Understanding DeepSeek R1

"Our generation's Sputnik moment is now." Popularised by President Barak Obama in a 2010 speech.

DeepSeek R1 has changed the AI

DeepSeek R1 has made such an impact you cannot help but notice the headlines in printed and digital media and on social platforms. As best I could, I tried to make sense of what the fuss was all about. I understand that DeepSeek R1, developed by a Chinese team, is an AI model. But what makes it different from what others have done in the AI space? This is what I have learned after reading articles and watching hours of YouTube videos. It has some characteristics that sets it apart from other AI models.

DeepSeek Performance

It was assumed that the USA had a significant AI lead over other countries including China. The USA has access to the latest computer chips, to skilled, experienced and talented people, and billions of dollars to invest. DeepSeek has performance levels on a par with Open AI’s Chat GPT O1, Googles Gemini, MS Copilot and other models from other tech teams. DeepSeek is opensource, to evaluate and verify the claims made by the developers is accessible to anyone who wants to verify the AI. DeepSeek can be downloaded from Apple or Googles Play Store. 

Tech details are often a mystery to me so do your own research. A maths comparison showed these results; DeepSeek open source 79.8%, Open AI Chat GPT 79.2%. Comparison explained in Big Technology Podcast.

Cost

CNBC reported Open AI budgeted $5 billion per year to build and maintain its Chat GPT AI models. According to Epoch AI, Google spent $191 million to develop Gemini Ultra. Anthropic used $100 million, according to Fox news.

DS Open Source 55c/million token input $2.19c/million token output

Open AI Chat GPT $15/million token input $60/million token output

Comparison explained in Big Technology Podcast.

DeepSeek out-performed USA models on a budget of $5.6 million. CNBC.

Resources

The USA banned China from importing the latest GPUs. DeepSeek used what was available, the GPUs were not the latest Nvidia GPUs. There is some speculation, China had a small supply of the most recent GPUs. It is not proven. Fox News reports that DeepSeek was built by a small team in less than two months. One wit on the France24 Debate on DeepSeek said, “China has more smart kids than Europe has kids.” Perplexity CEO Aravind Srinivas commented, “Necessity is the mother of invention.” Both statements implying that China made the best possible use of its resources while being imaginative and innovative to achieve the breakthrough with DeepSeek.

Innovation

DeepSeek used distillation to be as streamlined as possible. Not using all the features of the big AI models but enough to function effectively on devices with less computing power. It is also the basis of its self-reasoning ability.

Impact

The resources required to build an AI model like DeepSeek is not as huge as the billions of dollars being poured into the big tech companies to build an AI model. China has shown the way in building more with less. The reduced value of shares in most of the AI companies has shocked the market. Investors in Nvidia have decided that the projections made are unlikely to be achieved, as reflected in the drop in its share value. Shown in the table above the cost of an intelligence unit has dropped dramatically due to DeepSeek. Open source is its strength, already over 500 derivatives have been built upon the DeepSeek open source. It is the number one downloaded app in the USA.

My View

This is a Sputnik moment. A contributor to France24 called DeepSeek a gift to humanity. It can also be used for hybrid wars or in other dangerous ways. The fear of your data being controlled by the CCP is another concern. Despite all that the world is better with DeepSeek than without it. The opportunities for all of us to have a better life, be it in wealth creation, in better allocation of resources through informed decision making, in health outcomes, in self-fulfilment, in having a secure society. Benefits that cannot be ignored.