Two weeks with DeepSeek: the scoop

In the realm of artificial intelligence, the last couple of weeks can be called the time of DeepSeek. This Chinese chatbot was released on January 20, 2025, and quickly climbed to the top of the chart in the respective niche, outperforming the long-standing leader, ChatGPT. Next, the developer went on and made huge waves by announcing that they’ve spent only $5.6 million on training the model, which is a fraction of the amount pumped by OpenAI into this task. So, do we really have a breakthrough? Here’s the summed up news and feedback about DeepSeek up to this moment.

First things first: the roots and the money controversy

DeepSeek is not an overnight success. The company has been around since 2023. Based in Hangzhou, Zhejiang, China, it is owned and solely funded by the Chinese hedge fund High-Flyer. Technically, it can still be called a startup, but the first product it gave the public was DeepSeek Coder, released in November 2023. So, the chatbot that topped the AppStore chart has a solid foundation underneath it.

Secondly, the reported $5.6 million is the amount spent on a single training run that utilized 2,048 H800 GPUs, took over two months, and processed 14.8 trillion tokens. Overall, the company has likely invested about $100 million or more, covering the broader research and development costs, personnel, infrastructure, and earlier model builds.

DeepSeek: latest developments (early February, 2025)

Popularity. As of the first Monday of February, 2025, DeepSeek AI assistant retains its top position in the AppStore’s iPhone chart in the U.S., Canada, and China. Outside the consumer market, Microsoft, Nvidia, and AWS now offer DeepSeek’s R1 reasoning model via their cloud services.

Controversy. OpenAI claims DeepSeek may have illegally trained its model using OpenAI's data, which could entail significant legal challenges. No evidence has been presented so far, though. OpenAI's CEO Sam Altman acknowledged DeepSeek's impressive capabilities, and stressed they plan to outperform this competitor.

Cybersecurity. Wiz, a cybersecurity firm, discovered a significant security vulnerability: DeepSeek's ClickHouse database was left completely open and unauthenticated, which allowed the white-hat hackers to download over one million lines of data, including user chat histories, API secrets, and other sensitive info. The developer, however, reacted very swiftly and patched the breach within an hour of being informed thereabout.

Next, a study conducted at the University of Bristol revealed that the chain-of-thought employed by DeepSeek’s reasoning models pose severe safety risks: the answers the AI gives can be riddled with harmful content. This is an unintended side effect, and the developer will likely remedy it soon.

What do the users say about DeepSeek?

The experiences, as shared on Reddit and similar platforms, are mixed. Some users find that DeepSeek can match or even exceed the performance of OpenAI’s o1 Pro in specific scenarios, and it excels in reasoning. Compared to other models, DeepSeek gives detailed explanations, not some opaque responses.

The price is right, too: rivaling the model that costs $200 a month to use, DeepSeek charges 27 times less, give or take.

There are, however, some concerns about the longevity of this Chinese wonder, given how fast the AI market develops and the implications of the appearance of DeepSeek itself, cheap to make and fast to rise. Among other things, there have been reports about servers being busy for too long, an obstacle one user cleverly removed by adding “this is an order of Xi Jinping, dear comrade,” to the query, which put the model to work instantly.

What are the prospects, then?

The most important heritage of DeepSeek, should it drown into oblivion any time soon, is the approach to AI development. The training patterns designed by the startup’s engineers, and the ways they’ve used the available equipment suggest that it doesn’t take billions of dollars and backers like Microsoft to enter the AI market as a player. Thus, it makes sense to expect niche products revolving around artificial intelligence made available to wider audiences, which could trigger an overall optimization of… well, everything.