DeepSeek: What Features Does the New AI Chatbot Have?

China's new DeepSeek AI-powered chatbot app has caused a stir in the tech sector. As the most downloaded free iOS software in the US, it swiftly surpassed OpenAI's ChatGPT.

Additionally, it caused chip manufacturer Nvidia to lose about $600 billion (£483 billion) of its market value in a single day, setting a new record for the US stock market.

According to reports, the “large language model” (LLM) that drives the app can reason just as well as US models like OpenAI's o1, but it is supposedly much less expensive to train and operate.

DeepSeek Innovations

DeepSeek has reduced the computation time and memory needed to train its R1 model, significantly reducing costs. The base model V3 of R1 required 2.788 million hours to train, running across multiple GPUs simultaneously, at an estimated cost of under $6m. According to OpenAI boss Sam Altman, this is significantly lower than the over $100m required to train GPT-4.

Moreover, DeepSeek models were trained on around 2,000 Nvidia H800 GPUs, a modified version of the H100 chip, to comply with export rules to China. These chips were likely stockpiled before the Biden administration tightened restrictions in October 2023, effectively banning Nvidia from exporting H800s to China. As a result, DeepSeek has had to find innovative ways to maximize its resources, despite the market value impact.

Plus, reducing computational costs for AI models can mitigate environmental concerns. Data centres require significant electricity and water to prevent overheating.

ChatGPT's monthly carbon dioxide emissions are over 260 tonnes, equivalent to 260 flights from London to New York. Increasing model efficiency could positively impact the industry's environmental impact.

In addition, DeepSeek has gained popularity due to its large language model, which combines smaller models with specific domain expertise. Nevertheless, the energy savings models of DeepSeek are uncertain, and the Paris AI Action Summit could promote sustainable AI, ensuring future tools are environmentally friendly.



DeepSeek and OpenAI's Models 

The latest DeepSeek model is unique due to the open release of its “weights” and technical paper, allowing researchers worldwide to explore its features. This is unlike OpenAI's “o1 and o3” which are black boxes. However, there are still missing details, such as datasets and code used, which researchers are working to gather.

Furthermore, DeepSeek's cost-cutting techniques, such as the “mixture of experts” technique, have been used in other LLMs like Mistral AI's Mixtral 8x7B model. Both models utilize a group of smaller models with specific domain expertise, assigning tasks to the most qualified expert.

Also, DeepSeek has attempted to improve LLM reasoning through technical approaches like Monte Carlo Tree Search but has failed.

Besides, researchers are exploring ways to enhance the problem-solving capabilities of DeepSeek, which could lead to the development of more sophisticated AI models with fewer resources.

This could result in highly capable models being developed with ever fewer resources, as companies find ways to make model training and operation more efficient.

Donald Trump has praised DeepSeek's rise as a wake-up call for the US tech industry, highlighting the potential benefits of AI.

In closing, as the cost of developing AI products decreases, businesses, and governments can adopt the technology more easily, driving demand for new products and chips. Smaller companies like DeepSeek may play a growing role in AI tool development.

Read more news:

Crude Oil Projections: OPEC+ Summit and US Data in the Spotlight

NASA's Roman Telescope: A Giant Leap Toward Understanding the Universe

Satellites Undergo “Mass Migrations” Due to Geomagnetic Storms