Science News

Article originally written by Sejal Mackey ’27.

AI has taken over our daily lives. Whether it’s schools using it (or banning it), artworks being created with it, and hearing of the dangers of it online, it’s impossible to avoid any mentions of AI. Part of what makes AI so interesting is the hidden, paywalled, tightly-controlled limitations of working with it: making it much more mysterious since most people do not understand the code behind it. However, on January 27, a small Chinese start-up changed this – launching an open-source AI model that opens new possibilities for AI: DeepSeek. DeepSeek took over the world quickly, shocking the world (even hurting AI-related stocks for companies such as Nvidia) and bringing the world closer to achieving a more eco-friendly AI.

There are already many different AI models, such as GPT-4 or Google Gemini, so what makes DeepSeek so different from the rest? Well, AI companies normally spend billions on training models, relying on massive supercomputers to hold so much data, and using high-end, expensive computer chips to create bigger models. However, DeepSeek wanted to work more with efficiency, striving to show that AI can be made without technological brute force. DeepSeek’s model uses a technique called Mixture-of-Experts (MoE), which is more like a team rather than a single model. So when the AI receives a prompt, only the “team members” that are needed to answer the question work to respond, while the rest are idle – highly reducing computing needs. Also, they designed their model to work on less powerful yet more widely available computer chips, which are much cheaper. Using a program called PTX, DeepSeek was designed to allow developers to control how AI interacts with the chip at a much lower level. This way, a lot of performance could still be achieved on a less powerful chip. Also, the training was optimized so there were smaller teams of human reviewers. This process was automated with DeepSeek, using reinforcement learning – the AI learns from experience, not by human oversight.

The company was founded in late 2023, but it quickly changed the world after its first language model was released: DeepSeek-R1. This was very sudden and unexpected, as a small startup fully changed the industry after barely a year. Since the US had imposed trade restrictions on the more expensive chips that the other AI models used, DeepSeek adapted to the more available, cheaper chips – optimizing its model to run on lower-end hardware. This launch helped the world see that smarter software, with open access and optimization, could be used for AI, not just large and expensive computers.

Most AI models are restricted to paying customers, but their technology and training data largely are kept private. These companies control how the models are used to prevent unauthorized changes. Some AI models are publically available, but not for commercial use or for large-scale purposes. However, DeepSeek has no restrictions. anyone can change the model and use it without any agreements with the company. This gives everyone access to AI without fees, encouraging global AI development. This can have some risks though, since DeepSeek’s model cannot be monitored for harmful content: such as misinformation, fraud, abuse of the model, or even cyberattacks.

DeepSeek can seem to be possibly harmful, as security and misuse can be concerns of the model. The U.S. Navy, Texas, Taiwan, and Italy have all restricted or banned DeepSeek, with South Korea, France, Ireland, and the Netherlands reviewing its practices. The U.S. faces a challenge of AI regulation – can they encourage progress while preventing security risks?

On the other hand, Deepseek brings exciting opportunities to the entire world, opening new possibilities for AI development. Also, since DeepSeek is open source, it can redefine how AI is developed globally in every field. DeepSeek poses new questions to our world, changing how AI development was typically limited to large corporations with massive computing powers. Will open-source AI change the industry or introduce new risks to our country?

Sources:

https://www.thedrum.com/opinion/2025/01/27/how-deepseek-ai-just-changed-the-game-llms

https://www.weforum.org/stories/2025/02/open-source-ai-innovation-deepseek/

https://dev.to/nitdgplug/the-deepseek-revolution-the-ai-game-changer-you-need-to-know-about-1f57