AI Engineering, Modern PHP, and the Perils of Git as a Database

The AI Engineering Guidebook

TLDR: A free, 350+ page guidebook offers a deep dive into the engineering fundamentals of Large Language Model (LLM) systems. It covers everything from model architecture and training to prompt engineering, Retrieval-Augmented Generation (RAG), and deployment.

Summary: This comprehensive resource provides a structured and in-depth exploration of the practical side of AI engineering. It moves beyond the hype to detail the core concepts required to build robust LLM-powered applications. The guide covers a wide array of topics, including the underlying architecture of models, the nuances of prompt engineering, and the implementation of RAG systems to ground models in factual data. It also delves into advanced techniques like fine-tuning with LoRA, the development of AI agents, and strategies for optimizing model performance. For architects and teams, this guidebook serves as an essential manual for navigating the complexities of the LLM landscape, offering a solid foundation for designing, building, and deploying effective AI solutions. It appears to be a very practical, hands-on resource that consolidates a vast amount of information into a single, cohesive document.

Key takeaways:

• A comprehensive, 350+ page guide on LLM system engineering is available.
• It covers model architecture, training, prompt engineering, and RAG systems.
• Advanced topics like fine-tuning (LoRA) and AI agents are included.
• The guide focuses on practical implementation and deployment considerations.

Link: The AI Engineering Guidebook

Why PHP in 2026?

TLDR: Modern PHP has evolved into a powerful, mature language with a robust type system, JIT compilation, and a streamlined development workflow that requires no build steps. Its simplicity and direct deployment model make it a strong contender in 2026.

Summary: This post challenges the common perception of PHP as an outdated language. The author argues that modern PHP has undergone a significant transformation, incorporating features that make it highly competitive. Key improvements include a strong type system comparable to TypeScript, Just-in-Time (JIT) compilation for better performance, and features like enums and readonly immutability. One of PHP's most significant advantages, according to the author, is its simplicity of deployment. Unlike the complex build pipelines often required for JavaScript frameworks, PHP allows developers to write code and deploy it directly, drastically simplifying the workflow. While the ecosystem has powerful tools like Laravel and Symfony, the core language itself remains straightforward. The author seems to be glossing over the fact that the JS ecosystem's build steps solve real problems of bundling, transpilation for older browsers and tree-shaking, which PHP doesn't have to deal with in the same way.

Key takeaways:

• Modern PHP has a strong type system, JIT compilation, and other advanced features.
• It offers a simplified development workflow with no required build or transpilation steps.
• The ecosystem includes powerful frameworks like Laravel and Symfony.
• PHP's direct deployment model is a significant advantage over complex JavaScript toolchains.

Link: Why PHP in 2026?

Salesforce Regrets Firing 4,000 Experienced Staff and Replacing Them with AI

TLDR: Salesforce executives have publicly admitted that their decision to lay off 4,000 customer support staff and replace them with AI was a mistake. The AI systems could not handle complex customer issues, leading to a decline in service quality.

Summary: This article serves as a cautionary tale about the premature replacement of human expertise with AI. In 2025, Salesforce reduced its customer support workforce by over 40%, from 9,000 to 5,000, with the expectation that AI could fill the gap. However, the company's leadership has now acknowledged that they overestimated the capabilities of their AI systems. The automated solutions struggled to handle the nuance and complexity of real-world customer problems, which experienced human agents were equipped to manage. For technology leaders and architects, this is a critical lesson in understanding the current limitations of AI. While AI is a powerful tool for augmentation and handling routine tasks, this case demonstrates that it is not yet a wholesale replacement for experienced human professionals in complex, customer-facing roles. It highlights the importance of a phased, strategic approach to AI integration rather than a disruptive, cost-cutting-driven replacement.

Key takeaways:

• Salesforce admitted that replacing 4,000 support staff with AI was an overestimation of AI's capabilities.
• The AI struggled with complex customer issues, leading to service quality problems.
• This serves as a cautionary tale about replacing human expertise with AI prematurely.
• Highlights the need for a strategic, not just cost-driven, approach to AI integration.

Link: Salesforce regrets firing 4000 experienced staff and replacing them with AI

Package Managers Keep Using Git as a Database, It Never Works Out

TLDR: Multiple major package managers, including Cargo, Homebrew, CocoaPods, and Go modules, initially used Git for their package registries but eventually migrated away due to severe performance and scalability issues as the number of packages grew.

Summary: This article explores a recurring pattern in the history of package management: the temptation to use Git as a backend for a package registry, and the eventual, inevitable failure of that approach. The author details how several prominent package managers started with Git-based indexes, attracted by its built-in versioning and distributed nature. However, as these registries scaled to thousands and then millions of packages, the performance of Git degraded dramatically. The file system-based nature of Git, which is not optimized for the query patterns of a database, led to slow operations and a poor user experience. Cargo, Homebrew, CocoaPods, and Go modules all had to engineer new, custom solutions to solve the performance bottlenecks they encountered. For architects, this is a classic example of choosing the right tool for the job. While Git is excellent for source control, it is not a database, and using it as one in a high-scale system is a predictable path to technical debt and performance problems.

Key takeaways:

• Using Git as a database for package registries is a recurring anti-pattern.
• Major package managers like Cargo, Homebrew, and Go modules have all abandoned this approach.
• Git's performance degrades significantly at scale when used as a package index.
• This serves as a lesson in choosing the right tool for the job and understanding the limitations of technologies like Git.

Link: Package managers keep using git as a database, it never works out