Ray Poynter, 25 January 2025
Caveat, whilst I have researched the information in this article, much of it comes from China and I can’t crosscheck it with alternative sources.
DeepSeek is a Chinese AI startup founded in 2023 as a spin of the hedge fund High-Flyer. DeepSeek’s AI is causing shockwaves in the AI community because of the speed at which it has developed, and it appears to have developed models that rival or exceed the performance of established industry leaders like OpenAI and Google but at a fraction of the cost. Its latest breakthrough, DeepSeek-R1, is a ‘reasoning model’ akin to ChatGPT o1 (the latest model released by OpenAI).
DeepSeek-R1
Released in January 2025, DeepSeek-R1 is a large language model (LLM) that uses a “chain-of-thought” (CoT) reasoning process to tackle complex problems step-by-step. This method is claimed to be similar to human reasoning inasmuch as it allows the model to backtrack and evaluate its approach when solving tasks in fields such as mathematics, coding, and scientific research.
The key features of R1 that are catching attention are:
- Performance: figures have been published that suggest that on some tests, R1 outperforms OpenAI’s o1 model. There are a lot of stats about the claimed performance on the DeepSeek home page: https://www.deepseek.com/
- Cost: It is claimed that R operates at 1/30th of the cost of o1. For example, a task costing $300 on o1 would cost $10 on R1. It is also claimed that the cost of developing R1 was about $6 million, the cost of Meta’s Llama 3.1 was about $60 million.
- Open-weight: R1 shares the weights that underpin the model, unlike OpenAI’s models. This allows engineers to evaluate and tweak the model more than they can with OpenAI. However, R1 does not share its training data, so we would not call it open source.
- Distilled Models: A distilled model is one that is created from R1, but is much smaller. Being smaller, a distilled model designed for a specific task can run on a much smaller computer.
- API model: Like ChatGPT, DeepSpeek has an API option for DeepSeek that allows programmers to make calls to DeepSeek.
Why is DeepSeek Causing so much Fuss?
The USA has been the home of most recent AI developments that have hit the market, with OpenAI, Google, Microsoft and Amazon all at the fore. The US Government have sought to hold back China by limiting the technology, especially key chips, that can be sold to China. Therefore, the rapid development of an AI platform that rivals the best US platforms has come as an unwelcome surprise to the USA.
DeepSeek’s success seems to stem from its resource-efficient approach, driven by its limited access to high-end hardware due to US export controls. The key to success appears to be the use of smarter designs and better algorithms to compensate for not having the brute force available to US companies. For example, The Economist claims that while DeepSeek V3 required just 2,000 high-end chips, Llama 3.1 used 16,000.
A Chinese Context (and Threat?)
The first point to note is that DeepSeek is just one advanced AI system; others are being developed in China, for example, by ByteDance, Alibaba, and Huawei. The emergence on the Western world of DeepSeek is likely to be followed by many others.
The security concerns that many in the West have about TikTok, Huawei’s routers, and information being gathered by Chinese electric cars will be just as relevant using a Chinese AI platform. Will it read your information, will it use it for learning, will it share it with the Chinese Government?
Because DeepSeek conforms to Chinese regulations, there are topics that it won’t cooperate with. For example, when I asked it to “Tell me about Tiananmen Square”, it replied, “Sorry, that’s beyond my current scope. Let’s talk about something else.”
Another concern that has been expressed is about General AI. General AI refers to an artificial intelligence system with the ability to understand, learn, and perform any intellectual task that a human can, across a wide range of domains. General AI can adapt to new situations, solve unfamiliar problems, and exhibit reasoning and creativity similar to human intelligence. General AI is causing concern amongst the AI community, regardless of who controls it, because of its potential. However, the recent advances in China could mean that the Chinese Government has General AI before any other country, which is an additional concern for many of us. (It might also be many years before we see General AI, anywhere.)
Hope for Other Countries?
The fact that China is building advanced models that take less money, energy, and resources than the USA uses could be good news for other countries and regions. Along with the tendency to make parts of the process transparent, it could enable countries like Japan, France, Canada, Argentina, etc, to develop AI platforms that are not controlled by the oligarchs.
You Can Try DeepSeek
You can visit DeepSeek.com and use the web version. You will need to create a free account. In the UK, that means using a phone number or email address or signing in via Google. A disposable email address and an incognito browser is probably the most secure way of doing that.
My recommendation for trying DeepSeek is to open DeepSeek in one browser and your preferred AI (e.g. ChatGPT, Copilot, or Garmin) in another. Then do the same exercise in both windows, comparing the speed and usefulness of the responses.
Ray’s Summary
At present, I will continue to play with DeepSeek, and I will use it for some non-confidential work, e.g. researching my blogs, cross-checking statistical solutions, and experimenting with my open-source projects (such as the NewMR studies into AI and Insight Health). However, I won’t use it for commercial work. I am more prepared to believe that OpenAI and Microsoft with abide by their contracts than with High-Flyer (the Chinese hedge fund behind DeepSeek). I figure that if OpenAI or Microsoft were to be break their contracts and privacy rules, they would be sued by some pretty big players, very quickly.
At this stage, I think the rapid development of AI platforms in China is a good thing. First, it can’t be a good that all of the leading global AI platforms to be owned by American oligarchs. Second, the rise of these platforms may open the way for more AI platforms to be developed by other countries, improving access and oversight.
