1002: The Real Pricing of LLMs

Wes and Scott dive into the recent shift to usage-based pricing for AI models, especially highlighting GitHub Copilot’s new model multipliers which significantly increase costs—up to 27 times more expensive for certain models. They discuss the implications of these costs, the previous subsidization by large tech companies, and how models are evolving to be more expensive over time due to tokenization and model improvements. The conversation considers whether AI might become less affordable, leading to a shift toward smaller, more targeted models that are cost-effective for specific workflows.

A story from Adib Hanna about a near-hack through a fake recruiter interview emphasizes the increasing dangers of malicious code and security breaches. The discussion highlights concerns over malicious scripts, untrustworthy apps, and the importance of using decompiler tools and cautious practices when dealing with code repositories or untrusted software. The conversation underlines the need for vigilance in securing development environments and understanding the security pitfalls of new AI-powered tools.

Wes and Scott reflect on the challenge of keeping up with fast-moving technologies, especially after taking recent time away from active development. They suggest focusing on results rather than just the latest tech trends, and emphasize the importance of understanding native browser APIs—like CSS features and new web standards—that often go underutilized. They advise prioritizing meaningful projects and learning about fundamental capabilities, such as the improvements in CSS and browser APIs, which remain relevant and powerful.

The hosts discuss the tension between using React component libraries (like those for accordions and dropdowns) versus native HTML and CSS solutions. They argue that native solutions often suffice and can be more performant and accessible, provided developers explicitly seek out and leverage these APIs. They critique complicated component libraries (such as the elaborate accordion implementations) that add unnecessary complexity, advocating for simple, semantic HTML with minimal JavaScript whenever possible.

The podcast covers new tools for improving CSS quality, including efforts from Project Wallace and CSS kit.rs. These tools aim to analyze and lint CSS more effectively, helping developers avoid the pitfalls of unmaintained or poorly written CSS. The discussion emphasizes the importance of guardrails in CSS to prevent common mistakes like unnecessary styles or overuse of certain properties, and highlights ongoing innovations in web tooling.

Scott and Wes revisit how to evaluate developer skills given the rise of AI-assisted coding. They recommend using lower-powered, faster models during interviews to better gauge a candidate’s problem-solving process and approach. Practical advice includes asking architectural questions and discussing system design—areas where understanding process and reasoning reveal more than simple coding tasks. They also debate the importance of assessing candidates' ability to think strategically rather than just rote memorization.

The hosts share their preferred strategies for handling Node.js versions and package managers, discussing tools like core pack, N, and Vee+ (V-plus). They express a preference for using Vee+ because it manages Node versions and package managers dynamically, simplifying workflows and reducing headaches related to environment setup. The conversation emphasizes the importance of having an all-in-one tool that automates switching versions per project to avoid common issues with version mismatches.

Scott and Wes share their recommendations for hardware, lighting, and practical tips to improve their work setup. Scott mentions the Amaran Halo 100x light for its standard attachments and Bluetooth control, while Wes highlights the Zeiss lens cleaner for screens and glasses and shares cleaning routines for maintaining equipment. These insights serve as affordable, useful upgrades to optimize their video and workspace quality.