Summary

极蜂科技 is betting not on “industrial AI that can chat,” but on digital workers that can predict operating conditions and directly control production lines. 王筱圃 draws a clear line between 2 types of foundation models: large language models are built for generation and conversation with people; time-series foundation models follow the logic that “the past affects the present, and the present affects the future,” making them tools for prediction and decision-making—perhaps more suited to “talking to the future.” The core investment question is not just model scale, but whether the same combustion, electrochemical, synthesis or polymerization mechanisms can be reused across industries.
The waste-incineration case offers the most complete commercial validation so far: 1 Agent can fill a workstation that previously required 3 to 4 veteran operators working shifts. 极蜂 entered with its combustion-focused G1 baseline model and, despite having no prior waste-incineration experience on the team, put its first digital worker on shift in under 3 months. Agent uptime exceeded 99.9%, main-steam flow per ton of waste rose by 5 percentage points, and the gain translated into roughly RMB4M–5M in additional annual revenue; the customer subsequently bought more digital workers. “4 veterans used to rotate through the chair; now the chair is empty” (“原先有 4 位师傅轮流坐在椅子上，现在椅子是空的。”).
The most tradeable business model is not an amorphous Result as a Service, but billing by the digital worker’s hours on duty. A 5% capacity gain or 10% energy reduction can be disputed as a function of feedstock, load and other variables; logged online hours are an absolute, auditable figure. As long as a digital worker costs less than the 3 to 4 human workers it replaces, procurement becomes a familiar labor and staffing decision rather than a technology retrofit. That lowers customers’ upfront investment while giving the vendor recurring revenue and higher LTV.
Whether 极蜂 can reduce the project-based nature of the business depends on generalizing the same process across industries, not repeatedly customizing for 1 industry. The team has just over 20 people—later specified as 23 full-time employees—yet covers 7 industries and 15 core process sections. The key is to break workflows into configurable atoms and abstract metallurgy combustion, waste incineration and other settings at the level of first principles so they can be served by the same foundation or technical model. 王筱圃’s entrepreneurial correction is that vertical depth remains necessary, but the company “cannot focus entirely on 1 industry,” or a startup’s risk becomes tied to a single industry cycle.
For industrial Agents, the first deployment barrier is trust; model capability comes second. The product begins in “shadow mode,” observing the same data and making the same judgments as veteran operators without touching the equipment. After the switch, operators can inspect the chain of thought, control outputs and operating conditions before gradually moving from watching the screen to handling other work away from the post. The reasoning process, decisions and execution actions are all logged, with safety-instrumented systems, equipment interlocks and post-event attribution providing the backstop; as of the interview, no serious production incident had been caused by a digital worker on duty.
极蜂 is choosing to compete around the automation giants, cloud vendors and general-Agent players by focusing on the last mile “inside the workshop, before the control loop.” Siemens, Honeywell, Schneider Electric, SUPCON and Hollysys extend advanced process control; Huawei and Alibaba Cloud are better positioned for compute, model platforms and full-stack transformation of corporate operations; general Agents are stronger in knowledge bases, office work, supply chains or industrial design. 极蜂’s narrow opening is direct interaction with equipment, electrical automation and production processes, with a promise of faster deployment, lower cost and quicker payback.
If single-workstation validation can expand into multi-Agent coordination, the endgame shifts from saving labor to reorganizing factories and even supply chains. Once digital workers take over high-attention, repetitive process control, human technical workers can move into design, supervision and innovation. Provided industrial networks and production safety are secured, upstream and downstream Agents could also adjust plans and capacity in sync with end demand. 王筱圃 says that is what could qualify as an “Industrial Revolution-level change,” but millisecond response times rule out multiple Agents voting in real time; a more stable architecture would pair a central coordinating Agent with execution Agents.

Deep dive

1. 极蜂 Bets on Industrial Digital Labor with a Team of Just Over 20

王筱圃 was born in 1988 and graduated from the University of Science and Technology of China. Before starting the company, he worked in industry on the R&D and delivery of AI-enabled industrial digitization products, and also conducted foundational research in computational neuroscience and brain science. He summarizes that period as being “a happy wage earner” in industry, before eventually leading an R&D team.
极蜂科技 closed a seed round in the tens of millions of RMB when it was founded and is preparing to launch a Pre-A round; current orders have reached the tens of millions of RMB. 王筱圃 says margins are in line with, and relatively high versus, comparable AI companies, but R&D spending is also heavy. The team has just over 20 people, with 23 full-time employees mentioned later.
The product is not a platform sold to factories for customers to configure themselves. It is “dispatching our Agent digital workers” to replace or assist human technical workers in managing and controlling continuous production lines. The wording also foreshadows the charging model: the company is selling verifiable digital labor, not software seats.

2. The Job of a Time-Series Foundation Model Is to Make the Future Predictable, Manageable and Optimizable

王筱圃 explains time series through stock prices, running routes and estimated navigation arrival times: observations have an order in time, and there is a causal structure in which “the past affects the present, and the present affects the future.” The model uses what has already happened to predict what comes next, then helps people make the best or near-best decision.
Time-series research did not begin with the recent foundation-model boom. He traces the field’s recognized starting point to the 1970s. Navigation systems that estimate arrival times accurately and route around congestion are an everyday example of time-series-model capability that ordinary users already rely on without necessarily recognizing it.
Large language models and time-series models are both foundation models, but their data and training objectives differ. Large language models learn from corpora and public information, with generation as the goal; time-series models learn from data structured by time, with prediction and decision-making as the goal. “Large language models are for talking to people; time-series foundation models may be more about talking to the future.”

3. Generalization Connects Energy Dispatch, Marketing Forecasts and Industrial Control into One Market

In power grids, recurring patterns such as residents turning on air conditioning at night and offices consuming electricity during the day allow time-series models to forecast demand and dispatch energy. In healthcare, the same class of models can monitor electrocardiograms. 王筱圃 also cites transportation, finance, supply chains and marketing, emphasizing how broad the application base already is.
The marketing example is worth preserving as a chain of logic: a company needs to identify which products may see higher demand as advertising increases exposure; demand forecasts can then support supply-chain decisions and factory scheduling. If mint-scented shampoo is expected to sell best, the company increases orders for it and cuts back on slower scents—“using a view of the future to support decisions in the present.”
Time-series foundation models perform the same class of tasks as traditional time-series models. The difference lies in scale and the ambition to generalize: large parameters, large datasets and large-scale training are meant to produce a pretrained model rather than requiring a separate small model for every domain. That is why 王筱圃 connects the category to the AGI path.
According to 王筱圃, an overseas example is a UK startup whose name is apparently Applied Computing. Its Orbital foundation model targets process management in energy, chemicals, oil and petrochemicals, and it has reportedly just completed a relatively large seed round of close to RMB100M. In China, Huawei’s Pangu forecasting foundation model, or Pangu time-series foundation model, has been applied to steel, nonferrous metals, cement and weather. 王筱圃 believes investor interest also reflects the possibility of improving industrial efficiency while reducing emissions and carbon intensity.

4. The Core of an “Industrial Jarvis” Is Not Designing the Armor, but Making It

王筱圃 uses Iron Man’s Jarvis to distinguish several interfaces in industrial AI. Designing the armor based on a requirement is AI-assisted industrial design; controlling the armor in autonomous combat resembles motion control for embodied robots; operating production equipment to actually manufacture the design is industrial production-process management and control.
极蜂 chose the third interface: driving production-line equipment to continuously convert raw materials into the products a company requires. It does not design the armor’s shape or control its movements; it produces the iron, aluminum, copper and other basic materials needed to make the armor, turning process design into a stable industrial product.
That positioning requires the product to enter the live production loop rather than stop at knowledge Q&A or an analytics dashboard. The Agent’s output ultimately has to reach equipment controls and intervene in control loops, making stability, response speed, observability and accountability part of the product itself rather than post-delivery add-ons.

5. The Challenge in Process Industries Is Not Scheduling, but Holding Stability, Efficiency and Quality Around the Clock

The host initially described the setting as continuous production. 王筱圃 further distinguishes discrete from process manufacturing. The chassis, four doors and two hoods, motors and other assemblies produced by Xiaomi Auto, NIO and Li Auto are examples of discrete manufacturing: workflows, orders, BOMs and raw-material organization are complex, but individual production processes are relatively simple.
Metallurgy, chemicals, new energy and materials more often involve process manufacturing: raw materials flow in continuously and products flow out continuously, with the line generally not stopping once it starts. The central task is therefore no longer order planning, but continuous management of stability, efficiency and quality.
Process operations combine physical reactions, chemical reactions and energy flows, making them inherently unstable. Traditional automation relies heavily on rule-based logic, so a change in operating conditions can cause quality to deteriorate, energy consumption to rise or raw materials to be wasted. The flexibility missing from rigid production lines has historically been supplied by technical workers.
Veteran operators can combine textbook knowledge, master-apprentice experience and what they see, hear and feel on the floor to respond in real time. That is why process manufacturing has long depended on technical workers to participate in production, manage production and control production: equipment can execute mechanically, but struggles to handle complex changes in operating conditions on its own.

6. Digital Workers Should Preserve Human Value While Removing Fatigue and the Apprenticeship Bottleneck

王筱圃 does not build the replacement narrative on the idea that people have no value. He instead argues that human flexibility, tool use, synthesis and communication must first be captured, then amplified by Agents. “We are not denying human value; we are saying that people are highly valuable.”
The first human limitation is sustained attention. A technical worker may stand watch for 7 or 8 hours in a shift, and a momentary lapse in a chemical plant can trigger a serious safety incident. The team has observed that inefficient operation occurs more often during lunch breaks and overnight shifts because people get tired and cannot remain highly focused at all times.
The second limitation is the replication of experience. A veteran operator’s decade-plus of internal process modeling can be passed on only through one-to-one oral instruction, which is slow, lossy and difficult to verify. A digital worker aims to turn knowledge, experience and real-world responses into an organizational asset that can be reused and continue to improve.
The replacement scope is not limited to a single equipment control point. It covers operations, maintenance, safety, planning and energy dispatch, following a path “from assistance to replacement, and from replacement to surpassing.” 王筱圃 emphasizes that “local optimization is not global optimization”; the real goal is coordinated optimization across the full process from raw materials entering the workshop to finished goods leaving it.

7. A Qualified Digital Worker Must Close the Loop from Observation to Thought, Decision and Execution

The first defining trait of a human worker is observation: reading production metrics in the control system as well as camera feeds and video streams. Some equipment even uses high-definition cameras designed for high temperatures and corrosive environments to film the reaction directly. A digital worker must have an equivalent observation and sensing interface.
When temperature drops sharply or equipment speed becomes abnormal, a worker draws on first principles, academic knowledge and a master operator’s experience to diagnose the cause, then decides which parameter to adjust, which equipment to stop or which process to activate. 王筱圃 describes this as thought and decision-making, not a simple threshold alert.
The final link is execution: “One important human trait is the willingness to think and act.” After reaching a decision, the worker must influence production through supervisory software, lower-level electrical control systems and equipment control loops. If the model can only make recommendations but cannot act, it has not replicated the full behavioral logic of the job.
极蜂 assigns the different parts of this brain to different models. Large language models handle knowledge decomposition, information parsing and natural-language generation; time-series foundation models handle pattern recognition, causal reasoning, numerical response and future-condition forecasting; an Agent workflow then packages them into a controlled, explainable loop with memory, state and explicit steps.

8. The 合谷 Platform Serves 极蜂’s Engineers First, Rather Than Asking Factories to Build Their Own Agents

The 合谷 industrial-intelligence platform configures Agents and the overall workflow, while the industrial time-series foundation model serves as the underlying model for some workflow nodes. Their relationship resembles that between a general-Agent platform and a foundation model, except that the system executes industrial production processes.
王筱圃’s view of current To B Agents is that end customers are better served by ready-configured intelligent applications than by building them from scratch. Intervening in production requires specialized process knowledge, parameter tuning and stability safeguards. Inside a factory, it is also unclear whether the chief engineer, process expert or frontline worker should be the user responsible for configuration.
The platform’s other key value is serving 极蜂 itself. With 23 full-time employees supporting 7 industries and 15 core process sections, the business would fall back into project-by-project customization unless industrial processes could be atomized, configured and recombined quickly, then integrated through a highly generalizable model. “If we cannot put this digital worker online quickly, none of it matters.”

9. Veteran Experience Matters, but the First Product Cannot Be Built on Interviews

王筱圃 says technical workers may account for only 6.5% of the overall factory workforce. They are typically better educated and are themselves process or electrical-automation experts. They are not incapable of understanding new technology; once the product logic is explained clearly, most veteran operators will openly assess whether it actually solves a problem.
极蜂 deliberately avoids relying too heavily on interviews because information is lost when it is expressed through language, and different veteran operators may give conflicting answers about the same operating condition. If oral accounts are converted directly into rules, the result is likely to be an electronic operations manual whose foundation remains error-prone rule-based logic.
The first version of the digital worker therefore rests on process first principles and the idea that “data does not lie”: the physics and chemistry of combustion, electrochemistry, synthesis, polymerization, hydrogenation and cracking, combined with what veteran operators have done in the past and how the line responded. Individual experience is gradually introduced after deployment as short- and long-term memory, where it can correct the system.

10. 极蜂 Moved from Black-Box Foundation Models to Agents to Gain Control and the Ability to Improve

The team initially pursued the route of “putting the model inside existing software”: the time-series foundation model and large language model received multiple inputs, directly generated multiple outputs and executed them. Like early attempts to have foundation models generate PPTs or translate text directly, the engineering results were unstable, decisions became black-boxed and there was no reliable workflow.
王筱圃 summarizes 3 problems: lack of control, lack of observability and difficulty making adjustments. At the beginning, there was not even a CoT to inspect. An Agent can instead expose, node by node, “what it thought of, what it relied on, what it did and what it may do next,” allowing every connection’s input and output to be located.
An industrial digital worker also cannot “start one way and stay that way forever.” 极蜂 retains recent key events, then decides which short-term memories affect near-term performance and which should be promoted to long-term memory. At this stage, the capability is “not that mature”; most of the work involves adapting mature open-source frameworks to time-series data flows.
The team cares more about effective memory and explicit management than about network stability, fault tolerance or concurrency, because “I’m afraid it will remember the wrong things.” They need to know what the Agent has stored. Multi-Agent coordination likewise cannot rely on real-time voting—production lines require millisecond response, and “the daylily would be cold by then”—so a central coordinating Agent paired with execution Agents is more likely.

11. Before Entering a Factory, the Digital Worker Carries “Long-Term Memory from Having Been in the Trenches”

The host’s key question was this: people grow through practice, but a digital worker has never practiced in the factory before delivery, so how can it have any initial capability? 王筱圃’s answer is to reduce factory-level cold starts through process-level generalization rather than begin with a blank model and wait for on-site data.
Digital workers for combustion, metallurgy, electrochemistry, synthesis and polymerization carry long-term memory accumulated in other identical or similar settings. 王筱圃 compares this with a medical student who studies multiple disciplines at school and rotates through multiple departments during an internship: the parameters of a new factory may be unfamiliar, but the worker brings experience from “the same process in different settings and different industries.”
This also eases the shortage of industrial data. Companies cannot be crawled for data the way internet corpora can. If every waste-incineration plant or synthetic-ammonia plant had to accumulate 6 to 8 months of data before training, the economics would be untenable. Abstracting combustion data from different settings at the level of first principles and using it to serve the same foundation or technical model is the way to maximize scarce samples.

12. “Shadow Mode” Lets Veteran Operators Build Trust Without Changing Their Work Habits

The host did not avoid the opposing question: if veteran operators know the model may replace them, why would they cooperate? 王筱圃 first reiterates that the product does not depend on them “pouring out their innermost oral experience,” but acknowledges that deployment is impossible if the end users resist it.
His product judgment is that many To B AI products are still delivered by educating the user: customers are asked to learn prompts, knowledge bases or Agent workflows, creating a high barrier. A genuinely good To B product should fit naturally into the existing workflow—“make the product adapt to the customer,” rather than first asking frontline workers to adapt to the product.
Before going live, the digital worker runs in shadow mode: it sees and hears the same information as the veteran operator and reasons alongside them, but does not control the equipment directly. Operators can judge its performance by whether its decision direction matches theirs and whether the results are credible. Common evaluations include: “It’s definitely not as good as me, but basically reliable,” and “It’s close enough.”
After the formal cutover, veteran operators initially continue watching the screen, checking the chain of thought, control outputs and actual operating conditions, but no longer operate the equipment by hand. If the system remains reliable, people naturally reduce their attention, do other work or even leave the post. User stickiness comes from making the job easier, not from extensive training or persuasion.

13. Industrial Accountability Has Clearer Boundaries than Autonomous Driving, but Safety Still Requires Full-Chain Logging and Hard Backstops

The host asked who would be responsible if the system made a mistake during the early cutover period. 王筱圃 contrasts industrial production with autonomous driving: a road accident may involve another party’s violation, complex road conditions or a driver who failed to keep their hands on the wheel, making the boundary unclear. Industrial production has a stricter management system that can distinguish raw-material, equipment-control and other causes.
The Agent’s reasoning process, judgment and industrial-actuator actions are all written to logs for post-event attribution. The site also retains hard safeguards such as safety-instrumented systems, equipment interlocks and emergency procedures. The model records the decision process; traditional safety systems provide the production-safety backstop.
As of the interview, 极蜂 had “not had a single relatively serious production problem caused by a digital worker being on duty.” 王筱圃 does not claim future risk is zero. If an incident occurs, clear responsibility boundaries and traceability will identify the cause, after which the product can continue to be improved.

14. DeepSeek Opened the Factory Door but Also Created Suspicion of “Showcase Projects”

After DeepSeek went mainstream during the 2025 Lunar New Year, large numbers of factories began looking for domestic large-language-model solutions. All-in-one machines combining Huawei Ascend computing platforms with DC or other solutions were briefly hot sellers. Many of 极蜂’s customers had already bought such machines for knowledge graphs, meeting-minutes generation or financial-chart production.
The problem was that although these applications had “used AI,” customers could not yet feel a direct financial return. 极蜂 was therefore also lumped in with another large-model vendor selling all-in-one machines or knowledge graphs. Decision-makers worried that it was merely a “showcase project, a leadership project,” rather than a system entering production for real.
王筱圃’s sales strategy is to downplay the identities of “large-model company” and “Agent company,” first explaining which production problem needs to be solved and “turning the logic of a technology purchase into the logic of an investment”: when will the customer recover the cost of the digital worker, and how much additional profit will it generate after payback?
At the general foundation-model layer, 极蜂 has used DeepSeek and Qwen mainly to process unstructured documents such as process packages and daily work reports, converting them into structured decision information while ensuring that professional Chinese generation does not produce irrelevant hallucinations. Edge-side VRAM supports only 4-bit or 8-bit quantized versions; Qwen uses a 14B-parameter model, while the source text elsewhere refers to “DPC possibly being single-digit,” without clarifying which model that refers to.

15. Industrial Systems’ Long Lags Forced 极蜂 to Train Its Own Time-Series Model

王筱圃 says time-series foundation-model research began receiving broad academic attention in 2021. He cites Informer, Google’s TimesFM, Lag-Llama and Timer-XL from a Tsinghua team in early 2025. Their common direction is Transformer-based networks for time-series forecasting, zero-shot transfer and optimized decision-making.
Off-the-shelf models struggle to cover industrial data because complex physical systems are often “long-lag systems.” Setting an air conditioner to 24°C does not produce an instant change; the temperature moves gradually from 26°C to 25°C and lower. By the time it reaches the target, the person may already feel cold and turn it back up. These cross-time relationships are stronger and more complex in industry.
Industrial data also intertwines multiple variables and covariates. The model must identify features and relationships across time series. 极蜂 acknowledges that it will “stand on the shoulders of giants” and reference mature designs, but still needs to train a model from scratch that adapts to industrial lag characteristics, while continuously adjusting the network and structured memory.

16. The Waste-Incineration Project Validated Generalization, Labor Replacement and Higher Output at Once

王筱圃 explains that municipal solid waste is now generally incinerated rather than landfilled, with its calorific value converted into electricity through steam and turbines. The problem is unstable waste sorting: food waste has low calorific value, other waste streams differ, and the energy content of the same weight of feedstock can fluctuate sharply from day to day.
The plant wants both to “burn it all the way through” and extract more heat from limited waste, and to reduce the labor required for a 4-shift, 3-rotation schedule. A workstation is typically covered by 3 to 4 equipment operators, energy managers or safety personnel rotating through it, because continuous 24-hour production cannot lose supervision or intervention.
极蜂 arrived with its combustion-focused G1 baseline model, first using a small amount of new data to adapt to the site’s parameter ranges, then fine-tuning for the greater volatility of solid waste versus natural gas or coal gas. No one on the team had previously worked on waste incineration, yet it put its first digital worker on shift in under 3 months and had it stably handle both normal and temporary operating conditions.
The workstation Agent achieved more than 99.9% online uptime and was essentially no longer operated by a person. Main-steam flow per ton of waste rose by 5 percentage points, translating at the grid-connection tariff into roughly RMB4M–5M in additional annual revenue. The customer saw both “the chair is now empty” and the profit, then repurchased more digital workers to cover additional posts.

17. 4 Hours of Manual Control at a Chemical Plant Was the Most Direct Test of the Agent by a Veteran Operator

During a shadow-mode deployment at a chemical plant, a veteran operator suddenly asked a 极蜂 employee with a computer-science background and no chemical expertise to control the equipment manually based on the Agent’s reasoning. The employee sat nervously for 4 straight hours without drinking water or using the bathroom. The veteran initially supervised from behind, then went outside for a cigarette, prompting the employee to ask for help on the internal IM system: “The master operator left. It’s just me now. I’m terrified.”
The employee ultimately completed the operation smoothly based on the Agent’s reasoning, until the next-shift operator took over. 极蜂 later explicitly prohibited this during shadow mode, but 王筱圃 viewed it as an informal test: the veteran eventually left to smoke, and the episode showed how the digital worker’s judgment gradually earned professional trust.

18. Hourly Billing Solves Attribution, while Vertical Control Capability Avoids Platform Giants

王筱圃 agrees with the direction of Result as a Service, but says industrial results are difficult to attribute strictly. After deployment, a 5% capacity increase might be credited by the customer to better raw materials; a 10% year-on-year decline in energy consumption might simply reflect lower monthly output. If pricing is tied entirely to such results, the parties can end up arguing indefinitely.
The digital worker’s advantage is that “the worker-use logic is labor logic; it is staffing logic.” Online service hours cannot be denied. As long as the digital worker meets labor standards, fulfills the post’s KPI and remains reliably on duty, it can be billed by the hour. Its monthly cost must also be lower than the wages of the 3 to 4 human workers in the same post, combining low upfront investment for the customer with continuous collections and LTV growth for the vendor.
The first group of adjacent competitors includes Siemens, Honeywell, Schneider Electric, SUPCON and Hollysys. Their direction is to upgrade traditional electrical automation into advanced process control, with the management boundary still centered on the existing automation interface. 极蜂 instead uses “replicating 1 person” as the product unit, covering observation, judgment, coordination and execution.
Huawei and Alibaba Cloud are better suited to provide compute, model platforms, ecosystem partners and broad transformation of enterprise operations. Fourth Paradigm, MiniMax, StepFun, Zhipu Qingyan and other general-Agent players are stronger in knowledge bases, office work, supply chains or industrial design. 极蜂 occupies the position inside the workshop, interacting directly with production lines, equipment and electrical automation, while emphasizing faster deployment, better cost performance and shorter payback periods.

19. Digital Workers Will Ultimately Rewrite Job Boundaries, and Startups Must Rewrite Their Scaling Logic Alongside Them

王筱圃 expects human technical workers to move from operators into design, supervision and innovation after digital workers take over mechanical, repetitive dynamic control that requires sustained attention. The example from a hazardous-waste customer is that once a veteran operator is freed from routine monitoring, they can study how to extract iodine from waste when iodine prices rise, reconfigure the production line and capture new profit opportunities.
The more distant vision is for multiple digital workers to coordinate inside the plant and, once industrial networks and production safety are secured, connect across the supply chain. Changes in end demand could be transmitted in real time to upstream factories, which would adjust plans and capacity in sync. 王筱圃 believes this leap in response speed is what could reach an “Industrial Revolution-level” impact.
Cross-disciplinary programs such as “intelligent chemical engineering” have already emerged, combining chemical processes, big data and artificial intelligence. If he could return to university, 王筱圃 says the skill he would train earlier is defining requirements: whether writing a prompt, managing employees or conducting research, every CEO must clearly articulate the result, the underlying need and the first principles.
Entrepreneurship has also changed his view of products. An Agent should fit into the customer’s workflow and co-evolve with the customer, becoming “more tailored the more it is used,” rather than preserving a rigid buyer-supplier relationship. A company should focus on solving 1 type of scenario problem while using foundation-model generalization to cover the same process across different industries, avoiding dependence on a single industry cycle. Internally, the company has fully embraced AI tools such as Canva, Figma and Copilot, turning its team of just over 20 into a group of versatile generalists—“keep your feet on the ground while looking up at the stars”—as job boundaries become increasingly blurred.

Verification Notes

The reference transcript renders the relevant model as “DPC” without clarifying what it refers to; it has therefore not been attributed to DeepSeek. The uncertainty around the name Applied Computing has also been preserved as “apparently.”

50. Six Months After CES Went Viral: Ropet on Weathering the Storm, Copycats, and the Meaning of Companionship Inside the White House Tech Dinner, Weak Jobs Report, Tariffs Court Challenge, Google Wins Antitrust