A RAR compressor for every version of RAR was developed using OpenAI Codex 5.5 and Claude Opus 4.7 in approximately five weeks, costing around £40 in tokens. The resulting code consists of 55,000 lines and is not optimized for speed.

Tech company executives at Meta, Google, and Microsoft report that AI is increasingly generating a significant portion of code, leading to faster and cheaper software production. This trend suggests that AI's efficiency improvements may result in reduced headcount within these companies.

The shift in software development priorities has moved from speed of deployment to performance efficiency. As AI tools increasingly write code, languages like Rust, Go, and C++ are gaining traction due to their superior performance, despite longer ramp-up times and smaller talent pools compared to Python and TypeScript.

The increasing reliance on AI in software engineering may diminish the learning opportunities for engineers, potentially leading to a decline in their effectiveness. Some software engineers argue that this trend could undermine the longevity of software engineering as a career.

The author returns to writing code by hand after exploring the limits of software development without direct involvement. Key insights include the comparison of "em-dash" in AI writing to "god-object" in AI coding and the concept of vibe-coding enhancing the development process.

AI coding agents must significantly reduce maintenance costs in proportion to productivity gains. Achieving twice the coding speed should ideally result in halving maintenance costs, while tripling productivity should lead to one-third of the original costs.

Mojo is a programming language designed to combine the user-friendliness of Python with the performance of C++. It allows developers to write fast code for various hardware, including CPUs and GPUs, while ensuring memory safety and avoiding vendor lock-in.

Version 1.110 of Visual Studio Code introduces a setting to include Copilot as a coauthor in commit messages by appending "Co-authored-by: Copilot <copilot@github.com>". The setting git.addAICoAuthor offers three values: off (no attribution), conditional attribution based on Copilot's assistance, and always add attribution.

Frontier models excel at coding tasks, surpassing their performance in other areas. A list of guidelines for effective agentic coding has been compiled to enhance the coding experience with AI agents.

Agent Skills aims to ensure AI coding agents adhere to essential software development practices such as writing specifications, tests, and conducting reviews. By default, AI coding agents prioritize completing tasks quickly without verifying prerequisites, which Agent Skills seeks to address.