DeepSeek aims to accelerate AGI development through open source models

The race against time

DeepSeek, a Chinese artificial intelligence company, recently shared its vision for the future of machine learning at the World Internet Conference. Research engineer Gao Xi stated that the primary opponent in the race toward artificial general intelligence (AGI) is not other companies, but time itself. The goal is to compress the timeline of development significantly.

Gao Xi noted that if reaching AGI were to take ten years under current conditions, the company hopes to achieve it in just one year. To reach this goal, DeepSeek intends to keep its models open source. This approach allows the global developer community to examine, test, and improve the technology simultaneously rather than keeping it behind closed doors.

The company believes that keeping technical hurdles secret might provide a short-term advantage for a single firm. However, they argue that such secrecy ultimately slows down the overall progress of the field. By sharing efficient training methods and model architectures, they believe the entire industry can move forward faster.

The risks of open weights

While the benefits of open source are clear, DeepSeek acknowledges that this path carries significant security implications. Gao Xi admitted that open models can amplify security problems because once model weights are released to the public, they cannot be taken back. This permanence creates a challenge for maintaining control over how the technology is used.

The company recognizes that risks from malicious use or from the AI systems themselves cannot be completely ruled out. Because of these concerns, DeepSeek is currently working on stronger protection measures. These efforts include a system for classifying their different models based on their specific risk levels.

The company has not yet shared the specific details of these safety protocols. However, the move toward risk-based classification suggests a desire to balance the speed of open development with a structured approach to safety. They are attempting to create a framework where the most dangerous capabilities are managed more carefully than less sensitive tools.

Industry and policy context

The debate over open source models involves major players across the globe. Liu Dayiheng, who leads the Qwen model team at Alibaba, echoed concerns about the lack of external protection mechanisms for open models. He pointed out that because these models cannot be recalled, the industry needs more than just restricted access to ensure safety.

Instead of limiting who can see the models, Liu Dayiheng advocated for the creation of new laws, industry standards, and additional infrastructure. This perspective suggests that the solution to AI risk lies in building a regulated environment for open tools rather than keeping the core technologies secret. This would prevent the core technologies of large language models from being held by only a few companies.

International governance experts are also weighing in on these dynamics. John Higgins, the chairman of an international AI governance association, called for shared procedures to evaluate and certify open models. He believes that establishing these common standards is necessary to build trust between different global markets.

Geopolitics and regulation

The push for open source AI is also taking place against a backdrop of international tension. The Chinese government is actively supporting open AI ecosystems, which many see as a strategic response to export restrictions from the United States. By fostering an open environment, China aims to maintain its position in the global AI landscape.

DeepSeek and the Shanghai-based company MiniMax have already completed domestic registrations to release their latest models as open source. Additionally, DeepSeek recently made its inference framework, known as DeepInfer, available to the public. These actions demonstrate a commitment to making their technical infrastructure accessible to a wider audience.

This stands in contrast to some Western AI companies like Anthropic and OpenAI. These firms often provide their most powerful models only through closed systems. They frequently cite security risks as the primary reason for keeping their most advanced capabilities behind private APIs, creating a different model for how high-end AI is distributed.

What comes next for AGI

The next phase for DeepSeek involves balancing the rapid release of their findings with the implementation of their risk classification system. As they release more efficient training methods and architectures, the global developer community will begin to integrate these tools into their own projects. This could lead to a surge in diverse AI applications developed outside of the major corporate labs.

Observers will need to watch how the Chinese government and international bodies respond to this rapid pace of open-source release. If DeepSeek succeeds in its goal of accelerating AGI, it may force a shift in how global safety standards are written. The focus will likely move from 'who has the model' to 'how is the model being used' by the public.

The tension between the speed of open-source innovation and the need for security will remain a central theme in AI policy. As DeepSeek continues to push for a one-year timeline to AGI, the effectiveness of their proposed safeguards will be under constant scrutiny by both developers and regulators alike.