Man Vs Machine

Millions of repetitive clicks leading inevitably to carpal tunnel. Millions of taps on the spacebar to repeat a command. Left click execute, repeat until what is in my head is on the screen. Initially, in your early 20s your mind needs the time to think through the order of operations to draw a building. Mid-20s, after a few thousand hours of mind-numbing arrays, fillets, copy-rotate, mirror, and offsets, my brain is successfully faster than my hands.

I'll never forget my first site plan at my first job in 2015. A variation of it is on the corner of Main and Jamboree in Irvine, California. My boss at the time laid out some flair tip pen lines on trace paper. His 25 years of experience in solving site plans clearly on display: in eight strokes, here's half the building mass. Four more, there's the parking garage. Eight more, an open courtyard. Liner units here. Weird condition on the back. Pool goes here, elevators there. He didn't need a scale. Didn't need a computer. He handed me the trace paper, and I left on my merry way to validate his assumptions. Count them: eight, four, eight. Twenty strokes for the big moves. What he handed me was intent, not a checked plan.

That first "validate his assumptions" assignment, turning those strokes into a drafted plan that proved whether they actually worked, took me two weeks to do, granted it was my first time doing it. I thought to myself, never again will this take me two weeks. I built dynamic blocks with labels for my units. That eliminated annotations, kept my tabulations accurate, but at the end of two years of refining the process, I couldn't get my time below one work day to do a site plan, 3D massing model and tabulation. Notice what both numbers time: one person drafting and checking one scheme. Two weeks was a beginner's calendar time; one day was my floor with my own tools. The final workflow I landed on before working on TestFit full time? Excel. Just model the whole thing in Excel first (to scale — 1 pixel = 1 foot). Then draw the plan to match. Commodity buildings are math problems inside a spatial container.

I became John Henry from folklore: the hammer strikes the same spot with the force of a locomotive, defeating the steam shovel. For one project in Arizona I did nearly 20 different iterations to solve the deal, and that project is built today. (It's The Cameron in Tempe. In a 2024 LinkedIn post I finally did the math: 502 stalls in the garage and maybe 100 on the surface, so about 600 stalls per iteration. Twenty iterations, 12,000 stalls drawn, one real building.) Iterations matter. Coordinated iterations with the finance and construction team mattered significantly more.

Ideas to get others to solve site plans like math problems fell on deaf ears. Excel isn't a design tool, they said. Just use trace, it's faster. (It isn't.)

This early in my career, I was already contemplating what it would take to unseat my ability to draw site plans — what would be a better solution to all of these clicks, all of this noise, on the 95% or so of my work that would simply be thrown out. The idea to build a better machine isn't new. The idea to solve a quality site plan in less than 30 milliseconds, while providing all of the math, so I can focus on building massing? That's new. Be precise about that number, though: it is machine time for one solve. It does not include anyone typing in the site, or anyone deciding whether the answer is any good.

The Impetus To Write This Down

On April 2, 2020, I sat down and decided to draw a site plan the traditional way, the way that most designers draw site plans: AutoCAD. But that wouldn't be very scientific. I invited guests to join to simulate a typical office environment, but to add pressure, we livestreamed it.

For those of us in the soft-skilled AEC community we spend thousands of hours on {click, spacebar, rotate, array, mirror, array, select all, copy, rotate} to get our work done. The machine wins on repetition. The human wins on intent.

Four Clocks, Four Different Jobs

Each time figure above measures a different job. The man-versus-machine debate gets dumb when they're compared as if they were the same one:

• Sketching intent. My boss's twenty strokes on trace. Fast, and loaded with 25 years of judgment, but no scale and no math. In a 2018 LinkedIn comment I described a design "that you sketch on trace paper in 5 minutes, but it isn't a bim model, it has no data, and its static."
• Drafting and checking one option by hand. Two weeks the first time; a floor of one work day after two years of dynamic blocks and Excel, for a site plan, a 3D massing model and a tabulation. This is the clock AutoCAD runs on.
• Generating one option in software. Milliseconds of compute. In that same comment I explained why a site submitted to us took 3 to 36 hours to come back when the solve took milliseconds: the window was so I had time to get to it, and the work inside it was me entering the site by hand and adjusting the parameters. Setup and queue are human clocks even when the solve isn't.
• Reviewing and coordinating iterations. Deciding which of twenty options the finance and construction team can live with. Nothing in this record times it, and it's the one I said mattered most.

So compare per iteration, not per drawing. Had each of those twenty Arizona iterations cost my one-day floor, that's twenty working days (four work weeks) of drafting for one building. Twenty solves at 30 milliseconds each is 0.6 seconds of compute. The machine doesn't erase the work. It moves the work from drafting to input and review, which is where the judgment was supposed to go in the first place.

What the Livestream Record Shows

The stream was announced on LinkedIn as "Man Vs Machine: TestFit Live Stream," scheduled for April 2 from 10:30 to 2:00 Central. That three-and-a-half-hour window is a broadcast slot, not a drafting time. My comment on the post that afternoon: "It was great to look at general purpose software vs a point solution. I think we will spend some time thinking about this and make a solid write up." By that description the matchup was AutoCAD, a general-purpose drawing tool, against TestFit, a point solution built for this one job.

This essay, published twelve days after the stream, is that write-up, and it stops before the scoreboard. The recording, the plan drawn on stream, the AutoCAD time and the TestFit output it was compared against are not in the record, so this page claims no winning time.

Here's where I land. The twenty strokes on trace stay human; that's intent, and no solver handed it to my boss. The two weeks, and later the one day, between those strokes and a checked tabulation should not stay human. Once the drafting clock drops out, the scarce clock is review. My bet: the teams that win are the ones that can judge twenty options as fast as the machine can draw them.