Li Guangmi on AI Task Engines, Model Competition, and Agents in 2025

Zhang Xiaojun asks Li Guangmi to review the fourth quarter of 2024 in large language models and consider what products might come next. Their year-end discussion is full of predictions rather than settled outcomes.

Guangmi suggests that the next major internet product may be an entry point for completing tasks, not just a search box or a chat window. His phrase “the next Google” refers to potential scale and position, not a copy of Google’s form.

He traces the evolution of internet gateways from portals that aggregated information to search engines that responded to keywords, and then recommendation systems that distributed material according to interests. Each changed the relationship between people and information.

Products such as Xiaohongshu combine search and recommendation. Guangmi believes models could add interaction to both: users can find information, ask follow-up questions, generate material, and advance a concrete task.

Such an entry point could cross boundaries between applications. Given a goal, the system would need to decide what to research, which tools to use, how to organize the result, and when to return a decision to the user.

He uses investment research as an example: collecting material, searching, and forming an initial analysis can be parts of one workflow. The product’s value comes from completing more of that workflow, not merely producing one clever answer.

Guangmi speculates that a competitive company might integrate models and applications closely enough to optimize experience, inference cost, and task quality together. He offers this as a possible industry path, not a settled organizational model.

Discussing growth, he estimates from the information available at the time that model-based products already showed broad demand. Different measures and frequencies of use, however, complicate comparisons.

Monetization may also differ from traditional search. Comparing advertising with subscriptions, he uses hypothetical user counts and payments to illustrate that total inference spending can grow with usage even as the cost of an individual answer falls.

Many searches are navigational: the user already knows the destination. Others require understanding, comparison, and continued interaction; Guangmi sees more room for models there, while noting that the product still needs workable costs and pricing.

If a system can take on complete tasks, pricing might shift from queries or monthly subscriptions toward the value of work performed. This is a proposal contingent on reliable delivery, not a description of a proven market.

Zhang asks how a chatbot becomes a task product. Guangmi regards conversation as a transitional interface: future products may also need to show task status, results, and next steps instead of putting everything in one chat window.

He imagines a system that understands a user’s goals and context over time and offers suggestions when useful, rather than requiring a new prompt for everything. It would need to balance initiative against interruption.

Personal context might come from a user’s files and workflows. Guangmi thinks it can improve results, while Zhang points out that access to private data and the risks users will accept are unavoidable product questions.

Guangmi extends the argument to AI software: if models understand requirements and write or modify programs, some workflows may rely less on fixed interfaces. The practical scope still depends on reliability and cost.

He is optimistic about coding’s effect on productivity, but describes a chain of capabilities: understanding a problem, generating code, running checks, and correcting errors, not merely generating faster once.

For him, “the next Google” means reorganizing how people obtain information and complete work. Even strong foundation models would require new product and distribution design.

On OpenAI, Guangmi reads its direction from its search and coding products and its team at the time. These are outside observations, not a statement of internal plans.

From an investor’s perspective, he thinks model improvements can attract users and build brand advantages. Leaders must still fund continued investment, face competitors, and retain users.

He distinguishes several potential moats. A lead in pretraining may change, while user recognition, product experience, and sustained innovation also affect competition; one technical lead does not guarantee lasting victory.

Frontier model training requires sustained coordination of capital, compute, and research teams. Guangmi says access to resources, talent organization, and turning research into products matter alongside any individual release.

He also discusses activity, retention, and frequency, arguing that these reveal more than registrations about whether a product has entered daily life. The figures mentioned are his estimates at the time, not verified company disclosures.

There is also a tradeoff between user data and model intelligence. Guangmi hypothesizes that a more capable system with little initial context could learn a user’s needs through interaction, while access to data alone cannot compensate for weak models.

Google has distribution through search, browsers, and phones, as well as accumulated data and infrastructure. Guangmi also argues that established products and business models can shape the way and speed with which new features are introduced.

They discuss whether interface advantages from the smartphone era will carry into AI. For Guangmi, the question is not only adding chat to existing products but embedding model capabilities into real tasks.

Graphical interfaces let users operate controls, while conversation lets them express intent. The two can coexist in products combining natural language, visible state, and human confirmation when needed.

Guangmi treats post-training and product design as important beyond base-model competition. Whether a model can do what users need, and how its results are presented, both shape the experience.

He does not assume that competition must produce only one winner. Google, OpenAI, Anthropic, xAI, Meta, and cloud providers have different combinations of models, distribution, compute, and product assets.

For search products such as Perplexity, he discusses room for independent growth and speculates that larger companies might be interested. This was a competitive and acquisition scenario, not a completed deal.

He analyzes cooperation and possible tension between OpenAI and Microsoft: a cloud platform serves different models and customers, while a model company may want more product and infrastructure options. Their future relationship remained uncertain at the time.

On developer products, Guangmi mentions Claude Sonnet’s coding reputation and explains its appeal through his own observations. Popularity and performance vary by version and task.

He highlights Anthropic’s Computer Use and MCP: the former explores models operating computers, while the latter provides a protocol for connecting tools and data. Both point toward models participating in work as well as answering questions.

Connecting a model to tools does not by itself complete a task. A usable workflow still has to handle permissions, feedback, errors, and the division of work between people and the system.

Discussing xAI, Guangmi considers opportunities in compute and multimodal models. Outside observers, he notes, cannot know each company’s internal progress precisely.

In his framework, Perplexity is not merely a box returning links; it could become an interactive information assistant that searches and synthesizes around a question, then follows the user’s next inquiry.

He is optimistic about the coding-product direction represented by Cursor. Combining a development environment with models can support more of the coding process, rather than just completing a line.

Devin leads him to consider longer, multistep tasks in which a system acts in a browser or development environment, observes the result, and adjusts from feedback. Such loops expose the reliability problem in long tasks.

He makes unverified guesses about other model companies’ strategies. His underlying question is how much resource sustained frontier training requires and how different companies choose a competitive position.

Looking toward 2025 and 2026, they consider whether agents could complete longer-horizon work. Guangmi thinks better capabilities may enable new interfaces and business models, while longer tasks make accumulated errors more consequential.

One possible approach, he argues, is to construct verifiable tasks in particular domains and improve models with expert feedback. General capability matters, but so do good feedback environments and clear evaluation criteria.

The conversation does not establish how quickly models will improve or where their ceiling lies. Guangmi is willing to bet on further progress while acknowledging that different methods and tasks may face different bottlenecks.

He considers whether synthetic data can expand training material and asks how much new knowledge can arise from recombining existing knowledge. These are open research questions rather than settled conclusions.

If coding capability continues to improve, software creation and use could change: users describe needs, systems create or modify tools, and execution supplies feedback for corrections. Whether this can be done reliably remains to be demonstrated.

Building such products requires more than a model: compute, talent, design, and a sustainable business all matter. Guangmi sees a large technical opportunity but expects constraints across those areas to shape progress.

On the pretraining data wall, he notes that high-quality public text is finite and that simply adding more data may become harder. Better data efficiency and new feedback sources could be part of further progress.

Rewards and feedback are not always as easy to define as a single correct answer. Different tasks need different evaluations, and generalizing behavior learned in one environment to another remains difficult.

Guangmi sees the reasoning capabilities shown by o1 as worth watching, but does not treat an experiment as proof of the method’s ultimate ceiling. The relationship among training, test-time compute, and task design still needs exploration.

Scaling feedback requires more real tasks and raises boundaries around data use. Zhang and Guangmi discuss work materials, personal information, and model services: deeper access to workflows makes consent and privacy questions more concrete.

Looking back on 2024, Guangmi singles out June’s Claude Sonnet as a catalyst for developer interest in AI coding. This is his choice of a consequential development from a product and investment perspective.

He watches user and developer adoption after releases and the compute demand associated with actual use. For him, demonstrations need to be considered alongside sustained use to assess impact.

He revisits earlier views on OpenAI and open models, saying that Meta’s progress caused him to adjust his assessment of open-versus-closed competition. Capability, distribution, and cost can all change the landscape.

He also predicts falling inference costs will make models accessible to more applications. Lower costs, however, do not automatically solve reliability, product design, or customer acquisition.

On public markets and company selection, Guangmi shares his personal preferences and reasoning at the end of 2024. They are historical opinions, not guarantees of future returns.

Guangmi groups long-term opportunities into three directions: a “next Google” for completing tasks, new products around interactive content, and AI for science. Each would need further product and technical breakthroughs.

For knowledge workers, he is especially interested in coding assistants and agents. If tools can handle more preparation, execution, and checking, people may spend more time on judgment, communication, and higher-level design.

Asked for a word, Guangmi associates 2024 with coding and offers agents as his expectation for 2025. He mentions using Claude, ChatGPT, and Perplexity, reflecting his habits at the time of the interview.

Zhang and Guangmi close their year-end discussion, leaving the predictions to be tested over time.