Core Insights - Meta Platforms Inc. is shifting its focus towards commercializable AI models, with CEO Mark Zuckerberg personally involved in daily operations to build one of the most expensive tech teams in history [1][8] - A new model, codenamed "Avocado," is expected to be launched in spring next year as a "closed" model, marking a significant shift from Meta's previously promoted open-source strategy [1][8] - The company has faced disappointment with its open-source model Llama 4, leading to a strategic pivot and the recruitment of top AI researchers with lucrative compensation packages [1][8] Group 1: Strategic Shift - The new AI model "Avocado" will be strictly controlled and accessible for purchase, aligning with approaches taken by competitors like Google and OpenAI [1][8] - Meta's new Chief AI Officer, Alexandr Wang, supports the closed model approach, indicating a departure from the open-source philosophy [1][8] - The company has marginalized some personnel involved in the Llama project and is now focusing on a team called TBD Lab, which is closely monitored by Zuckerberg [1][4] Group 2: Training and Development - The TBD team is utilizing various third-party models, including Google's Gemma, OpenAI's gpt-oss, and Alibaba's Tongyi Qianwen, to train the new model [2][9] - Zuckerberg's previous concerns about national scrutiny of models have shifted, as he now calls for U.S. government support for domestic tech firms to maintain a competitive edge in AI [2][9] Group 3: Leadership and Personnel Changes - Alexandr Wang, a 28-year-old talent from Scale AI, is seen as crucial for the success of Meta's upcoming AI products, with Zuckerberg positioning himself as Wang's mentor [4][12] - The departure of Yann LeCun, a prominent figure in AI research at Meta, reflects internal challenges and a lack of resources, as well as a shift away from the open-source strategy [7][15] - Following the release of Llama 4, the next generation model was initially codenamed Behemoth but has been abandoned in favor of new directions [15]
Meta人工智能方向转向闭源模式 据悉训练模型的工具包括通义千问