Why does DeepSeek make Western AI practitioners feel insecure?

Author: Carl Franzen; Translator: TechFlow

Just a few days ago, only the most dedicated geeks (I say this as one) had heard of DeepSeek, a Chinese AI company that is a subsidiary of the equally uniquely named High-Flyer Capital Management, a quantitative analysis firm founded in 2015. However, in the past few days, it may have become the most talked-about company in Silicon Valley.

This is mainly due to the release of DeepSeek-R1, a new large language model (LLM) that is capable of "reasoning" similar to OpenAI's current best model o1 - taking seconds or minutes to reflect on its own analysis in a step-by-step or "chain of thoughts" manner when answering difficult questions and solving complex problems.

Not only that, DeepSeek-R1 scores as well or better than OpenAI’s o1 on various third-party benchmarks (tests that measure the performance of AI in answering questions on a variety of topics), and it reportedly cost only about $5 million to train, using far fewer graphics processing units (GPUs) than are strictly prohibited in the United States (OpenAI’s home base).

But unlike o1, which was only available to paid ChatGPT Plus-level subscribers ($20 per month) and higher tiers like the $200 per month Pro tier, DeepSeek-R1 was released as a fully open source model, which explains why it quickly climbed to the top of the list of the most popular and active models on the AI ​​code-sharing community Hugging Face.

And, since it’s completely open source, people have fine-tuned and trained the model in many ways for different specific tasks, like making it small enough to run on mobile devices, or combining it with other open source models. Even if you want to use it for development purposes, DeepSeek’s API costs more than 90% less than OpenAI’s equivalent o1 model.

Most impressively, you don’t even need to be a software engineer to use it: DeepSeek offers a free website and mobile app for US users, and its R1-powered chatbot interface is very similar to OpenAI’s ChatGPT. However, DeepSeek once again surpasses OpenAI by connecting this powerful reasoning model to web search — something OpenAI has not yet achieved (web search is currently only available on the less powerful GPT series of models).

An Obvious Irony

Given OpenAI’s original goal of democratizing AI for the masses, there’s a rather interesting, if disturbing, irony here. As Nvidia senior research manager Jim Fan said at X: “We’re living in a timeline where a non-US company is continuing OpenAI’s original mission of truly open, cutting-edge research that empowers everyone. It doesn’t make sense. But the most interesting outcomes are the most likely to happen.”

And as X user @SuspendedRobot said (citing reports that DeepSeek appears to be trained on ChatGPT-generated question-answer output and other data): “OpenAI steals data from the entire internet to make themselves richer, while DeepSeek steals from them and gives it back to the masses for free, which reminds me of a British folk tale.”

Meta in crisis, falling behind in open source Llama?

But it’s not just Fan that has noticed DeepSeek’s success. Based on my conversations and reading with various engineers, thinkers, and leaders, the open source availability of DeepSeek-R1, its high performance, and the fact that it seemed to “come out of nowhere” to challenge the previous generative AI leaders has sent shockwaves throughout Silicon Valley and beyond. If it’s not “everyone” going crazy about it, as my exaggerated headline suggests, it’s at least a hot topic in tech and business circles.

A message posted on Blind (Silicon Valley’s anonymous gossip sharing app) is circulating widely, suggesting that Meta is in crisis over DeepSeek’s success because it has so quickly surpassed Meta’s own efforts to become the king of open source AI through the Llama model.

“This changes the rules of the game”

X user @tphuang made a convincing point: “DeepSeek has commoditized AI beyond the top tier. The first picture made it clear to me. R1 has much lower labor costs than the US, which means many jobs will be automated in the next 5 years.” He later pointed out why DeepSeek’s R1 is more attractive to users than OpenAI’s o1:

“o1 has 3 huge problems:

1) Too slow

2) Too expensive

3) Lack of control for end users/over-dependence on OpenAI.

R1 solves all of these issues. Companies can buy their own Nvidia GPUs to run these models without worrying about extra costs or slow/unresponsive OpenAI servers.”

@tphaung also posed a thought-provoking analogy: “Will DeepSeek be the Android of LLM?”

Web entrepreneur Arnaud Bertrand was blunt at X about the astonishing impact of DeepSeek’s success: “It’s hard to overstate how much of a game changer this is. This is not just about AI, but also a huge irony of the US’s misguided attempts to block Chinese technology development, without which DeepSeek might not have existed (necessity is the mother of invention, as the saying goes).”

Censorship issues

However, there are concerns about DeepSeek

As a member of the news media, I certainly value freedom of speech and expression, and it is one of the most fundamental principles I firmly support.

However, I also feel compelled to point out that OpenAI's models and products (including ChatGPT) also refuse to answer a range of questions - especially those involving human sexuality and adult/NSFW content, even if these questions are mundane.

Of course, this is not an apples-to-apples comparison. For some, resistance to relying on foreign technology will make them skeptical of DeepSeek's ultimate value and usefulness. But its performance and low cost are undeniable.

In an era when the US imports 16.5% of its goods from China, it’s hard for me to warn against using DeepSeek-R1 based solely on censorship concerns or security risks — especially when the model code can be downloaded for free, used offline, run on devices in a secure environment, and fine-tuned at will.

I do detect some existential crisis thinking about the “decline of the West” and the “rise of China” in the heated discussion around DeepSeek. Some have linked this to American users joining the Xiaohongshu app when TikTok was briefly banned, and being amazed at the quality of life in China shown in the videos shared there. The emergence of DeepSeek-R1 takes place in the context of this narrative — one in which China looks (and on many metrics is) rising, while the US looks (and on many metrics is) declining.

The first, but definitely not the last, Chinese AI model to rock the world

It won’t be the last Chinese AI model to threaten the dominance of Silicon Valley giants—even if those giants, like OpenAI, are raising more money than ever before for their efforts to develop artificial general intelligence (AGI, programs that outperform humans in most economically valuable tasks).

Just yesterday, another Chinese model, Doubao-1.5-pro, from TikTok’s parent company ByteDance, was released, performing on par with OpenAI’s non-inferential GPT-4o model on third-party benchmarks, but at 1/50 the cost.

Chinese models are getting so fast and so good that even people outside the tech industry are taking notice: The Economist just published an article about DeepSeek’s success and other Chinese AI efforts, and political commentator Matt Bruenig posted on X: “I’ve been using Gemini, ChatGPT, and Claude for NLRB document summarization for almost a year now. Deepseek is better than them all. Its chatbot version is free. And the price of using its API is 99.5% lower than OpenAI’s. [shrug emoji]”

How is OpenAI responding?

No wonder OpenAI co-founder and CEO Sam Altman said today that the company will bring its yet-to-be-released second-generation inference model series, o3, to ChatGPT, making it available even to free users. OpenAI still seems to be forging its own path with more proprietary and advanced models—setting the industry standard.

But here’s the question: How long can OpenAI keep its lead in producing and publishing new cutting-edge AI models as DeepSeek, Bytedance, and other Chinese AI companies follow in its wake? And if it does fall behind, how fast and how severe will its decline be?

OpenAI does have another historical precedent to draw on, though. If DeepSeek and Chinese AI models do what Google’s open-source Android did to mobile — taking over a large portion of the market for a while — you only need to look at how the Apple iPhone took over the high end of the market with its closed, proprietary, all-in-house approach and steadily expanded downward from there, especially in the United States, to the point where it now has nearly 60% of the domestic smartphone market share.

Still, for all those who are paying big bucks to use AI models from leading labs, DeepSeek shows that the same functionality may be available at a lower price, and with greater control. In an enterprise setting, that may be enough to win.