DeepSeek decoded: 5 myths and realities about the Chinese AI startup’s rise

The excitement around DeepSeek is giving way to scrutiny with questions emerging about how its AI models were developed and various interpretations about their broader impact. As the discussion unfolds, let’s break down some of the common myths surrounding DeepSeek’s rise.

1. 01

   Myth #1: DeepSeek’s AI models signal AGI is within reach

   Reality: DeepSeek’s AI models are a significant improvement in efficiency and cost, but they don’t necessarily indicate a leap towards artificial general intelligence (AGI).

   AGI is a term used by the tech industry to describe an AI model capable of equaling or surpassing human intellect on a wide range of tasks. No one has declared that they have developed such an AI model yet. However, OpenAI and some of its rivals have said that they are eagerly working towards reaching the AGI milestone.

   In 2023, DeepSeek reportedly evolved from the AI research unit of a Chinese hedge fund, High-Flyer, to an AI company. The firm was established by hedge fund manager Liang Wenfeng with the objective of developing large language models (LLMs) on the path to AGI.

   OpenAI CEO Sam Altman has repeatedly expressed confidence that the ChatGPT-maker will achieve AGI. In his response to the buzz around DeepSeek, Altman again shifted focus to AGI while hailing the R1 model as ‘impressive’.

   Although R1 marks an inflection point in the race for AI supremacy, DeepSeek has not introduced entirely new technology. “Getting to AGI probably requires five or six more breakthroughs and the company or country that can ramp up those breakthroughs first may win,” Gary Marcus, New York University (NYU) professor and AI expert, told CNBC.

2. 02

   Myth #2: DeepSeek’s breakthrough shows export controls don’t work

   Reality: US export restrictions on the sale of advanced GPUs may continue to have a significant impact on China’s AI development.

   DeepSeek’s breakthrough has been viewed as the unintended outcome of US export controls that restricted Chinese tech firms from purchasing advanced GPUs to scale their AI models. Without access to Nvidia’s top-of-the-line chips, DeepSeek researchers were reportedly forced to come up with clever ways to make AI models more efficient in their consumption of raw compute power.

   Critics have argued that US export controls backfired, but DeepSeek reportedly stockpiled 10,000 of Nvidia’s older generation A100 GPUs before the trade restrictions were imposed.

   Miles Brundage, an AI policy expert who recently left OpenAI, has suggested that export controls might still slow China down when it comes to running more AI experiments and building AI agents.

   “DeepSeek was forced through necessity to find some of those techniques maybe faster than American companies might have. But that doesn’t mean they wouldn’t benefit from having much more [GPUs]. That doesn’t mean they are able to immediately jump from o1 to o3 or o5 the way OpenAI was able to do, because they have a much larger fleet of chips,” Brundage said in a recent podcast interview.

   In addition, Dario Amodei, the CEO of Anthropic, the company behind the Claude series of AI models, has said that DeepSeek’s results “make export control policies even more existentially important than they were a week ago.”

3. 03

   Myth #3: DeepSeek is a grave threat to Nvidia

   Reality: DeepSeek’s R1 model may not be as concerning for Nvidia as some might think.

   The buzz around DeepSeek caused panic among Nvidia investors, which resulted in its shares dipping by 17 per cent and wiping out nearly $600 billion in market value on January 27. The chip giant’s stock recovered from the sharp slump on January 28, it fell another 4 per cent on January 29.

   While DeepSeek’s R1 model may have diminished the requirement of vast arrays of special purpose AI hardware from the likes of Nvidia, it does not exactly spell doom for the chip giant.

   Microsoft CEO Satya Nadella pointed out that DeepSeek’s impact could, counterintuitively, increase demand for advanced GPUs. “Jevons paradox strikes again!” Nadella wrote in a post on X.

   Jevons Paradox is an economic theory which suggests that when technological progress makes the use of a resource more efficient, overall consumption of that resource tends to increase.

   Tech investor Andrew Ng also said that it remains to be seen if DeepSeek’s results will reduce the demand for GPUs and compute power. “Sometimes making each unit of a good cheaper can result in more dollars in total going to buy that good,” he said in a post on X.

4. 04

   Myth #4: DeepSeek R1 is a fully open-source model

   Reality: DeepSeek R1 can be downloaded, modified, and reused for free, but it may not be considered truly open source.

   The impressive results of DeepSeek R1 has been interpreted by many as a sign of China pulling ahead of the US in the race for AI supremacy. But beyond the geopolitical angle, DeepSeek’s success is also being celebrated as a win of open-source AI over closed AI.

   Echoing this sentiment, Meta’s chief AI scientist, Yann LeCun, said, “DeepSeek has profited from open research and open source (e.g., PyTorch and Llama from Meta). They came up with new ideas and built them on top of other people’s work. Because their work is published and open source, everyone can profit from it. That is the power of open research and open source.”

   R1’s underlying model architecture and weights (numerical values used to indicate how an AI model processes information) has been made publicly available under a permissive MIT licence. This means that the model can be deployed without restrictions.

   But R1 doesn’t fit the widely accepted definition of ‘open-source’. According to the Open Source Initiative (OSI), a truly open-source AI model must provide access to details about the data used to train the AI, the complete code used to build and run the AI, and the settings and weights from the training.

   The data used to train R1 has not been made available. The training code and other instructions for training have not been provided either. Open-source AI developers are generally wary of releasing training datasets as it may invite copyright infringement lawsuits.

5. 05

   Myth #5: DeepSeek’s AI models carry extra privacy risk

   Reality: DeepSeek’s AI poses the same risk to privacy as other LLMs

   DeepSeek’s meteoric rise has been accompanied by data privacy concerns among users and authorities. Some of these concerns have been fueled by the AI startup’s Chinese origins while others have pointed to the open-source nature of its AI technology.

   In its privacy policy, DeepSeek unequivocally states: “We store the information we collect in secure servers located in the People’s Republic of China.”

   However, tech industry figures such as Perplexity CEO Aravind Srinivas have repeatedly sought to allay such data protection worries by pointing out that DeepSeek’s R1 model can be downloaded and run locally on your laptop or other devices. Running local instances means that users can privately interact with DeepSeek’s AI without the company getting their hands on input data.

   According to Srinivas, Perplexity is hosting the R1 model in data centres located in the US and European Union (EU), not China. He also claimed that the Perplexity-hosted version of R1 is free from censorship restrictions.