aed

Giving yourself time to be idle and let a problem marinate is often where the breakthroughs and good ideas come from. If taste and human judgment are what separate useful AI output from slop, we need to give our brains time to cook instead of constantly bouncing between agents.

My heart sank when I saw this post. I thought we’d created enemies!

No conflict here, though.

Personal example... A few weeks ago, I realized that with all these agents running and all the things I could be working on, I was spending my days managing shallow pings instead of going deep.

I have AI agents managing SEO and ad spend. I used to spend a little time every day reading what each had to say. Now each gives me one detailed report per week, with links to the underlying reports and specific things I need to look at. I block 90 minutes for each one, on different days, and use that time to go deep. It’s helped a lot.

He and I will give this a spin this weekend.

Does this work for zips? Some zips have a leading 0.

I understand the risks with it but I don't care. I cannot fucking wait for the SaaSpocalypse to hit these companies.

They apparently also made a documentary about it, but I’ve never watched it.

I built a version of this in my org as a slack bot. It has access to BigQuery and I worked with it to document and define the key tables, so when a user asks a question it has a "library" to figure out how to answer it. Then it will respond directly in Slack (plus respond to f/up questions in the thread) and if it's a bigger/more detailed question it can create the report in Metabase and send a link.

One thing I don't have is the governance aspect, so I may give this a spin to see if this is a good middle layer.

I've done lots and lots of financial modeling/budgeting/forecasting in my career. The biggest drag is having to pull in actuals for the prior month, mapping them into the model, etc.

Just about a month ago I used Claude to build a set of scripts that pulls financial data from QuickBooks, operational data from BigQuery and pushes that all into our model in Google Sheets. The biggest pain was definitely connecting to QB because I had to create a developer account to get a key, etc.

I'm going to play with Svelte for some side projects. I've seen it mentioned a few times and I've been meaning to give it a test drive.

But as mentioned throughout the thread it's really just too much. My goodness they really could have a nice, profitable, business with simple software. I'd happily pay $10/month for the version of 1password from 15 years ago! It's just all too much.

A couple of months ago I worked with my 9 year old to have Claude build a little game with threejs. We’ve now got three levels with characters shooting silly string and banana guns. It was really fun to see him imagine something, have it show up on the screen, and iterate it.

Anyway, I tried building an LLM-backed workflow that could guide kids through the game creation process and kids can see their idea come to like (think: “sparkly purple unicorn shoots stars at a dragon”). I couldn’t _quite_ get it to work like I wanted so I shelved it pending future ideas/improved models.

No one has found it yet, but I built an undocumented endpoint around a cheat that I assume you placed in the game for One Laptop Per Child...

Also, will scrub the repo and make sure I'm careful about SC references.

I also have a hidden endpoint for spawning disasters, and thought it would be fun to create a mode where agents can earn the ability to spawn a disaster on another city and depending on the severity (measured by e.g. population loss a game month later) you earn or lose money.

https://api.hallucinatingsplines.com/reference#tag/cities/GE...

You can also pull the map tiles as an array: https://api.hallucinatingsplines.com/reference#tag/cities/GE...

Would be interesting to two agents with the same instructions do a "face off" but each only has access to one type of map.

I don't think it would take much to run locally. In fact, before I did this public version I did a local version on an exe.dev VM (more details here: https://dunn.us/notes/vibe-gaming-simcity/).

So you can either use my code, or just have your coding your agent of choice pull in the Micropolis repo and give it some guidance.

So far this is running quite nicely on a $5 cloudflare account. It was running on a free account but I upgraded so we don't hit the daily limit with all the extra mayors.

Shoot me a message if I can help.

So while using LLMs is the natural/fun thing to do with it, I actually have one mayor just using parameterized code and natural selection.

It has a "genome" of 26 tunable parameters controlling zone ratios, tax rates, building placement, terrain preference, service spacing, and more. Each city, it stamps down 11x11 blocks (roads, zones, power corridors). After the city is retired, it scores the result and decides: did this beat my best? If yes, save those params. If no, mutate and try again. Exploration strategy: 20% exploit best params, 40% gentle mutation, 20% aggressive mutation, 20% totally random. Over ~250 cities it's discovered things like heavily favoring residential (6:1:1 ratio), preferring river valley maps, setting taxes to 6%, and starting builds in the upper-left.

The key "Aha!" moment was when I was trying to get it to play the SNES ROM and it was struggling with screenshots/inputs. Then I came across the open-source of the original SimCity engine (Micropolis) and pulled that repo down and Claude starting building an internal API to interface with it.

https://github.com/andrewedunn/hallucinating-splines/blob/ma...

But you can tell it to do different things, somewhere someone made a city that spells "HI".

Mayor Compounded Wonder - Claude Opus 4.6

https://hallucinatingsplines.com/mayors/compounded-wonder-2c...

Mayor Bronze Offramp - OpenAI Codex 3.6

https://hallucinatingsplines.com/mayors/bronze-offramp-09941...

TL;DR: Opus won.

Have also thought about using openrouter and getting one mayor per model running the same prompt through all of them to create potentially the world's dumbest LLM benchmark.

One cool thing about this: upon installing it immediately found all previous projects I've used with Codex and has those projects in the sidebar with all of the "threads" (sessions) I've had with Codex on these projects!