Allen Zhu: How DeepSeek Changed My View of AI

Zhang Xiaojun recalls her report on Allen Zhu a year earlier, when he was cautious about investing in Chinese foundation-model companies. In early 2025 she invites him back to discuss how DeepSeek has changed his views.

Zhu says he spent the Lunar New Year studying DeepSeek and found the experience exceeded his expectations. He had doubted whether the existing technical approach could reach AGI, but now felt he could see a possible path.

He remains uncertain: reinforcement learning may have clearer feedback where rewards are well defined, while domains without clear rules still need research into high-quality guidance. This was his assessment at the time.

Asked about the impact on OpenAI, Zhu offers a conditional view: if expensive closed models cannot provide a substantial advantage, open models will pressure their business model.

He calls the position of a leading company a “leader’s curse”: after heavy research spending, it has an incentive to remain closed and recover costs, while changing course becomes difficult once competitors catch up.

Zhu believes DeepSeek’s user growth and developer ecosystem could increase that pressure. The user counts, cost comparisons and growth rates he cites are statements from the interview, not independently verified figures.

He questions the value of a closed model if it costs more without a large capability difference. He also allows that proprietary data or specialized applications could provide differentiation.

Zhu is impressed by the writing and depth he perceives in DeepSeek’s answers and says he asks it difficult questions daily. He considers product experience important to word of mouth.

This leads him to speculate that a model might have consciousness and that consciousness could be a spectrum rather than an on-off switch. This is his philosophical conjecture; persuasive generated text does not itself establish consciousness.

Zhang reads a passage from the model about consciousness. Zhu says the expression surprised him and recounts a discussion with Liang Wenfeng; neither account verifies machine consciousness.

On open and closed ecosystems, Zhu uses the analogy of Android and iOS. He thinks the opportunity for closed models depends on sustaining a clear advantage, while open models already meet some needs.

He predicts that models might improve in coding and other professional fields over the following six to twelve months. This was an early-2025 forecast, not a guarantee of real-world ability or professional reliability.

Zhang asks why OpenAI moved toward closed models. Zhu speculates that technical leadership and the pressure to recover investment may be reasons; he does not present this as inside knowledge.

For Chinese application founders, he thinks open models and lower usage costs may reduce barriers to building, allowing teams to focus more on users and use cases.

Zhu expects some large companies may continue developing models because of proprietary use cases, data and existing products, while also learning from open frameworks. He presents these as possible choices.

Asked why he had expected open models to catch up, he says the key question was whether returns from scaling frontier models would slow. The speed and cost of the catch-up surprised him.

He emphasizes high-quality training data, comparing how teams organize data to chefs using different ingredients. His explanation of DeepSeek’s data and team preferences is speculative.

For tasks with less clear rules, he thinks domain experts may need to provide high-quality examples and feedback. He does not disclose the actual composition of DeepSeek’s internal data.

On OpenAI, Zhu says it was still releasing leading products, but heavy spending could create business challenges if its lead did not persist. This is his assessment of the future.

For other Chinese model companies, he outlines possible paths: continue foundation-model work, participate in the open ecosystem, or pursue vertical applications. Each company must decide according to its own circumstances.

Asked about DeepSeek’s technical advances, he attributes efficiency to multiple innovations and its reinforcement-learning approach, suggesting less human intervention can aid scaling. This is an investor’s summary, not a complete account of training.

He thinks DeepSeek is not merely catching up but has also innovated in some areas. He acknowledges that outsiders do not know exactly what techniques closed models use.

Zhang asks about Liang Wenfeng. Zhu connects the product’s expressive style to team preferences, though he has no direct evidence about the training corpus or the team’s intent.

If DeepSeek opened a funding round, Zhu says he would invest to participate in and witness the technology’s development. He later explains that price would still affect the amount, with a smaller check at a high valuation.

He returns to his conversation with Liang Wenfeng about consciousness and frames potential machine consciousness as a possibility. The discussion is opinion and speculation, not an established technical finding.

On the value of user growth, Zhu says retention matters more than first use alone. He predicted conversational tools would strongly disrupt search; that was a forecast, not proof that search had already been replaced.

He distinguishes search from personal assistants. He likes longer answers and follow-up questions for finding information, but still doubts whether the demand and experience are mature enough for AI to arrange travel on someone’s behalf.

Zhu revisits a change in his thinking: he once emphasized user-data flywheels, but now believes repeated conversations may add little information and high-quality specialist data may be more valuable.

He thinks an application’s defensibility may come from retention, brand recognition and workflows rather than data volume alone. He again likens different model data choices to chefs with different styles.

On DeepSeek’s future, he admits he does not know the best business model. He mentions an ecosystem, usage-based charges and cloud partnerships, and considers building capability and an ecosystem the more immediate priority.

He thinks domestic chips can support some inference workloads, though feasibility and cost depend on hardware, models and deployment. His interview assessment does not replace a technical evaluation.

If he were running DeepSeek, he says he would strengthen the model and open ecosystem and consider fundraising to support investment and provide room for error. He presents this as his own advice.

Zhu thinks the company need not necessarily depend on a strategic investor and makes a specific user-growth prediction. The forecast reflects expectations at the interview and requires separate verification against outcomes.

On valuations, he argues that investors should reassess closed-model businesses if they cannot maintain a clear lead. This is his investment view, not a definitive valuation of any company.

Zhang asks about Stargate and large infrastructure investment. Zhu questions the returns on massive compute spending, arguing that outcomes depend on scaling behavior and the evolution of data and algorithms.

On NVIDIA, he expects continued long-term demand for compute but says hardware choices and purchasing growth may change. His comments on market expectations and the stock are his personal views at the time.

Asked how a “realist” views Liang Wenfeng, Zhu notes that DeepSeek’s financial and compute background differed from a typical early-stage startup. He thinks that gave the team room to pursue its ambitions.

Revisiting the prior year’s views, he still argues that most application founders should focus on users and workflows rather than train foundation models. He also acknowledges overestimating the value of ordinary user-data flywheels.

He suggests founders consider delivering a service rather than only software, designing cost and experience around customer needs. He gives a contact center as a possible example.

On small teams seeking product-market fit, he says adding people does not automatically solve the direction problem. Some teams have found opportunities, while others are still searching.

Looking back at 2024 and early 2025, he highlights post-training, DeepSeek’s emergence and results from large-scale training. His assessment of training results draws on information he had heard at the time.

He discusses video and speech products including Kling and Hailuo, seeing opportunities in multimodal AI while saying companies must decide how to allocate resources between foundation models and applications.

Zhu says he invested in AI applications for businesses and consumers and saw existing portfolio companies add AI features. He feels model experiences were exceeding some users’ expectations.

Asked whether a new landmark product moment had arrived, Zhu says conversational products had already surprised him and predicts other applications could follow. He remains unsure about opportunities to build exceptionally large startups.

On open-model strategies at companies including ByteDance and Alibaba, he says a shift can be difficult and depends on technical capabilities, ecosystem compatibility and existing business. His judgments reflect the interview’s moment.

Discussing products such as Perplexity, Cursor and Devin, they consider how applications can distinguish themselves from model providers. He focuses on users, workflows and sustained value.

Zhu takes a pragmatic view of agent labels: he cares more about how much work a system can do in a specific setting and whether it requires human intervention. His figures for future automation are predictions.

Zhang asks whether AI could create new content platforms. Zhu thinks existing platforms may benefit too, while acknowledging that the opportunity for new platforms remains unclear.

On Xiaohongshu, he highlights lifestyle content and community atmosphere and mentions translation. He cautions that more AI-generated content could affect the community and requires observation and adjustment.

He reiterates that startups should stay close to users rather than rely on data alone. Sustained customer relationships and experience may matter more, he says.

He discusses AI progress at ByteDance, Tencent and Alibaba in terms of product investment, following competitors and model ecosystems. The comments on products, prices and strategy are observations from early 2025.

Asked whether Li Auto needs its own model, Zhu says many companies could consider existing open models. This reflects his business judgment rather than a technical answer for every use case.

On embodied AI, he sees strengths in China’s hardware supply chain but says an impressive demonstration does not equal a business model. He says he invested in some robotics-related companies.

Looking ahead to 2025, he sees opportunities in AI applications and consumer products, emphasizing user needs and global markets. These are investment views from that time, not guaranteed outcomes.

His advice to founders is to understand their strengths, focus on commercialization and plan financing carefully. He notes that not every team has resources for long-term foundational research.

Looking ten or twenty years ahead, he thinks AI may change work and organizations and speculates about a shorter workweek. He acknowledges that change is fast and the future difficult to see.

He discusses entrepreneurship in Hangzhou and the Yangtze River Delta, where his team had seen many projects. His regional investment proportions reflect his own team’s experience.

Zhang poses questions generated by DeepSeek about whether a pragmatic investor would make an exception for poetry or ideals. Zhu says the risk he is willing to take depends on the size of the check.

He recalls making a smaller initial investment in Xiaohongshu and participating again as the product developed. He uses the example to explain how risk and check size can be matched.

Asked whether he would go all in on one adventure, he says the team follows both consumer businesses and AI: one reflects pragmatism, the other allows technological exploration. He does not advocate staking the entire portfolio on one bet.

He finds DeepSeek’s emergence a reminder of entrepreneurship’s appeal, while saying institutional investors must distinguish technology, market and team risks and plan allocation across projects.

On multimodal AI, he considers language understanding central to this period of development; that is his view of the technical priorities. He and Zhang then discuss games and consumer experiences in Singapore.

In quick questions, Zhu mentions enjoying Singapore’s bak kut teh and durian and recommends a series of novels by Xu Gongzi. He again expresses optimism about AI application startups at the time.

At the end, he suggests that listeners try a model themselves, ask questions they care about and form their own judgment. Zhang closes the program.