I can’t think of the last time a Chinese company made so many headlines in the United States.
The DeepSeek story has put a lot of Americans on edge, and started people thinking about what the international race for AI is going to look like. I got an intro to speak directly with a staff from Deepseek and got the inside story.
But we don’t always have to be in competition all the time. There’s a way to promote collaboration and unity in this important journey that we’re taking, and in fact, it just might help us to get greater success in adjusting to life in the AI age.
So here are some of the things I learned as I talked with someone with direct experience helping businesses to adopt DeepSeek open source models. Karl Zhao has a lot of industry experience – we talked broadly about where things are headed, and what strategies helped the firm to stand out at an inflection point in the industry.
1. It’s All Very New
One of the things that our conversation returned to, again and again, is that people are still trying to understand the ramifications of new open source models like DeepSeek R1.
Zhao said he often recommends an “ecosystem approach” for B2B or B2C applications. That means taking the whole techstack into consideration and looking at the context of what you’re trying to build.
2. Corporate Buy-In is Happening
Another related insight is that some of the biggest American tech companies are embracing open source AI and even experimenting with DeepSeek models.
That’s significant, because left to their own devices, a lot of these companies would probably shy away from utilizing Chinese products. Microsoft and Amazon are two companies that are reportedly using DeepSeek, and hosting these models stateside, which helps other businesses to feel more comfortable with adoption.
3. The Money is Secondary
Another thing I’ve learned is that DeepSeek isn’t really looking to monetize its models directly.
It’s much more of a research thing. Of course, end users are going to use this for enterprise, so people will be making money off of using the DeepSeek models. But according to Zhao, monetization is not the priority. It’s done “more on the back end.”
4. Engineering Principles are Important
Let’s talk about engineering optimization for a second.
People have been asking what DeepSeek did to make their model more efficient.
First, there’s taking full advantage of reinforcement learning,and skipping the supervised fine-tuning that’s often a part of the process.
But there’s also the mixture of experts or MoE approach, where DeepSeek used multiple agents to formulate those LLM processes that make its source model work. Zhao emphasized that this is going to be part of the industry moving forwar
In addition, there’s also playing around with data types (fixed point versus block floating point) operations and removing unnecessary computations from the pipeline, partially by working in assembly language instead of at the C code level.
We also talked about using alternatives to the Nvidia Cuda method. The idea is that if companies can get around the Nvidia CUDA API made for the company’s GPUs, there’s more versatility in play.
5. Multiple Models Work
If individual users or businesses are taking advantage of an ensemble approach, it stands to reason that not everyone will use the same mix of models.
That means there might be room for not only DeepSeek, but Meta, OpenAI and others in a sort of melting pot approach so the right tool is used different jobs. For his part, Sam Altman has recently said friendly things about open source as a concept, and “Open AI maybe on the wrong side of history.”
As a corollary point, open source is almost by nature not proprietary or provincial in certain ways. We really need to stop positioning this whole situation as a China vs US competition. By sharing their code, weights and training methodology Deepseek is helping advance the entire AI industry globally. The next versions from all the labs will likely be incorporating their advances.
Also, its important to point out Deepseek isn’t a state sponsored or funded project – it’s privately funded by an individual. And although the DeepSeek model is censored in the version hosted in China, according to local laws, Zhao pointed out that the models that are downloadable for self hosting or hosted by western cloud providers (AWS/Azure, etc.) are not censored. They are reinvigorating the open source AI movement globally by making a true frontier level model available with full open MIT license.
Those are some things to think about as we move forward in analyzing what happened with DeepSeek’s announcement, and how it impacts things like the U.S. stock market, as well as what it means for geopolitical interactions.
“Due to the extreme high costs of pretraining frontier models the last few years, academic institutions have been for the most part excluded from the innovation process in advance AI, but with the gift of Deepseek making such an advanced reasoning model available to the world with full source, weights, methodology and free MIT license, we now enable hundreds of thousands of researchers in small university labs or even at home to partake in bringing progress to the field. This will help bring on a new renaissance of invention that’s now accessible to all parts of the world.” – says author/tech executive Alvin W. Graylin
