Kentucky AI.

The open-weight frontier moved again: GLM-5.2 ships MIT-licensed, with a 1M-token context window and a 743B-parameter mixture-of-experts core (256 experts, DeepSeek-V3.2-style sparse attention), posting frontier-class agentic scores — Terminal-Bench 2.1 81.0, SWE-bench Pro 62.1%, as reported. A 4-bit MLX build for Apple Silicon already exists. The honest read on the 128 GB envelope: a 743B MoE is ~400 GB even at 4-bit and wants 256–512 GB of unified memory — open-weight is not the same as fits-on-the-desk. So the play isn't to host it locally; it's to use a model like this as a teacher — distil its judgment into a student that does fit the building — and watch expert-pruning (REAP) close the rest of the gap.

Song & Zheng reframe distillation as iterative correction, not single-pass imitation: instead of copying a teacher's flawless prefixes, the student generates and the teacher grades what the student actually produced — closing the exposure-bias gap that compounds over long outputs. The read for us: this is exactly how you transfer a frontier model's judgment into a small, locally-served student that has to hold up over a full planset, not a one-line prompt.

Apple's own measurements: the M5's GPU neural accelerators cut time-to-first-token by up to ~3.6× over the M4 under MLX, and 4-bit quantization fits a 30B-class model inside 24 GB of unified memory. Confirmation of the bet we build on — serious, reasoning-capable models now run sovereign on the desk, not only in someone else's datacenter.

Behrouz et al. give recurrent models a memory that grows with sequence length — caching checkpoints of the hidden state instead of crushing the entire past into one fixed vector, a tunable dial between an RNN's flat memory cost and a Transformer's growing one. Their variants land close to Transformers on in-context recall. The inference-infrastructure read for us: reading a full planset and its specs is a long-context recall problem, and a model that recalls without Transformer-scale KV blowup is exactly what locally-served, sovereign serving on our hardware needs.

Raza, Bilberg, Malik & Ribeiro da Silva map digital-twin implementation across manufacturing systems — integration levels, which components actually close the loop. Directly relevant to the spatial-knowledge-graph thread: a building, like a shop floor, is only as smart as the twin that stays current.

A PRISMA review of 392 studies mapping AI across cost, time, and safety management. The academic baseline for AI-first estimating — and a useful map of where the literature is thin (causal claims, field validation).

Zacua Ventures' read matches ours from the other side of the table: AI compounds only when paired with standardized data, repeatable process, and engineering oversight. 67% of ConTech investors are increasing AI exposure — the capital is following the data layer, not the demo.

Bluebeam's survey of 1,000+ AEC professionals (via ASCE): only 27% use AI day-to-day, but 94% of those are scaling up in 2026, and 68% of early adopters have saved at least $50k. Adoption is early and accelerating — exactly the window to set the standard.