.png)
February 4, 2025
The AI landscape is in constant motion. New partnerships, innovative technologies, and fundamental research are shaping its development. A current example of this is the strategic cooperation between the Japanese technology group Softbank and the US-American AI company OpenAI.
Softbank and OpenAI have announced the formation of a joint venture called "SB OpenAI Japan." The goal of this collaboration is the development of an AI-based work environment for companies under the name "Cristal Intelligence." Softbank plans to invest three billion US dollars annually in this project. The partnership also includes the implementation of ChatGPT Enterprise for Softbank employees, starting with the subsidiaries ARM and Softbank Corp. in Japan.
As part of the announcement, Softbank CEO Masayoshi Son expressed optimism about the imminent development of Artificial General Intelligence (AGI). His assessment suggests accelerated development, possibly based on a changed understanding of the AGI concept. This positive prognosis contrasts with earlier statements by OpenAI CEO Sam Altman, who had downplayed the significance of AGI.
The Chinese AI company DeepSeek impresses with the deployment of enormous resources for the development of its language model V3. According to reports, DeepSeek has around 60,000 Nvidia graphics processors, including high-end models like the H100. These investments in hardware suggest a significantly higher financial commitment than the officially stated training costs of 5.6 million US dollars.
Particularly noteworthy is DeepSeek's alleged access to 10,000 H100 chips, which are actually not allowed to be exported to China due to US sanctions. DeepSeek remains silent about the hardware for the even more powerful model R1. Speculation suggests that AI accelerators from the Chinese manufacturer Huawei could also be used here.
OpenAI CEO Sam Altman admitted errors in the company's open-source strategy in a public forum. He hinted that OpenAI is internally discussing a new approach that focuses more on open source. At the same time, Altman dampened expectations for OpenAI's future market leadership and acknowledged a dwindling lead over the competition.
Concrete plans for GPT-5 were not mentioned, but updates for the GPT-4o series are to follow. The already announced native image generation for GPT-4o is still in development, according to product manager Kevin Weil. In addition, OpenAI plans more transparency in the thought processes of its reasoning models, whereby a balance must be found between user interests and protection against competitors.
The French AI startup Mistral AI has released its new language model Small 3 with 24 billion parameters under the Apache 2.0 license. This model achieves comparable performance to larger models from Meta, Qwen, and OpenAI, but is optimized for low latency and is therefore particularly suitable for local use. Mistral plans to release further models with improved reasoning capabilities.
A study by Tencent has shown that reasoning models like OpenAI's o1 or DeepSeek's R1 tend to abandon promising approaches prematurely and switch between different strategies. This "underthinking" occurs particularly in complex tasks. The researchers are working on methods to correct this behavior and improve the performance of the models.
Scientists at the University of California, Berkeley have developed a robot that can use a whip to knock individual Jenga blocks out of a tower without it collapsing. By combining reinforcement learning and human correction, the robot achieves impressive precision and surpasses human players through its consistent performance.
