rate, lance height, and endpoint carbon content were all recorded. The large model ingested this data, identified the rules for "which parameters to use under which conditions," and then provided the optimal control plan before each furnace began smelting.

The results were directly reflected in the numbers. The qualification rate rose from 75% to 97%, the steel material consumption per ton of steel decreased by 4.5 kg, the smelting cycle was shortened from 35 minutes to 30 minutes, and the cost per ton of steel was reduced by over 20 RMB. Dai Ziwei calculated that every percentage point increase in the endpoint carbon qualification rate corresponded to less heat and raw material waste. "Digital and intelligent transformation isn't the old path of burning money; it's a new path to making money."

The Hebei steel industry has begun replicating this capability to more factories. HBIS Tangsteel (Hebei Iron and Steel Group Tangsteel) used an integrated scheduling large model, reducing raw material inventory turnover time from 10 days to 5 days, generating over 10 million RMB in annual benefits. HBIS Hansteel (Hebei Iron and Steel Group Hansteel)'s quality large model increased the qualification rate of key products by 8%, reducing annual losses by 12.6 million RMB. Shougang Qian'an Iron & Steel's AI production control model saves 70 million RMB in costs annually and reduces CO2 emissions by 40,200 tons. Data from the Hebei Provincial Department of Industry and Information Technology shows that steel enterprises across the province have applied AI large models to varying degrees, with over half deepening their application in industrial intelligent control. From January to November 2025, the total profit of Hebei's steel industry was 28.135 billion RMB, a year-on-year increase of 16.8 times, and the profit per ton of steel was 26.25% higher than the national average.

Translating 'identifying faults by sound' into technical parameters.

Shandong Haihua's transformation began in February 2025. Partnering with Inspur Digital Enterprise, it built an intelligent control large model for the salt chemical industry based on the Inspur Haiyue Large Model Ch1 version, creating three intelligent agents: predictive maintenance for equipment, process optimization, and intelligent inspection. The deployment process of the predictive maintenance agent was itself a process of "experience extraction." When first deployed at the end of May 2025, accuracy was only 60% to 70%. Wu Mingfu, Deputy General Manager of the Platform Products Division at Inspur Digital Enterprise, said engineers repeatedly consulted with master craftsmen, converting their crafts like "identifying faults by sound" and "judging wear by vibration" into technical parameters. Through optimization using a combination of large and small models, accuracy improved to 90% by August and later exceeded 95%.

The master craftsman's listening rod retired. The large model now monitors equipment operating status in real-time, accurately predicts maintenance times, and achieves fault identification accuracy exceeding 95%. Unplanned shutdowns of the chlorine compressor dropped from 4 times in 2024 to zero in 2025. Annual maintenance costs fell from 800,000 RMB to 200,000 RMB. The process optimization intelligent agent addressed another difficult problem. The electrolysis process is influenced by the coupled effects of multiple parameters like current, temperature, and acid addition rate, making it nearly impossible for manual calculation to achieve an optimal balance between indicators. The agent tracks data changes in real-time, achieving minute-level autonomous optimization of the process path. The electrolyzer process optimization model at the chlor-alkali plant saves 4.5 million kWh of electricity annually, extends the lifespan of the ion exchange membrane from 4 to 5 years, and generates comprehensive benefits of nearly 10 million RMB. Yan Guohui's ledger shows: Phase 1 investment of 32 million RMB, with an estimated benefit of 23 million RMB in 2025. The process indicator stability rate improved by 55%, the key indicator self-control rate increased by 61%, and the frequency of manual operations decreased by 68%. Business processes were streamlined from 1,803 items to 400, a reduction of 77.8%.

6 hours becomes 3 minutes.

Guangxi Huasheng's alumina large model, called "Zhisheng," was developed based on Chinalco's "Kun'an" large model, in collaboration with Chinalco Intelligent Technology and Central South University. It adopts a "four horizontal, four vertical" technical architecture, optimizing the entire process from data perception to intelligent decision-making. The most impactful effect is testing efficiency. The large model outputs prediction results within 3 minutes, compared to the 6 hours required for manual testing. The prediction accuracy for key indicators exceeds 80%, and the accuracy for the dissolution aK indicator surpasses 90%. Main control operator workload is reduced by 85%, and sampling/testing efficiency is improved by 30%. The annual benefit for a single plant is in the tens of millions of RMB. After full promotion within the Chinalco Group, the annual benefit is estimated at 130 million RMB. This project was also selected as a typical case of manufacturing digital transformation by the Ministry of Industry and Information Technology in 2025, the only project from Guangxi to be selected.

Zhongtian Technology's "Tianji" large model took a different path. In 2025, it established an AI Research Center and independently developed an industrial vertical large model. After the "Smart Fiber Routing" model was deployed at the optical cable general factory, only one operator was needed, and daily fiber routing was compressed from 6 hours to 40 minutes. Raw material inventory fell from over 50 million RMB to less than 10 million RMB. After being promoted to the transformer division, strip material inventory dropped from 120 tons to 75 tons, and the turnover period shortened from 45 days to 20 days. Tianji also achieved results in other scenarios. Spool remaining quantity detection responds in 100 milliseconds, saving over 80% of labor. Mandrel extension precision improved from 2 mm to 0.5 mm, surpassing Japanese imported technology. The ultra-complex image-text intelligent model parses vast document files in seconds, reducing the number of technical staff by 25% and the error rate by 74%. After introducing AI to the RF factory, labor costs were reduced by 80% and quality improved by 28%. Xue Chi, Chairman of Zhongtian Technology, stated: "AI is not a vague concept; it's a core tool for deepening the core business, improving quality, and increasing efficiency."

Four steps to transfer experience into the model.

Looking at the approaches of these four companies, the path is strikingly similar.

Step 1: Collect historical data. Yongnian Special Steel collected 200,000 furnace smelting records. Shandong Haihua worked with master craftsmen to convert their experience into technical parameters. Huasheng Alumina expanded the detection parameters for processes like evaporation, digestion, and sedimentation. Zhongtian Technology identified 50 typical scenarios covering 9 production links. Where does the data come from? From decades of sensor records, laboratory reports, and operation logs accumulated on the production line. Without this historical data, the large model has nothing to learn from.

Step 2: Train an industry-specific large model. General-purpose large models cannot understand industrial data. Yongnian Special Steel partnered with Northeastern University to develop an intelligent steelmaking large model. Shandong Haihua used the Inspur Haiyue large model. Huasheng used the Chinalco Kun'an large model. Zhongtian Technology independently developed the Tianji large model. The names differ, but the approach is the same: use a general-purpose large model as the base, inject industry data and process knowledge for fine-tuning, and teach the model the industry's language.

Step 3: Embed into the production line for real-time decision-making. The trained model isn't left in the lab for benchmarking; it's connected to the production line's control system. Yongnian Special Steel's large model processes tens of thousands of sensor signals per second, adjusting the blowing curve in real-time. Shandong Haihua's intelligent agent monitors equipment vibration data in real-time, automatically issuing alerts for anomalies. Huasheng's model provides prediction results in 3 minutes, allowing main control operators to adjust parameters accordingly. Zhongtian Technology's fiber routing model directly outputs the path plan.

Step 4: Master craftsmen transition from operators to annotators. The case of Shandong Haihua is clearest: engineers repeatedly consulted with master craftsmen, converting their crafts like "identifying faults by sound" and "judging wear by vibration" into technical parameters. The master craftsmen's skills were not discarded; they were translated into numbers the model could understand. The process of accuracy climbing from 60% to 95% was one of continuous experience extraction and calibration. The key point is that the master craftsman's experience never needed to become text; it only needed to become data. The judgment accumulated over 30 years of observing flames cannot be written into an SOP. However, the sensor records from 200,000 furnaces have completely captured the causal relationship of "when to stop blowing under what conditions." Text is the most convenient data format for feeding models, but it's not the only one. Temperature curves, vibration waveforms, chemical compositions, current parameters—these structured signals naturally present on the production line are more precise than text and easier for the model to digest. This is why a craft like "observing flames and listening to sounds," which seemed least likely to be optimized by AI, has been the first to succeed: it doesn't lack data, it just lacked someone to feed the correlation between the data and the results to the model.

From resistance to reliance.

The reaction of the master craftsmen is an unavoidable issue. Having a skill honed over thirty years suddenly taken over by a machine inevitably causes psychological turmoil. The deployment timeline at Shandong Haihua offers some clues. It started in February 2025. The first deployment in May had an accuracy of only 60% to 70%—at this stage, the master craftsmen were probably the most skeptical. If the model is less accurate than a person, why trust it? By August, accuracy had improved to 90%, and the model began to outperform humans. As it climbed past 95%, the master craftsmen's attitude likely shifted. This process has no shortcuts. When accuracy is below 60%, the model is just a reference. At 90%, craftsmen start comparing the model's results with their own judgment. Above 95%, craftsmen might actively look to the model for guidance. Huasheng Alumina's prediction accuracy exceeds 80%, and its dissolution aK accuracy surpasses 90%—manual testing takes 6 hours, but the model gives a prediction in 3 minutes. Accuracy is still improving, but the real-time capability is already overwhelming. Zhongtian Technology's approach is a "pilot-verify-rollout" closed loop. Once a plant succeeds, it's evaluated and quickly replicated to other plants. This ensures that master craftsmen in every factory have the opportunity to participate in the model's debugging and validation, rather than passively accepting a system imposed from above.