The recent developments in artificial intelligence (AI) have seen significant contributions from various global players. One such noteworthy contribution is from DeepSeek, a Chinese AI model that has caught the attention of major AI companies in the United States, including OpenAI and Meta. Despite initial concerns, DeepSeek’s emergence does not signal the end for these US-based companies. DeepSeek’s model, V3, was trained on only 2,000 H800 GPUs from Nvidia, costing about $5.5 million, a fraction of what OpenAI typically spends. Furthermore, the model’s release as open source is unusual for China, known for its strict regulation of its AI industry.
This efficiency in training likely stems from previous work with less efficient large models, which used heuristic methods and training data to inform the training of smaller, more efficient models. Known as “Distillation,” this method is common in the AI industry. The training of DeepSeek’s model was possible before the tightened export restrictions on AI chips to China. Chinese companies, including DeepSeek, had previously stocked up on Nvidia’s A100 chips, estimated between 10,000 and 50,000 units.
China’s AI development policy aims to enhance economic efficiency through AI, alternating between phases of strict regulation and greater openness. After initial enthusiasm and massive investments, the focus shifted to economically beneficial models, explaining the open-source release of DeepSeek R1. However, this openness could change if the government fears losing control.
Recently, DeepSeek faced cyberattacks, prompting the company to take countermeasures. As reported on their status page, since Monday, January 27, there has been “reduced performance.” Initially, neither login nor API or web interface were accessible. After initial investigations, services were restored at reduced speeds. On Tuesday, DeepSeek disabled new account registrations, while existing accounts remained operational.
In parallel, OpenAI has launched a version of ChatGPT specifically for US government agencies. This “ChatGPT Gov” version aims to help government bodies deploy the latest AI models in their own cloud environments while meeting security and privacy requirements. Over 90,000 users in 3,500 agencies already use ChatGPT.
Meanwhile, a study by the Stifterverband für die Deutsche Wissenschaft and consulting firm McKinsey reveals that 79% of German companies lack the skills to handle AI. Currently, 86% of businesses do not fully exploit AI’s potential. The study, which surveyed over 1,000 executives from various-sized German companies, indicates that most lack a strategy for developing AI-related skills. Only a quarter of respondents have a defined approach. Additionally, 54% of executives note insufficient investment in AI training, citing low AI usage, budget constraints, and uncertainty about required qualifications. More than half of the companies do not provide AI learning opportunities for their employees.
Alibaba has introduced new open-source language models capable of processing large context windows of up to one million tokens. The models, Qwen 2.5, come in two variants with 7 and 14 billion parameters. While large context windows are unique in the open-source space, paid services like Google Gemini already offer this. Experts debate the practical benefits of such large context windows, suggesting more efficient alternatives like RAG architectures with smaller windows of about 128,000 tokens. However, the Qwen models perform well in benchmarks for long texts, with the Qwen2.5-14B-Instruct-1M model achieving over 90 points in the RULER benchmark, surpassing GPT-4o mini in several tests.
In Berlin, the German Museum of Technology has opened an interactive exhibition as part of the national “Mission AI” initiative. Supported by the Federal Ministry for Digital Affairs, this project aims to build trust in AI. Visitors and school groups can experience practical AI applications, including a skin examination app, a robot controlled by a controller, and software for facial recognition and anonymization. Digital Minister Volker Wissing believes that trustworthy, certified AI from Europe will be in high demand in the future.
In another AI application, the Max Planck Institute for Solar System Research (MPS) in Göttingen used AI to analyze data from NASA’s Dawn spacecraft, revealing that hydrocarbons on the dwarf planet Ceres likely originated from external sources. The spectrometer on Dawn had detected signs of organic molecules in several craters on Ceres.