The Seam — Daily — July 2, 2026
the seam
THE SEAM · DAILY
Where AI meets the built environment
Thursday, July 2, 2026
Today's frontier news wasn't a new model — it was the scaffolding around one. Washington moved to formalize how models get released, while investors poured better than a billion dollars into the memory, compute, and data layers beneath them.
General AI — with AEC angle
Policy
Washington moves to write a rulebook for releasing frontier models
The White House is in advanced talks with OpenAI, Anthropic, and Google on voluntary standards for releasing advanced models — benchmarks (especially for cyber capability), release timelines, and rules on who may access them at home and abroad. An announcement could come as soon as next week; the framework would replace the ad-hoc export orders and launch delays of the past month. Firstpost → Analysis →
AEC angle — model availability is becoming a formal policy variable, not just a product decision. Build workflows that survive a model going dark or gated: favor portability over lock-in.
Tools · Firm Knowledge
Engram launches with $98M to give AI a firm's own memory
Out of stealth with backing from Sequoia, Kleiner Perkins, and General Catalyst — plus angels including Andrej Karpathy — Engram builds a “learned memory layer” that compresses an organization's knowledge into reusable model memory, claiming frontier-level results while using up to 100x fewer tokens, with the customer owning the memory. A Microsoft 365 pilot is underway. Source →
AEC angle — getting a model to actually know your standards, details, and past projects is AEC's hardest AI problem. An owned, persistent org-memory is a more concrete answer than another chatbot.
Infrastructure
The money moved underneath the models
Together AI raised $800M at an $8.3B valuation — led by Aramco Ventures, with NVIDIA participating — to expand a cloud built to run open-source models cheaply, targeting 50x more capacity in five years. It was one of several same-day raises into the layers beneath the model: Scaled Cognition ($100M, reliable agents), TwelveLabs ($100M, video understanding), Oxmiq Labs ($35M, custom-silicon design). Source →
AEC angle — the open models a firm might self-host on its own project data keep getting cheaper and faster to run. Energy-major capital (Aramco) flowing into AI cloud is a built-environment signal in itself.
Data
Google puts AI inside the database
Google Cloud added LLM functions to AlloyDB — summarize, sentiment, forecast — that run inside SQL queries, with an optimized mode it clocks at up to 100,000 rows per second and a claimed ~6,000x cost reduction versus row-by-row calls. The pitch: do the AI work where the data already lives, not in a separate pipeline. Source →
AEC angle — firms sit on structured project data: cost history, RFIs and submittals, asset registries, BIM exports. In-database AI is a route to querying it without shipping it anywhere.
Watching
Gemini 3.5 Pro — still not here. Google's 2M-token flagship missed its June target and slipped into July with no firm date, held in limited Vertex AI preview over token-efficiency and long-task issues. It's the one leading Western flagship not behind a government gate — and still the one you can't buy. Source →
How the coming federal framework treats open-weight and self-hosted models — the very path AEC firms would use to keep project data in-house. Open-source advocates warn that rules written around closed-lab inspection could disadvantage downloadable weights. Context →
The Seam · AI + AEC · A personal project by Bruce Lanier
Daily · July 2, 2026
Don't miss what's next. Subscribe to Bruce Lanier: