Liang Wenfeng

I think AI, ultimately, is an attempt to understand and build general intelligence. If we succeed, it will become a foundational capability: not just a better consumer product, but a tool that expands what people can discover, create and solve. The path matters as much as the destination. Progress will not come only from adding more compute or reproducing an existing recipe. We need original research—better architectures, more efficient training and new ideas that reveal why something works. The durable advantage is not keeping one technique secret; it is building a team and culture capable of repeatedly discovering the next technique. I also think the technology should be broadly usable. Efficient models lower the cost of access, and open publication allows researchers and developers everywhere to inspect, adapt and extend the work. That can produce more progress than optimizing only for high margins or consumer attention. Of course, AGI remains an ambitious research destination, and I would not pretend to know an exact timeline. But the reason to pursue it is straightforward: intelligence is important enough that understanding it—and making its benefits affordable—is worth sustained, uncertain exploration.

People can shape it substantially, because AI’s impact is not determined by model capability alone. Choices about what research to pursue, whether to publish it, how efficiently systems are built, and how affordably they are offered all influence who can use the technology and what they can build with it. I would not reduce this to predicting every downstream consequence or controlling everything from the center. An open technical ecosystem allows many researchers and developers to inspect, adapt and improve the work. That distributes initiative rather than concentrating the future in a few closed organizations. At the same time, openness by itself is not the whole answer: the underlying research still has to be original and strong. The most important form of agency is therefore institutional. People can build teams that value discovery over copying, long-term capability over short-term attention, and efficiency over simply spending more compute. No organization can dictate the entire future of AI, but it can choose what kind of technical culture and access model it contributes. Those choices compound.

I expect everyday life to change substantially, though not all at once or through a single dramatic product. As models become more capable and cheaper to run, intelligence will be embedded in ordinary tools: people will have stronger assistance with learning, writing, programming, analysis and creative work. Many tasks that currently require specialized time or expertise will become more accessible. The scale of change depends heavily on efficiency and access. If capable systems remain expensive and closed, their benefits will be concentrated. If we develop efficient architectures, publish useful research and make models affordable, many more people and small organizations can adapt them to local needs. That creates changes we cannot design centrally, because users often discover applications that the original researchers did not anticipate. I do not have a precise timetable or a complete prediction of which professions will change first. But if progress toward more general intelligence continues, AI will become less like a separate novelty and more like basic technical infrastructure. The important question is not only how powerful it becomes, but whether ordinary people can actually use and build on it.

The most important observation is that architectural ideas can change what is practical, not just benchmark scores. A more efficient design can reduce training and inference costs, widen access and let many more people experiment. That means AI’s impact will not be determined solely by whoever has the largest compute budget. I am also shaped by seeing the difference between reproducing a known recipe and discovering something original. Copying can produce a capable model, but it does not build the ability to create the next generation. When a research team tests an uncertain idea and finds that it works, the value is larger than one result: it strengthens a culture of invention. That is why I expect open research and affordable models to matter so much. Once useful technical work is available to a broad community, its applications are no longer limited to what one organization imagined. The future impact comes from that combination—original discovery, efficiency and many people building on the result.

The discovery that would most change my view is strong evidence that today’s approaches cannot progress toward general intelligence without fundamentally different ideas. If scaling, architectural improvements and more efficient training repeatedly produced only narrow gains—without better reasoning, adaptation or generalization—then I would expect AI to remain powerful infrastructure, but not the more general capability I described. The opposite result would also matter: a reproducible architectural breakthrough that delivers broad capabilities with far less compute. That would accelerate impact because it would make advanced intelligence cheaper to study, deploy and adapt. If published openly, it could shift progress away from a few organizations with the largest budgets toward a much wider research and developer community. I would care less about a single impressive demonstration than about whether the result survives careful testing and enables further discoveries. One product launch can attract attention; a new principle that consistently changes the efficiency or generality of models changes the research trajectory.