A Conversation with 张帆: From 智谱 to 元理智能 and AI’s ToB Opportunity

Summary

• 张帆 is betting on ToB not because he is bravely charging into a known tiger’s den, but because he believes the competitive structure of the AI era has changed. In the early internet, online space was empty and startups could claim territory at speed; today, “offline is full, and online is full too.” Even if a consumer Agent offers a superior experience, OTA may catch up within a year, while a startup could need 10 years to replicate the supply chain and service infrastructure. The conclusion: the interface “above the waterline” is not a moat. That is one reason he is betting on ToB and aiming to rebuild value deep inside enterprise workflows.
• Base models have evolved from roughly “60-70 IQ points” in 2023 and 2024 to the “100 or 110” of what he calls “today” or “this year,” shifting the opportunity from building brains to building an AI society. 张帆 uses a comparison with humanity: brain capacity has barely changed in 5,000 years, yet productivity has risen at least 1,000x. The real leverage came from “education, division of labor, tools and organizational collaboration.” Through 元理智能, he wants to train foundational intelligence into professional experts and close the enormous gap between models and productivity.
• The fundamental difference between AI ToB and SaaS is that the former benchmarks against the labor market, not the software market. 张帆 reduces enterprise value to traffic multiplied by conversion capability, with conversion capability made up of people and tools. Traditional SaaS can create a zero-sum conflict with existing teams; if AI is treated as a “colleague,” it directly improves conversion capability and creates a positive-sum relationship with the boss. “Chinese companies are not unwilling to pay; they will pay for results and productivity,” as shown by BPO demand far larger than the software market.
• 元理智能 deliberately capped its early customer base at 6 companies, validating value through financial and business outcomes rather than feature delivery. The team interviews 20-30 employees across departments and levels, maps business flows and cost structures, identifies 30-40 AI improvement points, then screens for opportunities in the upper-right quadrant of “business value × technical maturity.” The goal is not to deliver a knowledge base but to visibly change key operating metrics and even the financial model; if a production step becomes 5x cheaper and faster, the strategy of the entire upstream and downstream chain will change.
• “Commercial reinforcement learning” is 张帆’s core bet for scaling enterprise Agents: define productivity by occupation and optimize continuously against real business feedback. Math, coding and GUI tasks have clear answers, while a salesperson may spend 100 exchanges before learning that the customer did not buy, leaving rewards sparse and attribution difficult. 元理智能 envisions using roughly 5,000 sales logs to fit an enterprise environment, allowing an Agent to train “day and night” in simulation. What gets customized is not software code, but the company’s own environment and reward function.
• 张帆 rejects the idea that “industry best practices” are the endpoint for vertical Agents, arguing that every company needs its own individual best practice. Manner, M Stand and Luckin serve different customers and operate in different environments; Sequoia, IDG and BlueRun would also struggle to use one investment framework. “Asymmetry is not a defect of the model; it is a property of the model.” The startup opportunity is to find a general learning mechanism that adapts a general model to every dynamic and distinctive business environment.
• The biggest break in enterprise AI today is not on the supply side but on the demand side: compute, models and capital are scaling rapidly, while examples that actually change core business remain scarce. Looking back on 智谱’s deep work with more than 1,000 customers, 张帆 acknowledges that many projects delivered “emotional value rather than business value.” A POC can demonstrate a highlight, then face a “single-vote veto” on compliance, security, cost or reliability. His analogy: do not merely bring wires and light bulbs to a tribe and ask it to use more electricity; invent “plug-and-play appliances” such as refrigerators, televisions and washing machines.
• A durable moat is not a lighthouse built on a rising model sea, but a boat that rises with the sea level. 张帆 advises companies to keep “model content” at roughly 50%: half the competitive advantage should come from the existing business and half from model amplification. Jasper’s copywriting capability was quickly absorbed by ChatGPT, while Notion AI has continued to compound value by embedding itself in an established office system. Data should not be treated as a static crop either; the real moat is “a field” that keeps producing new data as the real world changes.

Deep dive

1. The contrarian case for ToB comes from the fact that old consumer-internet moats still work

• Koji opens by asking about the “physiological fear” Chinese founders and VCs have of ToB. 张帆 does not deny the difficulty; he questions the assumption, drawn from internet success stories, that ToC must inevitably be bigger.
• In the internet era, offline systems were already built while online space was nearly empty, allowing companies such as 携程 to rebuild supply chains and establish long-term strategic depth. Today, “offline is full, and online is full too,” with no new device capable of clearing the old structure.
• Travel Agents make the asymmetry clearest: a startup may quickly deliver a sophisticated experience, but 携程 might need only a year to catch up on the interface, while the startup could need 10 years to replicate its supply chain, service and operating system. “The interface above the waterline is only a small part.”

2. The first “Siri-like” wave taught him that without threshold-level generalization, product visions cannot be delivered

• 张帆 was not a conventional sales-oriented COO. He studied artificial intelligence at Paris Diderot University, researched machine translation at France’s National Centre for Scientific Research, and after returning to China in 2010, worked on Siri-like products at 搜狗 and 腾讯.
• After Siri launched, every major company believed natural-language interaction would become the next interface and even expected a 《Her》-style experience. But the methods at the time still involved manually defining dozens of categories, then recognizing, parsing and calling APIs separately for each one.
• After roughly 2 years, 张帆 concluded that the problem was harder than expected and, even if solved, would become a battleground for major platforms. He shifted to “AI plus vertical industries” and began his first startup in 2014.

3. 妙计旅行 proved the demand but lost because it could not deliver the full experience

• 妙计旅行 tried to upgrade 去哪儿-style “first-order metasearch” into “second-order metasearch”: users would not need to specify the order in which they visited Paris, Madrid and Barcelona; the system would jointly optimize price and experience across a much larger search space.
• The product connected flights, hotels, rental cars, private cars, attraction tickets, day tours and city transport, automatically stitching them into a complete itinerary. Koji believes this bottom-up demand still has not been properly solved, and 张帆 agrees that the demand itself is real.
• The failure came on the delivery side. The team treated travel as a fancy product problem, then spent 5 years and raised tens of millions of dollars before realizing that “travel is fundamentally a supply problem, a supply-chain problem.”
• The experience pushed him from idealistic technical researcher toward an entrepreneur’s perspective: sophisticated information on the interface has no commercial value if the supply chain and operations cannot fulfill it. The goal of a startup has to be “a business that can close the loop.”

4. Financing, media and angel customers can jointly train the wrong reward model

• In 2014, capital was plentiful. 妙计 quickly raised substantial funding from top-tier funds and turned down offers from major platforms; industry figures, early supporters and the media reinforced the team’s sense of correctness with narratives such as “the barbarian outside the gate” and “a bull charging into a china shop.”
• 张帆’s retrospective assessment is blunt: “Financing has never been the right reward model.”
• The real rewards to watch are retention, satisfaction and problem-resolution completion. Only after shifting attention from external applause back to operating metrics could the team see whether a project was truly improving or merely receiving emotional positive feedback.

5. 大搜车 and his second startup turned “industry × technology” into the team’s signature

• After 妙计, 张帆 realized that elite technical teams were not rare. What was rare was “a technical team that had spent tens of millions of dollars deeply grinding through an industry.”
• He then joined 大搜车 as group CTO. The company covered 17 business lines, including new cars, used cars, secondary dealers, finance, insurance, logistics, software and managed operations, with an industry team of nearly 1,000 people.
• 大搜车’s data-center team had roughly 100 people and built an end-to-end system spanning instrumentation, big-data scheduling, data governance, metrics and BI/CDP. The team also explored AI for used-car valuation, automated inspection and dealer matching.
• When he started his second company in 2022, he focused on midsize companies and small players in vertical markets, using standardized e-commerce as an example. Roughly 50 systems across 淘宝, 京东, 抖店, 有赞, 微盟 and various ERPs could already cover a large number of merchants. Data-platform costs could be reduced to a few dozen-ths of the normal big-tech cost, with RMB100K enough to provide a data middle platform.

6. After ChatGPT appeared, he abandoned an initially validated PMF in 1 hour

• The AI+Data company had operated for roughly 6 months, completed its angel round and acquired several angel customers. It was preparing for Series A when ChatGPT launched. 张帆’s first reaction was: “This is not incremental innovation; this is disruptive innovation.”
• Introduced by 张阔, former chief scientist at 搜狗, he faced 2 paths: layer AI onto the existing business and amplify the fundraising story, or abandon the existing PMF entirely and join a model company that had not yet sold a single order.
• He thought for roughly 1 hour and chose to join. He did not need 唐杰 to persuade him at length because his first 2 experiences had already convinced him: “The year in which you start a company is something you can control.”

7. Trends can invert a 10x capability gap into a 10x difficulty gap

• During 妙计, the team had only algorithm, product and small-team management experience, yet raised more than $10M without users. During the second startup, its capabilities and preparation were 10x stronger, but the business became far harder because of the Shanghai lockdown and changes in the fundraising cycle.
• 张帆 distilled the lesson into a deliberately sharp line: “When it comes to major events, your capabilities are fundamentally unimportant.” He does not mean ability is useless; he means the timing window can make the same task 10x harder or easier.
• He uses the example from 《Outliers》 that Bill Gates and Steve Jobs were both born in 1955 to explain trend dividends. You cannot control your birth year, but you can control your startup year; therefore, it is better to give up local certainty and enter a larger technological inflection point.

8. He left 智谱 because he saw model maturity as the starting point of AI’s second half

• 张帆 joined 智谱 in February or March of the year after ChatGPT launched and worked there for more than 2 years. By the time he left, 智谱’s technology and development were progressing smoothly in his view and the company was pursuing an IPO, so outsiders naturally expected him to wait for the listing.
• Koji notes that BlueRun invested $8M in the new company and calls his choice brave. 张帆 rejects the label: “Bravery” means doing something despite knowing it cannot be done, whereas he believes the new direction itself is more worth betting on. “If it is your mission, do not wait.”
• In 智谱’s early days, he often held 7 customer meetings in a day and replied to 200-300 messages after midnight. That intense exposure to model progress and enterprise demand let him see both models approaching the capability threshold and the gap in business adoption earlier.

9. Once model intelligence crosses the threshold, education, division of labor and collaboration offer more leverage than more intelligence

• 张帆 compares models in 2023 and 2024 to 60-70 IQ points: they already exhibited human-like behavior, but practical deployment remained difficult. By what he calls “today” or “this year,” he believes models have reached 100 or 110, changing the market paradigm.
• His analogy is that human brain capacity and basic intelligence have not fundamentally changed in 5,000 years, while productivity has risen at least 1,000x. The change came from “education, division of labor, tools and organizational collaboration,” not further increases in brain capacity.
• This also defines his People-Mission Fit. He does not believe he can materially advance the intelligence of foundation models, but he combines model understanding with product, business and industry experience. His mission is therefore to build the social form that takes AI from foundational intelligence to productivity.

10. 元理智能 is limiting customer count first, then pursuing repeatable lighthouse outcomes

• The company was registered roughly 3 months ago, and 张帆 set a clear ceiling: no more than 6 customers by Q2 of the following year, with no floor, and the 6 customers coming from different industries.
• Cross-industry deployment is a generalization test. Only if the same framework works across different customers can the team confirm that it has built general capabilities rather than project-based delivery in new packaging.
• Early customers receive “full-spectrum service” as co-creation partners, but the goal is not to turn the company into a customization shop. 张帆 wants 元理’s intervention to produce significant changes in key operating metrics and even financial models, making these customers templates for traditional companies entering the AI era.

11. Enterprise AI adoption must be reverse-engineered from the operating model, not forward-built from a feature list

• The team first interviews 20-30 employees across departments and levels to map the core business flow, then analyzes the cost structure, because “the financial model actually represents the key elements of the operating model.”
• In customized travel, for example, the business can be broken into customer acquisition, production and supply chain. Ad buying, content production and customer communication belong to acquisition; itinerary creation and modification belong to production; negotiation, ordering and confirmation belong to the supply chain.
• A company can typically identify 30-40 AI improvement points, but it will not implement all of them. The team uses business value as the x-axis and technical maturity as the y-axis, selecting only the upper-right quadrant: supply-chain bargaining has high value but immature technology, while daily reminders are easy to implement but limited in value.
• Engineers then enter the logs and service workflows to determine how to train the model and build the Agent, feeding performance back into business decisions through a PDCA loop. If production time and cost fall 5x, the impact is not limited to one expense line; it reshapes the entire upstream and downstream chain and business model.

12. Heavy service is not anti-scale; it is the necessary cost of finding the boundary of standardization

• Koji’s challenge is practical: if the first step alone requires interviews with 20-30 people, how can this become a large company? 张帆’s engineering answer is “divide and conquer,” splitting the problem into business understanding, model optimization and Agent optimization.
• Business understanding still requires human-to-human communication. On the model side, data, SFT and reinforcement learning adapt the model to the setting; on the Agent side, Prompt, RAG, Memory and Workflow are added. The latter 2 layers can eventually be unified and fully automated through commercial reinforcement learning.
• Even half of the top-layer consulting work could be assisted by Agents, including selecting interviewees, generating outlines, conducting interviews automatically and codifying SOPs. 张帆 estimates this could reduce the threshold and cost of professional services by 5x.
• The next step is not for 元理 to consume everything, but to let independent consultants, senior advisors and holders of industry know-how complete the last mile. Salesforce, SAP and 地平线 all heavily co-created with angel customers in their early days. “Without going through the process once, there is no basis for generalization.”

13. 元理 wants to be a “training institution” for models and is deliberately ruling out 4 categories of business

• 张帆 defines the company’s position in the ecosystem as “a converter between consulting and foundation models”: foundation models provide increasingly intelligent people; 元理 trains them into professional experts and sends them into enterprise work.
• The company therefore explicitly does not build foundation models, operate a MaaS platform, make SaaS or take on customized consulting. What it wants to accumulate is a general training capability spanning business environments and professional models.
• This boundary also explains its relationship with 智谱. 元理 remains downstream in its ecosystem, manufacturing “appliances” that turn models into business capabilities and ultimately expand the effective consumption of tokens.

14. AI is not better SaaS; it is the enterprise’s new colleague

• 张帆 approximates enterprise value as “traffic × conversion capability.” When conversion capability is roughly constant across an industry, companies can only buy advertising, media spend and storefronts to expand traffic. Conversion capability itself is made up of people plus tools.
• Traditional SaaS can enter a zero-sum game with employees. A company buys software worth RMB1M, and the boss immediately asks whether it can cut RMB1M of labor. People are mandatory; software is not, so the organization naturally resists.
• If AI is treated as a colleague, it competes not for the tool budget but against conversion capability and the labor market, turning the relationship between enterprise and vendor from zero-sum to positive-sum. “AI is not a tool. AI is your colleague.”
• Koji asks why the United States, Japan and Europe can accept SaaS. 张帆 does not attribute it to human nature, instead emphasizing generational differences in digitization: many Western managers worked in CRM from Day One, while Chinese users are gradually becoming people who “cannot work without 飞书.”

15. Accepting hallucinations means switching from a code coordinate system to an intelligence coordinate system

• 张帆 argues that “as long as machines still hallucinate 1%, humans have value” uses the old coordinate system. Code should execute strictly, but humans make mistakes too, and human hallucinations may exceed 1%.
• If AI is required to make zero errors like code, its generalization ability will disappear at the same time. The right question is not how to eliminate hallucinations completely, but how to manage them as we manage human error.
• His warning is: “Do not use the old coordinate system to search for a new map.” Enterprises do not need to understand the internal details of the Transformer; they need to understand the behavioral boundaries of a model as a form of intelligence.

16. The supply side has piled up light bulbs; the demand side lacks appliances that change people’s lives

• 张帆 observes that capital, talent and opinions have poured almost entirely into the supply side: semiconductors, chip clusters and foundation models keep receiving positive feedback, while investment in how the demand side consumes intelligence is clearly inadequate.
• In model commercialization roles, the objective is often to make customers consume more tokens, with price cuts and expanded business development as the standard tools. But if customers do not know what intelligence can solve, even free intelligence cannot create effective demand.
• His analogy is bringing wires and light bulbs to a primitive tribe and asking people to “use more electricity.” Once the bulbs are everywhere, there is no incremental demand. What needs to be invented are plug-and-play appliances connected to everyday life: refrigerators, televisions, washing machines and air conditioners.
• Industry forecasts often optimize the denominator of cost-performance, assuming that a 10x cost reduction will trigger an explosion. 张帆 asks whether the same result could come from a 10x increase in the numerator of business value with costs unchanged. 元理’s decision to “model productivity” is a bet on that numerator.

17. Occupation is the smallest standard unit for converting foundational intelligence into productivity

• A high-school student can read, calculate, reason and apply common sense, but is not yet productivity. Only through education and division of labor, becoming a doctor, lawyer or delivery worker, does intelligence acquire an evaluable identity.
• 张帆 therefore treats “occupation” as the first modeling medium: first define what work the model will perform, then put it into a specific company’s real environment to learn, rather than continuing to pursue abstract general intelligence.
• 元理 coined “commercial reinforcement learning” for the second layer of its method. It serves the commercial objective first and reinforcement learning second; each occupation has a different reward and cannot be measured with one universal score.

18. The challenge of commercial reinforcement learning is sparse feedback and incorrect attribution

• Math problems, coding and GUI tasks have clear answers, so reinforcement learning improves rapidly. Sales is different: after a salesperson talks for 1 hour and says 100 things without a purchase, it is impossible to simply label all 100 statements as wrong.
• Rewards in the commercial world are often sparse and difficult to attribute. Algorithms must be combined with business judgment to build an experimental setting that approximates the real environment.
• 张帆 still believes the path works because human salespeople also grow from junior to senior in the physical world. The point is not to hand-write every correct script, but to let digital occupations evolve continuously through feedback.

19. Agents are moving from rule-based workflows toward trainable native planners

• First-generation Agents were built around Workflows: a booking system would ask for departure, destination and passenger count in sequence. If the user answered “I want to go to Shanghai” when asked “Where are you departing from?”, the flow could loop indefinitely.
• Connecting every node into a spider web can add branches, but the result is a bloated rule system with incorrect routing. The second generation lets the model act as the Planner: specify what information a booking requires, and it asks for whatever is missing.
• Koji notes that many Agent-as-a-Service companies in A16Z Speedrun chose verticals such as customer service, sales, HR and user research. 张帆 believes these products use the model’s native generalization effectively, but still rely too heavily on the foundation model’s shared planning capability.
• He is betting on a third generation: beyond adjusting Context, it will intervene in the model itself so that it natively understands a particular company’s and occupation’s knowledge mappings, action decomposition and Planner. Commercial reinforcement learning is the proposed training method.

20. Industry best practice is a compromise; individual best practice matches intelligence’s asymmetry

• SaaS often sells “industry best practice,” essentially handing common methods to every customer. 张帆 contrasts portals with 今日头条 and argues that intelligent systems will ultimately move from uniform editorial selection toward individualized optimal choices.
• Even as coffee companies, Manner, M Stand and Luckin have different customer bases and operating environments. It is equally hard to imagine Sequoia, IDG and BlueRun buying the same service to help them decide whether to invest.
• His underlying judgment is: “Asymmetry is not a defect of the model; it is a property of the model.” A model can solve an Olympiad problem yet miscount the R’s in Strawberry; human geniuses also adapt to different objectives and cannot be ranked on a single scale.
• Enterprise environments also keep changing. A sales method that worked 2 years ago may fail today. A truly general product does not provide one universal answer; it provides a general learning method that finds the optimal solution for each environment.

21. Environments can be defined individually while the learning process still scales

• Koji points out the paradox: if every company has a unique environment, does the business ultimately fall back into custom software? 张帆 distinguishes 2 types of customization. In the past, understanding requirements meant writing code; today, it means defining the environment, objective and reward, with the environment itself also understandable as a Prompt.
• He uses AlphaGo as an analogy. AlphaGo trained on roughly 30 million endgames from human history and defeated Lee Sedol; 1 year later, AlphaGo Zero relied on no human priors and surpassed its predecessor after only a few days of self-play. “Human knowledge is sometimes noise too.”
• In sales, a company can provide roughly 5,000 sales logs, allowing the system to fit an experimental environment quickly. The Agent repeatedly communicates, receives feedback and learns inside it, emerging adapted to that company without software being rewritten for every customer.
• 元理 wants to standardize the ability to build the Go board and reward, not to label “30 million endgames” for each customer. This is the key boundary in 张帆’s shift from customized delivery to general learning infrastructure.

22. Enterprise AI is still in a chaotic phase; most POCs buy emotion rather than results

• The enterprises 张帆 encounters present “two extremes of fire and ice.” One side is intensely anxious, worried that its existing business will disappear; the other believes AI has already worked because an internal team built a demo.
• The supply side sees all-time highs in U.S. stocks, growth in the mega-seven’s earnings, and rising ARR and daily active users at AI companies. On the demand side, reports say roughly 95% of POCs fail. 张帆 says this broadly matches his review after serving more than 1,000 enterprises.
• He admits that 智谱’s services were reasonably good, but very few customers actually became AI-oriented companies or saw clear improvements in core metrics because of them. Many projects over the past 2 years provided “emotional value rather than business value.”
• A POC needs only to demonstrate a highlight at 1 or 2 points; a production system must pass compliance, security, cost and reliability simultaneously, and “each one is a single-vote veto.” Moving from demo to a closed-loop business is enterprise AI’s hardest leap.

23. The entrepreneur’s required course is not model technology but “modelity”

• 张帆 cites the view that “enterprise competition is competition in cognition.” People can be hired, money can be raised and technology can be bought, but an entrepreneur’s understanding of trends directly determines strategy. The cloud-computing era already demonstrated this divergence.
• He asks entrepreneurs not to outsource AI as a technical problem: “AI is a business problem.” A boss does not need to understand every Transformer algorithm, but must understand hallucinations, asymmetry and behavioral preferences as one understands human nature.
• He calls this body of knowledge “modelity.” Entrepreneurs should be the first people inside the company to understand how AI connects with their own business, because neither vendors nor model companies can understand that unique environment better than the operator.
• The failure of multi-Agent bargaining is a warning. When Seller and Buyer negotiate, both may close as soon as they reach the budget range because RLHF gave the models a “people-pleasing personality.” Layering multiple models can amplify the bias further.

24. The real moat is a “boat” with business and models each contributing half

• 张帆 compares foundation models to a sea that keeps rising. If a company builds a lighthouse on the surface and sea levels rise 100 meters every 6 months, its accumulated height will soon be submerged. “Do not build a lighthouse; build a boat.” The boat rises alongside model capability.
• He recommends keeping “model content” at 50%. Too much means direct competition with foundation models; too little means falling out of step with the era. Half should come from the existing business moat and half from model amplification; only the combination is durable.
• Jasper reached tens of millions of dollars in ARR during the GPT-3 era, then collapsed rapidly about 6 months after ChatGPT launched. Notion AI, by contrast, relies on an established office-software system. They represent the lighthouse and the boat respectively.
• Data should not be worshiped as a static asset. Directly training on tens of millions of chat records may leave a model “trained and useless.” Enterprises should instead build business settings that continuously produce new data—“treat it as a field, not a fixed crop.”

25. The data flywheel at the Agent layer may be easier to build than at the foundation-model layer

• 张帆 believes foundation models offer a latecomer advantage: training 3 months later may halve costs through cheaper compute and distillation. The early vision of an RLHF data flywheel also failed to become a sufficiently strong moat.
• Search engines could once extract unique data from behavior—for example, “荷塘月色” clicked as a residential compound in Beijing and a restaurant in Tianjin. General feedback on large models is often too ambiguous for users to determine which of 2 long answers is better.
• Once intelligence enters an occupation, evaluation becomes clear again. Better doctors, programmers and salespeople can all be measured by business outcomes. If a platform covers roughly 50 common occupations, greater usage will accumulate stronger shared knowledge and reduce the cost of adapting the model for each company.
• 张帆 therefore does not believe the opportunity necessarily belongs to foundation-model giants. A startup can build its moat by “modeling learning,” using one unified mechanism to adapt a general model to countless independent tasks.

26. First-principles reasoning has replaced single pain points, while synthetic data became his latest change of mind

• Koji observes that 妙计 began bottom-up from concrete user pain points, while 元理 reasons top-down from a technical threshold and first principles. 张帆 acknowledges the change: surface phenomena keep changing, and mature founders should look for certainty at the underlying level.
• He cites 黄仁勋’s early, long-term investment in CUDA. At the time, there were almost no application scenarios and only a handful of scientists used it; 10 years later, AI arrived and triggered geometric growth. 元理 is similarly trying to identify the underlying direction of “how learning can scale” before searching for the surface product.
• 张帆 once rejected synthetic data: if 100 rules generate 10,000 pieces of training data, why not just use the rules? He later reversed the question. A large model’s trillions of parameters can be understood as an enormous number of latent rules; 1 Prompt or a small number of examples may invoke “10,000 rules to generate 100 pieces of data,” without necessarily causing overfitting.
• This also explains AI gameplay in commercial reinforcement learning. Giving 2 models different Prompts and small adjustments is equivalent to making 2 sets of knowledge rules interact, generating new knowledge. “We are always, consciously or unconsciously, trapped in our original cognition.”

27. His final investment choices still revolve around how machines can learn more efficiently

• Given a hypothetical $3M to invest across 3 companies, 张帆 does not name 3 founders. He allocates by technology stack: one portion to a major platform, one to a model company and one to Thinking Machines Lab.
• He particularly values Thinking Machines Lab’s understanding of business and AI, believing its research focuses on “how to make machines learn better more efficiently,” aligned with 元理’s own judgment, though it has more capital and is moving more aggressively.
• That closes the conversation’s arc: from 妙计’s supply-chain lesson to 智谱’s model commercialization and 元理’s commercial reinforcement learning, 张帆 has repeatedly searched not for a flashier interface, but for how intelligence can be trained, delivered and ultimately change enterprise business outcomes.