When the public gained access to the Chinese AI company DeepSeek’s R1 large-language model (LLM) in late January, the media generally saw it as a major breakthrough. Pundits claimed it demonstrated the country’s innovation and ability to produce its own AI technological advances at 5% of the cost of US companies, also using open-source code and sharing its own code.
The headlines made it sound like that was the end of US dominance in AI and Big Tech. Global technology stocks, including OpenAI, Google, Meta, Nvidia and others, lost an aggregate $1 trillion in market value on the news.
Only a few days later, we learned that none of the company’s claims were true.
After examining the AI engine, OpenAI discovered that it appears to have been illegally trained on its ChatGPT4 version. DeepSeek prompts elicit almost identical information in many cases. Further investigation uncovered that the creators of DeepSeek used 50,000 H100 computers, relying on an AI chip that it wasn’t supposed to possess because of US export restrictions. A former Nvidia executive working for the company helped smuggle them in.
While DeepSeek claimed it only cost $5.5 million to create its LLM, more recent information has shown that number only represents the last stage of training, and that its costs were comparable to that of US companies. So no, the company didn’t innovate nearly as much as it would have us believe.
Once again a Chinese company misappropriated American hardware and technology to get ahead. Wall Street, investors and US AI companies stopped freaking out once they began to catch on to the reality of the situation.
DeepSeek did two unique things. It put a front end on the model that explains its reasoning. Second, it created its model using design innovations that US companies were also starting to use. DeepSeek’s model is based on a combination of multiple individually optimized language models, making it more efficient. It employs two innovative strategies, group-relative policy optimization (GRPO) and multi-head latent attention (MLA).
Using GRPO training can be described as training an AI to do a task by exposing it to how multiple LLMs do it. Compare it to a dog being trained on an agility course, observing how other dogs perform the same activities and learning from their successes and failures.
MLA helps AI process information more efficiently, especially long sequences of data. It works like having multiple spotlights focused on different parts of a large text document, allowing the AI to derive the overall meaning more effectively.
For China, necessity is the mother of invention because its companies don’t have access to the extensive research and development that US companies have already built.
DeepSeek has made some of its code available after using many of the open-source resources that are already available, like Mixture of Experts and Q-STaR. But it has not made it truly open-source in that it hasn’t published its training data, processes, data pipelines or other elements that feed the model. Some experts think DeepSeek withheld its training data and related processes because that makes it impossible to validate actual training costs and efficiencies.
So, again, the company has been deceptive about how it operates.
With DeepSeek, users who want to keep their data out of Chinese hands are instead giving them easy access. Sure, OpenAI admits that it trains on your queries, but the company that operates DeepSeek has a user agreement that gives it access to your data in a far more compromising way. A lot of curious people downloaded the free AI app to their phones and computers without reading the user agreement; they may want to rethink that decision and delete it. TikTok is far less intrusive. his is a major part of why several countries and many branches of our and other governments and militaries have already banned DeepSeek from their computers. At this writing US congressional leaders are proposing a law to ban DeepSeek and similar Chinese AI models to protect our privacy.
China’s government funded eleven “foundational” AI models to try to dominate the field globally. Because it has total control over its AI companies, they will always be suspect in terms of privacy.
NotebookLM
In writing this article I used a free AI tool called NotebookLM, which I’d call a game-changer. You can do research and feed the program your background links, upload a document with your summary, and tell it what you want it to do, such as create a podcast or an article, and it will whip one up in no time. I wrote this article from scratch, but I borrowed some analogies from the NotebookLM podcast that my husband Andy generated for his AI company, based on his white paper about DeepSeek’s R1 LLM. I was so impressed with how it synthesized information that I plan to keep it as a tool in my arsenal for research and writing, along with Grammarly, ChatGPT4 for background information and image generation, and Otter.ai for transcribing recordings.
