←
back to michaelgruen.com

Now 🗞

August 24, 2025

Summer, usually slow, has been anything but.

In early June, Althea and I traveled to Europe (London and Portugal). I took 1,200 photos but only managed to edit one of them. It’s been a low-output summer for me.

On the input side, well…

Between work and tinkering, I have gone deep on what’s popularly called “Agentic AI”, especially on its use in software engineering. I have about 3 half-baked dispatches I’m trying to finish While generative AI and LLMs could be helpful here, I refuse to use them here except to proof-read my missives. Future LLMs will need someone who can still write without spellcheck and autocomplete. (Maybe.) with one I’m targeted for Labor Day that goes into greater detail.

In the meantime, some general observations:

  • Talented software engineers are getting really good at orchestrating fleets of coding agents. Watch this space.

  • Properietary foundational LLMs (e.g., Claude Sonnet 4, ChatGPT 5, et al) yield less marginal utility every release. While I have no insider knowledge here, I’d be shocked if Anthropic wasn’t training a foundational software engineering model that eschews most general knowledge in favor of high-quality programming knowledge. When that’s out (or someone else beats them to it) everything we do today will look extremely hacky and computationally wasteful by comparison.

  • Speed to delivery is increasing, but not nearly at the rate as advertised by talking heads and AI companies (who are in the business of selling you on their future). In essence: computer code exists because it’s the most efficient way for humans to express complex, specific logic that a machine can run reliably. Human languages, like English, are exceedingly long-winded and imprecise… and while AI-enabled tools and agents can make software development more accessible to more people, it does not—and can not—reduce the burden of having to express oneself with precision or novelty. It turns out figuring out what to build against what works continues to be difficult. Dave Farley nails it here.

  • Similarly, live experimentation is now cheaper. This is a good thing.

  • Consequently, “AI Slop” is now everywhere. This is NOT a good thing. Our lizard brains are unable to parse the deluge. (Put in LLM terms, our own brains’ context windows cannot keep up, and reading summaries from a compact command is a chore. Our brains like to think about stuff and get into the flow, not swivel-chair all day.)

Also, I may have fibbed about the low-output summer.

In truth, I’ve been building a few software tools—some solo, some with collaborators—that should be ready in the fall.

Here’s one project that, except to untangle some design and UX mistakes, I built without writing a single line of code by hand: HANEWA. My codegen process has garnered some interest from colleagues, so a write-up will follow.

The Usuals

Upcoming Travel

  • Hudson Valley, NY (although if it’s weekend trip under 2 hours away from home, is it really considered travel?)

Vocational Notes

  • Fractional continues to suit; I have about ~5hours/week capacity remaining. Current foci:
    1. Environmental Impact and Sustainability
    2. AI-enabled Workflows
    3. Anti-social Networking and Mass Customization
  • Team, Assemble! still needs a product/technical lead. (Maybe this is you?)
  • Like many founder-types, the idea list is becoming longer than the time-to-execute budget. Working on a solve for this that trades money for execution time and will list opportunities there. (Domain names have been purchased, this relates to number 3.)
  • Looking to support another non-profit following Literacy Trust’s wind down. Currently thinking about what that should look like and actively soliciting ideas.

Avocational Notes

  • Lifting 3-4 times per week. It has been a cardio summer. Lifting resumes this week.
  • Electric Go-Karting is still the best. If anyone is into motorsports and lives in Manhattan, do reach out…
  • Abandoning this small project until hands-on coding work settles down. (Even with generative AI, I can only stomach coding so many hours in one day.)

Content Consumption (Abbreviated)

Television/Movies

Books

  • Babel: An Arcane History - R.F. Kuang. I’m halfway through, but this is such a compelling read that even if it falls apart in the back half I’ll be still glad I found it. H/T Island Bound.

Games

  • Chess
    • Elo ~1100 (Daily). Write me if you want my chess.com username.
    • I’m steadily getting better worse.
  • Dorf Romantik - I cannot play this before bedtime as everything starts to look like a hexagon that need to be properly places (like the Tetris/tetromino phenomenon) and my legs get restless.
  • Blippo+ - The Play Date is such a gem.

Sundry Notes

Earlier this summer I learned of a city named Utrecht. Notably, the road networks were designed so that it’s easier and faster to walk and bike than it is to drive.

As a New Yorker and cycling advocate, it’s proof of what’s possible when you build for micromobility. Not only because it’s clear how much more agency and freedom of movement children have (unburdened by having to always ask for a ride from an adult), but of how much nicer it is to live in a place where car is not the default.

People are generally supportive.

Alas, invariably someone will proffer:

“But these delivery guys / Citibike bros on their e-Bikes. They go too fast and reckless… one nearly killed me the other day!”

Unfailingly, similar anecdotes become the conversation in the way that shark attacks make for better headlines than heart disease. In the US, sharks kill ~1 person each year. Heart disease? ~1,000,000.

The thing is: in the US, more people drive cars than ride bikes. And because driving is more familiar, people are more willing to forgive drivers—and tend to forget any incidents—where they were “nearly killed”, which affects how they think about micromobility policies at their own safety and our collective peril.

A friend calls this “car brain” because when you look at the data, it’s not even close.

This chart shows what happens to pedestrians in NYC vehicular collisions since 2012 (when the city started tracking this stuff in earnest). It does not include the fates of any persons involved except pedestrians. It’s not that I don’t care about what happens to everyone else, but the charge is often levied at the dangers riders pose to pedestrians, not the other way around.

Vehicle Type Pds. Injured Pds. Killed Total Fatality Rate % of All Casualties
Large/Commercial Vehicles 3,974 254 4,228 6.0% 3.6%
Other/Unknown 2,218 58 2,276 2.6% 1.9%
Van 2,156 40 2,195 1.8% 1.9%
Motorcycles 1,345 20 1,365 1.4% 1.2%
Passenger Vehicles 95,082 1,170 96,252 1.2% 82.1%
Taxi/Livery 7,056 55 7,111 0.8% 6.1%
Bicycles/Scooters 3,826 25 3,850 0.6% 3.3%
All Vehicles 115,657 1,621 117,278 1.38% 100.0%

Bicycle ridership has gone up 60% over the last 5 years. Collisions per bicycle mile involving pedestrians have overall declined, down by nearly half since a high in 2022.

Are e-bikes more dangerous to pedestrians than pedal bikes? Absolutely:

Bike Type Total Pds. Injured Pds. Killed Fatality Rate
Traditional Bicycle 1,421 1,414 7 0.5%
E-Bike 943 932 11 1.2%
E-Scooter 562 558 4 0.7%
All Bicycles 2,926 2,904 22 0.8%

…but they still do not harm pedestrians as much as traditional bikes, or—oh, right!—cars do… which account for 90% of pedestrian casualties and 50× the fatalities.

In short, the shark discussion about e-bikes detracts from the larger heart-disease issue: we need to invest in better public transportation and reduce the amount of large, vehicular traffic into the city.

If you don’t believe my analysis, here’s another take.

Hat tip to nownownow.com.

• April 28, 2025 →