Raschka has built a reputation as one of the strongest hands-on educators in the field, and this course targets a real gap: most public material on reasoning models stops at prompting tricks, while the actual training pipeline — RL, distillation, test-time scaling — remains locked inside frontier labs. Implementing it in PyTorch from scratch is exactly the kind of transparency the field needs now that reasoning has become the industry's main battleground. The timing looks right to us: after DeepSeek-R1 demonstrated the recipe can be replicated outside big labs, demand for engineers who understand these mechanics will only grow. The usual caveat applies — courses age fast in this field — but the fundamentals covered here should hold up well.

A small model, but the strategy behind it is anything but. NVIDIA post-training a Google Gemma base and shipping it under its own banner shows the company is fighting for the software layer of edge AI, not just the silicon underneath. The ~4GB VRAM target aims squarely at the fastest-growing demand pockets — always-on voice agents, game NPCs and robots — where cloud round-trips kill latency and cost. On its face this is an incremental release rather than a breakthrough: the open small-model space is already crowded with Qwen and Mistral variants, so benchmarks will decide whether Bolt is best-in-class or mostly ecosystem glue for Jetson. Either way, expect more of these — NVIDIA clearly wants its stack to be the default brain for embodied AI, and readers watching the robotics market should treat this as a positioning move, not a headline model.