#41

1. I’ve been playing with OpenClaw a bit over the past month. As someone who is still largely non-technical, I find it a bit scary. I don’t know enough to know what I might be doing wrong. Computer use in Claude Code is more my speed, and I’m looking forward to seeing more of that in Codex, or whatever replaces it, from OpenAI too. Macrohard also sounds interesting.
   
   At Toronto Standard, I’d love help managing my inbox, replying to emails, saving attachments to Drive, sharing attachments from Drive, and keeping Drive organized. With today’s capabilities and the right setup, I think that I could save 4 to 6 hours per week while responding to people much faster, which is pretty meaningful. This will be one of my Q2 projects.
   
   More broadly, it’s becoming clear that AI agents will soon be able to do most of what people do on a computer, especially where those tasks are repetitive and supported by lots of training data. That opens up a lot of interesting business opportunities. I’ve written before about my desire for an AI-assisted Figma for architecture. Computer-use agents may make that less necessary in the short term, or at least change what needs to be built, if they can simply use Revit directly.
   
   Here’s another idea. Someone should clone Revit and offer it free to every architect. Given how good AI coding has become, that is a much smaller lift than it would have been even a year or two ago. The tradeoff would be explicit: every action taken in the software is recorded for agent training. How many architects, and clients, accept that deal? I bet the number is a lot higher than zero.

2. On a similar note, I think a lot about how AI progress generally, and computer-use agents specifically, could reduce the cost of government.
   
   Take the Province of Ontario. It spent about $69.9 billion last year on salaries and benefits. I think that 2029-era AI and computer-use agents could cut that in half.
   
   That’s about $35 billion in annual savings that could be redirected to any number of priorities, including deficit reduction and tax cuts. For reference, the Province collected about $27.8 billion in corporate income tax revenue in 2024-25. We could just delete that tax.
   
   If I were Premier, I would already be installing screen-recording software across the public service and collecting training data. By the next mandate, we could have an enormous corpus of real workflows to train on. I would then spend the first 100 days implementing a serious layoff and productivity plan. And deleting the corporate income tax. We could then really be Open for Business.
   
   The City of Toronto, for its part, spent about $3.79 billion last year on salaries, wages, and benefits. ChatGPT’s best guess is that about 60 percent of that is attributable to people who work primarily on computers. If we could again cut that share in half, that would be about $1.13 billion in annual savings, which is well above what the City has ever collected in Development Charges. You get the idea.

3. AI in construction is still my white whale, and will no doubt be for many years. That means both specialized robotics and humanoids.
   
   In my last letter, I did some very high-level math to approximate the amount of data we may need to train humanoids for construction. Modern systems learn from a mix of modalities, including simulation, teleoperation, egocentric video, and robot-generated data. All of these will matter, but egocentric video stands out because it’s the most scalable. Joel Jang from Nvidia has a good piece on this here.
   
   The idea is simple. Workers record first-person video of themselves doing their jobs. That footage can then help train systems on how work is actually done. Labs are already sourcing this kind of data for household and warehouse tasks. Construction is harder.
   
   Construction sites are live, messy, and safety-sensitive environments. You cannot just send people in wearing consumer hardware. Anything worn on-site needs to actually work there, and often needs to be certified PPE.
   
   We are working on something in this space, and I’d love to chat with robotics researchers and interested construction managers.

4. I was at Toronto East York Community Council this morning, where our Zoning Bylaw Amendment application for our midrise project at 1423 Dufferin St. was approved. Next is a City Council vote, after which the site will be zoned.
   
   In the meantime, we’re getting started on Site Plan Control, where the finer points of the design will be proposed, reviewed, and ultimately approved by City staff. Our hope is to get through to Site Plan Approval in time to submit building permit applications and start construction in Q4.
   
   This project is especially exciting for our team because it is our largest yet, and because it is very well located near Geary St., which we like to call the next Ossington.
   
   We’d like to do more projects like this. I think that 40 to 60 units is a sweet spot, especially in neighbourhoods where there is strong demand for nice apartments without parking. Toronto could use a lot more of these.

5. We are now just over a month away from the 2026 Missing Middle Summit. Long-time readers will know that this is the Summit’s fourth year, and the first at SixtyEight, a very cool event space on the sixty-eighth floor of Scotia Plaza.
   
   One nice thing about getting to the fourth year of an annual event is that marketing gets easier. In 2024, at this point, we had sold 33 tickets. In 2025, 75. This year, 76. And I think that I’m doing less regular promotion on social media each year.
   
   I’ve hosted many one-off events over the years, and am only now really learning the value of building a recurring series and improving it each time. We should all do more things that compound.
   
   In any case, this year really will be the best yet. Best venue. Best speakers. Best food. Best crowd. If you haven’t already, get your ticket at missingmiddlesummit.com.