AI
Seven Numbers, No Memory, 96.5%: The 14-Byte Brain That Shames Your Render Farm
MINIMIO's 14-byte neural net solves 96.5% of unseen mazes with no memory — and quietly teaches AEC teams when reactive beats heavy tooling.
A Fields medallist quit maths for AI safety — the lesson is where the gate sits
Fields medallist Jacob Tsimerman left maths for OpenAI's AI-safety work. The lesson for AEC offices: put the guarantee in the wiring, not in trust.
DeepSeek-V4 ships a million-token context at a tenth the KV cache
DeepSeek-V4's preview cuts KV cache to 10% and FLOPs to 27% at a million tokens — the infrastructure move that makes long-context AI affordable for AEC offices.
Kimi K3 Hits the Studio: a 2.8-Trillion Open Model That Reads Your Screenshots
Moonshot's Kimi K3 is the first open-weights 3-trillion-class model, with 1M-token context and native vision that names CAD. A PAZ hands-on for BIM desks.
Cache-augmented generation: the cheapest token is the one you never re-read
How CAG caches a frozen corpus to cut RAG cost — the O(n²) attention bottleneck, Redis TTL traps, and where a Swiss BIM studio should actually reach for it.
Vector Databases: The Nearest-Neighbour Geometry Under Every RAG Pipeline
Vector-database adoption grew 377% in 2025. Strip the marketing and it is nearest-neighbour search — a Voronoi partition of meaning-space. The maths, explained.
Opus 5 Rebuilt a Part From Pixels Alone — Read the Provenance
Claude Opus 5 reconstructed a machine part from a drawing it couldn't see. What that means for AEC reconstruction — and why every agent output needs a source stamp.
Open weights are a provenance instrument — pin the checkpoint before you ship the drawing
NVIDIA put 650 open models and 250 datasets on Hugging Face. For architecture offices the payload is not speed — it is a checkpoint you can cite.
One Endpoint, Every Model: Routing Your AI Coding Agents Through a Local Proxy
Weave's open-source router picks a model per request using an on-box embedder, not a prompt. Setup, a routing histogram hack, and the trade-offs.
Inkling and the Fine-Tune Question Every AEC Office Will Face in 2026
Thinking Machines' open-weights Inkling (975B/41B active) shifts the AEC question from which model to which data. What an architecture office should do Monday.
