Manus recently withdrew from the Chinese market, cleared its domestic social account content, and moved to overseas markets with all its strength. The official explanation was that the reason was mainly based on the adjustment of operating efficiency and international layout. On July 19th, Beijing time, Manus co-founder Ji Yichao published a technical blog, responding for the first time from a technical perspective, summarizing the experiences and lessons learned in Agent R&D and training since the company was founded.
From a technical perspective, Ji Yichao said that Manus will focus on context engineering and achieve rapid product iteration with the help of structural "memory" and processes. It mainly includes betting on context, no longer training models, emphasizing the significance of KV-Cache (Key-Value Cache, a caching mechanism) hit rate, not dynamically adding tools, and using the file system to host persistent context. The core is to save the training cost of the underlying model and focus on improving the training efficiency.
In large models, context usually refers to the collection of information that the model refers to when processing tasks or generating output content. It can help the model enhance its understanding, improve task performance, and enhance output coherence. Previously, Dark Side of the Moon Kimi founder Yang Zhilin emphasized the importance of context in an interview. He said that the ultimate value of Ai-native (product form defined by AI) products is to provide personalized interaction, and lossless long context (LosslessLongContext) is the key to achieving this goal. He judged that fine-tuning of the model should not exist in the long run. The interaction history between the user and the model is the best personalization process, and long context technology can better record and utilize these interaction histories.
In addition, the KV-Cache hit rate is crucial, mainly because a high hit rate can improve inference efficiency, optimize resource utilization, and reduce computing costs. Based on this, KV-Cache is often called the efficiency core of the Transformer model inference phase.
Choosing to improve training efficiency from the above aspects rather than starting from the underlying model is a lesson Ji Yiguo has learned over the years. He said that when he started his last company (Peak Labs), the team decided to train models for open information extraction and semantic search from scratch, but soon afterwards, OpenAI's GPT-3 and Google's Flan-T5 models appeared, and the internal models developed by the team from scratch became irrelevant overnight. "Ironically, these models mark the beginning of contextual learning and a new way forward." Ji Yichao said.
Based on previous lessons, after starting Manus, the team no longer invested in base model research and development, but chose between using open source basic models to train end-to-end agents, and building agents based on the context learning capabilities of cutting-edge models. Although the lessons learned from Peak Labs made the Manus team realize the importance of context, it was not easy. It took four Agent framework adjustments to achieve the local optimal solution.
However, it should be noted that this strategy still has limitations, especially when faced with the ChatGPT Agent just released by OpenAI. The core reason is that ChatGPT Agent relies on OpenAI's dedicated model and adopts end-to-end training, which can better handle complex tasks. Although Manus can improve efficiency, it still relies on external multi-model combination and engineering optimization, and is slightly inferior in task execution consistency and accuracy.
In addition, when Manus entered the international market, OpenAI brought the Agent industry to an inflection point with its underlying model advantages, attracting more developers and users to major manufacturers' platforms. Although startups have room for survival in vertical fields, they still inevitably face the challenge of competing for market share. Especially when agent products currently face difficulties such as serious homogeneity, unclear business models, and high costs, highlights in contextual engineering and other aspects are not enough for startups to stand out. The team still needs to continue to optimize technical strategies and explore differentiated development paths.