I’m excited about SKY‑001, the Engineering World Model built by a Toronto lab (ex‑Maluuba, now Microsoft Research). It treats a product as a hierarchical tree – body → assemblies → brackets/plates → fasteners – and generates designs by recursively splitting the tree, sketching each component in 2‑D, lifting to 3‑D, and stitching everything into a Fusion file.
- SKY‑001 benchmarks claim 100‑125× faster runtime and 150‑210× cheaper cost on CAD/BIM tasks compared to GPT‑6 Astra, with higher IOU.
I decided to adopt the tree‑based decomposition because flat token‑cost reasoning blows up on large designs; the hierarchical approach mirrors how Toyota engineers sketch cars and gives the massive speed/price gains reported.
I also promoted AgentCloak, a free browser extension by Peter Yared that swaps real data for fake data before requests hit ChatGPT or Claude. The launch video – four dead Mac Mini shells bought for ≈ $65 each, smashed on camera – pulled 2.2 M views in total and ≈ 200 K first‑hour views. Since launch the extension has 3 000 downloads and 8 700 bookmarks.
- AgentCloak metrics: 2.2 M video views; ~200 K views in the first hour; 3 000 downloads; 8 700 bookmarks.
Finally, I laid out a six‑level Model‑PM career roadmap. Each level ends with a concrete “build” artifact – from a pre‑training explanation (Level 1) to a release‑notes migration guide (Level 5) – and the whole path is informed by 46 open model‑PM roles (none require hands‑on ML research) and a survey of 316 PMs (only 7 have PhDs).
- Model‑PM roadmap facts: 6 levels; 46 open roles; 316 surveyed PMs (7 PhDs).