The Rapid Development of AI: Insights from Artificial Analysis Reports
Introduction
Hello friends, welcome. The field of artificial intelligence is evolving at an astonishing pace, especially the competition between China and the United States. Many wonder about the gap between them and the latest breakthroughs. Today, we'll explore reports by Artificial Analysis, an independent AI standard test and survey provider. These reports offer valuable insights into China's AI landscape and the DeepSeq R1 model.
Overview of the Discussion
Before delving in, this is an audio podcast from WOW's YouTube video channel. You can find and subscribe to it on YouTube by searching "addwow.insight". The video content is often more comprehensive.
Today, we'll first look at the recent contrast in AI models between the US and China, focusing on intelligence. Then, we'll examine the open-source model field, where China has taken the lead, and the role of DeepSeq. Next, we'll explore important trends in the AI field, such as reasoning models and multi-model abilities. Finally, we'll analyze the DeepSeq R1 model, especially its latest update.
The Narrowing Gap in AI Model Intelligence
Artificial Analysis' Q2 2025 report shows that the intelligence level gap between top AI models in China and the US has significantly reduced. The gap, which was over a year, has shortened to less than three months. This conclusion is based on the Artificial Analysis assessment system, which combines seven difficult assessment standards, including MLLU Pro, GPQA Diamond, AMME Mathematics, Humanities Last Exam, and Live Code Bench.
Using this index, they evaluated different models and found that the latest OpenAI model (O3E) and DeepSeq's May 2025 R1 version had similar scores, with a gap of less than three months when calculated in time. This challenges the common perception that the US, especially OpenAI, is far ahead in the field.
Driving Forces Behind the Gap Reduction
Both the US and China are making progress. In the US, the potential in the front-end intelligent level mainly depends on OpenAI, with its O1 to O3 models setting the standard. In China, two major driving forces are DeepSeq AI and Alibaba's Damo Yuan. Their results have quickly pushed China's top-level smart level up.
Impact of the Shrinking Gap
The rapid disappearance of the gap has several deep impacts. Firstly, it intensifies the competition in the global AI field, which has now entered a white-hot stage. Secondly, it shortens the innovation period, forcing all participants to speed up development and delivery. Finally, it increases the uncertainty of the future pattern, affecting investment, talent flow, and technology strategies.
China's Lead in the Open-Source Model Field
In the open-source model field, the situation has reversed. China gained a leading position in November 2024 when Alibaba released the QLVOQ model. Its 32B parameter preview version exceeded the intelligence index of Meta's flagship open-source model, Lama 3.1405B.
The report attributes this to a strategic difference. Chinese top AI laboratories like DeepSeq AI and Alibaba tend to open up their most powerful or flagship models, sharing the core structure and knowledge of the model. In contrast, US giants like OpenAI, Anthropic, and Google usually protect their most advanced models as exclusive assets.
The Role of DeepSeq in the Open-Source Revolution
DeepSeq R1 is a milestone in the open-source revolution. The version released in January 2020 was the first open-source model that could compete with OpenAI's O1 in terms of reasoning ability. The subsequent R1052 version in May 2025 was the most intelligent open-source model in the world, according to Artificial Analysis.
DeepSeq R1's Significant Update
The May 2025 update of DeepSeq R1 brought a significant ability leap. The model's score on the Artificial Analysis AI index increased from about 60 to 68. This improvement is comparable to OpenAI's major upgrade from O1 to O3.
The report listed specific score improvements on key standard tests. For example, it improved by 21 points on the AMI-E math test, 15 points on the LiveCodeBench code generation test, 10 points on the GPQA scientific reasoning test, and 6 points on the Humanities Last Exam comprehensive reasoning test.
Reasons for the Performance Improvement
The underlying model structure of DeepSeq R1 did not change. The huge performance improvement is mainly due to the post-training stage, specifically the optimization of reinforcement learning (RL). RL allows the model to learn through trial and error and reward mechanisms, making the output more in line with human preferences and improving performance on specific tasks.
The report also mentioned that the expansion of RL is more effective and cost-efficient than the expansion of pre-training. It provides a path for laboratories with limited computing resources to achieve significant performance improvements.
Other Factors Affecting R1's Performance
The report also observed an increase in token consumption by the new version of R1-0528. More consumed tokens usually mean the model is thinking deeper or generating more intermediate arrangements before giving the answer. This increased thinking time is considered an important factor in improving the model's performance in complex reasoning tasks.
Model Generalization and Data Stream Analysis
Sam Page's analysis based on output fingerprint is also interesting. It compares the output styles of different models. The first-generation R1 version's output fingerprints were closer to OpenAI models, while the updated R1-0528 version's fingerprints were closer to Google's Gemini model. This leads to the speculation that DeepSeq may have changed its training strategy from relying on OpenAI's output to Gemini's output.
Other Key Trends in the AI Field
The report also summarizes other important trends. The reasoning model, which thinks more internally and consumes more tokens to solve complex problems, is now the leading force promoting intelligence potential. The combination expert (MOE) architecture is becoming more popular due to its efficiency. Multi-modal AI, which can process text, images, and videos, is also a focus.
In terms of cost, the cost of intelligent reasoning is rapidly decreasing, but the actual total calculation needs may increase due to the generalization of reasoning models and the innovation of applications like AI Agents. AI Agents, which can understand goals, design plans, and use tools to complete complex tasks, are considered an important direction for the next step of AI development.
Conclusion
For those keeping up with the AI wave, the most important things are the speed of development and the generalization of technology. The competition between China and the US and major laboratories is driving rapid progress. The innovation of open-source models and more efficient post-training methods are making powerful AI technology more accessible.
Finally, the question of AI model training data sources, especially the potential sharing of output from other powerful models, is an important issue with both technical and ethical implications. It's worth our continued attention and discussion. Thank you for listening.
{# 
{# 
{# 
{# 
{#