Recently, Anthropic and OpenAI released Fable 5.1 and GPT-6 Astra respectively, but the former quickly lost its momentum, while Astra continued to dominate the headlines and maintain high popularity. Some opinions suggest that Fable 5.1 is not only expensive but also has many restrictions, leading to customer dissatisfaction; Astra may lag slightly in coding performance, but it excels in overall capabilities, has higher token efficiency, and actually reduces overall costs. Its application spans various fields, and netizens are actively exploring ways to use it to change existing work methods.

Liang Wenfeng: Not focusing on video or world models, focusing on LLM is the main line

The recent surge of Astra has also prompted a reflection from Professor Shota Imai, a visiting professor at Japan's JAIST University. His perspective is interesting—he mentioned a statement made by Liang Wenfeng, the founder of DeepSeek, in an internal speech several months ago.

DeepSeek previously had limited support for multimodal features, which relates to Liang Wenfeng's views. He believes that video generation AI, world models, or physical models are not the main path to AGI. Instead, he emphasizes focusing on the development of large language models (LLMs), and then using these large models to solve other problems. Imai believes that after seeing Astra's full capabilities, Liang Wenfeng's statement has become more convincing.

According to Liang Wenfeng, DeepSeek's real core goal is AGI, and other tasks can be handled easily—by aiming for higher technical goals and then applying them to lower-level technologies, it often becomes a "dimensional advantage." He also stated that DeepSeek will not focus its main efforts on those AI directions. This isn't to say those areas have no value, but they are not the key bottlenecks limiting AGI.

OpenAI's strategic adjustment pays off, domestic large models need to look at the next generation

OpenAI had adjusted its strategy due to competition from Anthropic and Chinese open-source large models, abandoning video generation models and focusing on core AI capabilities, including coding. This reform now appears very successful. After the release of GPT-5.5 and now GPT-6, OpenAI has returned to its peak. With the release of Astra, OpenAI confidently stated that it will usher in the AGI era. NVIDIA CEO Huang Renxun also agreed, and he even mentioned the role of his company's GPUs in training Astra, saying that only 100,000 GPUs were used so far, and the next time it would be 400,000.

The article also points out that although it's boring for American companies to market with "achieving AGI," Astra's capabilities indeed represent a significant leap forward. At the same time, it must be acknowledged that domestic AI large models have fallen further behind, especially given the existing gap in computing resources. The flagship large models of American companies have already been released, and the next generation of domestic large models will depend on Kimi 3.1, GLM-5.5, MiniMax M3 Pro, Qwen4, and DeepSeek V4.1.