The year 2026 is a strange year. Whilst making dinner, I had Claude write me a language model using PyTorch in around 100 lines of code, and trained it on a few open datasets in about an hour. After dinner, I was running inference on the worst language model I’ve ever used (obviously) but having the most fun I’ve had with one (maybe).
My favorite thing about this technology, I think, is that it’s a simple (enough) thing to make—if you’ve got the data and the compute, at least—but it produces something so complex that we still don’t understand how it really works. To make even a toy model on your (very warm, fans at full speed) MacBook is an almost spiritual experience.