Machine-Learning on Phil Wheat

Prompt Engineering

Sat, 17 Oct 2026 11:57:00 -0500

Muhammad is an AI/ML expert with a PhD who has led ML teams for eight years, and the channel’s stated goal is sharing the learning “without all the fluff and hype.” Recent uploads are exactly that: distilling a 4B Qwen model to catch the 27B, a 31x-faster NCM architecture explainer, raw head-to-head app builds across Kimi K3.1 Pro and Opus 5.5, and self-improving agents without training. 252K subscribers, uploads every few days — one of the better technical (non-hype) AI channels.

PyTorch

Sat, 17 Oct 2026 11:33:00 -0500

The official PyTorch Foundation channel (Linux Foundation-hosted), covering the framework plus the portfolio around it — vLLM, DeepSpeed, Ray, Helion, Safetensors. Recent talks are serious inference engineering: tier-aware KV-cache routing for vLLM at scale, Trainium’s journey to native PyTorch, masked-diffusion language models on AMD XDNA2 NPUs, and API parity across heterogeneous accelerators. The primary-source feed for open-source AI infrastructure; weekly uploads.

Welch Labs

Sat, 17 Oct 2026 10:56:00 -0500

The best visual explainers in AI: what’s inside the world’s smartest robot brain (VLA models), Yann LeCun’s $1B bet against LLMs in two parts, “the most cited paper of the century is a brilliant hack,” and whether humans can make AI any better. 8.74M subscribers, uploads every couple of months — each one a production. The companion “Illustrated Guide to AI” book is free on the channel.