101: Talking AI ✖️ Healthcare with 王小川: Toward the “Mathematical Principles of Life”
Summary
The differentiation of Baichuan M-1 Preview is not simply reproducing O1, but combining general reasoning, vision, search and evidence-based medicine. 王小川 says the model can search PubMed and other literature to support diagnostic hypotheses. He cited a patient in Inner Mongolia who had been issued a death notice: M-1 proposed 3 possible diagnostic directions, while doctors at Peking Union Medical College Hospital proposed 4 after the patient was transferred there, 3 of which overlapped with the model. This remains a single case rather than systematic clinical validation; the real value depends on whether it can reliably reproduce performance “far beyond the level of a prefecture-level hospital.”
王小川’s healthcare bet is not a retreat into a vertical niche, but a choice of a high-value field that general models cannot swallow and that will rise with model capability. Multimodality, long-term memory, embodied intelligence, literature search and hallucination reduction all have applications in healthcare, while medical data and new medical paradigms create barriers general models lack. His core formulation is that “healthcare is the jewel in the crown of foundation models” and “building a doctor is equivalent to AGI.” The investment implication is that Baichuan is not competing with the majors for the general-purpose entry point, but trying to establish the “super doctor” as a control point beyond the majors’ reach.
Commercialization starts with pediatrics, primary-care general practice and out-of-hospital management; under the plan described in the interview, Haidian residents could have access to an AI doctor’s assistant connected to regional medical records as early as 1Q25. Beijing Children’s Hospital has proposed “building 1 million pediatricians,” and 王小川 says roughly 80% of pediatric issues may eventually be handled outside hospitals. The “one big, four small” product architecture covers households, communities, city-level hospitals and children’s hospitals, supported by a cough-audio model that distinguishes upper- from lower-respiratory infections. He summarizes the future form as “Hospital at Home”: early screening, diagnosis, chronic-disease management and follow-up should move as far forward as possible into homes and primary-care settings.
Baichuan has designed 3 payment paths—G, H and C—but near-term revenue still depends on its ability to navigate complex government and hospital implementation. The G side covers family doctors, primary-care strengthening and regional healthcare projects; the H side ultimately points toward reimbursement; the C side may evolve from individual payments into multilayer commercial insurance. Overseas deployment will require local hosting. 王小川 only committed to seeking progress in developed countries during the year and did not name a first market. Early customers will require heavy in-house delivery from Baichuan; once the model is proven, deployment can be handed to partners. The execution risk is that even after “one word from the provincial governor,” every layer still has its own KPIs, procurement process and tendering requirements.
Healthcare models will cost more than low-priced general-purpose token services, but 王小川 believes the high value of life and health, the scenarios themselves and recurring services can cover the cost. Baichuan needs control of pre-train, post-train, reinforcement learning and RAG, while also calling on stronger external models when useful. Once AI doctors are deployed, the data path moves from “enrollment upon admission” toward “enrollment at birth,” turning lifelong care trajectories into life-science research data. The furthest-term moat is not a single consultation, but a clinical-research paradigm in which “AGI comes first, then life sciences.”
China’s model catch-up has narrowed from “one step behind” to “half a step behind, or very close,” but Baichuan is open-sourcing a 14B medical model to build an ecosystem rather than joining the general-model price war. 王小川 calls distillation an industry “open secret” and says some rapidly released O1 replicas merely have “similar looks” (“品相长得比较像”). He particularly recognizes DeepSeek’s performance, while noting that it already has compute and does not face the financing and delivery pressures typical of a startup. Baichuan’s 14B version is said to score higher on medical benchmarks than Qwen 72B and can run on a single 4090, with the goal of letting hospitals, checkup centers and researchers fine-tune it themselves and become partners.
The technical center of gravity remains language: vision handles interaction, reinforcement learning turns “fast thinking” into “slow thinking,” and the next paradigm shift may be AI moving from calling tools to creating them. 王小川 expects an AI “doctor friend” to handle some difficult and rare-disease cases within 2025, though surgical robots remain out of reach. Over the longer term, he believes AI will compress the division of labor, redefine learning and human value, and perhaps lead to human-machine integration. The upside is enormous, but so is the uncertainty: the next 10 years will be earth-shaking, yet he admits he still “can’t figure out” how people will express their value once supply explodes.
Deep dive
1. M-1 Preview first differentiates Baichuan through “evidence-based” medicine
王小川 set M-1 Preview the dual objective of placing its general capabilities in the first tier while making its medical capabilities distinctive. The model therefore combines general reasoning, visual reasoning and search reasoning rather than simply boosting its medical-exam scores.
He defines “evidence-based” as the combination of search and reasoning: the model does not merely learn papers, but actively searches sources such as PubMed so its answers are supported by the literature. “What evidence-based really means, most importantly, is support grounded in papers.”
程曼祺 asked why Baichuan was not the first Chinese company to reproduce O1. 王小川’s answer was that reproduction is not the objective; every release needs to establish the mental model of “general capability plus medical enhancement.” Chasing general capability first and building healthcare separately would consume “double the effort.”
2. A PUMCH referral case shows the model’s value—and the limits of the evidence
王小川 recounted the case of a patient in Inner Mongolia diagnosed with cerebral infarction. After roughly 2 weeks in hospital without improvement, the hospital had issued a death notice. The family refused to give up, handed the medical records to Baichuan while transferring the patient to Beijing Peking Union Medical College Hospital, and uploaded them into M-1.
M-1 produced 3 possible diagnostic directions and the tests needed to validate each. After the patient arrived at PUMCH, doctors proposed 4 directions, 3 of which “hit exactly” against the model’s output. 王小川 said this showed the model was already far beyond the level of a prefecture-level hospital.
His counterfactual judgment was that if the local hospital had continued diagnosing along those lines and then treated the patient, “the patient would have been saved immediately.” But the interview provided only this case and did not detail the final diagnosis, treatment outcome or any large-scale evaluation.
3. The 14B open-source model is a gateway into the medical ecosystem, not a traffic play
Baichuan previously open-sourced Baichuan 1 and Baichuan 2, but not Baichuan 3 or Baichuan 4. M-1 is again being offered in a 14B open-source version because healthcare cannot be completed by Baichuan alone: hospitals, checkup centers and research teams need local deployment and secondary fine-tuning.
王小川 says the 14B model scores higher in healthcare than Qwen 72B and can be deployed on a single 4090; larger models are materially harder to deploy. The aim is to lower the industry’s entry barrier and build hospital trust through transparency.
程曼祺 asked whether the move was inspired by DeepSeek and the open-source wave of January 2025. 王小川 repeatedly denied that, calling it a pre-existing plan. Concentrated releases by peers merely diverted attention, he said, while the healthcare sector could still identify Baichuan’s “unique contribution.”
4. China’s O1 catch-up relies on distillation, with some “looks-like” replication
王小川 does not avoid the subject of distillation, calling it an industry “open secret.” Data that once required human labeling can be generated by a stronger model; DeepSeek also distilled large models into a series of smaller ones, avoiding the need to retrain every size from scratch.
He also believes some quickly released models merely resemble O1 in form while remaining materially behind in capability. “Releasing quickly” therefore cannot be equated directly with achieving an original breakthrough at the same level. On Doubao 1.5 Pro’s emphasis that it used no data from other models, he said this reflected a major tech company’s independent choice on originality.
On DeepSeek, he revised the gap between Chinese companies and OpenAI from “one step behind” to “half a step behind, or very close.” But he stressed that DeepSeek already has compute and does not face the financing and commercial-delivery pressures of a typical startup. “It is not a typical startup,” and its operating environment is difficult to replicate.
5. Baichuan uses healthcare deployment to turn “one step behind in ideals, 3 steps ahead in execution” into reality
王小川’s timeline was: complete fundraising and team-building in 2023; convince people in 2024 that healthcare is the future and attract medical talent; begin healthcare deployment in 2025. He does not see healthcare as a difficult scenario that must be unlocked last, but potentially as one of the first areas to generate material value.
His technological idealism does not exclude industrial outcomes. Technical breakthroughs must connect to life sciences, while a startup’s next financing round cannot be built around the model alone. “Now you have to land in the industry. Everyone wants to hear your industrial story.”
Baixiaoying is a self-criticism worth retaining. When released in May 2024, it was still a broadly oriented chatbot. 王小川 said the company “had to release it—the pressure to release was coming from everywhere,” making it an action driven by external pressure. Only afterward did the company continuously raise healthcare’s internal priority.
On the major tech companies, he consistently stresses that organizational capability must be respected, describing DeepSeek in particular as “the most core player at the table.” Baichuan’s strategy is not to assume the majors are slow, but to ensure that the company “always knows it has to be outside their range.”
6. Foundation models are not recreating the physical world; they are using mathematics to deconstruct humans
程曼祺 placed technological idealism alongside scientific skepticism. 王小川 distinguished the two: faith is the broader term, while technological idealism is a subset; believing in science means believing experimental verification can help humanity discover the unknown.
He believes the old scientific paradigm relied too heavily on physics and mathematical formulas, using mathematics to model the objective physical world. Large models begin with language, and the modeled subject becomes the human being. “In the past, mathematics was used to deconstruct the physical world. Now it is being used to deconstruct people.”
In this framework, knowledge is not the physical world itself but humanity’s projection of and “reflection” on it. Language carries communication, knowledge and thought. Its inability to define every concept precisely is not a reason to abandon it; it shows instead that a new mathematical representation need not amount to a compact formula.
7. The question of the “mathematical principles of life” began with gene-sequencing research in 2000
王小川’s graduate thesis in 2000 focused on an algorithm for assembling gene sequences. Even then, he was drawn to the question: “What are the mathematical principles of life? How does life work?” He gradually came to believe that traditional mathematics and physics were insufficient to explain living phenomena.
He did not learn traditional Chinese medicine in order to call it science, repeatedly emphasizing that “traditional Chinese medicine is a philosophy.” His logic is that when a theory cannot explain an existing phenomenon, one should not simply declare the phenomenon wrong, but search for a new theoretical framework. Baichuan is still conducting research related to traditional Chinese medicine, though he disclosed no details.
He uses the fertilized egg as an example: a single cell can develop into a human being who resembles its parents, which seems natural but is mathematically “inconceivable.” The three-body problem and long-range weather may both be uncomputable, yet life generates order from even greater complexity. He therefore juxtaposes a precise physical world characterized by rising entropy with a complex living world that appears to reduce entropy.
He also retains doubts about the claim that the universe as a whole is subject to increasing entropy. Entropy increase requires a closed system with no exchange of matter or energy, but humanity cannot verify that the universe satisfies those conditions. “You cannot derive the infinite from the finite”; to him, this is reason continuing to move upward, not a retreat into intuition.
8. AlphaGo and the 魏则西 case jointly pointed to a language breakthrough and more doctors
王小川 sees AlphaGo in 2016 and the 魏则西 case as a paired contrast: on one side, AI capabilities were astonishing; on the other, ordinary patients still had no solution to their healthcare predicament. Two conclusions followed: “Only when machines master language does strong AI arrive,” and the bottleneck in healthcare is not appointment access but a shortage of doctors.
Registration platforms, Chunyu, Haodf and similar models solved connection, but did not increase the supply of doctors. Academicians and top physicians focus on frontier research, while ordinary people need trustworthy primary-care doctors. His proposition is therefore not for AI to assist only elite research, but first to ensure that “everyone has a doctor to use.”
During his Sogou years, he invested in more than 10 healthcare companies spanning CRISPR, organ transplantation, gut-microbiome drug development, internet healthcare and medical AI during its trough. Sogou’s 2021 farewell letter set out a 20-year plan around “life sciences and public health”; in 2022 he founded Wuji Health.
9. ChatGPT turned years of language accumulation into an executable AGI path for the first time
Input methods were already doing “predict next token,” but in the past they could predict only short words and phrases. Search engines had also long tried to move from retrieval to question answering, but could return only article indexes. 王小川 therefore sees his judgment on ChatGPT not as a sudden cross-industry leap, but as a continuation of accumulated work in input methods, search and question-answering systems.
In 2023, he entered only a few questions before concluding that this was not an incremental upgrade to old AI but “the world has changed; history has turned the page.” He immediately declared that “the era of strong artificial intelligence has arrived; AGI has arrived,” based on the fact that machines had genuinely mastered language for the first time.
By 2024, he further reframed the “Fourth Industrial Revolution” as a shift “from the scientific era to the intelligent era.” AI was not merely raising production efficiency; it was beginning to change the paradigm of scientific research, making “industrial revolution” an insufficient description.
Baichuan did not publicly emphasize healthcare at its founding because the financing narrative first had to benchmark against OpenAI and focus on general models. Once funding was secured and the team understood the scenarios, he explained the Bio meaning embedded in “Bai” and elevated medicine to the company’s most important direction.
10. Healthcare is the “jewel in the crown” because every AI capability can compound there
王小川’s first message to the team was that “healthcare is the jewel in the crown of foundation models.” Multimodality can read medical images; embodied intelligence can be applied to healthcare; long context can retain a lifetime of medical history; search can retrieve papers; and models must continuously reduce hallucinations. Almost every AI capability has a healthcare application.
He rejects the idea that applications are merely “eggs laid along the road” as models develop, preferring the image of rising water lifting all boats: the stronger the model, the more powerful the healthcare application, rather than the application being absorbed by the underlying model. Healthcare has a high enough ceiling to support a “super-application above the super-model.”
He cites Demis Hassabis’s view that biomedical progress over the next 5 to 10 years could exceed that of the previous 50 to 100 years. The reason is not merely that AI can summarize existing experimental data, but that it is beginning to participate in experiment design and create new data. Healthcare thus extends from diagnostic applications into AI for Science.
11. “Building a doctor equals AGI” turns abstract intelligence into a testable target
王小川 sees a contradiction in the industry: people say AI can do anything, but when someone proposes having a model act as a doctor, they say it is too difficult. If a model is still inferior to a doctor, one cannot casually claim that intelligence has been achieved. He therefore sets general-model capability equal to doctor capability.
程曼祺’s objection was that the natural sequence should be to build a “human” first, then doctors, lawyers and other professions. 王小川 countered that “human” is too broad—writing advertising copy is also a human capability—while a doctor is one of the professions with the highest level of complexity and intelligence, making it a better yardstick for AGI.
He uses the fact that there are as many natural numbers as even numbers as an analogy. Doctors may appear to be only a subset of humans, but infinite sets can be mapped one-to-one; similarly, the capability set required to build a doctor may be as large as that required for general intelligence. “Turn the proposition around this way, and building a doctor is AGI.”
12. The real challenge in shifting from a general team to a healthcare organization is defining the problem together
The strategy was proposed at an internal strategy meeting around July. 王小川 estimates that the team initially understood only “50% of it.” Engineers asked what they could do; over time, the organization moved toward “one medical textbook per person,” with every technical team required to connect its work to healthcare.
Baichuan already had a medical product division and more than 30 doctors, and had acquired a roughly 40-person medical-engineering company researching capabilities such as cough-audio analysis. Medical staff define the problems and build evaluation systems; technical staff must understand medical data, task definitions and evaluation standards.
王小川 admits the integration of the 2 talent pools is still being worked out, but says the team is beginning to develop combat capability. He cites the new medical director, who spent 8 years in a Tsinghua-PUMCH joint program, earned a master’s degree from Johns Hopkins University and worked 6 years in general internal medicine at PUMCH before turning to the development of AI doctors.
The capital logic has shifted with the strategy. Early on, it was difficult to raise money without talking about general models; now “you can’t raise money by talking only about models.” Healthcare deployment gives local governments real, open scenarios and has become an outcome Baichuan must prove for its next financing round.
13. Pediatrics is the fastest commercialization scenario because of its supply gap and out-of-hospital demand
The president of Beijing Children’s Hospital proposed “building 1 million pediatricians,” becoming the first hospital leader 王小川 encountered whose thinking closely matched the “build doctors” concept. Compared with academicians who insist that AI can never become a doctor, the shortage of pediatricians makes the idea easier to accept.
The hospital believes roughly 80% of pediatric issues will not need to occur in hospitals in the future. Repeated visits create overcrowding, cross-infection and work stoppages for entire families; younger parents are also more willing to try new technology. Pediatrics therefore combines policy value, user acceptance and potential out-of-pocket demand.
Baichuan and the hospital are planning “one big, four small”: one super-large model covering households, communities, city-level hospitals and children’s hospitals. Simple issues are handled by parents and the model at home; complex cases move into hospitals, while AI also serves as an assistant to human doctors.
The accompanying cough-audio model attempts to determine upper- versus lower-respiratory infection from a single cough. The former can be observed at home; the latter requires hospital care and intervention. 王小川 said the products were originally expected to take shape in 1Q25 and see further releases in March and April.
14. “Hospital at Home” shifts the center of care from hospitals to homes and primary care
王小川 uses “Hospital at Home” to summarize the future of healthcare: early screening, early detection, early diagnosis, chronic-disease management and full-course care should mostly not be confined to hospitals, but should continue in the home environment.
This aligns with the policy direction of strengthening primary care while creating incremental commercial value. Baichuan is not trying to take existing work away from human doctors inside hospitals; it is helping them detect risk earlier and manage patients outside the hospital, then returning patients to hospitals when surgery or critical-care resources are required.
Out-of-hospital care does not mean leaving the healthcare system. At home, “parents work with the super model”; in hospitals, the same model becomes a doctor’s assistant, linking follow-up, repeat visits and out-of-hospital management. 王小川 believes a product is fully usable only when it connects consumers, hospitals and government at the same time.
15. The Haidian pilot connects an AI GP to real medical records and family-doctor services
In general practice, Baichuan positions the AI GP as a primary-care AI doctor, with Haidian District as the key pilot. Under the plan described in the interview, Haidian residents could receive an AI doctor’s assistant in 1Q25, with the model connected to the regional healthcare system and existing medical records.
王小川 points out that medical records nominally belong to patients, but patients often cannot obtain complete data, and even when they do, they struggle to use it. The value of deep government cooperation is to create a “dual-doctor model” connecting records, the AI doctor and community doctors, rather than another standalone chatbot.
The plan aims to address the practical bottleneck in the family-doctor program: primary-care doctors are already saturated, while residents lack trust, making it difficult to expand through contracting alone. AI adds supply and continuous service; complex cases still move into hospitals.
On cooperation with Ningbo, he cooled expectations. The deal was still being signed, local approaches differed and the Haidian model could not be copied directly. The interview contained relatively complete information only on the Beijing Children’s Hospital and Haidian projects.
16. 3 classes of payers jointly support the high service value of medical AI
Baichuan’s first priority is not to choose between To C, To H and To G, but to build the “super doctor.” Once the core doctor capability exists, products for patients, medical students, hospitals and governments can all expand outward.
G-side revenue comes from family doctors, primary-care positions and regional public-hospital reform. The H side is purchased by hospitals and may ultimately enter reimbursement. 王小川 noted that AI healthcare services were beginning to be classified as “doctor services” rather than medical devices, although at the time this mainly applied to imaging projects; he expects the scope to expand.
The C side can initially be funded by individuals, but he believes users with the same condition will eventually aggregate into multilayer commercial insurance. The third path is therefore not permanent subscription revenue, but individual services gradually integrating with commercial insurers.
International expansion is the more important source of incremental growth. Chinese healthcare companies historically served almost exclusively domestic users, while models may globalize. Baichuan plans to meet data requirements through local deployment and seek a result in a developed country during 2025, but he did not commit to a first market or specific partner.
17. To B and To G are not project sales; they are about keeping AI doctors continuously “on duty”
王小川 acknowledges that Baichuan must deliver heavily in-house for its early customers. Once the first few reference cases work, deployment and surrounding IT can be handed to partners. The company’s core value must remain the model and AI doctor, rather than sliding into custom outsourcing.
Shenzhen is a priority region outside Beijing. It embraces innovation, was an early explorer of medical alliances, and has a mismatch between per-capita wealth and healthcare supply. Its need for AI to supplement medical capacity is therefore more direct than Beijing’s, where resources are abundant.
The halo around foundation models has made it easier for Baichuan than during the Sogou years to reach party secretaries, provincial governors, district heads and health authorities. But 王小川’s counterintuitive finding is that higher-level leaders tend to understand innovation better, while lower levels are more constrained by existing KPIs. The evaluation path must be redesigned together.
Healthcare projects still involve tenders and competition from companies that call themselves AI firms. 王小川 says that after Baichuan explains its philosophy and capabilities, partners “without exception feel Baichuan should be selected.” The distinction he emphasizes is that traditional project vendors leave after delivery, while Baichuan’s doctors operate over the long term and remain accountable for regional performance.
18. A medical super-model costs more, but can recover the expense through clinical value and research data
Compared with general models, medical models require more specialist data, dedicated tasks and long-term service, so costs can only be higher. 王小川 rejects the idea of supporting the business through low-priced tokens because the unit value of life-and-health services is high enough to cover greater inference and operating costs.
The capability mix is “50-50.” When stronger external super-models are available, Baichuan can call them; missing capabilities such as cough-audio analysis and smooth, emotionally intelligent human-machine communication must be trained internally. Baichuan still needs control of pre-train, post-train, reinforcement learning and RAG rather than merely wrapping external APIs.
The longer-term moat comes from AI doctors becoming research workers. 王小川 compresses the research-hospital model into “enrollment upon admission”: once a patient enters the hospital, diagnosis, intervention, outcomes and ongoing observation should all leave structured records rather than having clinical data discarded when treatment ends.
The extreme form is “enrollment at birth.” An AI doctor would accompany an individual from birth and create a complete health trajectory, driving life-science research. “AGI first, then life sciences” is the long-term direction he has designed for Baichuan.
19. The data flywheel depends on top-tier medical tasks, not indiscriminate case collection
Baichuan obtains usable data through hospital cooperation, medical-journal tasks and key laboratories in Beijing. Data from Beijing Children’s Hospital is not handed directly to the company; it is shared and studied through a jointly established laboratory entity.
王小川 distinguishes pre-training data from high-quality “gold standards.” The former requires scale; the latter must come from top hospitals and leading departments. Having an ordinary city hospital hand over all its data may not materially improve the model.
On whether a data flywheel makes the model automatically better as users increase, 程曼祺 pressed him on the different situations of MiniMax and Doubao. 王小川 gave no definitive answer, saying only that “there will be some effect, more or less,” while admitting he had not yet found sufficiently good dimensions for defining it.
This reservation matters. Baichuan believes long-term care data can drive both model improvement and research, but it does not equate every increase in users with model growth, nor did the interview offer a unified formula for data quality, authorization and the feedback loop.
20. The super-application will be a “doctor friend,” not merely a low-frequency diagnostic tool
王小川 has not lowered his threshold for a super-application, still measuring it at a minimum of 30M to 100M DAU. China sees roughly 8.4B medical visits a year, or more than 20M a day. He also cites Baidu Health’s roughly 50M DAU and more than 400M requests as evidence that health demand itself is high frequency.
A single diagnosis cannot support a super-application. Chronic-disease management, long-term care and continuous health consultation are what could reach that scale. The C-side form is a “doctor friend” that is always available, remembers personal history, thinks from the patient’s perspective and can even analyze whether a human doctor was reliable after a visit.
On the hospital side, the same capability becomes a doctor’s assistant covering out-of-hospital follow-up and full-course care. The super-application need not be only a mobile app; it can be called a service or an application. The key is that it “has to do some work for you.”
He even imagines a physical “Baymax,” but unlike industrial robots focused on picking up and moving objects, Baichuan wants a machine that understands people: one that reads social cues, understands intention, knows a person’s health status and guides intervention.
21. Healthcare, language and life data form a boundary that general models cannot simply erase
Facing the view that the strongest general model will consume every vertical application, 王小川 uses Einstein as an analogy: even the smartest person may not know how to practice medicine without medical data and training. Life sciences will also generate new data and paradigms that do not exist in today’s general models.
He divides the technology era into 3 stages. The scientific era turned physics into mathematics; the intelligent era begins by turning language into mathematics; the symbiotic era will turn life into mathematics. Life sciences and embodied intelligence lie beyond today’s AI capabilities and therefore will not simply be covered by existing general models.
The major tech companies could still enter, but he says Tencent and Baidu were reducing healthcare investment at the time. Healthcare requires both high innovation and substantial resources: it is too heavy for a startup and not a top priority for a major company, leaving Baichuan a window.
Ant Group’s acquisition of Haodf.com is a clear counterexample. But 王小川 believes its advantages remain in payments, hospital connectivity and business-model design—a complete strategic deployment rather than a risky breakthrough in AI doctor supply. It changes production relations more than it changes supply along the “build doctors” path.
22. Slow thinking, tool creation and organizational redesign will determine whether Baichuan reaches its end state
In M-1, vision mainly handles doctor-patient interaction and does not represent the core of intelligence. The real technical leap is O1 pushing language from “fast thinking” toward “slow thinking.” Reinforcement learning for healthcare is harder than for mathematics, programming or Go because the feedback is not always clearly right or wrong and must also capture differential diagnosis, medical tasks and communication logic.
王小川 insists that “there is no intelligence in the video.” Vision may contribute data and interaction, but language is not a disposable “crutch”; his minimal concession is to call language the “scaffolding” that constrains the overall structure. On Moravec’s paradox, his response is that machines previously could neither use language nor perceive; this time, language was simply mathematized first.
After reinforcement learning, he sees 2 steps. First, models call tools, as with Anthropic’s Computer Use and OpenAI’s Operator. Next, they create tools, especially by writing and running code themselves to solve their own tasks. “Building tools” could be the next paradigm shift because language and tools are classic boundaries between humans and animals.
On timing, he expects the “doctor friend” to answer some complex questions during 2025, including difficult and rare diseases. If general AGI arrives in 2027, as some predict, AGI doctors should arrive around the same time. Surgical robots remain a later stage; further out, human-machine integration may emerge. What continues, he stresses, is “human civilization,” not an unchanged physical body.
23. The end state is not selling the company, but making “building doctors” a sustainable institution
Compared with the Sogou years, this startup has no parent company: it must define itself while rapidly organizing hundreds of employees with different goals. 王小川 says he no longer believes only in technical elites, but must think simultaneously about industry, capital, medicine and organization.
Baichuan’s first end state is to solve healthcare’s “impossible trilemma” so everyone can use a good doctor. The second is to establish a new medical-research paradigm through continuous clinical data. An acquisition is not the core issue; “building the doctor” is. He accepts that he may not personally see the mission completed, but insists Baichuan must take responsibility for its position.
The technical team is not young enough, by his own diagnosis. In 2023, the company needed to hire experienced people quickly; later, it needed people 2 or 3 years out of school and even interns to take on challenging work. Healthcare, meanwhile, still depends on experienced doctors. The organizational challenge is getting these 2 “smells” to genuinely blend.
Entrepreneurship has not improved his health and cannot simply be called happy, but it repaired his understanding of “who I am.” He confirmed his vision in life and health and feels lucky to have participated in the AI wave. The company still has to move from exciting technology to results; technical idealism without scenario awareness is no longer enough.
24. The biggest option in AI doctors is “biological freedom”; the biggest unknown is human value
王小川 borrows the phrase “biological freedom” to describe the long-term outcome: like financial freedom, people would no longer be continually burdened by disease. This is not the same as immortality. He does not want unlimited life himself, believing that after one’s mission is complete, continuing to exist could be little more than “garbage time.”
The goal is to move from “enrollment upon admission” to “enrollment at birth,” reducing illness, detecting it early and shifting from disease intervention to health research. 程曼祺 asked whether individuals would be willing to surrender their full data history. 王小川 responded that patients answer doctors honestly, and many people are more willing to ask online about skin disease or sexually transmitted disease, but he did not elaborate on an institutional authorization framework.
On mental health, he distinguishes psychological, psychiatric and neurological issues. Neurological problems are relatively easier; psychiatric problems are harder. Depression also requires distinguishing pathology from emotion, and some identification tasks may not rely on foundation models alone but call on other technologies.
He no longer calls this change an industrial revolution because previous industrial revolutions made the division of labor increasingly granular and turned people into screws. AI may eliminate large numbers of occupations, compress the division of labor and create an explosion of supply. 王小川 looks forward to a world being “flattened” and most people living better, but admits he “can’t figure out” how people will express their value once supply surges.
His view of the next 10 years is both certain and restrained. Human-machine relations, education and value creation will all be transformed, but many specific outcomes remain unclear. Technology alone cannot solve politics, climate or institutions. In 2025, he is most looking forward to AI healthcare unlocking real-world scenarios; outside AI, the event he is watching is “seeing what Trump gets up to.”