The China-based AI company DeepSeek sent shock waves through both Silicon Valley and Wall Street by releasing an AI model that competes with the best US ones but was made at a fraction of the cost. Although the news has sparked a sell-off of tech stocks and prompted venture capitalist Marc Andreessen to describe the results as 鈥淎I鈥檚 Sputnik moment鈥�, this open-source AI is not as revolutionary as it seems.
Over the past year, AI models from companies such as OpenAI, Google and Meta have become capable of ever more complex tasks. To achieve this, tech companies have spent tens of billions of dollars on AI development. But on 20 January, seemingly out of nowhere, DeepSeek made its R1 model freely available and published a paper demonstrating the AI鈥檚 impressive performance.
顿别别辫厂别别办鈥檚 performed comparably to OpenAI鈥檚 o1 model on several maths and coding benchmarks. That鈥檚 no real shock, but it achieved this while being 20 to 50 times cheaper to train and using significantly less computing power. Like OpenAI鈥檚 o1, it is a so-called early reasoning model, according to , an AI policy researcher formerly at OpenAI. These systems can tackle more complex tasks than other large language models, because they break problems down into chunks and tackle each piece separately. This also means they consume more computing power and energy, which makes practical applications costly.
Advertisement
A more efficient way to achieve the same results 鈥� namely, 顿别别辫厂别别办鈥檚 model 鈥� could help more organisations beyond the richest companies benefit from AI services. And the fact that it is open source and freely available could force US competitors to drive down their own prices.
But it鈥檚 still unclear exactly how well 顿别别辫厂别别办鈥檚 model stacks up against the leading US ones. DeepSeek has been fairly transparent about publishing its results on various benchmark tests. Meanwhile, we don鈥檛 know whether US tech companies have achieved even more impressive results with other leading systems 鈥� such as OpenAI鈥檚 latest o3 model 鈥� while keeping them secret and unpublished. Furthermore, the benchmarks tested so far do not represent the full potential range of an AI model鈥檚 capabilities.
鈥淧eople are reading too much into the fact that this is an early step of a new paradigm, rather than the end of the paradigm,鈥� said Brundage.
There may also be other limitations to 顿别别辫厂别别办鈥檚 model. It has certain built-in biases from its training that are meant to conform with Chinese government restrictions, such as avoiding discussion of the treatment of the Uighur ethnic minority and toeing the Chinese Communist Party line on the political status of Taiwan.
And though it was efficiently trained, DeepSeek still had that could have benefited the model鈥檚 development, at the RAND Corporation, a think tank based in California, and at the University of Oxford wrote in a blog post.
They point out that DeepSeek has access to one of Asia鈥檚 largest clusters of AI chips, as well as Chinese and foreign computing resources stored in the cloud 鈥� which are not subject to the US export controls that have sought to limit US chips from being used in China. Such resources probably helped DeepSeek discover more efficient techniques by first generating synthetic training data and then allowing for ample trial-and-error experimentation.
Despite those US export controls, the company also had access to a recent generation of AI chips. The most recent export limitations were implemented in October 2023, so they have not yet made their full impact felt on Chinese AI development efforts, said Heim and Huang. DeepSeek has trained its AI models on NVIDIA H800 chips that were designed to circumvent older and less effective US export controls, they pointed out.
鈥淐hina is still running pre-restriction data centers with tens of thousands of chips, while US companies are constructing data centers with hundreds of thousands,鈥� the pair wrote. 鈥淭he real test comes when these data centers need upgrading or expansion鈥攁 process that will be聽聽but challenging for Chinese companies under US export controls.鈥�
If DeepSeek is not going to completely disrupt the US-based AI world, why does it have the US tech market and investors panicking? It鈥檚 not so much that it has changed the AI race for good, but it has certainly punctured unrealistic expectations in the industry. An AI hype bubble has built up in recent years, leading to sky-high valuations for tech companies. The shift in stock prices could reflect more realistic estimates of AI鈥檚 monetary value.
The original Sputnik moment was one in which the Soviet Union surprised the US by racing ahead to launch the first artificial satellite into space; however, the US was actually on par with the USSR in space technology and well aware of Soviet progress. It was only the public that was unaware, due to US government secrecy.
There is one way this moment reflects the Sputnik era: this could be a situation where a sensationalised story will cause public overreaction and spur more US spending on AI.
