The US tech world sat up and took notice this week of an AI model from a Chinese company that was able to achieve state-of-the-art performance without expensive computer chips. That company is DeepSeek, and its model R1 has the potential to reshape how we use chatbots.
The R1 model is a reasoning model that rivals ChatGPT and other viral chatbots. It has a low cost to run and is open source.
What is Deepseek?
Deepseek is a Chinese start-up that grabbed the attention of the AI world this week with a state-of-the-art reasoning model that is shockingly cheap to operate. The company’s R1 model reportedly operates at a fraction of the cost of similar models from competitors. It also requires far less computing power. The company has already seen its AI assistant app overtake ChatGPT on the US App Store charts, spawned hundreds of open-source derivatives, and is onboarded by Microsoft, AWS, and Nvidia AI platforms.
The company is credited with leveraging reinforcement learning (RL) to develop its new models. RL is a training method that uses trial-and-error to train a model by taking actions in a simulated environment and receiving feedback in the form of rewards or penalties. The model then learns to optimize its behavior by analyzing the results of each action to improve its performance. The team then built a system that allows the model to “wake up” only those parts of the algorithm that are most relevant to a given prompt, significantly cutting down its computational needs.
This breakthrough is being widely hailed as AI’s Sputnik moment. It’s forcing Silicon Valley giants to rethink their strategies and take a closer look at the costs and energy demands of their own algorithms. It’s even having ripple effects on other industries, with the stock prices of companies that make components used in the running of AI systems plummeting. This includes Nvidia, whose GPU chips are used to run most of the world’s AI-powered apps, as well as companies like Vistra and Constellation that build the data centers that power most AI models.
Is Deepseek good?
Despite the massive hype around Deepseek, there’s little to distinguish it from other AI models. It can understand natural language and generate outputs based on user input, like any other model in use today. What’s sparked buzz is that it was developed much more cheaply and with better performance than its competitors. That’s a big deal for companies that need to scale their AI capabilities, such as Nvidia.
DeepSeek’s R1 model is a reasoning model, meaning it breaks prompts down into smaller pieces and considers multiple approaches before generating a response. It also has a unique workflow that it uses to improve efficiency. Its Mixture of Experts (MoE) model has 256 “experts” that each specialize in different aspects of the response. For example, one expert may handle logic, while another focuses on music or visual description words, and yet another might be best at proper names or numbers.
MoE also has a gating system that selects the right experts for each query based on its context. This ensures that a single expert doesn’t become overloaded with tasks. It also provides more robust performance by avoiding over-reliance on particular experts, which can result in “model drift”.
The other major difference between Deepseek and its U.S. rivals is that it’s a largely open model, allowing developers to tinker with it and deploy it in a variety of workflows. This is a big deal in the AI world because it allows people to use cutting-edge models without paying for a proprietary solution.
Finally, the company is also transparent about its costs and training data. It claims to have trained the model for only $5.6 million. This is much lower than the billions that other AI companies have spent on developing their own models.
However, not everyone is convinced that the model really is as efficient as advertised. Some American AI researchers have cast doubt on the claim that DeepSeek is a cheaper and more effective alternative to its competitors. Others have pointed out that it still requires external synthetic data from models like GPT-4o to train, which can be computationally expensive. This external data generation could offset the cost savings that Deepseek claims to have achieved with its lean model architecture.
What are the advantages of Deepseek?
A key advantage of Deepseek is that it is open source, allowing users to analyze and adapt the model to their own needs. Additionally, the model can be used locally on the user’s device, minimizing privacy risks and avoiding entrusting sensitive information to a large tech company.
Deepseek uses Reinforcement Learning (RL) to train its logic engine, a type of machine learning that involves a reward system that encourages the model to explore and refine its own solutions to problem-solving scenarios. This approach enables the model to learn from its own mistakes and improve over time, making it a more robust solution than GPT-4o or Claude 3.6 Sonnet.
What sets DeepSeek apart from other reasoning models is its ability to generate holistic answers that take into account all aspects of a question. This is accomplished through the use of a special token in the model’s output, think>, which prompts the model to think about how best to answer the question. The content that follows the think> token is often like a long stream of thought, which can help explain how the model arrived at its final answer.
Another advantage of Deepseek is its ability to handle complex questions, such as those related to mathematics or programming. This is a significant improvement over other models, including Qwen and Claude 3.6 Sonnet, which are unable to handle these types of questions. Additionally, Deepseek is able to provide answers in multiple languages, which is not possible with other generative AI models.
However, the model has some notable disadvantages. For example, it is prone to producing incoherent answers and can become stuck in infinite loops during local reasoning, which limits its usefulness for some use cases. Furthermore, the model’s reliance on reinforcement learning can cause it to exhibit biases, such as censoring sensitive topics or prioritizing pro-Chinese narratives.
Finally, the model can be quite resource-intensive to operate, especially in terms of its training and parameter space. This makes it less suitable for edge devices. For this reason, it is important to consider the hardware requirements of your application when choosing a model.
What are the disadvantages of Deepseek?
DeepSeek R1 has caused shockwaves in the AI world. It has been hailed as the first model to deliver reasoning capabilities on par with O1 at a fraction of its cost and is available to everyone for free through a chatbot interface. It is also available to download and run locally for users with the right hardware, minimizing privacy risks as sensitive data doesn’t need to be sent over the internet.
It is incredibly powerful and fast. It can create text, answer complex questions, code, and perform math and scientific analysis tasks. It also excels at solving problems that require a combination of different skills, such as debugging software or writing an essay. This makes it ideal for automating repetitive development and data analysis workflows.
In addition to its performance, it has a number of other advantages that make it stand out from the competition. It is open source, which means that developers and businesses can customize it to their needs without paying expensive API fees. This allows them to have more control over their AI systems and minimizes the risk of vendor lock-in.
Another advantage is its scalability. Unlike other models, Deepseek R1 can handle multiple input streams simultaneously by activating multiple experts at the same time. This is made possible by the MoE architecture, which uses a load-balancing loss to evenly distribute work between the experts. This reduces the need for a large number of parameters, which is important when training the model.
One disadvantage of Deepseek is that it collects information about the user’s hardware, operating system, and keystroke patterns. This is necessary to enable the model to provide an accurate response, but some users may be uncomfortable with this practice. The company says that it will only share this information with its partners and with “third parties that are necessary to improve the model’s safety, security, and stability.” However, some users have expressed concern about this, arguing that it could be used to spy on them or sell their personal information.