Computer use 2025 wrapped

Zengyi Qin is the founder of OpenAGI. Last week, their computer use model, Lux, outperformed Gemini, OpenAI, and Claude by a significant margin on the Online Mind2Web benchmark. We wanted to know how he’s thinking about computer use in 2025, and figured you might too.

Why computer use models, Zengyi? Why is that an important problem to work on?

The first reason is that it’s a grand opportunity to change the interface and relationship between humans and computers. We’ve seen several revolutionary changes in the history of computers: from the mainframe, to the PC, to the cell phone, and then to the iPhone, where the keyboard was replaced by the touch screen. What if an AI model can use computers and understand human intent? Then the interaction paradigm may change yet again.

The second reason is that it’s a new opportunity to build a foundational model around computer use. At the beginning of this year, computer use models weren’t very good. They all scored quite low on the benchmarks. This is because, in the pre-training stage of those models, they used an extremely large amount of internet data to train, but internet data doesn’t have strong causality. What is causality?

Let me give you an example. If you go to driving school and your coach gives you a driving manual and videos to watch, you can remember and recite them. But if they don’t let you touch the car at all, you become a person who can answer a lot of driving questions but you still don’t know how to drive. However, if you actually turn the wheel to the left and you see the car going left, you see that this action introduces an observation, and that observation introduces your next action. This sort of causality is not contained in the current pre-training data of those large models. That’s why, early in 2025, those models could do gold-medal-level math and physics, but they were unable to use computers like a high school student. This highlights how special computer use capabilities are. It’s so special that even with an extremely strong base model, you still get things wrong. That means there is a rare opportunity to build a new foundational model around computer use.

Do you think DOM-based AI agents still have a meaningful role now that computer use models can understand and act on web pages visually?

Previous computer use models were not strong enough, so abstracting web pages into a DOM to use the text capabilities of LLMs was very good. But it was a shortcut. As computer use model capabilities increase, the advantages of the DOM-based approach will gradually disappear. We already see them disappearing right now. I don’t think the DOM-based approach will retain its advantage over the next six months.

When you look back at 2025, what actually changed in what people could do with computer use models? What were some of the meaningful improvements?

Because I come from a research background, I read a lot of papers on computer use, so I can share some perspective on the two biggest shifts in the last year. The first is synthetic data, and the second is multi-turn reinforcement learning.​

Over the past year, the field has converged on the idea that data is the main bottleneck for computer use models. Collecting high-quality human demonstration data is not very scalable because it is expensive, slow, and often noisy. As a result, people in both industry and academia have started to lean heavily on synthetic data generation to overcome these limitations.

The second big advancement is allowing models to use multi-turn reinforcement learning to discover good behaviors on their own, instead of relying on step-by-step human annotations. Rather than forcing the model to imitate a human, you give it a goal, verify whether it accomplished that goal, and let it gradually learn its own strategy through many trials. Human actions are not always the optimal or highest-reward actions, and with multi-turn reinforcement learning the model can uncover more efficient patterns.​

When you picture the first wave of successful computer use products, what are the very first industries and use cases you expect?

The first is Robotic Process Automation (RPA). Before computer use, process automation was only possible for some API-based workflows. With computer use, you can essentially operate anything on the computer, so almost any business operation. The second is software QA.

Final question: What are the open research problems you’re thinking about going into 2026?

The biggest one is reflection. You cannot assume the environment the model interacts with is static, so the model will frequently encounter software it has never seen before. After the model does something wrong, it needs to be able to recognize the mistake, reflect on its behavior, and eventually find the correct path. That will significantly improve the robustness of the whole system.

And that’s a wrap.

Thanks to Zengyi for sharing his S‑tier insights on computer use. This space is ever‑changing, so keeping up with the latest is hard. I hope this helped. If you’re experimenting with computer use models in production, say hello! We’d love to chat with your team.