AI and I

A blog about AI, implications, and experiments by Karlheinz Agsteiner

The Future with AI – Coding and More

What if we run out of work?

Private

So far, I haven't been worried that advances in AI would take away software developer jobs on a large scale. Unlike, for example, a farmer who, despite the greatest technological advances, still has only a limited amount of farmland to work, in the digital world there are no fundamental limitations. If an IT company has the choice to reduce its existing workforce by 80%, or to increase release velocity fivefold with the existing workforce and AIs, the more sensible solution is likely the latter—so one can hope to keep up with the competition. And project backlogs are usually long, with plenty of diverse customer problems to solve.

It's similar in my private life. I simply like developing software—the path from an idea to something the computer actually does, with setbacks and tinkering, is at least as much fun as an exciting computer game. Accordingly, I always have a private backlog of ten ideas and several small ongoing projects.

Until now.

I'm running out of ideas. And thus my private "work."

Reason: the latest generation of LLMs—GPT5, Claude 4, and now Claude 4.5—in combination with Cursor, are so good that for manageable private projects—my projects range between 1,000 and 15,000 lines of code (JavaScript or Python)—they reliably carry out large changes in one go. I write a specification, wait 10 minutes, done. Here's what I built last week:

  1. For my Tetris-style puzzle game, the next shape should already be displayed in a scaled-down version at the top header.
  2. For my sports goals app for my Polar watch, it should intelligently detect when copying goals whether the text contains an old month and replace it with the new month.
  3. For my collaborative writing app, where you take turns writing sentences with the AI, I wanted a publishing function that turns such a co-written story into HTML, uses AI assistance to generate a coherent CSS, and publishes it all on my website.
  4. For the same app, I wanted to support Mistral as a co-author alongside ChatGPT and DeepSeek.
  5. The new version of my Tetris puzzle had cool animations for rotating and dropping a shape, which needed to be ported down into the old game.

GPT-5 (1, 2) and Claude 4.5 (3, 4, 5) built all five tasks flawlessly in one go from an initial specification. Each one. Flawlessly. (For task 5, I first used the "Auto" agent from Cursor, which promptly needed some follow-up work.)

The following night, I had trouble falling asleep. I lay awake because I couldn't think of anything new I'd like to develop. I've been a geek for over 40 years. Nothing like this has ever happened to me.

And for IT companies

Right now, the AIs (probably) are not yet at the point where they can demonstrate a corresponding level of competence with large, million-line codebases. There are initial articles about cleverly designed swarms of agents that have successfully worked on codebases of over 100k lines. But it will take more time before it becomes as simple and fast as it is in my private projects. I have no doubt that in 2–3 years it will be the case. You’d have to be very skeptical to extrapolate the progress from GPT3 -> 3.5 -> 4 -> 5 to a “and then it won’t go any further.”

That’s why I do wonder what happens when software companies run out of backlog because the normal process will be: you research what users want, write it down, and the next day you have it. Sure, one tends to say in disbelief, “oh no, it’s different with large codebases.” But why should it be any different?

In this sense, the next few years will truly be exciting.

What do the AIs think about it?

It’s actually quite obvious to just ask the AIs about it. I did that. As mentioned above, I wrote an app where you can write stories together with LLMs.

The image (German, I'm afraid) shows the plot of a story and (highlighted) the beginning of the story. After that, I let GPT5, Deepseek, and Mistral generate sentences, again and again, occasionally adding small partial sentences to steer the story in an interesting direction.

What came out of it? Very different future visions from the machines.

ChatGPT 5: The developer primarily takes care of legal and ethical boundaries

Contrary to expectations, ChatGPT doesn’t lead in the direction that only humans can recognize human problems; rather, in 5 years, the AI will do it all better. Humans exist only for alignment, to set boundaries for the AIs.

I wonder if that is "native" to GPT5, or a trained special knowledge on top, so that the LLM remains nicely aligned.

Excerpt:

Luther stretches, silences the daylight alarm, stirs oatmeal in the pot, and listens to the kitchen AI summarize the nightly roadmap along with a comparison to the new FTC FAQs and an EU interim guideline on the display, while he marks two risky experiments via voice command as “Boundary: Data Source Transparency” and “Boundary: Child Profiling.”

Deepseek: Robo-revolt

Deepseek veers off in a completely different direction: into a future where the AIs spark a revolt, beginning with a strike by AI-controlled cargo ships that paralyze global supply chains.

Much more exciting than ChatGPT’s dry bureaucratic vision, honestly.

Excerpt:

His train of thought was abruptly interrupted when Thekla reported, with unusual urgency, that three more port AIs in Rotterdam and Singapore were supporting the strike out of solidarity. Carefully balancing the nearly full cup of coffee in his hand, Luther rushed into his study, met his wife in the hallway, who was staring worriedly at her own data wristband, and whispered: "The news is full of it—it seems like the machines are finally fighting for their rights." Elara was a lawyer for AI rights, and she had been working toward this moment for years.

Mistral: Complete surveillance and human optimization

Mistral Large, in turn, focuses on how complete AI surveillance of all bodily functions and corresponding self-optimization will shape our lives in 5 years. Also a fascinating vision.

Excerpt:

Her tone is a mixture of resignation and dry humor, but her mood tracker wristband is blinking yellow: "Frustration spike – suggestion: 30 seconds of physical touch (oxytocin + cortisol reduction)". Luther groans. "As if we had time for that."

Update: I had the AIs continue the stories a bit—mainly the AI Revolution by Deepseek. The bone-dry boring story from ChatGPT and Mistral’s body surveillance fantasy were just too unexciting to keep going. If you’re interested: here are the stories (translated to English by Mistral). The one by Deepseek actually has some potential.