Behind the MiniMax 10x Team, industrial AI isn't hitting a technological bottleneck—it's encountering the real-world chain of responsibility.

Author: Yan Jun

Source: 36Kr

Introduction: Large models are getting better at generating answers, but the real challenge on the ground is how those answers are adopted, explained, and held accountable. The significance of MiniMax’s 10x Team lies not just in hiring experts—but in model companies beginning to seek interfaces into the industry’s accountability chain.

Last year, I truly felt left behind by the times.

After twenty years of judgment and intuition, one day the ground suddenly vanished—not because anything was done wrong, but because the world had changed its scoring system.

Large models, agents, AI coding—one wave after another. Everywhere you look, it’s “tenfold efficiency gains” and “industry transformation.” At first, I was excited. But as the excitement faded, all that remained was a sense of emptiness.

So I started catching up—on AI, and on my own judgment. It wasn’t that I suddenly fell in love with technology, but rather that I realized I could no longer just stand on the outside looking in. Unexpectedly, at this stage in my life, I enrolled in a computer science master’s program, retook courses, read papers, and pushed myself to understand the technology and algorithms.

Both abstract and real, and it feels pretty good.

The more I use AI, the more I realize one thing: it can write, calculate, summarize, and is very skilled at handling problems with clear questions and well-defined boundaries. But in the real world, many problems don’t even have a clear question to begin with.

At the moment when a decision really needs to be made, AI’s advice is always—seems right, but lacks the “but.”

There is no such statement: "This isn't the right time—launching now would only make everyone look bad."

There is no such statement: "This risk is disclosed for compliance, but who takes responsibility if something actually goes wrong?"

There is no such statement: "You can't propose this plan like that—once you do, they'll know you don't understand who's in charge."

There is no such sentence: "This phrase is fine in the PowerPoint, but it will cause problems in the contract."

AI won't say these things—not because it isn't smart enough, but because it doesn't bear the consequences of being wrong.

So this article does not discuss whether AI will replace humans. What I want to probe further is: as answers become cheaper and cheaper, what kind of experience still holds value? When AI can write proposals, who decides whether a proposal can be delivered? And for those who once made judgments based on real-world experience, how can they still participate?

After seeing the message from the MiniMax 10x Team, I suddenly realized that the issue I had been pondering repeatedly over this period had found a real-world illustration within the industry.

This is not a new model, nor a funding announcement. Public information shows that the MiniMax 10x Team engages experts in industrial software, game engines, chip design, finance, and accounting, operating more like a “industry research partner” model: domain experts actively define problems, co-develop evaluations and workflows, and directly feed real-world industry insights into the model.

What truly matters is not how sensational this event is, but the signal it sends: industry AI must move to the front lines, requiring more than just stronger models—it must integrate real-world industry problem definitions, feedback loops, and accountability chains.

This is the break:

The cost of generating answers is falling rapidly. The cost of getting answers accepted, explained, and held accountable hasn’t decreased at all.

01 Why did AI get the answer right, yet the answer still didn't enter the chain of accountability?

AI has no real identity and no real losses. It doesn’t lose customers due to a misjudgment, isn’t held accountable for incorrect decisions, and doesn’t need to explain at review meetings “why it made that call.”

Without real losses, it cannot learn the kind of judgment that comes only from experience.

So consulting industry experts isn't just about filling knowledge gaps—it's about bringing real-world insights in: what questions are worth asking, what boundaries must not be crossed, what judgments can move forward in the process, and what consequences must be clearly stated upfront.

Experts are not patches for AI’s knowledge, but the nerve endings that connect AI to real-world industry applications.

In the past, writing proposals, making judgments, and taking responsibility were all bundled together. Now, large models have made the “writing answers” part inexpensive, while the abilities to decide whether answers are adopted, explained, or held accountable have become valuable again.

I call it the chain of accountability: the entire journey of an answer evolving from “seems right” to “someone dares to use it, submit it, sign it, and take responsibility for it.” In high-value, high-risk, and highly regulated domains—such as finance, healthcare, legal, industrial, and government—the chain is longer and harder to complete.

Only when the large model reached the front line did it truly understand what it means to be responsible.

02 Four scenarios: AI gets everything right, but gets stuck just outside the answer at every step

The issue is not that the AI gave the wrong answer. The issue is that the answer doesn't fit into the chain of accountability.

Scenario One: What regulators are truly asking is not "Do you have value?" but "Who do I hold accountable if something goes wrong?"

Once, our former employer faced regulatory conflicts simultaneously in multiple cities. Internally, we prepared extensive materials: user data, compliance certifications, legal provisions, and economic contributions. Today’s large models could certainly craft a compelling narrative—highlighting technological innovation, urban efficiency, and the platform economy’s release of social value.

These statements are all correct. But in that scenario, they are not critical.

Regulatory and law enforcement agencies don’t care about those commercialized value claims. The only thing they really want to know is: If something goes wrong, who can I hold accountable? How do I explain this to my superiors?

What the government truly cares about is: What if a mass incident occurs? Who is responsible if a safety accident happens? When the platform expands rapidly but regulation lags behind, who bears the responsibility?

The final step is not to submit more documents, but to reframe the platform’s capabilities: data can help identify anomalies, order records can aid in accountability tracing, and the technical system should not merely be a target of regulation—it must also become a tool for regulation.

Only then will the other party see a clear point of contact: if something goes wrong, I know who to reach out to; if there’s an issue, I know how to investigate it; if I need to report, I know how to explain it.

AI can organize the material flawlessly. But it may not know where the interface is, or why that’s the real leverage in the entire conversation.

This is not a materials issue. This is a regulatory responsibility interface issue.

Site Two: Whether reform advances depends not just on the plan, but on whether everyone has an escape route.

Once, I participated in a competition for a local reform pilot program. The competitor had more substantial funding, a more comprehensive proposal, and an airtight logic—but they were eliminated.

Because their plan overlooked a core issue not listed on any evaluation form: during the reform process, if problems arise, can each person here come up with a plausible explanation?

It's not about shifting blame—it's about maintaining dignity.

Many reforms aren’t lacking in understanding their value, but rather in willingness to take one more step for a solution with unclear accountability.

But eliminating fear alone is not enough. More importantly, each participating unit must see exactly what they will gain from driving this initiative—not vague phrases like “jointly advancing reform,” but concrete benefits: one department gains an additional pilot case to showcase externally, another unit secures a measurable, named performance achievement, and this manager gains another opportunity to be mentioned by their superior.

Something went wrong—I won’t get into trouble. What can I gain after it’s done?

Together, these two statements form the true action switch.

The local government is not reading a business plan. It is determining: who will lead? Which departments will cooperate? Where will the budget come from? How will success be measured? And who will take responsibility if problems arise?

This is not an issue with the plan, but rather whether each participant can explain why they are driving it.

Scenario 3: Even the most comprehensive business plan cannot replace business judgment and investment responsibility.

Once, an entrepreneur brought a project to meet with a fund—the business model was clear, the market opportunity was large, and all the materials were in place. Today, with AI, large models can quickly generate a well-structured, even internationally polished, business plan.

But what funds truly care about is often not whether the materials are complete.

That day, the investor flipped through a few pages and asked just one question: “Are your customers driven by genuine market demand, or are they merely pilot participants enabled by policy windows? Will they renew their subscriptions next year without subsidies?”

This question appears to be asking the customer, but it is actually verifying two things at once.

One thing is the founder’s business judgment: Do you truly understand where your revenue comes from, why customers are paying, and whether they will continue to pay next year? Are you confronting risk, or are you simply covering it up with polished materials?

Another aspect is the investor’s responsibility: If I bring this project to the investment committee, can I clearly explain what the revenue quality is, how dependent the policy is, where the renewal risks lie, and what supports the exit strategy?

The answers aren't missing from the material. It's just that often, no one knows which line holds the real key issue of the entire meeting.

The investor had already seen that line. He simply wanted to know: Have you truly thought about this issue, or are you using a polished presentation to avoid answering something you haven’t figured out yourself?

This isn't about finding fault with the materials, but about verifying whether two chains of accountability can hold: whether the founder can be held accountable for operational outcomes, and whether the investor can be held accountable for their investment decisions.

AI can organize everything flawlessly. But it doesn’t know that sometimes, a too-complete set of materials is itself a signal: not yet ready to be truly questioned.

Materials are never the core. What truly matters is whether the quality of revenue can be verified, whether risks can be explained, and whether business judgment and investment responsibility can both be upheld.

Scenario Four: When a trade gets stuck, the real conflict often isn't in the terms—but in the "two sets of responsibility systems."

Another time, a tech project was on the table, and all parties said they wanted to move forward. The technology had barriers to entry, the client quality was solid, due diligence was completed, and the terms were nearly finalized. On the surface, signing the agreement was just one step away.

But this transaction was unexpectedly paused. No one explained why.

The RMB fund said: “We need to take another look at the structure.” The USD shareholders said: “We need to confirm the subsequent rights.” The founder asked: “Is there still room for valuation adjustment?” Everyone was using safer language to express their true concerns.

RMB funds are driven by local industrial goals, investment attraction targets, reinvestment requirements, and compliance pressures from state-owned assets—they need the company to serve the local region in some capacity. But USD investors are not here to serve the local region; they seek efficiency, exit, and DPI.

These are two sets of accountability systems, and structural tension between them is inevitable within the same company.

Later, instead of asking either party to compromise, the structure was redesigned: USD shareholders remained in the top-tier structure to preserve overall flexibility and unimpeded exit pathways; RMB funds entered specific business lines through regional subsidiaries, with local state-owned capital’s reinvestment and investment attraction obligations fulfilled at the subsidiary level. Two distinct frameworks operate independently, each functioning seamlessly within its own layer.

For the RMB fund, this is a memo for inclusion in the investment committee—its purpose is not to prove “there is no risk,” but to enable them to answer: Why am I investing? What risks do I understand? And how are these risks being managed?

For dollar shareholders, the integrity of the top-level structure remains intact, and the exit path has not been altered.

No one gave in. But everyone got what they truly needed.

The essence of negotiation has never been persuasion, but rather a reallocation of interests.

03 Two public signals: AI can assist, but cannot take responsibility for humans

Looking back at these four moments, AI could get everything “right”—the information was correct, the logic was sound, and the terms were accurate. But each time, the crucial step that actually moved things forward happened outside of AI’s response.

This is the true boundary of industrial AI today: it’s not that it isn’t smart enough, but that it doesn’t bear consequences.

It doesn’t need to explain this decision at a review meeting three years later, or answer the investment committee on why it made that judgment at the time. Real-world decisions aren’t just about choosing an answer—they’re about choosing a consequence you’re willing to bear.

The judgment made after three seconds of silence in the meeting room isn't something the algorithm couldn't calculate—it's something the algorithm doesn't yet know: that during those three seconds, someone was worried.

Professional expertise is becoming cheap. Industry judgment is not.

The judicial context most clearly highlights this issue. The 2026 Supreme People's Court work report explicitly stated that it will actively and cautiously develop AI-assisted adjudication systems, maintaining an "auxiliary" role, with judges remaining the sole responsible parties for judicial decisions.

It’s not denying AI, but rather defining its role: it can assist, but cannot replace the person ultimately responsible for judicial accountability.

Another case occurred at the Tongzhou Court in Beijing. In a commercial dispute, the agent submitted an AI-generated “reference case” without verifying it personally; the court rejected it and criticized the action in its judgment.

This case is small but typical.

The issue isn't just the generation quality—it's that the intermediate verification and confirmation node was skipped. The problem isn't whether AI can produce content that looks professional, but rather who verifies, submits, signs off on, and takes responsibility for this content before it enters the real system.

04 Who Will Become More Expensive? Three Types of People and a New Ability

In the past, the value of industry services was often bundled together: data, relationships, experience, judgment, and responsibility—all charged as a package.

AI will break apart this mess.

Information depreciates first, expressions depreciate next, and general analysis also depreciates. What truly endures are the judgments that can enter the chain of responsibility.

This is also my understanding of the engineering judgment.

Engineering judgment isn't about feeding knowledge into a model—it's about breaking down "what can be approved, what cannot be signed, and what risks must be clearly disclosed upfront" into standards that systems can verify and organizations can adopt.

In the past, these judgments resided in the intuition of experts; in the future, they must be broken down into systems.

Behind this is a new capability: the ability to turn judgments into accountable workflows.

It’s not just about understanding the industry or knowing how to use AI—it’s about breaking down real-world boundaries, risks, counterexamples, accountability points, and acceptance criteria into processes that models can learn, systems can verify, organizations can adopt, and issues can be clearly explained.

Looking along this line, the three types of people who will likely become more valuable in the future are:

The first type consists of people who can break down experience into standardized components.

It’s not enough to just say “I have experience”—you must be able to clearly explain: what can and cannot be submitted; what risks must be disclosed upfront; what solutions look good on paper but will fail in practice. When someone can break down their experience into standards, counterexamples, evaluations, and checklists, they become the critical interface through which models enter real-world industry operations.

The second type: people who can simultaneously understand multiple accountability systems.

Governments, RMB funds, USD funds, and industry clients interpret the same set of facts in completely different ways. Those who can translate between these systems are not merely conveying messages—they are reassigning responsibility.

The third category consists of companies that can embed judgment into their workflows.

What is truly difficult to replace are systems embedded in customers' responsibility processes—knowing how a report goes through approval, how a risk is documented, and how a compliance judgment is adopted by the organization.

What the customer ultimately pays for is not "how well the AI writes," but whether this judgment makes me confident enough to use it.

The value of industry AI is short-term in models and tools, medium-term in vertical agents, and long-term in workflow systems that integrate into customer responsibility processes.

"Generating answers" will become increasingly like electricity and water—essential, but no longer a competitive advantage. The real opportunity for high profits lies in the industry workflow layer.

Conclusion: Those few seconds of pause carry not only logic, but also responsibility.

The real issue isn't about replacement, but about how humans and AI can redefine their division of labor. AI provides speed, structure, and scale; humans provide meaning, boundaries, and responsibility. Only when both are combined can a judgment move from “seems right” to truly become reality.

AI doesn't make real experience obsolete—it forces everyone to upgrade their experience.

Experiences that remain only in the mind are quickly diluted; those that can be broken down, articulated, verified, and iterated become the true fuel for human-machine collaboration.

Large models operate in a world of fixed rules. But the real world is alive—it rebounds, reinterprets itself, and transforms every "correct answer" as it is put into practice.

In this world, merely processing information is not enough. You must also perceive meaning—understanding who this matters to and why. You must also judge value—determining whether this answer is worth accepting, worth signing, worth entrusting.

This intuition is not calculated from data; it comes from repeated experimentation, facing consequences, and recalibrating within the industry environment.

So even as models grow stronger, they still need someone to tell them: what truly matters on the industry front.

This is not the transmission of knowledge, but the translation of meaning and value. Those who have made judgments on the ground, shouldered responsibility, and learned from mistakes are not merely sources of knowledge for AI—they are its interface to the real world.

This article is written not just for AI professionals, but for everyone who still makes decisions in the real world, has stumbled, and bears responsibility.

These experiences may be difficult to put on a resume or directly understood by models, but they are precisely the missing interface that industrial AI needs to enter the real world.

Large models will become stronger and faster. But on the front lines, the real world won’t automatically function just because an answer is logically consistent. It will respond, rebound, and make every decision carry consequences.

There will always be someone who pauses for a few seconds before clicking confirm.

In those few seconds, it wasn't just logic.

Also responsibility.