DeepSeek Engineer Says Open Source Accelerates AGI: The Rival Is Time

In the debate between open-source and closed-source approaches, DeepSeek has offered its own answer. According to a Bilibili video, DeepSeek research engineer Gao Xi said at the AI Wuzhen Dialogue of the 2026 World Internet Conference Wuzhen Summit that the core logic behind DeepSeek's firm choice of open source is to make AGI arrive sooner.

Open Source Is Not a Gesture, but an Accelerator

Gao Xi defined AGI as a form of intelligence more advanced than current AI and considered comparable to human intelligence. He observed that AI technology is developing at an exponential rate, with huge breakthroughs every year, from simple translation to deep reasoning. Facing such a rapid wave, Gao Xi admitted that DeepSeek^Ad's rival is not large model companies at home or abroad, but time.

In his view, no matter which company develops AGI, the technological transformation it brings will benefit all of humanity. Therefore, when AGI arrives, you will no longer ask which model company brought it. Based on this judgment, open source became DeepSeek's best choice for racing against time to accelerate the AGI process.

This view echoes industry discussions about the value of open source. The closed-source approach emphasizes commercial barriers and security, while the open-source approach bets on collaboration efficiency. Gao Xi's statement offers a more pragmatic perspective: open source is not a moral choice, but a strategic choice in a race against time.

AGI (Artificial General Intelligence) is a concept contrasted with current "narrow AI" (such as translation models and recommendation systems). It refers to intelligent systems that reach or exceed human levels across a wide range of tasks and possess cross-domain transfer and reasoning capabilities. There is no academic consensus on the criteria for determining AGI. Common discussion dimensions include: whether it can set goals autonomously, whether it performs well on new tasks it has not been trained on, and whether it has common sense and causal reasoning abilities. Gao Xi described it as a form "more intelligent than current AI and comparable to human intelligence," which is a relatively broad definition. There is debate over the path to AGI: one side believes fundamental breakthroughs at the architectural level are needed, while the other believes the current route centered on scaling and reinforcement learning can approach it gradually. DeepSeek's open-source strategy is precisely a bet on the latter—compressing the exploration time "from current models to AGI" through parallel trial and error by the community.

New Methods Are Made Public Immediately, Letting Peers Carry Forward Verification

Gao Xi gave an example of how open source works in actual R&D: if a new method is discovered during research, such as a lower-cost training method or a better model architecture, the most natural thing to do is make it public immediately and let more peers participate in testing and verification. If the problem is solved, everyone can move on to the next challenge, thereby advancing the overall process.

This "publish upon discovery" model essentially turns one company's R&D progress into parallel advancement by the entire community. For research teams with limited resources, external verification and reproduction can significantly reduce trial-and-error costs and avoid repeated investment in known dead ends.

Not Only Models: Engineering Tools Are Also Fully Open-Sourced

In addition to the models themselves, DeepSeek has also fully open-sourced the engineering tools used in internal R&D, including the communication library, DeepApp, the inference framework, and DeepInfer. These tools have been tested in real-world use, and after open-sourcing they can be reused by more developers and enterprises, jointly promoting the implementation and innovation of AI technology.

This point deserves attention: model open-sourcing often attracts attention, but the engineering pipeline for training and inference is the real bottleneck for implementation. The open-sourcing of the communication library and inference framework means that small and medium-sized teams no longer need to build underlying infrastructure from scratch and can focus their energy on application-layer innovation.

The communication library and inference framework mentioned in the article are core components of the large model engineering pipeline. The communication library is responsible for efficiently synchronizing gradient and tensor data during multi-GPU, multi-machine training and inference, directly determining cluster utilization. The inference framework is responsible for deploying trained models to users with low latency and high throughput, involving optimizations such as operator fusion, KV Cache management, and batching scheduling. This layer is critical because after model weights are open-sourced, users are often still constrained by "can't run it, or it runs too slowly." Storage and parallelism strategies on the training side, and memory and scheduling strategies on the inference side, account for the vast majority of implementation costs. Open-sourcing these tools that have been tested in internal real-world use essentially spills over the engineering efficiency dividends accumulated by big companies to small and medium-sized teams.

Closed Source Delays Society's Time Cost

Gao Xi emphasized that technological innovation requires open discussion and verification. If closed-source barriers delay society's access to technological dividends, the loss outweighs the gain.

This judgment elevates the open-source discussion from "whether to share" to the level of "time cost." A new balance must be struck between the benefits of the protection period brought by closed source and the cost of society's delayed access to technological dividends. DeepSeek's choice is to bet that the former is smaller than the latter.

The debate between open source and closed source is a fundamental divergence in the current AI industry. The closed-source approach, citing commercial barriers and security, advocates keeping model weights private and relying on APIs or productized charging. The open-source approach, selling points being collaboration efficiency and transparent reproducibility, allows anyone to download, modify, and use commercially. The former is conducive to short-term commercial returns and risk control for enterprises, while the latter may accelerate technology diffusion and lower the threshold for society as a whole. The two approaches are not absolutely opposed: some companies adopt a compromise model of "open weights but not open training data." Gao Xi's argument shifts the evaluation criterion from "enterprise revenue" to "social time cost," arguing that the cost of technological delay caused by closed source is higher than the benefits of its protection period. This provides a new quantifiable perspective for the open-source debate.

Open Source Is an Expression of Open Collaboration

In Gao Xi's account, DeepSeek chose open source to accelerate AGI development and let technological dividends benefit all of humanity. Open source is not only technology sharing but also an expression of open collaboration, reflecting responsibility and commitment to the future.

From an industry perspective, open source and closed source are not either-or. DeepSeek's practice offers a reference path: exchange speed for open source, exchange certainty for collaboration, and shift the focus of competition from "who keeps secrets first" to "who publishes first and drives verification." The endpoint of AGI is still unknown, but on the road to that endpoint, time is indeed the scarcest resource.