What are Software Devs doing in the AI era?

Posted on In Disqus:

Just a year ago, I couldn’t imagine that AI could have such a big impact on how we actually do our daily tasks. Every company is talking about and trying to embed AI into their products now. But what are the traditional developers doing in this AI era? This blog post is my observation during this time, and it might not be correct.

Disclaimer: The ASCII images in this post are generated by AI

1. The “trends”

1.1. ChatGPT

The year is 2022. OpenAI launched ChatGPT, free for everyone. The chatbot went so smoothly that we cannot use the word “impressed” to describe it. ChatGPT opened a lock that seemed impossible in the past. Suddenly so many text-based services were becoming obsolete, like Grammarly, writing assistants, etc. The effect didn’t stop there. It was like someone threw a big stone in the middle of the lake. We can still feel the ripple as this post is being written. Lots of research papers, ideas, tools, and of course, start-ups, were created. People found new ways to “play” with ChatGPT every day.

                o                    <-- ChatGPT drops, 2022
          .-----------.              ring 1: jailbreaks, role-play prompts
       .-----------------.           ring 2: courses, "prompt engineer" jobs
    .-----------------------.        ring 3: wrapper start-ups
 .-----------------------------.     ring 4: agents writing the code
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~    ... and we still feel it today

You might still remember those security researchers, hackers, or even a high schooler, telling ChatGPT to play the role of “insert_role_here”, and changing its pre-defined context, to do things that it was not supposed to do, and to expose the data that it should keep hidden.

This created a new job for the AI companies and their models: how to break, and fix, the vulnerabilities of their own model, to make it safer and more accurate. This is not a small task, more like a constant struggle between the builder and the breaker.

The software development industry watched these new tools closely. It was not a threat back then, but a new shiny toy that could spit out a code snippet meeting a specific requirement, if you prompted it properly.

1.2. Prompt Engineering

People did what people do best: optimize the chatbot, and extend it. ChatGPT, and all other chatbots at that time, were static chat models. They didn’t have memory. They calculated the probability of the next word, put them together to create a sentence, and proudly showed it to you as an answer to your question.

Naturally, by tweaking the wording, the result can be vastly different. The prompt, and how to prompt, became an “engineering” task, with all of its specialties. Companies were hiring “prompt engineers”, courses appeared online telling you how to create a good prompt, people shared tips and tricks for prompting on social media, etc.

A lot of “ChatGPT wrapper” style start-ups were founded during this period. With a specific, “proprietary” prompt, these wrappers could limit the answers to a specific topic, and sell that to users who didn’t know anything about these prompts.

  you                 the "product"              the actual work
   |                        |                          |
   v                        v                          v
question  ----->  [ hidden "secret" prompt ]  --->  [ ChatGPT ]
                                                        |
            answer  <-----------------------------------+

            ($20/month for a paragraph of text)

1.3. Agentic AI

Then, a new trend appeared in the wild. What if we keep extending the context window? What if we keep adding more parallel processing power, adding more parameters to the model?

The answer is agentic AI. AI evolved from just completing your code into writing the code itself. AI providers compete with each other by releasing AI models that keep improving benchmark results. People find new ways to harness them every day. Tips and tricks get shared on social media. Even those who are not tech-savvy are able to achieve things that seemed impossible.

2020            2022                 2024+
.------.        .------------.       .----------------.
| tab  |  -->   | copy/paste |  -->  | it does it all |
'------'        '------------'       '----------------'
completes       spits out a          plans, writes, runs,
your line       snippet              reads the error, fixes


     goal
      |
      v
 [ plan ] -> [ write ] -> [ run ] -> [ check ] -> you: review
      ^                                 |
      |          still broken           |
      +---------------------------------+

Take translation tasks for example. Years ago, you could only choose a professional translator, or use a machine translation service that read like a machine. Now with AI, it can translate the text just like a human would, even detecting word play and suggesting equivalent options in the translated text.

2. The limit

With new models released every day, the later the better, what would be the limit for these AIs? In other words, what tasks can these AI models not do, or not do well?

2.1. For now

The problems listed here could change in the future, when we have stronger boundaries, or better guardrails for AI agents.

2.1.1. Real data troubleshooting

 !!!         .---------------------------.
\   /        |    _____________________  |
 [x_x]       |   /                     \ |
/|###|\      |  |   ( ) ( )   PROD DB   ||
 |###|       |   \_____________________/ |
 |   |       |                           |
_|   |_      |    * * ACCESS DENIED * *  |
             '---------------------------'

    the one room the agent doesn't get a key to

There are a few stories about an intern, or a dev, who allowed the AI access to the database, and it wiped out all the data. With the right permissions, AI can wipe out the most resilient database backup. As a result, most companies won’t allow AI access to the real, production database. This makes troubleshooting just like it was before the AI era.

2.1.2. The last X percent of a complex task

+----------------------------------------------+------+
|##############################################|      |
+----------------------------------------------+------+
 <-------- AI: 95% of common patterns -------->   you
                                                  the 5%
                                                  that makes
                                                  it yours

AI can get you 95% done. For the last 5% (and this number will decrease over time), you will need to do it yourself. This is particularly true for complex tasks. It did that 95% well because it was trained on a large dataset: 95% of your task is a common scenario that the AI was trained on. The 5% left is where the fine details are, the ones you need to differentiate your results from others.

2.1.3. The same outcome everywhere AI touched

You create a website for yourself, without thinking much about its layout and design, then you tell the AI agent to redesign your site. It does the job. The result looks amazing, in a way that would easily have taken you 20 hours to do manually. It’s great.

+----------+  +----------+  +----------+  +----------+
| = LOGO   |  | = LOGO   |  | = LOGO   |  | = LOGO   |
|  HERO    |  |  HERO    |  |  HERO    |  |  HERO    |
| [ CTA ]  |  | [ CTA ]  |  | [ CTA ]  |  | [ CTA ]  |
| [][][]   |  | [][][]   |  | [][][]   |  | [][][]   |
+----------+  +----------+  +----------+  +----------+
  site A        site B        site C        site D

But hundreds of AI-created websites out there have the same look and feel. AI just uses the best “design” that it was rewarded for using during its training. Where is all the creativity? Where are all the weird UIs that make a site unique? They were all dropped early in the training stage of that AI model.

I think this will change in the future, as more models are released. Soon there will be a model that rewards good, unique design, not just for websites, but for anything related, like icons, banners, advertisements, images, etc.

2.2. For the future

There are some aspects where a human factor is needed that AI cannot replace, even in the future.

2.2.1. Innovation

            \    |    /
             \   |   /
         .-'''''''''-.
  ---  .'             '.  ---
      |        o        |
  ---  '.             .'  ---
         '-.........-'
             /   |   \
            /    |    \

nobody had ever seen one, until somebody imagined it

AI is exceptionally good at combining and summarizing stuff. But for something new, it struggles. What about something that doesn’t even exist yet? That’s what human imagination can do. We can create new concepts, new theorems, new definitions. Humans are curious. In the search for why, we explore and invent things that push humanity forward. All thanks to our biological brain, which consists of nearly a hundred billion neurons. I don’t think artificial neurons can mimic that. Even if they can in the foreseeable future, we outnumber them by billions.

2.2.2. Adaptation

                   .---------------.
round holes   -->  |  ( )     ( )  |====]
(this wall)        |               |          -->  flat pins
                   |      ( )      |====]          (that wall)
                   '---------------'

     same current, whatever shape the socket demands

Humans, from the very beginning of time until now, have been adapting to the surrounding nature. In pre-history, we adapted to be able to hunt bigger prey, and to plant food. Human adaptability enabled us to make more of it, make it better, and make it beautiful, whatever “it” is. This won’t change in the AI era. I’ve already seen some adaptation in software development. From writing code ourselves and painfully fighting with the framework, now we focus on the idea and the problem, and just reviewing the code is enough for many tasks.

2.2.3. Critical thinking

       AI                                human
.----------------.                        (?)
| option #1 ... x|                      -.[   ].-
| option #2 ... x|                         | |
| option #3 ... x|
| option #4 ... x|          "why are we opening
| ...9,996 more  |           this door at all?"
'----------------'

 every path that was              a question nobody
 ever written down                 wrote down first

I doubt that current AI models are able to think critically. They can be very good at brute forcing all possibilities, just because all these scenarios were already written somewhere on the internet. In a completely new situation, human creativity and critical thinking will be our advantage.

3. What are Software Devs doing?

Software developers lost their jobs due to AI. This is not new. It’s been happening every day, even before the first ChatGPT wave. Software development is one of the fields most impacted by AI, alongside financial managers and office and admin support. Pretty much any job that can be done in front of a computer screen can be replaced by AI in the future, if not now.

Writing software is hard. Writing good software is even harder. Devs are still needed, even when AI can write 100% of the code. AI-written code seems good at first glance, but who will confirm it is actually good?

There’s a recent saying in the wild comparing AI-written code to the assembly code produced by a compiler: you trust the compiler to produce the correct assembly code, so why not trust the AI now? I don’t agree with this argument. The compiler is a fixed set of instructions whose output you can predict and guarantee. You cannot say the same for AI. Code generated by AI changes on every iteration.

same input        fixed rules         same output, every single run
source.c  ---->  [ compiler ]  ---->  out.asm == out.asm == out.asm

same input        sampled             a different answer every run
prompt.md ---->  [   LLM    ]  ---->  v1  !=  v2  !=  v3

Unless we have a different model than LLM, I think devs are still needed.

4. Jobs

4.1. What jobs will be safe from AI

At the time of this post (2026-08-05), the jobs that cannot be done in front of a computer screen are safe. For example:

  • Healthcare (doctors, nurses, registered physicians, etc.)
  • Teachers of young students (I won’t be surprised if college and university students can learn from an AI teacher.)
  • Trade jobs (construction worker, auto technician, cook, electrician, welder, etc.)
  • Service industry (waiter, waitress, maid, janitor, security guard, driver, delivery, store clerk, etc.)

If in the future we can make smaller AI models and insert them into a robot’s CPU, then even these jobs won’t be safe from AI.

4.2. Jobs created due to AI

In every industrial revolution, a lot of jobs disappeared, then were replaced by even more jobs. The pattern is this:

  1. A job requires human capability due to its complexity and cost.
  2. A machine is invented to do it faster and better.
  3. People can’t compete with the machine, and lose their jobs.
  4. The machine breaks. It needs fixing. It needs maintenance. We need more machines. We need better machines.
  5. People adapt, and find jobs in one of those “needs”.

What does AI need?

  • Electricity => electricians, engineers, etc. are needed
  • Chips => microchip designers, mine workers, chip-making machine designers, etc.
  • Data => data collectors, data labellers, data organizers, data correctness judges, etc.
  • Algorithms => mathematicians, critical thinkers, developers (or anyone who can think the way a computer thinks)
  • Policy => researchers, philosophers, or even doctors

AI will grow the needs in the areas above, and keep creating jobs related to them. Of course, the jobs created are the ones that cannot be automated.

+--> (1) job needs a human
|             |
|             v
|    (2) machine invented: faster, cheaper
|             |
|             v
|    (3) people can't compete, jobs lost
|             |
|             | => WE ARE HERE
|             v
|    (4) machine breaks / needs fixing / needs more of it
|             |
|             v
+---- (5) people adapt into the "needs"