By Vlad Collak
It’s very difficult to miss the latest AI news about DeepSeek. If you’re not tracking this, a small Chinese startup released an AI model that could be on par with OpenAI’s best models in terms of reasoning capability but (as they claim) at a much cheaper cost to train. Also, the cost of the DeepSeek API as compared to OpenAI API is an order of magnitude cheaper. The initial stock market reaction was swift erasing about a $1 trillion of value on the fear that the world will need fewer GPUs (chips used to train and run models). We’ll know soon enough if the model is much cheaper to train, but at the moment, it seems that it’s pretty efficient in inference (running the model) and with an on-par performance compared to large commercial reasoning models from the likes of OpenAI, Google, and others.
However, I think there is a bigger and more nuanced story here than a temporary stock market dip. The existence of a better, cheaper, and faster model will mean more competition, which will lead to more innovation, and the world will benefit. Cheaper models will lead to more usage, which will in turn increase GPU usage anyway (Jevon’s Paradox). We must also remember that the current state of the art Transformer architecture coupled with RLHF (Reinforcement Learning from Human Feedback) will likely not be the last innovation, and the AI disruptions will only continue to drive the costs down and the level of intelligence up. What’s important is that the DeepSeek model is open-source (code and weights). With smarter, cheaper to run, and free to use models, we can create a much more intelligent world where individuals and organizations benefit immensely – whether that’s automating their processes via agents or providing insights and intelligence to their employees to increase productivity.
At NewTide, we’re excited about great open-source models as long as we can leverage them in a safe, secure, and private way. We are certainly intrigued by DeepSeek, but we don’t recommend using the DeepSeek API as it’s not clear whether your data will be used to train new models. Others are reacting similarly as shown by the immediate ban of the DeepSeek web and mobile apps by the US Navy.
However, using the open-source version should be fine as long DeepSeek has not violated anyone’s Intellectual Property as the OpenAI now alleges. To that end, we aim to be a trusted AI platform for the Fuels and Convenience industry that’s both model agnostic as well as safe and secure. This will enable us to both leverage the intelligence these kinds of models provide while de-risking customers from having just one type of model and leveraging the best model for a particular industry use case.
