AI Slop: An Experiment in Discovery

Solo Episode Reflection: I’m back behind the mic after about a year-long break. Producing this podcast takes more time than you might imagine, and I was pretty burnt out. The last year brought some major life events, including moving my family back to New Zealand from Denmark, dealing with depression, burying my father, starting a new business with my wife, and having a teenage daughter in the house. These events took up a lot of space.

The Catalyst for Return: Eventually, you figure out how to deal with grief, stop mourning the way things were, and focus on the way things could be. When this space opened up in my life, AI came into the picture. AI got me excited about ideas again because for the first time, I could just build things myself without needing to pitch ideas or spend limited financial resources.

On “AI Slop”: I understand why some content is called “slop,” but for those of us who see AI as a tool, I don’t think the term is helpful. We don’t refer to our first clumsy experiments with other technologies—like our first map or first lines of code—as slop. I believe that if we want to encourage curiosity and experimentation, calling the results of people trying to discover what’s possible “slop” isn’t going to help.

My AI Experimentation Journey

My goal in sharing these experiments is to encourage you to go out and try AI yourself.

Phase 1: SEO and Content Generation My experimentation began with generating SEO-style articles as a marketing tool. As a dyslexic person, I previously paid freelancers thousands of dollars over the years to help create content for my website because it was too difficult or time-consuming for me to create myself.

• Early Challenges & Learning: My initial SEO content wasn’t great, and Google recognized this, which is why those early experiments don’t rank in organic search. However, this phase taught me about context windows, the importance of prompting (prompt engineering), and which models and tools to use for specific tasks.
• Automation and Agents: I played around with automation platforms like Zapier, make.com, and n8n. I built custom agents, starting with Claude projects and custom GPTs. I even experimented with voice agents using platforms like Vappy and 11 Labs.

Unexpected GIS Capabilities: During this process, I realized you can ask platforms like ChatGPT to perform GIS-related data conversions (e.g., geojson to KML or shapefile using geopandas), repro data, create buffers around geometries, and even upload a screenshot of a table from a PDF and convert it to a CSV file. While I wouldn’t blindly trust an LLM for critical work, it’s been interesting to learn where they make mistakes and what I can trust them for.

AI as a Sparring Partner: I now use AI regularly to create QGIS plugins and automations. Since I often work remotely as the only GIS person on certain projects, I use AI—specifically talking to ChatGPT via voice on my phone—as a sparring partner to bounce ideas off of and help me solve problems when I get stuck.

Multimodal Capabilities: The multimodal nature of Gemini is particularly interesting; if you share your screen while working in QGIS, Gemini can talk you through solving a problem (though you should consider privacy concerns).

The Shift to Single-Serve Map Applications

I noticed that the digital landscape was changing rapidly. LLMs were becoming “answer engines,” replacing traditional search on Google, which introduced AI Overviews. Since these models no longer distribute traffic to websites like mine the way they used to, I needed a new strategy.

• The Problem with Informational Content: Informational content on the internet is going to be completely dominated by AI.
• The Opportunity: Real Data: AI is great at generating content, but if you need actual data—like contours for your specific plot of land in New Zealand—you need real data, not generated data.
• New Strategy: My new marketing strategy is to create targeted, single-serve map applications and embed them in my website. These applications do one thing and one thing only, using open and valuable data to solve very specific problems. This allows me to rank in organic search because these are problems that LLMs have not yet mastered.

Coding with AI: I started by using ChatGPT to code small client-side map applications, then moved to Claude, which is significantly better than OpenAI’s models and is still my coding model of choice. Currently, I use Cursor AI as a development environment, swapping between Claude code, OpenAI’s Codex, and other models.

• A Caveat: Using AI for coding can be incredibly frustrating. The quality of the code drops dramatically once it reaches a certain scale. However, even with flaws, it’s a thousand times better and faster than what I could do myself, making my ideas possible. Crucially, I believe that for the vast majority of use cases, mediocre code is good enough.

Success Story: GeoHound

After practicing and refining my methods, I decided to build a Chrome extension. Every GIS professional can relate to the pain point of sifting through HTTP calls in the developer tools networking tab to find the URL for a web service to use in QGIS or ArcGIS.

• The Impossible Idea Made Possible: I had pitched this idea to multiple developers in the past, who were either uninterested or quoted between $10,000 and $15,000 to build it.
• The AI Result: Using AI, I had a minimum viable Chrome extension—GeoHound—that filtered out common geo web services within 3 hours. It took a few days of intermittent work before it was published to the Chrome and Edge web stores.
• Current Use: GeoHound has thousands of users (my own statistics suggest closer to or over 3,000 users, compared to the 1,000 shown on the Chrome store). While not perfect, it is clearly good enough, and this was something that was impossible for me just six months ago.

My Point: Now is the Time to Experiment

AI is here, and it will lead to profound change. Experimenting with it is vital because it will:

1. Help you develop the skills and knowledge needed to meet the needs of the people you serve.
2. Help you better understand what is hype and what is not, allowing you to decipher which voices to listen to.

We are moving from a world where information is ubiquitous to a world where knowledge is ubiquitous. Now is the time to be making sloppy mistakes. Don’t let perfection stop you from learning how to make stuff that is going to be good enough.

If your work consists of repetitive tasks that follow step-by-step recipes, that’s going to be a tough gig going forward. Long-term, there will be new opportunities, but you need to be experimenting now to be in a position to take advantage of them.

Resources Mentioned

You will find a list of the tools I’ve been experimenting with in the show notes.

• Automation: make.com, n8n, Zapier
• Voice/Agents: 11 Labs, Vappy, custom GPT (MCP servers)
• Coding Models: Claude (current choice), OpenAI’s Codex, ChatGPT
• Development Environment: Cursor AI
• LLMs/Multimodal: Gemini (studio.google.com)
• Browser Extension: GeoHound (for Chrome and Edge)

https://chromewebstore.google.com/detail/nooldeimgcodenhncjkjagbmppdinhfe?utm_source=item-share-cb

If you build anything interesting with these tools, please let me know! I’d love to hear about your own experiments.

In Conversation

Why I Stopped — and Why I’m Back

If you’ve followed this podcast for a while, you may have noticed I stopped publishing for about a year. I was pretty burnt out — it takes far more time than you might imagine to produce this podcast — and if I’m honest, I didn’t really miss it until now. The last year was busy in other ways. Even though I was born and raised in New Zealand, moving back here from Denmark with my family hasn’t been without its challenges. There were a few major life events along the way: we now have a teenage daughter in the house, my wife and I started a new business, I’ve been learning to live with depression, and I buried my father.

There’s nothing exceptional about any of those things. Your kids grow up, lots of people start businesses, a high percentage of people experience some form of mental illness in their lifetime, and children are supposed to bury their parents. But these events took up a lot of space — space that used to be filled with something else. At some point you have to decide what to keep and what to throw away, and for a while podcasting needed to be put in storage.

So what changed? Eventually you figure out a way of dealing with grief. You stop mourning the way things were and start focusing on the way things could be — the way you want them to be. Before you know it, you have more space again, for the things you put in storage and for new things that catch your eye. This is where AI comes into the picture. AI got me excited about ideas again, because all of a sudden I could just make things myself. I didn’t need to pitch an idea to anyone or decide what to spend limited financial resources on — I could have an idea and build it. That excitement translated into the motivation to start publishing the podcast again.

On the Word “Slop”

Before I share my experiences, I want to address this idea of AI slop. I understand why people call some AI content slop. But for people like us — people who understand this is a tool and not just a shortcut — I don’t think it’s helpful. We don’t refer to our first clumsy experiments with other technologies as slop. The first map you created — was that slop? Your first few lines of code, “hello world” — were they slop? And what about all those lines of AI-generated code behind the new features in your favourite piece of software — is that slop too?

My gut feeling is that we’re more than happy with some AI content, and the stuff we don’t like, well, that is “clearly slop.” If we want to encourage experimentation and curiosity, pointing at the results of other people trying to discover what’s possible and calling it slop isn’t going to help. So today I’m going to share my AI experiments, in the hope it encourages you to go out and experiment too.

Where It Started: SEO and Content

For me it all started with generating SEO-style articles as a marketing tool. Search engine optimization sounds like a dirty word, and honestly it’s the only reason this podcast still exists. The idea is simple: you publish content relevant to the audience you’re seeking to serve, and if it’s good enough Google shows it to the people you’re trying to attract — and hopefully they discover the podcast.

I’m dyslexic. I’ve learned a bunch of ways of coping with it, but I’m still very dyslexic — if you’ve ever received an email from me, you’ll know what I’m talking about. Over the years I’ve paid freelancers thousands of dollars to create content for my website because I couldn’t create it myself, or it would take too long. My early AI experiments weren’t great — you can still find some on the website, and Google recognised they weren’t great, which is why they don’t rank in organic search. But through that experimentation I learned about context windows, the importance of prompting, and which models and tools to use for which tasks. I played with automation platforms like Make, n8n, and Zapier, and built my own agents — first custom GPTs and Claude projects, then voice agents using ElevenLabs and Vapi, and at one stage an MCP server. If you haven’t played with MCP servers, they’re incredibly interesting — I think they’re the apps of the future.

Discovering AI’s GIS Capabilities

It’s funny what you learn along the way. My goal was marketing content, but at some point I realised you can just ask ChatGPT to use GeoPandas to do data conversions — GeoJSON to KML or shapefile — to create buffers around geometries, or to reproject data. You can upload screenshots of tables from a PDF and have it convert them to CSV files. I wouldn’t blindly trust anything that comes out of an LLM for critical work — there’s no way — but it’s been interesting to learn where it makes mistakes, perhaps even why, and what I can trust it for.

Today I use AI on a regular basis to create plugins and automations for QGIS. I work remotely and I’m the only GIS person on some of the projects I work on — I don’t have colleagues to bounce ideas off — so I use AI as a sparring partner to help me understand and solve problems when I get stuck. The easiest way for me is to push the button on my phone and just talk to ChatGPT; it’s unbelievable. Gemini is interesting too: because it’s multimodal, you can literally talk with it. If you share your screen with Gemini while working in QGIS — and you’ll want to think about privacy here — it can talk you through the problem you’re working on. Is it perfect? Absolutely not. But it’s getting dramatically better with every model release.

Why I Switched to Single-Serve Map Apps

While all this was going on, organic search was changing. More and more people use LLMs as answer engines instead of searching Google, and Google introduced AI Overviews, which just give you the answer. These large language models no longer distribute traffic out to websites like mine the way they used to — and honestly, that’s what I do too. For the most part, asking a large language model is a better experience than sifting through multiple websites all trying to form a relationship with me. If it’s really important to get right, I’ll do the research myself — but even then, as a first pass, an LLM often does a better job.

Watching large language models turn into the answer engines of the internet, it became clear that informational content online is going to be completely dominated by AI — which meant I needed a new strategy. AI is great at generating content, and it can be exceptional. But if you need contours for your plot of land in the small town you live in on the top of the South Island of New Zealand, you don’t want generated data that looks reasonable — you want actual, real data. So my new marketing strategy is to create targeted mapping applications and embed them in my website. There’s a ton of open, incredibly valuable data out there; what’s missing is for someone to show up and use it to solve very specific problems. I call these single-serve map applications — they do one thing only, with one or two layers, not thousands. And they let me rank in organic search, because these are problems LLMs haven’t mastered yet.

Coding with AI, and “Good Enough”

To create these apps I started with ChatGPT to code small client-side map applications, then moved to Claude, which is significantly better than OpenAI’s models and is still my coding model of choice. Today I use Cursor as a development environment, swapping between Claude Code, OpenAI’s Codex, and the different models available — often with several running at once. For anyone in the geospatial industry who does anything with code, Claude Code and Codex are incredible. If you do anything that needs to be automated and you’re not an extremely skilled programmer, look into them. Recently they’ve added the ability to create agents — experts in different subjects that the environment hands off to as it works through a task.

It’s important to note these don’t always work, and using AI to code can be incredibly frustrating. Once your code reaches a certain scale, the quality drops significantly and the process slows dramatically — at times I’ve simply hit a wall and abandoned projects. But for me that’s all part of the learning experience, and no matter what, it’s still a thousand times better and faster than what I could do myself. It makes my ideas possible. A lot of people go out of their way to point out the flaws in AI-generated code, and there are flaws — it’s not the silver bullet the thread-boys on X claim. But the side of the argument that’s missing is that for the vast majority of use cases, mediocre code is good enough.

GeoHound: From Impossible to 3,000 Users

After experimenting with map applications and getting better — learning where I made mistakes and how to preemptively avoid them — I decided to build a Chrome extension. Every GIS person will relate to this: you see data you want to use in a web map, open developer tools, hit the networking tab, and start sifting through HTTP calls looking for the URL of the web service, then strip out the right parameters so you have something you can paste into QGIS or ArcGIS.

For the longest time I thought a browser extension would be a fantastic way to solve this. I’d pitched the idea to multiple developers — they were either uninterested or going to charge me $10,000 to $15,000 to build it. That’s not a critique of them, but it meant the idea was impossible for me to create. With AI, I had a minimum viable Chrome extension that filtered out common geo web services within three hours. It took a few days of on-and-off work before it was ready to publish. The extension is called GeoHound — there’s a link in the show notes — and you’ll find it for Chrome and Edge. The Chrome store shows about a thousand users, but my own bare-minimum statistics suggest the number is closer to, or over, 3,000. Is GeoHound perfect? Absolutely not. But it’s clearly good enough — and bear in mind this was impossible for me just six months earlier.

Why Now Is the Time to Experiment

So what can we learn from all this? AI is here. Even if it stopped improving now, it’s still here, and it’s going to lead to profound change. In my humble opinion, experimenting with AI does two things. First, it helps you develop the skills and knowledge you need to meet the needs of the people you’re seeking to serve. Second, it helps you better understand what is hype and what is not. A lot of people are capitalising on the AI hype right now, and most of it is completely self-serving — having your own experiences to draw on helps you decipher which voices to listen to, and form your own picture of what this AI-first world might look like and how you fit into it.

Now is the time to be making sloppy mistakes. Don’t let perfection get in the way of learning how to make stuff that’s going to be good enough. We’re moving from a world where information is ubiquitous to a world where knowledge is ubiquitous, and we have to decide what that means for us. In the short term, I think it means it becomes even more important to be known, liked, and trusted by the people we work for. And if your work consists of repetitive tasks that can be solved by following step-by-step recipes, that’s going to be a really tough gig going forward. Long-term, I’m sure there will be new opportunities — but if you’re not experimenting with AI, you might not be in a position to take advantage of them. And who knows, maybe you’ll learn something and have some fun along the way.