With the advent of DeepSeek, the balance of power between the two nations appears to be shifting.By Eric Schmidt, Dhaval Adjodah | 2025-01-28
The DeepSeek app’s user interface. (Greg Baker/AFP/Getty Images)
Eric Schmidt, former CEO and chairman of Google, is co-founder of Schmidt Sciences and chair of the nonpartisan think tank Special Competitive Studies Project. Dhaval Adjodah is co-founder and CEO of MakerMaker.AI.
It has become almost a cliché to say that the artificial intelligence landscape is changing fast. But in recent days, even those on the cutting edge of AI research were taken by surprise — by a Chinese company.
Last week, the AI company DeepSeek released its R1 reasoning model, which is on a par with OpenAI’s o1 (and much better than the ChatGPT models) across a variety of logic tasks, including math and coding. The cost of running it is also much lower, only about 2 percent of what OpenAI charges. And on Monday, DeepSeek released Janus Pro, a model small enough to run on your laptop that can generate synthetic images, which it claims outperform OpenAI’s Dall⋅E 3. DeepSeek’s speed of AI innovation is taking the world by storm.
What’s even more remarkable is that DeepSeek’s entire collection of models is open-source — which in this case means they have open weights that anyone can reproduce and build on top of.
It’s a peculiar moment when a Chinese company becomes the de facto open-source leader, while most major American firms, with the exception of Meta, continue to keep their methodologies tightly under wraps. In fact, this is a growing trend for Chinese AI companies — from start-ups such as Minimax to tech giants such as Alibaba — that are giving developers worldwide free access to their AI models.
Until now, closed-source models such as OpenAI’s o3 and Anthropic’s Claude 3 Opus were considered the industry standards with the most advanced capabilities — and they were built in the United States. Open-source and Chinese models were thought to be months behind. But DeepSeek’s R1 and Janus Pro show just how quickly the tides of technological supremacy can turn. The introduction of these models has roiled stock markets and caused U.S. tech stocks to plunge. The balance of power now appears to be shifting along two key axes: one between the United States and China, and another between closed- and open-source models.
Defenders of closed-source models are betting that they can preserve their capability gap by protecting their model weights and training methodologies. Open-source advocates, on the other hand, argue that transparency — allowing others to build on their work — can enable these systems to rapidly catch up with larger, closed models. If the open-source thesis is correct, this would turn the AI ecosystem on its head. Open-source models are generally cheaper to use, so when two equally capable models are available — one open, one closed — the open-source model is likely to gain wider adoption, giving it a strategic advantage.
The United States already has the best closed models in the world. To remain competitive, we must also support the development of a vibrant open-source ecosystem.
The race between open- and closed-source AI, as well as between the United States and China, does not yet have a clear winner. But there is clearly mounting pressure on America’s Big Tech players if DeepSeek can compete with them using far fewer resources. Export controls were aimed at choking off China’s access to the most advanced computer chips, impeding its ability to keep pace. But in fact, the relative dearth of high-performing chips in China might have pushed the nation’s companies and researchers to be more efficient and led them to uncover new methodologies that significantly reduce training costs. For example, DeepSeek demonstrated that large model training could be made more efficient by bypassing the traditional supervised fine-tuning stage. They even created R1-Zero, a model that omits this step in AI training, to challenge the research community’s assumptions about fine-tuning’s indispensability.
DeepSeek’s success has also called into question the importance of pretraining, which involves training ever-larger models that predict the next word based on vast amounts of text. This process requires enormous up-front investment in graphics processing units (GPUs) and data — so much data that OpenAI co-founder Ilya Sutskever recently noted we might soon exhaust all the data available on the internet.
But there is another, emerging way to improve models’ performance. Introduced with OpenAI’s o1 model in December, this approach enables models to perform reasoning through self-reflection, similar to how humans reason, using intermediate steps and self-correction to reach a final answer. The training recipe for this approach had previously been closely guarded by OpenAI. DeepSeek blew the lid off that by publishing a paper detailing how it works, allowing others to implement the process.
DeepSeek even demonstrated that you can do this much more cost-effectively by taking a publicly available base model such as Meta’s Llama 3 and teaching it to reason through reinforcement learning — a trial-and-error process with human-devised feedback and rewards. Over time, the models seem to spontaneously learn how to reason, backtrack when they hit dead ends and explore novel approaches. This method eliminates the need to expensively pretrain a new base model, and its implications for AI innovation are profound. Traditionally, even the top-funded university labs have struggled to contribute to AI research due to computing and data limitations. With DeepSeek’s breakthrough, the moat surrounding large, well-funded companies might be shrinking.
It is unlikely that American frontier model companies will change their business models anytime soon, nor is it immediately clear that they should. Open and closed competition will most likely find a natural equilibrium, with a range of different offerings and price points for different users.
But DeepSeek’s release marks a turning point.
The path forward for American innovation involves not just ramping up open-source development but also encouraging the sharing of training methodologies and increasing investment in AI research and development — exemplified by the White House’s recent announcement of the Stargate Project, which aims to spend $500 billion on AI infrastructure over the next four years.
America’s competitive edge has long relied on open science and collaboration across industry, academia and government. We should embrace the possibility that open science might once again fuel American dynamism in the age of AI.
This article was downloaded by calibre from https://www.washingtonpost.com/opinions/2025/01/28/china-deekseek-ai-us-supremacy/
Leave a Reply