The author required a tensor library to run a poker model in Web GPU but found that the desired library didn't exist, and ultimately, it wasn't needed. The project aimed to create an open-source, in-browser poker solver, which presented significant compute challenges.

Instead of relying on a pre-existing library, the author instructed Codex to build custom Web GPU kernels. This approach, leveraging Py Torch as a reference for correctness, successfully generated the necessary code and achieved a 10x speedup.

The experience highlights how, in the current LLM era, custom, verifiable code can outperform general libraries when the computation is well-defined and tests are robust. The author emphasizes that while LLMs can now generate complex code, human oversight remains crucial for planning and judgment.