Li Guangmi on Model Competition, Online Learning, and AI Startups at the End of 2025

At the end of 2025, Zhang Xiaojun and Li Guangmi hold their third year-end conversation, reviewing model competition and discussing online learning, robotics, and startups. He says the podcast connects him with researchers and helps him revisit his views.

Asked about an AI bubble, Guangmi prefers the metaphor of an intense “AI war.” He thinks major companies may keep investing for strategic reasons; that interpretation does not prove investments will pay off.

He imagines that if agents become gateways to services, value distribution among cloud, search, and intermediary platforms could change. The scenario depends on agent capability, user adoption, and platform responses.

Discussing infrastructure commitments and revenue, he acknowledges that near-term cash flow is difficult to reconcile. The episode’s figures reflect his reading of reports and future spending, not realized revenue or an independent audit.

He uses hypothetical subscription, advertising, commerce, and API revenue to estimate model-company business potential. Asked about less visible revenue, he adds agents and scientific applications, but their scale and timing are highly uncertain scenarios.

He argues that agents might create new economic value if they reliably complete end-to-end work, with scientific discovery a longer-term possibility. Both depend on capability, cost, reliability, and real-world adoption.

Guangmi is optimistic about advertising and commerce over the long term, but says monetization speed depends on company priorities. His description of any company’s internal goals is an outside inference.

He roughly frames late-2025 competition around Google’s TPU and other companies’ GPU ecosystems, comparing integration of chips, cloud, models, and products. This is an analytical frame; actual partnerships are not limited to two sides.

He speculates that the stronger a leader becomes, the more others may cooperate to balance it. Zhang returns to the models: could leadership alternate? Guangmi sees that as a pattern worth watching at the time.

Comparing GPT, Claude, and Gemini, he expects leading teams to catch up with one another in general capability while emphasizing coding, multimodality, or consumer products. Rankings and revenue figures are estimates from the interview.

He distinguishes alternating leadership from model differentiation: similar overall ability does not prevent teams from optimizing for different users and tasks. Data choices and organizational strategy can shape those differences, he says.

On new research labs, he explicitly says he can only speculate. Whether they pursue online learning or other approaches must be judged from published research and products, not inferred solely from staff backgrounds.

He compares foundation-model companies to broad e-commerce platforms: establish one use case, then expand to others. Application founders must consider which experiences, data, or workflows a platform does not cover.

On competition between Google and OpenAI, he is optimistic about both long-term opportunities. His views on valuations, users, and future share are personal forecasts, not investment advice.

He imagines chat products and search gradually converging, while acknowledging that search incumbents can integrate AI themselves. The effect on advertising and information distribution needs observation.

Zhang asks whether a third paradigm may follow pretraining and reinforcement learning. Guangmi proposes online or continual learning: can models learn more efficiently during interaction rather than relying only on large offline datasets?

He hopes for progress in memory, long context, parameter updates, and learning efficiency, but mechanisms, maturity, and effects remain uncertain. His proposed breakthrough timing in 2026 is a forecast.

Guangmi goes further, suggesting generalization and autonomous learning may matter more than separately extending robotics or world models. Zhang calls this a provocative view; it reflects his research priorities, not a verified conclusion.

He compares pretraining, expert data, and online learning to fossil fuels, renewables, and fusion to convey differences in resources and potential paths. The analogy should not be mistaken for proven technical progress.

He hopes more young researchers can obtain compute and long-term support, arguing that investment and talent development help explore new directions. More resources do not guarantee a particular breakthrough.

After studying data from expert work, he places more weight on agents’ difficulty with real workflows and long-tail tasks. Appropriate data and feedback matter, he says, and he has lengthened his timetable for fully general agents.

He thinks coding, office work, and research tasks may see substantial automation earlier, but that does not mean all knowledge work can be reliably replaced. Both value creation and who captures it remain uncertain.

On investing, he identifies leading models, computing infrastructure, and beneficiary applications as three areas to watch, saying technology changes too quickly to identify one certain winner. His portfolio is a personal hypothetical, not investment advice.

Looking to 2026, he watches online learning, proactive agents, and multimodality, and suggests an agentic web could change gateways and traffic allocation. Deployment depends on technology, users, and business conditions.

He summarizes his view as “the model is the product, and data is the model”: data choices for different tasks can shape experience. Teams also need to consider data quality, acquisition permissions, and real workflows.

His “horizontal and vertical” framework combines learning expert knowledge across industries with exploring the next learning paradigm. The paths can proceed in parallel, he says, but their results require experimental validation.

He describes new research labs and their possible interests in materials, collective intelligence, and world models. For teams without published results, his characterization and forecasts are outside observations from that time.

On robotics, he compares teleoperation, gloves, and other data-collection methods and says generalization and physical hardware remain difficult. Company claims about dataset size and capability should be understood in their own testing conditions.

He says robotics approaches and hardware had not converged, using language-model stages as an analogy for distance from maturity. His timetable is an estimate.

Asked about revenue growth among leading companies, Guangmi lists annualized revenue estimates for model and application firms and discusses valuations. He provides no item-by-item audit; the numbers and investment conclusions reflect his views at the time.

For Chinese model and application companies, he hopes teams find distinct product priorities and predicts continued competition in consumer chat. Future model performance must be judged by releases and user experience, he says.

He says the next stage of competition involves both end-to-end product strategy and potentially a new learning paradigm. Which approach works and who leads remain unknown.

For entrepreneurs of Chinese heritage, he recommends understanding global users and using their teams’ talent and engineering strengths, while acknowledging practical difficulties in cross-border financing, trust, and geopolitics.

He hopes more long-term risk capital will support original, high-risk exploration and says talent, compute, and an open innovation environment must come together. His numerical probability for future Chinese-heritage leadership is a personal guess.