The Chinese AI startup, DeepSeek, is in the news today due to its rapid rise and the impact that had on NVIDIA stock. Apparently, DeepSeek developed an AI model that has caused a stir. For one thing, it just became the top-downloaded free app in the Apple Store, surpassing ChatGPT. This led to a sharp decline in NVIDIA’s stock, as investors are supposedly concerned that DeepSeek’s AI model is more cost-efficient. More more cost-efficient means that it threatens NVIDIA’s dominance in providing hardware for AI processing. If fewer are needed to build a powerful AI agent, then NVIDIA’s products will be in lower demand.
But what do we know about DeepSeek’s mysterious claims that it can do more with less? They claim that DeepSeek achieved comparable results to Western AI models but at a fraction of the cost (using fewer NVIDIA chips). This spooked investors, and suddenly NVIDIA’s stock broke a record, suffering the largest single-day loss in history, dropping by nearly $600 billion in market capitalization, according to Forbes and just about every other news source there is (though I’m not sure how that number is accurate, since the market cap should equal Stock Price × Shares Outstanding, which is only about 60 billion lost).
How did DeepSeek do it? If it did in fact do more with fewer NVIDIA chips, what’s the secret? DeepSeek even claims that it achieved comparable results to Western AI models. What has the West been overlooking?
Multiple sources suggested today that DeepSeek’s AI model (V3) has displayed signs of identity confusion, often misidentifying itself as ChatGPT. (Neowin, 2024) This seems to imply that DeepSeek V3 may have absorbed and internalized elements of ChatGPT’s identity during its training, likely due to the inclusion of ChatGPT-generated content in the training data.
So does the use of ChatGPT content make it a knockoff? Doubtful. The point of learning is to build on existing knowledge, which makes the open source model as authentic as its teacher. If the reasoning ability can be developed and the knowledge base too, then it’s hard to see the downside. Rather than being problematic, what we’re seeing in DeepSeek may just be the rise of a technology that isn’t available only to a chosen few who can afford to horde powerful processors. The work that has been done by those powerful processors can be leveraged to give DeepSeek a head start.
But the initial knowledge base isn’t the only difference between the two models. DeepSeek’s creators thought outside the box:
With Monday’s full release of R1 and the accompanying technical paper, the company revealed a surprising innovation: a deliberate departure from the conventional supervised fine-tuning (SFT) process widely used in training large language models (LLMs).
The developers apparently sidestepped SFT and relied on reinforcement learning (RL) instead.
The core strength of DeepSeek-R1 lies in its use of RL to develop reasoning capabilities. Unlike traditional supervised learning, RL allows the model to improve autonomously through feedback, adapting its behavior and becoming better over time. (Nucleusbox)
RL is a powerful reasoning mechanism because it allows LLMs to self-correct to fix their mistakes “on-the-fly,” as explained by Kumat et al. (2024) and (DeepSeek-AI 2025).
Stated in a language that is fit for a regular human, the 3 main differences between the for-profit ChatGPT LLM and DeepSeek are (1) teaching & learning styles, (2) the ability to transplant skills and knowledge from one DS to a smaller DS, and (3) one is being previewed and tailored by those who can afford it, while the other one is being worked on and improved by anyone who can get their hands on it. That’s open source. The first one is easier to understand because we we all know that some approaches to teaching and learning tend to work better than others. The second one is sort of science fiction-like and very cool. The third is the most effective way to ensure the progress of science. That said, OpenAI will no doubt be complaining about property theft.
I leave you with the words of DS’ creator.
We directly apply RL to the base model without relying on supervised fine-tuning (SFT) as a preliminary step. This approach allows the model to explore chain-of-thought (CoT) for solving complex problems, resulting in the development of DeepSeek-R1-Zero. DeepSeek-R1-Zero demonstrates capabilities such as self-verification, reflection, and generating long CoTs, marking a significant milestone for the research community. Notably, it is the first open research to validate that reasoning capabilities of LLMs can be incentivized purely through RL, without the need for SFT. This breakthrough paves the way for future advancements in this area.
We demonstrate that the reasoning patterns of larger models can be distilled into smaller models, resulting in better performance compared to the reasoning patterns discovered through RL on small models. The open source DeepSeek-R1, as well as its API, will benefit the research community to distill better smaller models in the future. https://arxiv.org/html/2501.12948v1
