DeepSeek tops the App Store, reveals training costs as low as $30

On January 27th, DeepSeek became the No.1 download in China on the free list of Apple App Store China. In the past 24 hours, DeepSeek also topped the App Store Free Chart in the U.S., surpassing OpenAI's ChatGPT. Recently, a research team led by Dr. Jiayi Pan, a PhD student at the University of California, Berkeley, successfully reproduced the key technology "epiphany moment" in DeepSeek R1-Zero at a very low cost. It is reported that the team used the algorithmic framework of DeepSeek R1-Zero for the experiment, and the team chose the game "Countdown" as the experimental platform. The research shows that even for small-scale language models, through reinforcement learning (RL), the models can develop strong self-verification and search capabilities on their own. Notably, the team showed that the cost of training the model was less than $30. The "aha moment" mentioned above is a key technology in DeepSeek R1-Zero. According to the DeepSeek-R1 technical report on the technology, through the RL framework, AIs may spontaneously develop human-like reasoning abilities, even beyond the limits of predefined rules. And this is also expected to provide direction for the development of more autonomous and adaptive AI models, such as dynamically adjusting strategies in complex decision-making (medical diagnosis, algorithm design). A few days ago, a Meta employee posted on TeamBlind, an anonymous workplace community in the U.S., saying that Meta engineers are frantically analyzing DeepSeek and attempting to replicate some of its content due to the high performance and low training costs of its models. Earlier in the day, blogger scaling01 analyzed DeepSeek's model, which not only outperforms DeepSeek, but also costs only $5.58 million to train, with 671B parameters. In comparison, the computational cost of Meta's Llama3 series of models is enough to train DeepSeek-V3 at least 15 times. In addition, on the afternoon of January 26, the DeepSeek App briefly experienced a busy server and even crashed due to a surge in user traffic.