Posts

This article summarizes the lessons learned by indie developers on how to avoid zero revenue for two to three months without team or funding support. The core lies in shifting the development mindset from code-centric to market-centric, ensuring that even if the product fails, there is still something to gain through strategies such as reversing the development process, leveraging external forces, extreme constraint development, and asset compounding. The article provides specific action steps and practical tips to help developers quickly validate market pain points, obtain user feedback, and make efficient use of time and resources.

This article mainly introduces the Alibaba Cloud OSS client to implement singleton mode running account. In the original code, every user request for a file inside the bucket would be created and destroyed on the backend, causing a huge waste of performance. The method used to download files is: oss downloads files to the backend, and the backend transmits the file content to the frontend through the input and output streams, which also puts great pressure on the backend. The flow chart of the original code is as follows:

This article provides two solutions to the problem of GPU memory explosion during PyTorch training: first, monitoring memory usage step by step via `torch.cuda.memory_allocated(device)` to locate the bottleneck; second, analyzing the accumulation mechanism of the computation graph, pointing out that improper operations in loops (such as appending/accumulating loss) can cause historical variables to remain in memory for extended periods, and emphasizing the need to proactively release tensors using `detach` or `item` rather than relying solely on `no_grad()`. The article also demonstrates a custom Trainer class implementation, explaining how to avoid unintended retention of intermediate results such as `hidden_states`.

This article systematically reviews the core mechanisms of the Rust language, including ownership transfer (implemented through assignment and `into`), the single-threaded shared `Rc` smart pointer, the cross-thread-safe `Arc` smart pointer, the `Mutex` for multi-threaded data protection, the borrowing helper functions `as_ref`/`as_deref` for safe reference conversion, as well as the error handling patterns of `Option`/`Result` and the rules for lifetime annotations. It focuses on analyzing the implementation principles and application scenarios of these features in memory management and concurrent programming.

**AI Summary:** This article delves into the role of MCP (Function Calling Protocol) in enhancing the stability and flexibility of LLMs. By unifying the specification for function descriptions, it resolves the instability in calls caused by differing description styles; and by adopting a client-server model to dynamically retrieve parameter information, it enables seamless adaptation during interface upgrades. The article progresses from basic function calls to technical challenges such as multi-style compatibility and dynamic parameters, revealing how MCP establishes a standardized function-calling framework that provides a reliable foundation for interactions between LLMs and external systems.