Ethan Lin (@cuda_ethan) header
Ethan Lin (@cuda_ethan) profile picture

making models fit in vram they have no business fitting in. inference, quantization, kernels. dropped out to ship. gpu poor, e/acc

hayes valley, sf
Joined November 2026
Photo by Ethan Lin (@cuda_ethan)Photo by Ethan Lin (@cuda_ethan)Photo by Ethan Lin (@cuda_ethan)
You're so not ready for what AI just did to GTA Lucia and Jason face the old crew on a desert road at sunset. Then Tommy, CJ and the others walk off into the light. A goodbye before GTA 6 GTA 6 is the biggest traffic wave of the year, and this is how people are getting paid from it. Full setup below
1
13
517
A random number generator just snitched on DeepSeek How the GPU ends up inside the weights: - PyTorch fills fresh weights with Philox from one seed - The work is split across 2,048 threads per SM - H100: 132 SMs. B200/B300: 148. GB200/GB300: 152 - Same seed on a different chip puts the same numbers in different places - Rows for tokens the model never saw don't get trained, so the starting values ship in the release Match the layout and you know the chip. 49 to 1,385 sigma. A B300 rebuilt DeepSeek-V4-Flash's starting weights bit for bit, an H100 rebuilt none It proves which chip created the weights. Not where the chips were
SITUATION DETECTED: Former OpenAI researcher Will DePue says he has found strong forensic evidence that Chinese AI labs DeepSeek, Moonshot AI, MiniMax, and Ant Group all trained their newest frontier models on NVIDIA Blackwell GPUs, despite U.S. export controls.
10
40
529
Do NOT buy a GPU rack for local AI yet Harbor pools unified memory over Thunderbolt 5. Plug in, start, done. No CUDA, no rack Pooled Macs fit bigger models, and with RDMA over Thunderbolt 5 they run them faster too. exo measured it: - 1 Mac: 1.0x tok/s - 2 Macs: 1.8x - 4 Macs: 3.2x Apple pulled the app the same day
Apple is working on a new app called Harbor Harbor turns your Mac into a private AI inference server. Run open-weight models locally on Apple silicon—your data never leaves your hardware, and nothing reaches the cloud. Connect multiple Mac computers over Thunderbolt 5 to pool unified memory, enabling large models and higher concurrency that a single computer can handle alone. Set up in minutes, not hours Harbor automatically detects your devices over Thunderbolt 5 to form a cluster. Plug in, start, and you’re running—no manual configuration, no command line, no infrastructure expertise required. Download and run models directly Browse, download, and load open-weight models from within the app. Harbor handles model management so you can focus on the work, not the setup. Cluster for scale Connect two or more Thunderbolt 5-capable Mac computers and Harbor pools their unified memory into a single inference resource. Run models that exceed any single computer’s capacity, or serve more concurrent requests across your cluster. Harbor is currently in a private preview under the Claris brand.
10
27
378