Is Deepseek too good to be true?

                  Recently, a China-based artificial intelligence startup called Deepseek released its R1 model, claiming to have only spent USD $6M in training, a fraction of the cost of similar large language models. Deepseek R1 has shown comparable performance versus the most popular large language model, ChatGPT, and additionally boasts a Deepthink feature which provides the thought process behind the model’s response. Furthermore, Deepseek is completely open source, meaning that anyone with the proper resources can run its services for free. Deepseek immediately disrupted the stock market with the stock prices of many major tech companies like NVIDIA, known for backing the GPU hardware behind language models, crashing.

However, controversies arose when clear biases in the model were shown. For example, the model censors anything related to the 1989 Tiananmen Square Massacre. Although, what may be more concerning is the misleading nature of Deepseek’s claims that it took USD $6MM to train their latest model. While at face value this claim is true, this claim does not account for the great hardware and labor costs that go behind developing AI models. High-Flyer Quant, the hedge fund backing Deepseek, invested in around ten thousand NVIDIA A100 GPUs in 2019, the most advanced at the time. Furthermore, the server capital and operation costs for building the model may have reached around USD $1Bn according to Semianalysis, an AI research company.  The misleading nature of Deepseek’s claims are definitely concerning as it became a leading language model overnight and many are backing it as a leader in AI research.

Despite these concerns, Deepseek’s model still has a lot of potential as a research paper published by Deepseek researchers boasts new technology that can make the training process much more efficient. Furthermore, Deepseek set a precedent for open-source language models that can be run on any cloud server with the right technology and technological know-how. Deepseek also brought more competition to the AI market, mostly dominated by large American tech giants and should not be merely viewed as a deceitful model using trickery to sell itself.