The AI world is evolving at a rate faster than ever, with new technologies gracing the market for commercial and personal use every day. Given this, a certain level of expectation had been set for the advancements we could be seeing in AI in the near future. Which is why the launch of Deep Seek created such waves.
A few weeks ago we thought we knew all the major players in the AI development industry, OpenAI, Meta, Google, and so on. Deep Seek has quickly established a place for itself on that list, with its new AI Chatbot, the DeepSeek-R1, boasting performance as effective as the products their industry giants offer.
Being a Chinese Company though prospective customers, especially business owners, are bound to have doubts about its legitimacy and ethicality, especially with the allegations purported against them by agencies such as OpenAI of stealing code. So what is their story? Let’s being with;
Deep Seek: The Origins
Deep Seek’s founder, Lian Wenfeng, actually got his start in the stock exchange industry. Graduating as a student of Electronic Information Engineering from Zhejiang University, He founded High Flyer, a quantitative hedge fund that used AI-driven trading strategies. The gist of the way High Flyer functioned was using an AI model to predict market trends and function accordingly. The business became wildly successful, and Wenfeng, wanting to continue exploring the application of AI, opened up an AI Lab under the flagship of High Flyer, known as Deep Seek Labs.
Interestingly enough, in 2021, Wenfeng actually started stockpiling Nvidia GPUs for an unannounced AI project that was later revealed to be Deep Seek Labs. According to 36KR, a Chinese media company, he acquired 10,000 Nvidia A100 GPUs before the US restricted the sales of such chips to China.
Deep Seek Labs began research on their first version of the Deep Seek chatbot. They soon became incorporated, and with High Flyer as their investor and backer, became their own company. They soon after released their AI chatbot model that would cause waves in the industry, DeepSeek-R1.
Chat GPT-O1 Versus DeepSeek-R1, Where did the rivalry begin?
So what is with the arms race between Open AI and Deep Seek? Well;
DeepSeek-R1 achieved performance comparable to leading reasoning models on several widely used benchmarks. Math, History, Literature, and Code production are some of the few benchmarks on which it was tested against these models. As seen in the graph, the model could clearly hold its own.
Not to mention, financially, Deep Seek seems to be doing a better job at keeping the operating costs lower than OpenAi or Google. DeepSeek reported a training run cost of roughly $6 million, a paltry figure when compared to the training cost that OpenAI revealed for their latest model, which branches over $60 Million, though its total development costs were likely much higher. Meaning, their training costs for a model that met OpenAI’s model benchmark, was less than 10% of what OpenAI had to use.
Given this tough competition it is no surprise that tempers should rise. OpenAI even accused the team at Deep Seek for possible model distillation and potential misuse of OpenAI outputs, and consequently disabled some accounts for suspicious activity after launching an investigation into them. This turned political very quickly, leading to the United States and China butting heads over the AI race.
What does Deep Seek do differently?
Given that we have now established DeepSeek’s origin, as well as its standing on the world stage when compared to other AI models. What does it do so differently? Well to start with;
The MoE (Mixture of Experts) Architecture
Deep Seek uses a “Mixture of Experts” architecture, or MoE to perform its tasks. While not the first to use it (having been also used by Google, Meta, Mistral, xAI, and multiple research organizations in the past) one could argue that they had the heaviest hand in its popularization in the public consciousness.
This architecture allows the model to analyse information using specific “parts” of its architecture. Considering how we are sometimes told that we only use certain parts of our brain for specific actions, MoE works in a similar way.
The model contains different “Experts”, and as such, will only enable those Experts to “think” given the context of the question you asked it. That isn’t to say it only ever uses one expert at time, in fact multiple experts are often activated, with routing to them occurring at a token level. But overall this leads to lower operating costs and resource usage.
Open Source Deep Seek
DeepSeek has released model weights openly, allowing researchers and developers to inspect and use many of its models. This makes it a popular model for computer scientists, engineers, and most others who work in the IT industry.
It also builds confidence in the model because of the consequent transparency it provides. While we don’t know much about the company behind Deep Seek, the model itself is partly open to research and analysis.
This is in stark contrast to, say, OpenAI, who are very open about their company policies and practices, but keep much of the backend working of ChatGPT to themselves.
Pricing Structures and Audience Interaction
When it comes to customer-side financials, DeepSeek-R1 is completely free to use while other popular models, like Gemini, require tickets or subscriptions for access to their more complex and “smarter” models.
In response to DeepSeek-R1’s release though the companies behind these models have already begun working on free versions of their models to combat this competitive pricing. Consider OpenAI’s recent free release of GPT-o1-mini.
Privacy
In terms of privacy, there isn’t much difference honestly. All major AI chatbot models acquire user data, including prompts, usage logs, and meta data, which they claim to use for further training of their chatbots. Though opt out mechanisms do exist now, with separate privacy policies for you to go through and “enterprise plans” where data is not used for training.
Privacy laws between the two also differ based on;
- Data retention periods,
- Government access rules,
- Legal frameworks,
- Training policies,
- Enterprise protections,
- Compliance certifications,
It is also to be noted that DeepSeek does restrict information on certain political Chinese topics, such as the Tiananmen Square Massacre or the current political status of Taiwan.
In conclusion, DeepSeek may be a shiny new model for the industry, and yes it’s sudden appearance was a shock to the market, but all-in-all it performs with similar statistics to the current models on the market.
In terms of use, it is not dissimilar to models like ChatGPT-O1, but its chatbot being free-to-use, as well as the lower costs of self-hosting and enterprise deployments, does give it an edge in the market.
For those of you who still feel lost when it comes to deciphering AI products, feel free to reach out to us as Genetech Solutions, where our team of AI experts would be happy to help you out! But ultimately, choosing to use DeepSeek-R1 or ChatGPT-O1 comes down to personal preference. Which business would you rather support?
Let us know down in the comments!
