Articles
This article, as Shixiang AGI's quarterly observation report, reveals the trend of LLM routes diverging from general-purpose to vertical fields through interviews and analyses of Silicon Valley large model companies (such as OpenAI, Google Gemini, Anthropic, xAI, etc.), presenting two major directions: “horizontal platform approach” and “vertical integration”. The article emphasizes that while exploring the upper limit of intelligence, the importance of product and non-technical barriers is increasingly prominent, especially exemplified by ChatGPT's brand and brand recognition. At the same time, AI products face the challenge of shortening the innovation cycle, with giants accelerating the layout of Agent products, squeezing the space for startups. The article also proposes the concept of Level 4 Agent experience, believing that information search (such as ChatGPT Deep Research) and software development (such as Claude Code) have taken the lead in achieving it, and discusses the core capabilities required of future AI product managers.
This article provides an in-depth analysis of the Physical Intelligence (PI) team's exploration in general-purpose robotics. It elaborates on how the VLA (Vision-Language-Action) model evolved from the VLM (Vision-Language Model) to directly generate robot action commands, rather than merely understanding the scene. The PI team built an efficient robot data pipeline from scratch, significantly enhancing the model's generalization performance in unknown environments by collecting large-scale, diverse real-world operational data. To address the contradiction between efficiency and generalization ability in traditional model training, PI proposed an innovative "Knowledge Insulation" mechanism. This mechanism protects the semantic understanding ability of the backbone network by discretizing action sequences and truncating gradient backpropagation, while also increasing the training speed tenfold. The article also discusses in detail the three core challenges faced in deploying robots in the open world: data scarcity, performance instability, and the complexity of hardware platform migration. Finally, it envisions that future robot intelligence will move towards deep integration of software and hardware, general-purpose task "recipes," and the business model of "Robot Model as a Service" (RMaaS).