Chinese AI DeepSeek R1 Roils Up Tech World
On Monday 27 January 2025, tech stocks tumbled down following the release of DeepSeek R1 #AI Engine. Based on benchmarking tests, DeepSeek R1 Chat may be as good, if not better than the OpenAI o1 ChatGPT. The news sent shock waves to Silicon Valley as it was generally assumed that American technology was at least ‘a good few years ahead of Chinese copycats.’
What is even more remarkable is that the performance was achieved under stringent restrictions which prevented Nvidia from shipping its top of the line GPU to China. Therefore, DeepSeek had to rely on less powerful older-generation chips, unless it found a way to get around the embargo.
To those who remain skeptical, China went one step further and made the DeepSeek R1 code opensource, which means anybody can take a look at it, tweak it, and test it to verify the claims. In the name of transparency, DeepSeek even published a technical paper to explicit how its improved algorithm actually works.
From that technical paper, we can learn that the improvement in efficiency stems from three innovative aspects: reinforcement learning, distillation, and chain of thought (COT).
DeepSeek Founder LIANG Wenfeng
“Chinese firms often settle for copying and commercialization. We want to change that. DeepSeek is here to ignite ‘hardcore innovation’ across the economy. We’re driven by curiosity, not competition. Artificial General Intelligence (AGI) is the ultimate mystery, and we’re here to unravel it.”
Reinforcement Learning is a form of unsupervised learning in the sense that the system extracts knowledge from an unlabeled dataset as opposed to Supervised Learning where the system is trained on labeled dataset. In Reinforcement Learning, the system attempts to maximize cumulative rewards as it iterates. For instance, that is how a baby learns to walk: the more steps the baby can take, the higher the cumulative rewards, whereas a fall requires a correction and provides a lesson of what not to do.
Distillation allows a much smaller system to have similar performance as a much larger system. The savings is achieved by transferring key knowledge from the larger system to the smaller system. By analogy, this is how a old teacher would teach a young student. The guru will distill the sea of knowledge for the disciple to digest so that the disciple can perform satisfactorily in a particular subject matter.
The chain of thought actually gives the user a valuable insight into how the machine is thinking as it works towards an answer. #COT thus allows the user to monitor and fine-tune the line of reasoning should the AI be going off on a tangent or falling down into a cul-de-sac.
Analysts were divided on the impact of DeepSeek. Some highlighted that the prowess of Chinese AI spells gloom for the US tech companies. Other underlined the superiority of the opensource model will fosters innovation at a faster pace and a lower cost. By making the code #opensource, China has offered the world a gift for public good which imposes a paradigm shift and a rethink of how AI is developed.
One thing is for sure though. AI will never be the same as from now. To put things in perspective, whereas other AI models need data centers to train and run, DeepSeek’s AI model can run on a laptop, paving the way for personal AI. Some analysts compared the moment when Personal Computers (PC) started to challenge mainframes.
Nvidia, AMD, Microsoft, Google stocks all closed down with a deep dip at the closing bell on 27 January 2025, with a combined loss in value estimated at trillions of dollars. On its own, Nviidia lost a record-breaking $600 billion in market value on a single day of trading of Monday the 27th January 2025.
The rise of ‘the mysterious force from the East’ caused President Trump that it sounded ‘a wake-up call’ to Silicon Valley which will have to double down on innovation to keep up with the Chinese. The global race for #AI dominance is on!