I've been trying to come to terms with my inconceivable fear of AI recently. Certainly, the constant scrolling on tech Twitter - in search of some novel answer - did not help in the slightest.

The Crisis

Truth be told, I've always believed (and still somewhat believe) that AI is a bubble. But something about the most recent developments in AI has just felt different.

From high school to college, I knew machine learning was going to be the next breakthrough. MNIST, ImageNet, YOLO, Transformers, the progress was swift. But with GPT-2 struggling to produce proper sentences, I thought generalizable intelligence was decades away.

I still remember the day when my roommate Casey showed me, in the living room of our small townhouse in the University District, this thing he saw called "ChatGPT". I didn't believe it was more than "some form of heuristic or rule-based chatbot" like the rest that existed on the internet. But as my prompts delved deeper into obscurity, the bot replied with ease, albeit struggling on basic math problems (sometimes explaining how to hot-wire a car if I told it to role-play a pirate).

It felt like a gimmick at the time. Something that seemed like magic, but had no obvious real-world applications.

Soon the world changed, well, more specifically, my world. It first started with GitHub Copilot, where it was autocompleting lines of code when you pressed TAB. Then Google and Microsoft followed suit with their own AI chat. Soon the models started "thinking" - Chain of Thought (CoT) - something that resemble human behavior. Then we gave them tools, made them take multiple turns, made them plan, made them agentic. With it, the nature of my job evolved quickly. From coding by hand (I refer to this today as trad-coding), to having AI generate small sections, to now, multiple autonomous agents executing and collaborating on full projects where human review is mostly "optional".

As of September 2026, AI has been able to one-shot software tasks, convert million-line codebases to Rust, and produce (partial) solutions to Millennium Problems. We went from LLMs being unable to add two numbers to them maintaining production systems. How do we reconcile with this? And what's next?

What is AI really doing?

Before I get to my existential crisis, I want to first explain what AI is not. I remember a few years ago people started claiming that these intelligent systems are sentient. And even though AI does behave similarly to some behaviors in humans that we would consider "sentient", the underlying mechanisms of AI differ drastically.

It works like this:

1. We collect a massive set of human-generated data (possibly synthetic these days).
2. We train the AI to (hopefully) grasp the latent properties of the underlying distribution space - be it language, images, video, 3D space, etc.
3. We then generate some output by sampling that learned distribution.

Or in simpler terms, AI learns what "looks" like human output, and through some randomness tries to generate something that sounds like human output (through further training we can likely make it almost indistinguishable from that of humans).

Even when it is "thinking", it isn't actually thinking. It's just constraining the output distribution space to be more aligned with the response that we want.

How has AI changed my world?

It is a stretch to claim that AI has changed the world for the vast majority of living things on the planet, but it certainly has for me.

Work

If you told me 3 years ago that I would not be manually coding a single line in 2026, I would think you had psychosis. But this is how the software industry has evolved today (Fortune). Most engineers I know do not physically type out code anymore. We went from writing for loops to using Claude Code or Codex to generate full stack systems. Most people would say that their productivity has increased 10x; what used to take 2 weeks to do could now be done in a few minutes. This by itself isn't necessarily an issue; most people embrace it. But it is what follows that I'm concerned about.

Lack of Human Reasoning

We have started losing the ability to think critically. I like to think of these coding harnesses as an abstraction on top of code. When we initially programmed computers, we started out with punch cards. Then we abstracted to assembly, then early languages like FORTRAN, COBOL, to systems-level languages like B, C, and then object-oriented languages like Java, then more interpreted languages like Python/JavaScript, and finally we have arrived now at: natural language programming.

Each level of the paradigm at which we think was different. In C we think in terms of memory and pointers, but in Java it became objects. Natural language is different though.

The difference is determinism. In any programming language the behavior we define is exactly the one we type out. If we make a mistake, it is uncorrected until we fix the syntax, or fix some erroneous assumptions that we made. But today, it is trivial to tell Claude to design a system, see some part of it is slightly broken (without understanding why), and just re-prompt Claude again to "debug and fix the issue".

Natural language abstraction is okay, but abstracting away the process by which we reason, iterate, and learn will be incredibly hurtful in the long run. This also leads us to the next problem.

Producing Future Thinkers and Developers

Now there exists a gap between those who wrote code before the AI revolution and those who didn't. The former have a fundamentally different understanding than the latter. But I will note here that I'm not claiming developers who grew up in the AI age are somehow less productive or competent; the opposite at times is true. But that productivity comes at the cost of some deeper grasp of structure that I, with no logical reasoning as of now, believe is important in the long run.

One thing that worries people today is the development of future talent. Entry-level job openings have slowly been decreasing across the board since 2022 (Brynjolfsson, Chandar, and Chen, "Canaries in the Coal Mine?"; SignalFire, State of Talent 2025). Senior engineers can now do the job of 5 junior engineers. Employers are seeing less and less value in hiring new grads if a Codex agent can fully replace them.

And they are right. This might be controversial, but most of the time interns and even junior engineers are negative EV. But that was never the point of internships / entry-level jobs. Even though the staff infra deities that we all have worked with before feel incredibly out of reach technically, we must remember that at some point they themselves were juniors, and even interns.

The pipeline was designed to ease engineers into the space, introduce them to different tech stacks, opinions, and nurture them to be future seniors. I know that I would not have the (however little) foundation that I have without my internships and new grad roles.

Companies will claim increasing profit, and we won't feel the effects for a while. But alas, it will be detrimental to future talent.

Losing the Process

The paramount issue, and especially the one affecting me the most, is the loss of the process. Code used to never work on the first try (maybe it's a skill issue), and that was expected. The fun and learning came not from getting from A to Z, but stopping at B, C, D, and doing E twice and going back to C until we eventually arrive at Z.

Today, if my prompt doesn't one-shot the problem, I can just follow up. And the AI will triage, diagnose, and resolve the entire issue. I don't need to understand anything to move past the hurdle. I want to, but I don't need to.

And this isn't isolated to programming. Mathematicians are feeling the same way. Most of the learning and puzzling is done at the process level, not the prompt level. To the detriment of our understanding we are scrounging for scraps of productivity. Solving a Millennium Problem teaches us nothing if it was just brute-forced through millions of compute hours.

Creativity

AI can generate more than just language. We have now trained models that can produce high-quality images, videos, music, 3D models, etc. Most of the outputs nowadays are indistinguishable from real human creative work. Some could argue it is more "impressive" than human work.

Even besides the obvious issue with training on art without the consent of the artist (worth an essay on its own), I'm personally more worried about how art and creativity will be redefined.

Art is sacred for me because that is part of how I separate what is "human" and what isn't "human". One of the few moats that we have today as a species is our ability to create, and to think outside of the box. As AI trains on more examples, its ability to generalize is improving. Once it has surpassed humans at everything, how do we define creativity?

Will we have to define creativity as being "out of distribution" for AI? But then, at what point will all of humanity be in distribution?

And how does art evolve if the only thing that matters now is the process? The product can be produced by an AI model at the snap of a finger.

I have way more questions than I have answers.