July 30, 2026
All-in on AI, and the railbirds erupted
Solving poker in custom WebGPU kernels
AI built a free browser poker brain, and the comments instantly went all-in
TLDR: A developer used AI to turn his poker software into custom browser code, making a free in-browser solver possible and much faster than a basic version. Commenters were split between cheering the breakthrough, nitpicking poker details, and dropping the usual “nice, but try this other tool” energy.
A developer wanted to make a free poker helper that runs in your browser, without making people pay for pricey software. The catch? The usual building blocks didn’t exist for what he needed. So instead of waiting for a library, he had an AI coding tool generate custom browser code, checked it against his old Python version, and—plot twist—it apparently worked on the first try and then got more than 10 times faster after an overnight optimization binge. That alone was enough to make the replies split into two camps: “this is the future” and “hold on, that description is a little too neat.”
The hottest mini-drama came from the poker purists. One commenter jumped in with a classic well, actually, pushing back on how the post described a famous poker solver and insisting it doesn’t use the shortcuts the article suggested. In other words: yes, people were absolutely fact-checking the poker nerd details before the applause had even finished. Meanwhile, another reader gave the project the ultimate community stamp of approval, saying their favorite test hands were handled correctly—basically the poker-solver version of a standing ovation.
And then there was the comic relief. One disappointed reader showed up expecting old-school poker-themed computer science storytelling and left with a tiny existential sigh: “shoulda known huh.” Another person popped in with a handy alternative tool link, because no tech thread is complete without someone saying, “cool, but have you tried this other thing?”
Key Points
- •The article describes building an in-browser poker solver despite the lack of a general tensor library for WebGPU comparable to PyTorch.
- •The author used a PyTorch implementation as a trusted reference and had Codex generate WebGPU kernels that matched its outputs.
- •According to the article, the generated WebGPU implementation passed parity tests after one prompt and was later optimized to achieve over a 10x speedup versus a naive version.
- •The article argues that custom generated kernels can outperform general libraries when computation is well defined and verification tests are reliable.
- •The technical context includes poker equilibrium solving with counterfactual regret minimization, older abstraction-based tabular solvers, and newer re-solving methods that use neural networks at search cutoffs.