Open source tools and writeups hosted on GitHub.
Where in the 32-bit search space do Bitcoin's winning nonces actually land, and is it predictable? Six months of block headers (26,040 blocks) with models trained on everything but the most recent week, then tested blind on that held-out week. The placement is not random and each pool's fleet has its own signature, but the honest headline is that the best model beats the uniform baseline by 0.0075 bits per block: a real edge, and a tiny one. The lottery itself is fair.
AI-generated meshes land at around 300k triangles with 2K textures, which is far too heavy for a web build. The usual fix is Draco or KTX2, but plenty of importers (glTFast in its default mode among them) don't read those. So this slims GLBs using plain glTF features only: weld, decimate, resize textures. It also fixes the sharp edge that costs people hours, where a texture re-encoded as JPEG still declares itself PNG and the whole model silently fails to load.
Two dependency-light shell scripts that drive the Tripo generative API from the command line: text to a 3D model (GLB), and text to an image (PNG). The official flow is a multi-step REST dance (create a task, poll it, fetch the output) and these collapse it into one command, with a balance check that refuses to run at zero so a misconfigured key never silently burns credits, and polling that survives a network blip.