Friday, 15 May 2026

The Ghost in the machine: AI writing software for us

Applied Science Fiction

This is science fiction stuff! I call it applied science fiction.

I don’t know if it’s ultimately good or bad; for sure, there will be issues. It reminds me of when the internet first appeared. Some of us—poor souls—thought that people would finally wake up. We thought that because they could read anything, and because the excuse of "not having an encyclopedia handy" was gone, the info would be there for all to see and a clear picture of reality would finally rise...

Yeah, right.

The result? People now have supercomputers in their pockets—created by the pinnacle of human ingenuity, from the chips themselves to the ASML machines that build them—only to be zombified by meaningless videos and to post "evidence" proving the earth is flat.

So much for progress.

So, I’m very reserved about this AI thing. It is a genuine breakthrough, a revolution. But if we stop thinking for ourselves, we will simply become idiots.

On a more practical note, an AI system is already a better developer than most juniors. This means today’s junior developers may never gain the experience needed to grow into the seniors of tomorrow. We are facing a real grassroots issue here.

BUT, after that rant, I have to admit: AI is the most exciting technology I have encountered in my 40+ years in the industry. I remember reading about Perceptrons in the mid-80s, the Lisp machines, and all the promises that never materialized during the AI winters that followed.

But it seems the time has finally come. Massive advances in chip technology and relentless research led us to the Transformer... Self-attention is the key, they say. So simple, so majestic.

And it’s here, right now.

So, let’s jump on the science fiction bandwagon. Let’s witness the rise of a new age, where systems write the software that will eventually become part of their own existence.

Back to earth. Here, in front of a Linux prompt, we run the following command:

curl -fsSL https://opencode.ai/install | bash

This downloads and installs OpenCode—a software-writing robot.

Next, we configure the connection to our local brain by creating the configuration file:


Then we have to create this file:

$ cat ~/.config/opencode/opencode.json
{
  "$schema": "https://opencode.ai/config.json",
  "model": "rtx5090/qwen3.6-mtp",
  "provider": {
    "rtx5090": {
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "baseURL": "http://100.126.179.69:8080/v1",
        "apiKey": "not-needed"
      },
      "models": {
        "qwen3.6-mtp": {
          "name": "Qwen 3.6 27B Local",
          "limit": {
            "context": 180000,
            "output": 8192
          },
          "supportsImages": false,
          "supportsTools": true
        }
      }
    }
  }
}

And then, all it takes is a single command to wake it up:

nobait@Dev-Station:~$ opencode

The ghost in the machine has started.

It might not look like much on the surface, but this is applied science fiction. You tell it what you want, and it builds it. It might take minutes, it might take hours—but as long as you can describe the path correctly, it will walk it.

Well... more or less, to tell you the truth. But mostly more.

In order to make sure it works, I wil ask the system to write a program to find where th International Space Station is.

The prompt that will make OpenCode to do that is this:

The OpenCode Prompt

Task: Initialize a workspace and build a live satellite tracking tool.

Steps to execute:

  1. Environment: Create a directory named space_tracker. Inside it, set up a Python virtual environment (venv) and install requests.

  2. The Script: Write a Python script named iss_now.py that does the following:

    • Reach out to the Open Notify API ([http://api.open-notify.org/iss-now.json](http://api.open-notify.org/iss-now.json)).

    • Fetch the current latitude and longitude of the International Space Station.

    • Use a reverse-geocoding trick (or a simple coordinate check) to tell me if the ISS is currently over "Water" or "Land."

    • Display the coordinates in the console with a cool, retro-scrolling ASCII effect.

  3. Execution: Once the code is written and the venv is ready, execute the script using the venv's python interpreter so I can see the result right here in my prompt.

Make the console output feel like a transmission from a 1980s Lisp machine. Go.


The actual cherry on top is that you can use a chatbot just to help you craft the perfect prompt for the machine.

Do you truly feel the weight of that action?

We have reached a point where a human uses one machine to generate the message for the next. We are no longer writing code; we are managing a relay race of silicon intelligence. It’s a strange, circular evolution—using the 'ghost' to help us better command the 'robot.

The system takes the command and starts working:

The system writes the program we asked it to build. It makes a few mistakes, pauses to make its own corrections, and finally runs the code it wrote—all while maintaining the aesthetics of an 80s LISP machine.


It gives us the answer:


So, there it is. The ISS is somewhere over the Pacific/Antarctic region. Now we know.

I’ve just shown you the tool that will write the YouTube Anti-Clickbait app for me. I’ve presented the "ghost" that will do the heavy lifting in the next post.

But as I look at the console output, I’m left with a bitter taste in the triumph.

Are we happy?

Should we be celebrating the fact that we’ve caught the ball, only to realize it’s a hand grenade? We are standing among the last generation of true programmers. While the machine hums along perfectly, the juniors—the seniors of tomorrow—are being daisy-cut at the roots before they ever get a chance to grow.

I’ve waited forty years for this science fiction to become real. Now that it’s here, I’m just not sure if we’ve reached the summit or the edge of a cliff.

But for now, the prompt is waiting. See you in the next post.






From Terminal to Browser: Giving the Anti-Clickbait Tool a Face

Designing the YouTube Anti-Clickbait Experience

It’s time to design the actual YouTube Anti-Clickbait app!

While the interface is intentionally minimal, a clear plan is essential to ensure a seamless user experience. We want the focus to be on saving time, not navigating menus.

The Landing Page

Taking inspiration from the classic, distraction-free Google homepage, the user is greeted by a clean white screen. The central focus is our mission statement:

Stop Wasting Your Time on Empty Promises.

Fed up with videos that promise the world but deliver nothing? Let us watch them so you don’t have to. We’ll analyze the content, provide a concise summary, and deliver a verdict: Is it clickbait?

If the video is actually worth its salt, we’ll provide an elaborate breakdown so you can get the value you need in seconds, not minutes.

The Interface

Directly below this message sits a sleek, dark input field for the YouTube URL, paired with a prominent button: [ Watch it for me ]


The "Processing" Phase

Once the user submits a link, we provide real-time feedback. Instead of a boring loading bar, the app shows the system’s "thought process" through a series of status updates:

  • 🔍 Fetching audio track...

  • 🎧 Analyzing content...

  • 📝 Generating summary...

  • ⚖️ Passing judgment...

The Final Verdict

After the processing is complete, the app reveals the results in a clean, high-contrast layout:

  1. The Verdict: A clear "Clickbait" or "Not Clickbait" badge.

  2. The Quick Take: A 1-2 sentence "TL;DR" of the video.

  3. The Deep Dive: (If the video isn't clickbait) A detailed summary of the key points, allowing the user to skip the video entirely if they choose.


So... writing the actual Python script is... let's just say, 'straightforward.' We’ve completed the investigation and the design is ready to go.

But why do the heavy lifting ourselves? What if we ask the summary model to write the script for us?

That’s the mission for the next post!