AI is killing your career

I don't hate AI; I've been drinking the Kool-Aid for a better part of 3-4 years now since the initial public release of ChatGPT.

It’s a major productivity booster for sure when used correctly, and it has helped me a ton during this time. Although it’s not really an “it.” It’s crazy how easily people get bamboozled by all the marketing hype created by AI companies.

You know, the “model escaped its sandbox” or “AI is going to replace people” rhetoric we hear every few months. Let’s be real: it didn’t escape anything, and at this point, it cannot replace anybody. What’s replacing people are competent people who use AI and automation to get more work done.

A model has no consciousness; it's just code, just like any other piece of software. You prompt the model for something; it matches a piece of data in its training dataset and layers that with tool calls, other contextual data, plus some grounding to generate the final response.

The model doesn’t understand the question or know the answer in the way a human does. It predicts a response from patterns learned during training. When grounding is used, another system retrieves relevant information and gives it to the model, which then turns that information into a natural-sounding answer.

That conversational style makes it feel as though you’re talking to something that understands you, even though there may be no human-like understanding behind it.

Tokens IN, tokens OUT - that's all it is.

This is why models can't really learn anything; they don't learn in the same way humans do. We “receive input” from the outside world, interpret that information based on our worldview, and retain those memories.

We don't memorize every single book we've read throughout our lives; we simply interpret the message of the book and form a memory of it.

This is vastly different from how AI works. With every new model, you see the parameter count increasing, from 8B to 32B, 1T, 1.5T, and so forth.

We're just shoving more tokens into a pattern-matching algorithm, giving it more information to match against, and thus improving its retrieval and lowering the amount of hallucinations you would experience.

The model has no concept of what that data actually represents, no real understanding of it, and no intrinsic sense of its meaning. It’s just pattern matching, with some fancy polish added through a harness.

When these models “break out,” it’s not because they suddenly woke up and became intelligent beings. Nope! It’s because somebody wrote a prompt asking it to do that task, either directly or indirectly, and the model just invoked a tool that gives it access to break out of its sandbox.

For example, you could write a Python module that can SSH into remote servers and then expose its API to the model.

Pair this with so-called “loop engineering,” and all that’s really happening is that the harness repeatedly prompts the model until it achieves some sort of testable goal.

The model draws on patterns learned from similar information found somewhere in its training data, whether from GitHub, a book, a blog, or elsewhere; it can't just create something totally new. Everything is seeded and based on some knowledge source, either its training data or context data.

The model breaks the task down into checklist items and creates a plan, while the agentic harness follows that plan by constantly re-prompting the model, calling tools, and checking the results.

Sure, that is intelligence to some extent, but you could also argue that it’s simply the modern, far more sophisticated version of for loops and if statements.

The base model never learns anything new. It’s simply layered with contextual data and tools, so the model itself never gets smarter or evolves.

As humans, we’re constantly learning from the moment we become aware of the world. We’re always re-evaluating and refining our knowledge and understanding.

Why are AI companies so obsessed with AGI?

AGI is just a gimmick born out of anxiety. LLMs are becoming a commodity, and open-weight models improve with each new release cycle.

AGI is the only way for these companies to maintain their moat. They need to keep chasing funding in the hope that GPUs will become more abundant and efficient over time, allowing them to turn those subsidized plans into something more profitable, crushing the small guys and open-weight competition in the process.

The goal is to keep throwing more data at these models, following the logic that more parameters equal more intelligence. They have to keep fueling the fire to justify the cost; otherwise, users will simply jump ship to the closest open-weight model.

I would assume that they are fully aware that AGI is not possible with the current architecture; you would need to solve long-term memory, and you would need to solve continuous learning and also deep, nuanced understanding of the data itself, not just pattern matching.

Yet with every new model release, while there are many improvements from time to time, the major focus is solely on having the biggest model out there.

Have you noticed how, five years ago, “AI” usually meant the kind of intelligent being we saw in movies? Now, in 2026, “AI” has effectively become shorthand for LLMs, while the original idea has been pushed further down the road and relabelled as “AGI.”

You are destroying your career by blindly trusting AI!

This brings me to the point of this article as a whole.

The reality is that LLMs are giant prediction engines, and most of the magic behind agentic coding comes from really good tool-calling and planning harnesses.

By relying solely on the model to write your code, you are giving up control to something that has no critical thinking whatsoever, and you become only as good as the latest model.

It may generate code that works, but somewhere down the line, in those thousands of lines of code, it will miss something.

Maybe not today, but 2 years from now, you'll have to come back and fix a bug. By that time, you've been vibe coding for so long that you don't even know how to use a debugger anymore, nor do you even understand the code enough.

You're now completely dependent on a model to find the solution; you've forgotten how to code, how to debug, and you don't even understand the codebase you once added this piece of code to.

This is especially dangerous for new coders with < 3 years of experience; you've not given yourself the chance to learn about design patterns, how to structure applications, how to debug, etc...

Programming is about muscle memory, but you've never gotten to the point where you've trained that muscle for long enough.

I've been programming for 15-odd years or so before AI. I've worked on a wide variety of systems; I have done a ton of code reviews and debugged complex bugs over the years.

Yet still, when I look at a vibe-coded app, I find it difficult to understand what's going on sometimes. LLM-generated code tends to be overengineered and verbose at times; it'll follow some pattern it found on GitHub and basically overcomplicate things if you don't prompt the model correctly.

I've seen devs just give it a ticket and ask it to build out the feature or fix the bug, without questioning what's going on. They do a quick read of the code and push it up.

They don't provide direction and let the model decide the best route to take. They don't even manually edit the code or refactor it.

Eventually LLM's will reach their ceiling, or you'll have to work in a commercial codebase that is very large and custom, where AI will struggle. Then what?

How to use AI properly?

I can't remember the year, probably 2009, 2010 - somewhere then. DHH did that presentation on Rails. Wow! Blew my mind, because previously CRUD was such a chore.

Now with one command, I can scaffold an entire CRUD UI. As coders, we've always been excited about code generators; they make the grunt work more bearable.

You know those laborious tasks you've done three million times, and just can't anymore?

AI is rails g scaffold on steroids; it's remarkable, and used correctly, it can save you hours of work, but used incorrectly, it can slowly kill your career.

I'm still figuring things out, but here's how I use AI:

Use AI as a coder reviewer:

Here's a ticket https://ticketlink.com . I've implemented a bugfix to resolve this issue. Can you review my changes, and advise on anything I missed? Syntax errors, bugs or security issues created.

Break down tasks and set rules:

I've implemented the following class methods: putImage(image i) , getImage() -> image . I want you to implement these two methods, you need to check that the image is a valid PNG, JPG, GIF using the mime type. Do not trust headers, or the file extension. Make sure to only store the URL in the db, and store actual file in S3.

Research task:

I want you to use your web search tool, and documentation from stackoverflow, and Google's official docs. Can you research and break down in a bulleted list or short plan of how I should implement a Signin with google button. Give me the different approaches, and best practices. Please includes links to documentation sources.

Testing tool:

I've migrated this feature from our old dashboard. Here's the new dashboard link: http://127.0.0.1:8001/dash/quotes , and old: http://127.0.0.1:8000/dash/quotes . Use this test user: kevin, pass: 12345. I want you to login and create a quote in the old dashboard, note the steps taken, view the final quote. Then repeat the same process on the new dashboard, and point me to the final quote PDF for both and compare the results and code such that both codebases follow the same rules and generate the exact same result. Note: V1 uses PHP 5.6 and V2 uses PHP 8.3, thus the new version must use modern standards applicable to PHP 8.3 so the code maybe different but the rules and outcomes should match exactly.

What we're doing here is using the AI as a code gen tool and a code reviewer. We're using it as if it were a junior coder and a second pair of eyes.

We're giving the tool clear rules and making the important decisions ourselves. Paired with this, you would also manually do the testing, code review, and refactoring where necessary.

You're not just pushing up a PR; you're using AI purposefully and smartly.