I Spent 1.5 Billion AI Tokens on Absolute Slop

Summary

Chris Titus describes an experiment using an AI coding model and multiple workers to expand Win OneShot, a .NET counterpart to his existing PowerShell-based Windows utility. Over roughly 24 hours of compute and an estimated 1.5 billion tokens—which he guesses might represent around $10,000 or more—the AI was given an existing project, a specification, and a roadmap to add features such as Windows 11 creation and AppX removal.

The result added 169,310 lines of code to a project that was intended to need only a modest expansion. The first version looked promising, and some of its interface ideas, such as a sidebar, were useful. But testing and iteration exposed problems: slower startup, a missing category element, settings that sometimes reported the wrong status, and updates that did not work. Fixing and polishing these issues became difficult because the codebase was too large and hard to understand or maintain.

Titus argues that the main failure was not simply the model. He let the project run too long without reviewing its architecture or setting checkpoints. AI-generated tests and other implementation choices added further complexity, and relying on more prompts to fix problems became risky once the code was no longer readable. A more powerful model might have produced a better result, he says, but it would not have solved the underlying problems with planning, specifications, and oversight.

He demonstrates reverting the project with Git, cleaning leftover generated files and build artifacts, and returning to the simpler pre-AI version. The original project was around 10,000 lines, while straightforward interface changes might have required only 100–200 lines. The experiment did yield a few design ideas, but the large AI-generated implementation was discarded.

Key Advice and Practical Takeaways

• Use Git and understand how to revert changes. Titus emphasizes version control as essential protection against costly, uncontrolled changes.
• Work incrementally. Make one clearly scoped change at a time, review the code and behavior, and only then continue.
• Set checkpoints on large projects. Don’t let an AI agent make extensive changes without inspecting the structure and results.
• Keep code readable, testable, and maintainable. A working demo is not necessarily ready for release, especially if bugs cannot be diagnosed without repeatedly asking the model.
• Use AI as a programming aid, not a replacement for developers. He recommends it for code review, debugging subtle issues, and speeding up small tools or website work, while leaving architecture and accountability to people who understand software development.

Review or Guide Content

The video is a review of a failed AI-assisted software expansion. It also includes a practical demonstration of using Git to restore the project and clean generated files, followed by advice on managing AI coding work through incremental implementation and review.

Main speaker/source: Chris Titus, discussing his own Win OneShot project, AI-generated code, and Git history.