Move 37

In March 2016, an AI bot called AlphaGo (by DeepMind) played against Lee Sedol, one of the greatest players of Go, an ancient board game far more complex than chess. Go is deceptively simple: players take turns placing black and white stones on a 19×19 grid, trying to surround territory. But the number of possible moves is astronomical. For centuries, Go was considered impossible for computers to master because it requires intuition and creativity, not just calculation.

AlphaGo was doing well but still struggling when it made Move 37 in the second game. This move was different. It paused for ten to fifteen minutes, much longer than usual. When it finally placed its stone, experts were baffled. The move seemed wrong, appearing in only 1 out of every 10,000 games historically.

But that extra thinking time revealed something unexpected. The move was brilliant, and Go experts later recognized it as genuinely creative. Not just optimal calculation, but the kind of intuitive play they thought only humans could achieve. It changed how people thought about what AI could do. That extra time it took to think was what made the difference — it allowed the algorithm to explore, validate, re-think and finally derive an optimal solution.

The quick answer problem

Most AI interactions work like this: you ask a question, AI responds in seconds with a single pass through its knowledge. It’s like asking someone for directions without giving them time to check a map, you’ll get their best guess, but not necessarily the optimal route.

This single-call approach has inherent limitations. AI can only work with what immediately comes to mind, can’t verify its assumptions, and can’t explore multiple approaches. For simple queries, this is fine. But for complex, mission-critical questions where the answer is buried in thousands of pages, quick isn’t best.

What makes AI take longer (and why that’s good)

When AI takes time to process a request, it’s working harder to do things right. While there are many approaches it could take, the main ones typically include:

• Multiple data retrieval passes. Instead of relying on initial results, AI fetches relevant information multiple times, refining its search based on what it learned. It pulls from document repositories, specifications, historical project data, and regulatory requirements, then cross-references everything.
• Relevancy validation. AI evaluates whether each piece actually answers your question. Is this specification from the current revision? Does this detail apply to your specific building zone? Is this the right trade’s scope?
• Context interpretation. Long-running tasks analyze who you are, what project you’re working on, and what you actually mean versus what you literally asked. “What’s the fire rating on the third floor?” means different things for a structural engineer versus a finishing contractor.
• Answer validation and refinement. AI compares its draft answer against your original question, checks for contradictions, and rewrites for clarity. It’s the difference between “the fire rating is Type II” and “the third-floor corridor walls require 1-hour fire-rated assemblies per Detail 7.3 on Sheet A-401, using 5/8” Type X gypsum board over 3–5/8" metal studs.”
• Tool usage and integrations. Modern AI can interact with multiple platforms, pulling data from project management software, searching specification databases, accessing BIM models, and fetching current code requirements from the web. Each interaction takes time but dramatically improves accuracy.

Working in the background

These long-running tasks don’t require you to stare at a loading screen. They operate in the background while you do what actually matters, value engineering, attending coordination meetings, solving problems on site and more.

Think of it like sending someone to research a complex question. You don’t stand over their shoulder. You let them dig through documents and compile findings. AI works the same way, except it processes thousands of pages in the time it takes a person to read one.

Automatic triggers: Answers before you need them

Even better, imagine setting up automatic triggers. When a new submittal package arrives, AI can immediately start analyzing it against project specifications. When RFIs come in, AI can begin researching similar precedents and relevant contract language. By the time you sit down to review, the deep research is already done and the answers are waiting for you instead of you waiting for the answers.

Seeing the plan first

One (legit) concern about long-running AI tasks is losing control, what if it goes down the wrong path?

The solution: use tools that allow you to review the plan before it’s executed. Such tools use AI to first plan the task thoroughly, presenting their steps and strategy: “I’ll search these document sets, cross-reference these specifications, validate against these codes, and structure the answer this way.”

More advanced tools even allow you to intervene as an expert, edit and refine the workflow, keeping you in control while having AI do the heavy lifting and executing your joint plan.

The construction connection

This matters enormously in construction. Document sets aren’t blog posts, they’re comprehensive technical packages running thousands of pages. A mid-sized commercial project might include hundreds of drawings, a spec book as thick as a phone directory, and thousands of submittals, change orders, RFIs, and meeting minutes.

The details buried in those documents are mission-critical. The wrong fire rating, a missed structural connection, or an overlooked code requirement can mean rework, delays, failed inspections, or safety issues. These aren’t fluffy questions where “close enough” works.

Quick AI responses can’t dig through that volume with the required thoroughness. But AI that takes the time, that makes multiple passes, validates relevancy, cross-references specifications, and checks its work, can find those buried details with precision that would take a human days to match.

The trade-off worth making

Lee Sedol later said that Move 37 showed him something beautiful, a move no human had thought of in thousands of years of playing Go. It took AlphaGo’s willingness to pause and explore deeply to find it.

Your construction questions & outputs deserve the same depth. When the answer is buried in thousands of pages and the stakes are high, speed matters less than accuracy. Long-running AI tasks give you both: the thoroughness of deep research with the speed only machines can achieve.

The next time AI takes time to respond, remember Move 37. Sometimes the wait isn’t a bug, it’s the breakthrough.

‍