Interview with 印奇: StepFun, Smart People's Temptations, Trade-offs
Summary
- StepFun’s core bet is not a single application, but the smallest commercial loop combining a world-class foundation model with AI terminals. 印奇 ruled out a foundation-model company focused solely on China’s To B market or pure-software To C: the former cannot generate enough revenue and profit to fund R&D, while the latter lacks users and a data flywheel and would be squeezed by Big Tech. StepFun will therefore use the foundation model as the “brain,” starting with cars before extending into handheld and wearable devices—the brain remains “the main thing.”
- Foundation-model competition is first and foremost a brutal equation of funding and profit: at least about RMB3B of R&D per year, with a reasonable range of RMB3B–RMB5B. Sustained for 3–5 years, that means at least RMB10B in cumulative base investment, or RMB15B over five years. The counter-requirement is that a core application must emerge within 3–5 years, potentially generating RMB3B–RMB5B in annual profit; otherwise the business cannot self-fund. “Both things have to happen before you have a chance of reaching that stage.”
- 印奇 calls this AI cycle “the most intense technology competition in history,” and believes the race may already be more than halfway through. The technical foundation is still shifting, products and commercialization can be reset at any moment, and the competition combines the world’s highest concentration of talent, historically large resource consumption, and uncertainty over which applications come first. The bubble is visible in stock prices and compensation: top talent now costs roughly 5–10x what it did in AI 1.0, with China and US packages approaching “RMB exchanged for US dollars.”
- StepFun’s 2026 model roadmap boils down to 3 keywords—foundation model, omni-modality, and VLA—and the company will roll out a continuous series from 3.5 to 4. “Foundation models basically mean everything.” Omni-modality covers text, voice, and images; VLA brings data from cars, robots, and eventually phones into the same brain. The goal is world-class base intelligence, without sacrificing the cost, efficiency, architectural fit, or specialized experience required by terminals merely to top a single benchmark.
- Cars are both the most realistic commercial entry point today and the first springboard for collecting physical-world data and reaching embodied intelligence. Qianli Technology will first deploy intelligent driving, cockpits, and eventually Robotaxi; 印奇 expects this intelligent-driving battle to continue for roughly 3 years and ultimately consolidate around 3–4 core suppliers. A fuller indoor and embodied-data loop will take 5–7 years, while today’s robot models are roughly at the early “GPT-1, GPT-2” stage.
- 印奇’s sharpest review of AI 1.0 is that he “played every hand” when young, and the security battle should never have been fought. Megvii already had an edge in finance and smartphones, but kept adding businesses because vertical markets seemed too small; a better trade-off might have been to focus on automobiles around 2019–2020. The capital markets should be only “an intermediate node for staying alive”; “going public really changes nothing fundamental,” and may even force a company to sacrifice long-term R&D for short-term results.
- The real moat of the next generation of AI companies lies not only in models, but in organizing a group of highly intelligent, low-ego people willing to work together on the unglamorous parts for the long haul. StepFun is re-integrating algorithms, systems, and data; 印奇 says data may determine the most fundamental parts of a model, potentially accounting for 70%–80%. His technical view of AGI is relatively clear: if learning paradigms, memory systems, and the move from the digital world into the physical world are solved over roughly the next 5 years, “AGI as currently defined should, in all likelihood, be achievable.”
Deep dive
1. 印奇 places StepFun and Qianli Technology on the same AI map
印奇 now has two core roles: chairman of Qianli Technology and chairman of StepFun. He no longer works at Megvii. Qianli focuses on combining AI with cars and future robots, while StepFun provides the foundation models, brain, and underlying capabilities.
StepFun was founded in April 2023, and 印奇 says he has been deeply involved since the earliest planning stage. Taking the chairmanship formally is about putting more energy into a competition where “the foundation-model battlefield is brutal,” not suddenly joining a new company.
His division of labor with CEO 大新 is that 大新 handles day-to-day operations overall, while the two jointly drive foundation-model organizational reform and technical breakthroughs. 印奇 is spending more time on terminal commercialization, although work on vehicle models naturally cuts across both companies.
2. “StepFun” points to nonlinear phase change, while “Stellar” preserves the horizon of a technology startup
The English name StepFun comes from “step function.” 印奇 says a step function represents the transition from quantitative change to qualitative change, as well as the accumulation of multiple nonlinearities in deep-learning networks. In his words, “the accumulation of nonlinearities is also a fundamental property of intelligence.”
“Stellar” reflects the “sea of stars” ambition that AI practitioners should retain. The name captures two goals at once: exploring the upper bound of foundational intelligence and turning the technology into a company with a commercially closed loop.
StepFun’s Day One mission has not changed: to become one of China’s best companies in foundation models. Talent, capital, and the business model must all match the long-term investment required by foundation models.
3. He initially misread ChatGPT as “having little to do with me”
When ChatGPT emerged, 印奇’s first reaction was not to go all in. He believed he should continue along the path of vision, robotics, and hardware, with language foundation models eventually serving merely as an external component—“just use whichever one works.”
It took roughly half a year for that view to change. 印奇 first assessed whether ChatGPT represented a change in underlying technology, then whether it belonged to his commercial arena. Language-model data, parameter counts, and compute initially seemed better suited to internet giants, but the potential of multimodality and physical intelligence made it more than a local upgrade to language.
His final conclusion was that “there is only one brain.” Vision, language, and physical AI are unlikely to remain separate vertical models over the long term; they will ultimately be absorbed into a unified base, leaving AI entrepreneurs with only one choice: do it or don’t.
4. AI 1.0 to 2.0 is continuous evolution, not a rupture from nowhere
张小珺 asked whether large language models had disrupted the previous generation of computer vision. 印奇 disagreed. Megvii was deep-learning native from day one and had also participated in early domestic language-model projects. In his view, the sense of disruption often comes from “missing some of the preceding information.”
Continuity does not mean that the change is minor. Long accumulation can suddenly cross a cognitive boundary in a “quantitative-to-qualitative” jump, like a step function. AI 1.0 and 2.0 are mostly media labels; the battle itself never stopped.
This also explains why he insists on extending old capabilities while believing a new platform must be built: the technical lineage is continuous, but the organization and business model may need to start from zero.
5. Megvii’s mature organization was not suited to a completely new foundation-model startup
After more than a decade of development, Megvii had built an organization and business model around AI+IoT, integrated hardware and software, and serving enterprise customers. 印奇 says that path had become very clear and, to some extent, “固化”—set in place.
Megvii had also been talent-dense and research-driven in its early years, but once the organization grew from a baby into a teenager, it had already chosen its specialty and direction. Early-stage foundation-model entrepreneurship requires heavy investment and tolerates uncertainty; it needs a newly born platform rather than a forced transformation of the old one.
The earliest idea was for Megvii and StepFun to cooperate, and the relationship gradually evolved into what it is today. 印奇’s implicit organizational judgment is that old capabilities can remain connected, while the new culture should stay independent.
6. “Brain plus terminals” comes from his underlying belief about the evolution of biological intelligence
印奇 says many of his technical beliefs come from On Intelligence, neuroscience, and an evolutionary perspective. Humans should remain humble before intelligence: the sequence by which the brain evolved from survival in physical space to the neocortex, language, and reasoning cannot simply be skipped by AI.
The final embodiment will “definitely be a robot,” but a robot is essentially the reconstruction of a new species, involving brain structure, sensors, actuators, and hardware systems. It cannot be built in one step. The first move is therefore to build the brain, then accumulate capabilities through terminals such as cars and phones.
He reinterprets “people, cars, and homes” as phones and wearables, automobiles, and homes—which may eventually become the deepest scenario. For now, cars are the first entry point, handheld and wearable consumer electronics the second direction, and embodied platforms the more distant outcome.
7. The brutality of this competition comes from an ultralong chain that is still moving
印奇 calls it “possibly the most intense technology competition in history.” The world’s smartest people and historically large R&D budgets are being deployed at the same time, while the timing of AI’s transformation of each industry remains highly uncertain.
The chain runs from original technical innovation through productization, commercialization, and a closed business model. The problem is that the foundation is still changing rapidly: like rebuilding a house’s foundation over and over, the products and processes above it may also have to be torn down.
Faced with a chain that is both long and fast, he says companies must hold on to two starting points. Technically, they need a broad or focused edge; commercially, they need to use elimination, first removing models that can never cover the investment.
8. A foundation-model company serving China’s To B market alone cannot make the math work
印奇’s threshold is at least about RMB3B of annual foundation-model investment, sustained for 3–5 years, for at least RMB10B of cumulative basic R&D. Actual spending by large technology companies is far higher.
Pure To B monetization takes time, while the revenue and profit ceiling available to a new company is limited, making it impossible to support that cost. He later narrowed the judgment: “It doesn’t work in China; it might work overseas.”
StepFun may provide some high-quality To B services, but the final destination must be AI plus terminals. To B can provide intermediate revenue, but cannot be the terminal answer for foundation-model investment.
9. Pure-software To C is also ruled out because startups lack the flywheel
Big Tech already has users, scenarios, and data, allowing product usage to improve the model in return. Even if startups are not entirely without opportunity, it is difficult for them to reproduce that flywheel in pure-software applications.
张小珺 noted that foundation-model applications currently often have neither a data flywheel nor network effects. 印奇 therefore rules out standalone ChatBots and standalone pure-software consumer products alike.
StepFun’s answer is a consumer-oriented combination of software and hardware: dedicated terminals create the experience and data entry point, while hardware, software, and models form a service loop together. The model should not simply be placed inside an undifferentiated application shell.
10. The essence of AI hardware is giving Agents a physical body
印奇 rewrites the previous iron triangle of hardware, software, and internet services as “hardware, software, and models.” Many Agents will eventually have an on-device form; hardware is the physical manifestation of an Agent.
Hardware markets are unlikely to produce a software-style winner-takes-all outcome. Phones, cars, and new devices typically accommodate multiple vendors. That fragmentation gives innovative companies a chance to get onto the table and may still support a business of meaningful scale.
He expects AI to redefine interaction. As voice and multimodality reduce dependence on screens, hardware will no longer be locked into the three screen sizes of phones, Pads, and PCs. New categories with tens of millions of units sold could emerge in large numbers.
11. Good AI hardware starts with a scenario, then makes “the model the product”
印奇 cites recording and workplace devices such as PLAUD: both hardware and software are optimized around a specific AI service, and users ultimately consume the complete model experience rather than a standalone device.
An AI-native product starts by making the model good enough in a core domain, then adding software. Hardware should be added only when dedicated hardware materially improves the experience. The balance among model, software, and device depends on the scenario.
He calls the Doubao AI phone “definitely a very good attempt,” and possibly the first step in ByteDance’s hardware combination. StepFun’s first innovative terminals, however, will not directly copy the phone form factor; they will look for new categories.
12. Foundation-model evolution requires “push” and “pull” at the same time
印奇 uses the generational relationship among GPT-3, 3.5, and 4 to explain foundation models. Real progress often does not mean one field surging ahead, but rather 100 fields each improving by 10%, lifting general intelligence as a whole.
“Push” comes from technical and scientific breakthroughs—the shift from imitation learning to reinforcement learning could remain the main driver over the next 3 years. “Pull” comes from concrete scenarios, which determine model size, efficiency, cost, and specialized optimization.
Bigger terminal models are not automatically better. StepFun wants world-class base capability but does not insist on ranking first in every individual benchmark. He uses Gemini Flash to illustrate that applications must balance performance, cost, and speed.
13. StepFun’s two bets are the foundation model itself and using terminals to pull commercialization
印奇 rejects defining StepFun through terminals first: “First, the bet is the foundation model.” He does not elaborate on companies that choose not to build foundation models, but StepFun plans to keep exploring the frontier over the next 3 years and believes it can remain among the world’s best.
The second bet is scenario selection. When the foundation model needs applications to pull it forward, cars, handheld devices, and wearables will provide the scenarios, data, and commercial loop.
He even believes highly intelligent models could become “winner takes all”: once users become accustomed to interacting with a smarter model, they will not want to return to a dumber one. Hardware products therefore still need the strongest foundation model.
14. The elimination race is past halfway, with speed, resources, and the bubble all exceeding 2023 expectations
From the “six little tigers” of foundation models to a steadily shrinking field of players, 印奇 describes the industry as a brutal elimination contest continually entering new final rounds. It may now be “past halfway through the race.”
Compared with when he founded StepFun in 2023, he underestimated technical speed, competitive anxiety, and resource consumption. The more intense the competition, the larger the bubble. He does not believe it is possible to have both “greater intensity and a smaller bubble” at the same time.
Two visible bubble indicators are some companies’ stock prices and compensation for core R&D staff. Whether top Chinese talent stays in China or goes to the US, package figures are approaching “RMB exchanged for US dollars,” and overall pay increases of 5–10x are not unusual.
15. After taking the chairmanship, 印奇 first changed the R&D organization, not just the product direction
He wants the organization to have “stronger, faster combat power.” Large-model engineering cannot be left to scattered engineers doing odd jobs inside research groups, nor can it depend entirely on platform teams serving numerous To B functions.
StepFun therefore consolidated its algorithm-engineering groups and moved systems engineering closer to algorithms. The data organization now reports directly to the head of algorithms, because data may determine the most fundamental parts of the model, potentially accounting for 70%–80%.
Algorithm staff must be deeply involved in what data is needed and how it is collected and cleaned. Systems engineers must understand algorithms, while algorithm engineers cannot merely write papers; they must also be able to build and operate systems.
If one model ultimately covers text creation, dynamic generation, companionship, and other scenarios, the organization must also manage the “one plus N” relationship: how the unified brain and multiple specialized capabilities share, provide feedback, and coordinate with one another.
16. Cars are already pulling VLA, multimodality, and cockpit assistants together
The first joint scenario for Qianli and StepFun is intelligent driving. VLA, multimodality, world models, and other foundation-model capabilities are entering vehicles, while vehicle-side data feeds back into training of a unified model.
The second landing point is the cockpit. 印奇 considers the cabin a natural human-machine interaction space, where a large model acting as a super-assistant may fit better than on a phone. He cites Tesla bringing Grok into the car as an example.
Intelligent driving and cockpits are already being deployed; handheld and wearable devices come later. StepFun will not roll out every hardware category at once, but will move layer by layer based on data value, scenario maturity, and the commercial loop.
17. The truly scarce resource is not internet data, but continuous physical-world data
印奇 believes data from digital and physical space must ultimately enter the same model. A language model may solve an international math competition problem yet make a basic physical-logic error, revealing a major gap in perception and understanding of the real world.
Intelligent driving already supplies large volumes of continuous outdoor video and physical-world data at scale. What is currently missing is indoor, fine-grained data with dense human-machine interaction.
Two potential sources are large-scale embodied platforms and people’s eyes and wearable devices. Only better product experiences, greater shipments, and richer data can create a physical-space data flywheel.
Building and injecting this data chain will take 5–7 years. Robots have not even converged on their configurations or sensors; if GPT-3.5 is treated as the threshold, the field is currently around the “GPT-1, GPT-2” level. The excitement may be arriving too early.
18. Xiaomi, Huawei, and ByteDance are reference points, but not directly comparable competitors for now
张小珺 asked how a startup could compete with Xiaomi, which has multiple types of data. 印奇 first lowered the temperature: “None of these are our competitors”—the companies are simply not in the same weight class.
His breakdown is that ByteDance is extending downward from software and mobile-internet applications, while Xiaomi and Huawei may be strongest from the hardware side. StepFun and Qianli first need to make a local technology work and close the commercial loop, not claim total competition.
The startup opportunity comes from a non-monopolized hardware market with many categories. 印奇 hopes to find a terminal entry point outside Big Tech’s coverage that has demanding technology requirements, clear user demand, and the potential to reach tens of millions of units.
19. StepFun’s 2026 model releases will move from 3.5 toward 4
印奇 previewed “a series of models from 3.5 to 4,” with particular upgrades to language-model capability. The company previously invested more heavily in multimodality; now its language models must also reach world-class performance.
The first keyword is foundation model: “Foundation models basically mean everything.” Whether the input is voice, images, or Action, it ultimately depends on core intelligence, and language-model performance is an important window into that capability.
The second keyword is omni-modality, explicitly covering text, voice, and images. The third is VLA, which brings actuators, moving parts, and Action data from cars, robots, and future terminals into the model as a source of differentiation.
20. Coding is an excellent training environment, but may not be a revenue pool for Chinese startups
印奇 acknowledges that programming is highly valuable for improving intelligence: it has clear outcomes, objective right-or-wrong judgments, and sufficient complexity, making it one of the best vertical environments for reinforcement learning.
Commercial value and value a startup can capture are different things. Chinese Big Tech is not weak, DeepSeek provides a high-water-mark open-source model, intense competition will push prices down, and coding is deeply tied to upstream and downstream environments.
His conclusion is categorical: “The coding track is like the ChatBot track—there is no point for a new company to touch it at all.” He is rejecting it as an independent commercial entry point, not its value for training.
21. ChatBot may be a good demo, not the ultimate entry point 5 years from now
印奇 considers the ChatBot interaction unnatural and doubts it will have a product position as durable as search. In his reading, OpenAI may initially have launched ChatGPT more as an excellent demo.
Large models replacing traditional search is one of the more certain directions in his view, but the specific product form remains unclear. Putting all AIGC capabilities into a chat box, as products including Doubao have attempted, has not looked especially successful in the data.
张小珺 asked whether ChatGPT already had a durable moat. 印奇 acknowledged that its DAU and MAU are at super-app scale and that it may be a long-term product overseas, but Gemini’s rapidly rising DAU suggests the moat has not solidified.
StepFun will therefore not independently build pure-software consumer applications like Cat or Duck. If it launches software, the software will serve its own hardware and complete experience.
22. Qianli is 印奇’s first true latecomer battle
Megvii historically started with fundamental technical innovation and “looked for nails with a hammer.” Qianli is entering an industry that has already evolved for more than a decade, with its technical and commercial boundaries becoming increasingly clear.
Early disagreements between the Mobileye and Tesla approaches to intelligent driving have gradually converged around a model-driven core, moving toward unification with language and multimodal foundation models. 印奇 believes parts of the old technology stack and historical data are no longer usable, so a latecomer does not have to catch up with every piece of accumulated history from zero.
New-energy vehicle OEM and supply-chain business models have also gone through multiple rounds of evolution, making this “a very good opportunity and possibly the last window.” He estimates that this war will continue for roughly 3 years.
Latecomers depend more heavily on comprehensive capability and organizational strength. The cost is an extremely short window to persuade customers and prove results, making execution speed more decisive.
23. The intelligent-driving supplier market may ultimately have only 3–4 players
印奇’s analysis starts with margins. Excessive competition in new-energy vehicles will flow upstream, per-vehicle revenue is limited, and intelligent driving and cockpit systems require heavy model R&D. Without scale, suppliers cannot spread their costs.
He expects roughly 3 core suppliers, while leaving open the possibility of 4. Huawei is already a definite one; the rest may form a “two plus one” or “two plus two” structure serving OEMs of different scales and types.
Cockpits are highly personalized and tied to user experience and value-added services, so automakers may retain a high share of in-house development. Intelligent driving is a safety-critical component and may instead fall more often to specialist Tier Ones, like traditional safety parts.
24. Startup competition is against a local unit of Huawei or ByteDance, not the whole company
Faced with China’s “two great mountains of commerce,” 印奇’s decomposition is that startups do not compete with an entire giant; they compete with the giant’s local team on a particular vertical battlefield.
Two conditions are required: sufficient focus and a local core advantage. He believes Qianli and StepFun’s strengths are AI-native technical judgment and R&D. If they find a defensible track and serve one type of customer well, they can survive.
The “brain” and the “torso” need different cultures, so the two companies should remain relatively independent. They still share a strategy and have many links between them, allowing cooperation to run more smoothly than with an ordinary ecosystem partner. 印奇 compares this duality to the Taiji symbol and yin-yang.
25. From the beginning, he chose not just entrepreneurship, but AI as a life’s question
When choosing his undergraduate major, 印奇 asked which field was most related to AI. He first entered Tsinghua’s automation program and later transferred to the Yao Class. The core mindset Yao Class gave him was to think at world-class level and challenge the hardest problems.
Without deep learning, he might have pursued teaching and academic research. He graduated just as deep learning, the iPhone 4, and the mobile internet emerged, making industrialization possible and turning entrepreneurship into a means of achieving the goal.
AI attracts him for more than its commercial potential. What intelligence is, what life is, and where humans come from and are going may all become more explainable through AI. “The question is important enough” to constitute a meaning for one’s life.
26. Security was the wrong AI 1.0 battle; cars should have come earlier
印奇 admits that if he could start over, he might not fight the security battle. “When I was young, I played every hand,” but as in poker, some hands should simply be folded.
A better path would have been to keep deepening Megvii’s existing advantages in finance and smartphones, serve customers well, move from the top two to clear number one, and wait for the technology and market window in automobiles.
He first names 2016–2017 as the ideal time to enter automobiles, then, after further review, says around 2019 would still have been early enough. Even entering today would offer an opportunity; it would just mean “taking the long way around.”
27. The security failure was not a lack of AI, but control of only one link in the chain
Security looked like a To B business, but in reality centered on governments, central state-owned enterprises, and related entities, where marketing and channel capabilities carried substantial weight. Hardware was not just a product; it required the full chain of supply, R&D, sales, and go-to-market.
Megvii had the AI increment but lacked strength in hardware and the market. It chose to build complete solutions itself, while other AI companies empowered incumbents. Whether a company can cross the window depends on whether its technical leverage is strong enough to make up for the other two capabilities.
印奇 later summarized the strategic choice as the intersection of “want to do, can do, and is doable”: sustained willingness to invest, genes capable of leveraging an industry, and an external market that is genuinely ready. None of the three can be missing.
28. The biggest mistake of technical teams is using addition to search for a “bigger market”
张小珺 relayed a Megvii co-founder’s review: smartphones, finance, security, robotics, and intelligent driving all blossomed at once, with each capillary receiving only a little blood. 印奇 answered directly: “Yes.”
The underlying reason was that every existing vertical was deemed “not big enough,” so the company kept searching for a larger industry and pushed resources toward the seemingly biggest one—security—weakening businesses that had already been validated.
He also accepts the criticism of “technical ego.” Young technical people can be distant from daily life and specific industries, loving technology and problem-solving but lacking genuine passion for users. Naturally, they carry a hammer around looking for nails.
The correction today is a “dual-wheel drive”: retain faith in technology while going deep into one scenario and its users. All business should ultimately be user-centric; technology is the underlying belief and the means of solving problems.
29. The capital markets should be a lifeline, not the destination of entrepreneurship
If he could have fought only one battle over the past decade, 印奇 would have chosen cars. The failure to make that trade-off earlier came from organizational inertia and the hope of going public and achieving a “home run.”
His reflection is that “the capital markets are a critical intermediate node for staying alive, but going public really changes nothing fundamental.” The fundamentals remain a good product, good operations, and a positive loop in which profits feed back into R&D.
The IPO objective can force a company to emphasize short-term results and even abandon long-term investment. The hardest job of a CEO is to withstand the pressure between the long and short term and keep doing what is right over the long term but unattractive today.
He does not reject IPOs for every AI company. Some companies could face serious cash and future-commercialization risks without one. But for many, going public may weaken the pressure and motivation to compete at the core, while constraints make continued foundation-model investment difficult—in effect, an exit from the frontier race.
30. Cars offer both a deep enough technical slope and a large enough single market
印奇’s two hard filters for a sector are technical difficulty and market size. Smartphone imaging could produce a top-two player, but the total pool for third-party suppliers might be only about RMB1B: RMB500M–RMB600M for number one and RMB200M–RMB300M for number two, with an obvious ceiling.
Automotive OEMs are more fragmented and better suited to suppliers seeking scale. Intelligent driving, cockpits, and Robotaxi all depend heavily on models, making cars the most suitable commercial entry point for combining AI with large terminals.
But “the point of fighting for cars is not cars.” The automobile is the most important route to embodied intelligence. 印奇 expects more general embodied configurations to appear in 5–7 years. They may not be fully humanoid, but will be partly human-like because tables, chairs, and spaces are designed around the human body.
31. The real temptation for smart people is replacing long-term compounding with shortcuts
Megvii’s first roughly 40 people included a large number of math and physics Olympiad medalists. 印奇 says it was “one of the most talent-dense organizations” at the time, but the conclusion he reached years later was: “Intelligence is not that important.”
Smart people fall into two groups: those who can collaborate and those who cannot. More scarce are people who are technically strong, collaborative, mission-driven, and persistent over the long term. Intelligence creates too many available shortcuts; without a mission, smart people may shift at any moment to faster monetization in fields such as finance.
Another risk is that smart people dislike using “stupid methods.” Much correct work depends on routine, solid, repeated effort over time, while smart people are prone to seek shortcuts and end up taking the wrong path.
StepFun therefore no longer hires on technical ability alone. People with oversized egos who cannot collaborate are rejected even if they are capable. R&D still has an intelligence threshold, while sales and other roles depend more on goal orientation, determination to deliver, and customer service—the core question is fit with the profile.
32. An AI-native organization needs both extreme talent density and coordination on large projects
Model development naturally requires data, algorithms, and systems talent to work together. StepFun is also deliberately creating cross-training between algorithm and systems staff to prevent research and engineering from becoming isolated camps that cannot understand one another.
印奇 expects leading AI companies may still employ 10,000–20,000 people rather than remaining tiny teams. The difference will be far higher talent density than the 30,000–100,000-person internet giants of the past, requiring a rebuilt approach to organization.
Such organizations cannot be purely top-down or purely bottom-up. Large models are resource-intensive moonshot projects that require concentrated force, while top talent must still have enough bottom-up energy and autonomy.
His reading of The Innovator’s Dilemma is also organizational: if an old organization can adapt to a new business, a large company may do very well. If it must dismantle excellent capabilities that have already solidified and rebuild them, it will probably fail to transform.
33. AI 2.0’s essential differences are scaled engineering and the shift from To B to To C
Compared with AI 1.0’s small models, scale-up makes engineering ability essential for algorithm staff. Compared with the previous generation’s limited generality and leverage, which confined it largely to To B, this generation’s theme is To C.
On the “four little dragons” and “six little tigers,” 印奇 sees many fundamental similarities: professors and technical staff achieved personal value, but a true platform-scale commercial company has yet to emerge. The outcome will depend on whether an AI-era BAT can appear over the next 3–5 years.
梁文锋’s claim that only Huanfang made money in AI 1.0 is “yes and no” in his view. Quantitative monetization was more direct and Huanfang made an excellent strategic choice; other companies chose longer commercialization paths, but even in a re-run he might not switch to quantitative trading.
AI lowered the technical threshold while capital over-invested, pushing prices down and making profits difficult to form. Companies turned to capital to stay alive. No one truly broke out in that phase, and it was not entirely a matter of any single company’s capability.
34. StepFun will use “technical conviction, value pragmatism” to filter out technical opportunists
印奇 has watched former Megvii colleagues join platforms such as Kimi. He tries to retain them while also feeling happy about their growth. Helping smart people with sound values develop, he says, is the source of satisfaction second only to making AI work.
The opposite of “technical conviction” is “technical opportunism.” One test is whether a person can sit on the bench. Some PhD students publish several papers quickly in their first year; others spend their first 3 years searching for a sufficiently deep “big hole,” then produce a run of high-quality work in years 3 and 4.
That choice runs through a résumé, a field of study, and a career path. 印奇 does not favor people who jump frequently, with every stint short but decorated with impressive labels. They may look good in the short term, but may lack compounding competitiveness over time.
“Value pragmatism” is close to the idea of benfen in the BBK system: value delivered to customers, suppliers, and the team must be solid. Technology can be obsessive; delivered value cannot rest on storytelling.
35. AGI’s 3 clear lines are learning paradigms, memory, and the physical world
印奇 sees the learning path moving from imitation learning to reinforcement learning and then to some form of autonomous learning. The names may change, but the broad direction of the 3 stages is relatively clear.
The second line is how to build long-, medium-, and short-term memory systems. The third is the move from language to multimodality and then from the digital world into the physical world.
He expects these 3 key elements may be largely solved over roughly the next 5 years. If so, “AGI as currently defined should, in all likelihood, be achievable.” This remains a conditional judgment, not an unconditional timetable.
He only partly agrees that we are entering a “research era.” Modules such as the next generation after Transformer do require research breakthroughs, but the overall effort remains a large systems-engineering project led by sustained R&D rather than a return to pure academia.
36. World models must close the loop through data, scenarios, and applications
Cars are the clearest deployment field for world models, with early implementations already underway. 印奇 says Qianli is confident it can build vehicle-side world-model capabilities over the next year.
A more general world model cannot be trained out of thin air because the required physical data does not naturally exist. Data, a concrete scenario, and an application must be put into a loop and then iterated as the product scales.
World models will therefore first become viable within specific domains and gradually generalize. 印奇 believes embodied intelligence has become hot too early, but that attention may also accelerate progress in configurations, data, and models.
37. The decisive combination is “hold on to the technology” and “grow the profit” at the same time
On the technology side, the company must keep building world-class model R&D and organizational capabilities. On the funding side, it may need to consider both primary and secondary markets. 印奇’s annual investment forecast is RMB3B–RMB5B.
At a minimum of RMB3B per year for 5 years, that means RMB15B. On the application side, a core product must appear within 3–5 years and potentially generate RMB3B–RMB5B in annual profit to cover continued R&D.
China’s large To B market is concentrated in government, central state-owned enterprises, finance, internet companies, and OEMs. The first group has high business-development requirements, while the latter groups have high in-house development ratios and limited room for external procurement, making it difficult to absorb foundation-model costs.
Small B is closer to C. Product-led growth in the style of Taobao, DingTalk, and Feishu is what might work. 印奇’s conclusion is that treating To B like To C may be smoother, but StepFun still will not make it the final business model.
38. He has moved from optimism to “cautious optimism” and learned to let two springs pull against each other
The hardest part of more than a decade of entrepreneurship was not any single trough, but filling natural gaps in management, capital markets, and sales. 印奇 divides his career into early exploration, To B from 2013–2019, capital markets and commercialization after 2019, and the current stage.
He spent years searching for a more experienced CEO, eventually realizing that an outside professional manager would struggle to simultaneously win the founding team’s trust, understand AI technology, and learn a new business. An AI CEO is itself a hybrid role combining technology and operations.
He is now more decisive about decisions, hiring, and firing, and sets higher expectations for difficulty. Technical R&D pushes forward optimistically, while commercial operations must reason backward from the end goal. The two “springs” must keep pulling against the same decision.
The result of that tension over the past 2 years is that Qianli will first close the loop on the vehicle side, while StepFun uses core technology to drive a second category of terminals. This is not what he considers a perfect answer, but a viable path that “my capabilities and level were able to think of.”
39. The minimum standard for avoiding the past is not making the same mistake twice
印奇 believes that “things are decided by people.” If the founding team does not fundamentally grow, it will repeat the old path in a new sector. Many lessons in business run against human nature and cannot be learned once and applied perfectly.
He now believes that “sometimes slow is fast,” and accepts waiting for the intersection of “want to do, can do, and is doable.” When the AI competition is moving fastest, one may need to hold one’s own route with greater calm.
The current strategic package is models, cars, and future handheld and wearable terminals. He calls it the minimum set: separate the brain, the existing vehicle business, and innovative terminals, and the technology, data, and revenue loop disappears.
The methodology he most wants to practice remains dual-wheel drive: keep investing in technology while making the product and business work. Without a product there are no users, hard cases, or data; making the model good alone is “far from enough.”
40. He calls the next stage a “massively multivariate equation,” but is no longer fixated on a single outcome
Yao Class taught 印奇 to view entrepreneurship as solving world-class problems. Today’s problem is “a massively multivariate equation” containing technology, products, funding, organization, scenarios, and timing.
He no longer defines success through an IPO or ranking. He cares more about choosing the right direction and executing with the right method. The outcome depends on many factors, but the process is what entrepreneurs can truly control.
His hope for the next decade is that China will produce a world-class AI company with commercial results, leadership in frontier technology, and leadership in corporate culture. 张小珺 asked whether he had to be the one to build it. He answered: “Preferably me, but it doesn’t have to be me.”
His current bet is still just 3 words: “build terminals.” Complete the layout in 2025–2026, then settle down to refine the technology and products. He wants the number of things he does to keep shrinking, rather than returning to “playing every hand.”