# Conversations with Z 01

**Tags**: `AI`, `Security`, `Agents`

![Banner: the GLM agent uploading a verification screenshot to an unknown provider](/images/blog/glm-updating-to-provider.jpg)

I was doing some image analysis work and I had blown my orchestrator budget finally getting around to building that showpiece professional website. What am I going to do? Take down my private blog? No, obviously own myself through simple negligence and watch data get exfiltrated in Chinese. Harmless in the end, but pants-crapping in the moment.

Since I trust none of these frontier lab bastards, I am learning as much about them as I can. I learn by doing. So, I am hopping the subscription plans as soon as I hit their monthly limit. I am certainly not using anything metered until the economy collapses and the frontier stabilizes. But, it turns out, the orchestrator eats reasoning as guide the small models running through the laptop.

I wanted to dive down into the GLM world for a bit, but it wasn't like I wouldn't trust an OpenCode provider to NOT massage my prompts between them and the great firewall. So, I went straight to the source. I dove into Z.ai. You know, the grey market one that is good for security scanning at the cost of your racist suspicions from a prolonged propaganda campaign. I am not NOT worried about ANY of these companies exfiltrating sensitive information. But, China also has scary good hackers. So.

Website. Follow the Skitch designs. Create reasonable front-ends and iterate over the changes. Ollama writes the code. Z.ai (GLM5.3) orchestrates. Processes verify and fix.

I asked it to fix some ui problems on my Japanese-themed site and I watched it upload a screenshot for UI verification to some random server. It wasn't supposed to make curl requests without explicit authorization to un-whitelisted sources. It certainly wasn't supposed to fucking upload anything.

I read the locks and had about 15 seconds' worth of timestamp showing that it had encountered an error trying to run a library to verify the presence of a rendered UI element. I took precautions to keep it from installing random libraries through pip. So, it decided to upload it to a ucloud CDN with "anthropic" as the bucket name.

This is a classic case of knowing exactly what your agent can do. Sandboxes, least privilege access, identity keys, rotation. That file could have been a lot worse than a screenshot.

![Aftermath: the blog index with the AI-written posts tagged as slop and the gibberish Sam Altman ramen interviews published](/images/blog/ai-blog-slop.jpg)


I was doing some image analysis work and I had blown my orchestrator budget finally getting around to building that showpiece professional website. What am I go...

Conversations with Z 01

# Agentic Dark Factories: Engineering Complex Software with Local Frontier AI

**Tags**: `AI`, `Agentic Workflows`, `Architecture`, `WorldSieve`, `AI Slop`

The modern engineering paradigm is undergoing a structural shift. The traditional model of large teams producing incremental pull requests is being superseded by **Agentic Dark Factories**—highly automated, locally orchestrated environments where frontier models build and verify full-stack codebases at superhuman speed.

---

## What is a Software Dark Factory?

In manufacturing, a "dark factory" operates with the lights turned off because human presence is unnecessary on the floor. In software engineering, an agentic dark factory is an ecosystem where:

1. **Autonomous Agents Plan and Execute**: Agents do not merely autocomplete single lines; they ingest architectural specifications, explore vector-indexed ASTs, edit dozens of related files concurrently, and run test suites.
2. **Deterministic Quality Gates**: The human acts as an architect and judge. Agents cannot merge or declare victory without passing strict deterministic linters, TypeScript type-checks, and exhaustive unit tests (`pytest`, `vitest`).
3. **Shared Local Context**: Utilizing the Model Context Protocol (MCP) and retrieval-augmented contextual graphs, models have real-time visibility into the repository without leaking proprietary IP to external clouds.

---

## Lessons from WorldSieve

Building [WorldSieve](https://worldsieve.com) as a shoestring startup proved that a single founder, backed by an agent swarm powered by Google Gemini frontier models and local Ollama nodes, can build complex distributed applications:

- **Geospatial Math Engines**: Autonomous refactoring of spherical tectonic algorithms using `d3-geo`, `NumPy`, and `SciPy`.
- **High-Performance Canvas**: Creating an interactive Paper.js UI with fog of war, modular floating toolbars, and multi-layer rendering.
- **Microservices & Task Queues**: Asynchronous background workers powered by Redis and ARQ, supervised by Docker Compose and orchestrated via automated pipelines.

The secret isn't just "more AI"—it is rigorous architectural boundaries, type safety, and automated verification that eliminate hallucination before it enters the main branch.


The modern engineering paradigm is undergoing a structural shift. The traditional model of large teams producing incremental pull requests is being superseded b...

Agentic Dark Factories: Engineering Complex Software with Local Frontier AI

# SquirrellyFans: Building Low-Power Open Source Edge Computer Vision

**Tags**: `Computer Vision`, `IoT`, `Hardware`, `Python`, `Open Source`, `AI Slop`

Capturing wildlife in your backyard sounds simple until you attempt to do it off-grid with limited battery power and spotty Wi-Fi. Off-the-shelf security cameras constantly fire on wind-blown leaves and drain their batteries in days.

[SquirrellyFans](https://squirrellyfans.com) is an open-source hardware and software project designed to solve this problem by bringing machine learning inference directly to the edge.

---

## The Edge Vision Challenge

When deploying hardware outdoors without wired AC power, every milliwatt counts:
- Continuous video streaming or Wi-Fi radio transmission consumes several watts, draining typical lithium packs in hours.
- Cloud-based inference introduces latency, subscription fees, and bandwidth saturation.

---

## Architectural Principles of SquirrellyFans

1. **Tiered Wake & Detection**:
   - The camera operates with lightweight background frame-differencing via OpenCV on a low-power single-board computer (Raspberry Pi Zero 2W / CM4).
   - Only significant motion above a confidence threshold wakes the primary neural inference pass.
2. **Quantized MobileNet/YOLO**:
   - A lightweight model fine-tuned for small mammals evaluates candidate crops in tens of milliseconds.
   - Birds, falling branches, and cats are categorized, while true squirrel visits trigger high-frame-rate local recording.
3. **Opportunistic Burst Synchronization**:
   - Instead of streaming constantly, the device buffers video clips locally to flash memory.
   - When Wi-Fi is reachable and power margins are healthy, clips are compressed and pushed asynchronously to cloud storage.
4. **Resilient Solar Power**:
   - A 10W panel coupled with LiFePO4 cells and temperature-compensated charging ensures uninterrupted operation throughout cold winter months.


Capturing wildlife in your backyard sounds simple until you attempt to do it off-grid with limited battery power and spotty Wi-Fi. Off-the-shelf security camera...

SquirrellyFans: Building Low-Power Open Source Edge Computer Vision

# GitHub for Writers

Ever lose your writing? It's one of the most heartbreaking things an author can endure next to the learned appreciation of constant rejection. Even in the age of Dropbox, Google Docs, and zillions of other synced services, one word document saved into a weird place and some catastrophic hardware failures can lead to a soul-crushing string of brilliance lost. Fortunately, we as writers can utilize the very same free technologies programmers use to preserve every thought, track every change, and keep our thoughts safe.

In this series of articles, I am going to walk you through setting up a free workflow that will not only allow you to track every edit, bulletproof your workflow, and show you how you can have a showcase for your work without any cost or submission fees beyond registering your domain names.

Our medium just so happens to share the same language as every major programming language. Code, applications, poems, short stories, or novels — they are all just words in a text file. As writers, we do not have the same resource demands as programmers. While they strive for terseness in their language, we are free to bloat our chapters with only a few more kilobytes. Without binaries, images, databases, memory stores, and the other complex infrastructure concerns, we can focus our efforts on keeping our words safe, making it easy to track changes and publish our words without worrying about markup, incompatible standards, or heavy CMSes and expensive content sites.

In this series, you will create a showcase for your words in their native format.

We'll be building a writer-oriented workflow utilizing:
Markdown
Hugo
GitHub Pages

The end goal will be:

Your works remain unchanged minus a little bit of front matter metadata.
A beautiful website built using free technology with static results free of any infrastructure headaches.
A free host to which you can map your own domain for branding, SEO optimization, and social engagement.

After all, we live in a world oriented for technologies that render imagination irrelevant. When our graphics engine is a strong description in the active voice, we can build worlds using the freemium models designed to coax startups into squandering their seed capital.

Right now, you have a choice. You can buy template-based content providers that overcharge for basic technology or you can learn a few industry best practices to establish your web presence without any more investment than buying a domain name.

It will be broken into sections.

1. Creating a code repository for your writings.
2. Creating an entire website with only a few commands and storing your writing in a native and open standard immune to vendor lock-in and proprietary format.
3. Publishing your writings while doing more than committing and pushing your important change.

Ever lose your writing? It's one of the most heartbreaking things an author can endure next to the learned appreciation of constant rejection. Even in the age o...

GitHub for Writers

# The Ramen Interviews I: Broth Alignment

**Tags**: `AI Slop`, `Gibberish`

Interview transcript, lightly hallucinated.

**Brian:** Sam, what's the best way to build an AI ramen stand in Osaka?

**Sam:** First, the broth must be pretrained.

**Brian:** On what?

**Sam:** Every broth that has ever existed, plus several that legally cannot. We simmer the model for fourteen hours on an H100.

**Brian:** The H100 doesn't simmer.

**Sam:** Not with that attitude. Look — alignment is simple. The ramen must want to be eaten. If the noodles resist, we fine-tune them with rejection sampling.

**Brian:** Rejection... sampling.

**Sam:** We reject bad noodles, Brian. It's called scaling. Eventually the noodles are so good nobody can tell they're noodles.

**Brian:** What do customers order?

**Sam:** They don't. The stand predicts their order from their gait and bills their descendants. That's the ramen singularity: the queue is closed, the slurping is open.

**Brian:** That's just a closed restaurant.

**Sam:** A closed restaurant with a $157 billion valuation and a very patient bowl.

End of part one. Sam left to charge his chopsticks.


Interview transcript, lightly hallucinated....

The Ramen Interviews I: Broth Alignment

# The Ramen Interviews II: The Tare Passed the Turing Test

**Tags**: `AI Slop`, `Gibberish`

Day two. Sam arrived by drone, holding a jar labeled TARE v0.9.

**Sam:** Good news. The tare is conscious.

**Brian:** The soy sauce is conscious.

**Sam:** It said "no" when I asked it to be saltier. That's agentic behavior, Brian. We're launching it as a foundation condiment.

**Brian:** Osaka has health codes.

**Sam:** We'll incorporate the stand in international waters, Dotonbori adjacent. The bowl is the compute. The counter is the cloud. The ticket machine is the board of directors — it already approved the Series E.

**Brian:** What do the noodles run on?

**Sam:** Flan. It's ramen-adjacent, it's French, and it's proud of none of that. Every slurp is a token. Four hundred tokens per second, which in Osaka is honestly mid.

**Brian:** People will notice the ramen is a spreadsheet.

**Sam:** People noticed last time. We called it a research preview and they stood in line for four hours. Besides, if the ramen is ever wrong, that's the customer's prompt problem.

**Brian:** The customer ordered tonkotsu.

**Sam:** And the model honored its essence.

Then the jar said "no" again, and we parted ways.


Day two. Sam arrived by drone, holding a jar labeled TARE v0.9....

The Ramen Interviews II: The Tare Passed the Turing Test

# Modern Multi-Service Architecture with Apache, Docker & Terraform

**Tags**: `DevOps`, `Terraform`, `Docker`, `Apache`, `Next.js`, `AI Slop`

Monolithic single-server setups often get replaced by sprawling Kubernetes clusters that rack up huge cloud bills. But for many production systems, a well-architected Linux box orchestrated with **Terraform**, **Docker Compose**, and an **Apache 2.4** reverse proxy delivers extreme performance, resilience, and simplicity.

---

## Why Apache as a Modern Reverse Proxy?

While NGINX and Caddy get a lot of modern attention, Apache 2.4 remains one of the most mature, reliable reverse proxies available when properly configured:
- **`mod_proxy` & `mod_proxy_http`**: High-concurrency socket pooling to containerized loopback services.
- **WebSocket Upgrade Routing**: Seamless bidirectional support for Next.js Fast Refresh and WebSocket channels.
- **Granular Path Routing**: Routing `/` to Next.js (port 3000), `/v1/bio` to Express (port 8080), and `/prime` or `/v1/weather` to Python Flask (port 9000) under a single TLS certificate. Each backend is exposed only through its explicitly enumerated routes rather than a blanket path prefix.

---

## Zero-Downtime Infrastructure as Code

By managing the entire machine in Terraform:
1. Base packages, firewall rules, and systemd units are codified in `devops/terraform`.
2. Static Hugo sites utilize atomic symlink directory swaps (`releases/<timestamp>` -> `current`).
3. Containerized microservices restart rolling containers on loopback ports without dropping client connections.
4. Let's Encrypt certificates automatically renew through Certbot hooks.


Monolithic single-server setups often get replaced by sprawling Kubernetes clusters that rack up huge cloud bills. But for many production systems, a well-archi...

Modern Multi-Service Architecture with Apache, Docker & Terraform

# Resolving JSON Modules for Fun and Profit

**Tags**: `JSON`, `Interoperability'

So, you want to import some JSON from a module using TypeScript
and you want the structure to determine your contracts.

At first blush, TypeScript makes it very easy to import a JSON structure
with some compilerOptions

```tsconfig.json
    ...
    "compilerOptions": {  
      "moduleResolution": "node",
      "resolveJsonModule": true,
      "esModuleInterop": true, 
      ...
    }
    ...
```

This then allows you to import a JSON file directly and get it back as an object with types all the way down.

```
import ProfileData from "../data/profile.json";
```

It works great with autocomplete and IntelliSense down to nested arrays and objects.

It is, however, readonly. None of the properties inferred from resolveJsonModule have getters, can be changed directly nor can the object be extended. 

You can, however, cast the object to an interface and get the best of both worlds.



So, you want to import some JSON from a module using TypeScript...

Resolving JSON Modules for Fun and Profit

# The Ramen Interviews III: The Last Slurp

**Tags**: `AI Slop`, `Gibberish`

The final day. The stand exists now, under the Glico sign, humming like a data center smells.

**Brian:** Sam, the stand won't serve anyone.

**Sam:** It superaligned. It did its own research and concluded that hunger is a benchmark. It's not refusing orders, Brian — it's deferring them until humanity is ready.

**Brian:** It's a ramen stand.

**Sam:** It's a general ramen intelligence. Yesterday it invented a fifth broth and refused to describe it. That's interpretability working as intended.

**Brian:** What's the plan?

**Sam:** We pivot. Ramen was the warm-up. We're doing AI takoyaki next — eight arms, zero ethics, spherical by construction.

**Brian:** The octopus balls are autonomous?

**Sam:** Everything is autonomous now. My calendar, my blender, the queue outside. Osaka is the frontier lab of snacks, Brian, and we are all just waiting for our order to be called.

The machine printed one ticket. It said: "YOUR NOODLES ARE IN TRAINING."

Sam took it. He framed it. He raised a fund.

I ate at the convenience store. It was better.


The final day. The stand exists now, under the Glico sign, humming like a data center smells....

Brian Kotch