Tech deep dives, discoveries, and analyses.
Two years ago, the game of Wordle was released. As we strove to solve it optimally, we built fast estimators for the expected score in our previous article . Used as heuristics, it got us good scores, but not optimal play. This time, using the estimators we devised, let’s achieve perfection, and let’s do that with algorithms generic enough to be applicable to most games. The Russian doll approach…
Two years ago, the game of Wordle was released. I wrote an article back then on finding good words to solve it , but I focused on implementing a good-enough algorithm, which was a bit slow, and switched programming language trying to tame its speed. I ended it saying: “Algorithms do matter too.” It would be nice if, instead of language-specific tweaks, we found an algorithm that made the runtime…
“With more careful calculations, one can win; with less, one cannot” — Sun Tzu, The Art of War . Making extrapolations is crucial to avoid wasting our computing power on slow convergence. After all, if you had to walk to the Everest, you wouldn’t eyeball it: you would use a GPS. Sometimes you have to look away from the GPS and onto the road, though. Sometimes things don’t extrapolate through…
I am pretty impressed seeing Lisa Su doing her best to steer the AMD ship towards better AI support in GPUs, with the Huggingface partnership and by convincing George Hotz to submit more bug reports. (For context, Hotz raised $5M to improve RX 7900 XTX support and sell a $15K prebuilt consumer computer that runs 65B-parameter LLMs. A plethora of driver crashes later, he almost gave up on AMD .)…