From AI Skeptic to AI Convert (Infrastructure Edition)

Cold open: three months, total reversal

In literally three months, I went from a hardcore anti-AI skeptic who wrote almost every line of code I was responsible for to someone who has essentially decided I won't be wasting my time writing code anymore. I don't want to write it. I want to create it. I know that sounds like a word game, but I mean it literally: the mechanical act of fingers on keyboard typing tokens into files is now, for me, the least efficient way to produce most of the code I ship.

This is a story, not a hype piece. It's about why I was a skeptic, how that started to crack, the epiphanies that changed my mind, and the concerns I still have.

Where I come from (and why it matters)

I am not an application developer. Titles vary wildly from company to company, but I've always thought of myself as an infrastructure engineer more than a software developer. Yes, we write software --- lots of it --- but we write tooling to manage infrastructure. Sometimes that tooling looks like a containerized app with a REST API and a database behind it. From a distance, it looks like application development. But that's a small part of what we do.

Writing the code is the easy part. The hard part is managing, scaling, testing, deploying, and operating that code in production while the world around it is constantly changing. External dependencies evolve. Vendors change behavior. Scalability boundaries show up where no one predicted. This work feels less like math and more like a living organism that you're keeping alive while it mutates. That's the lens I bring to this.

Why I was a skeptic

My skepticism came from watching people misuse these tools. Over the last few years --- especially when ChatGPT hit and a lot of people argued it could replace Google or Stack Overflow --- I watched people with no expertise ask for instructions, copy the output, and present it as their work without contextualizing it to their environment or the problem they were actually solving.[1] That blind trust is dangerous.

I also watched corporate mandates appear: "every business unit must have generative AI in their plan." That wasn't about using AI as a tool. It was about bolting AI onto user-visible features whether or not they were useful.[2][3] A lot of consumers see AI features as nice-to-have rather than must-have, and public concern about AI is rising even if sentiment remains mixed.[4][5][6][7] That's a big part of why the backlash exists.

My interest wasn't in building infrastructure that depends on AI at runtime. It was (and is) in using AI as a cognitive amplifier to help people build robust, well-tested, stable infrastructure. It can amplify expertise, but it doesn't create it. And for infra, I want to avoid runtime dependence unless the risk controls are strong.[8] Think of it like a compiler: you use the tool to build the product, but you don't ship the compiler inside the product.

The first cracks in my skepticism

Late 2025, I started using ChatGPT for basic problem solving and debugging. I had it analyze large volumes of vendor docs to hunt for clues and patterns. I used it mostly through the web interface for about a month. I started seeing just how much of a cognitive amplifier it could be.

One of my weaknesses is writing. I write great design docs and code, but I've always struggled with documentation and prose. AI filled that gap. I could dump raw ideas and get back clear lists, well-formed docs, and strong structure. With iteration, I produced design documents in a fraction of the time.

At the same time, I watched a junior engineer flail with Copilot. It generated code that ignored critical requirements. The tool didn't help them because they didn't know what they were doing. That was a lesson: AI can clarify and speed work, but it also amplifies errors when judgment is weak; it can boost expertise, but it can't create it.[9][10][11]

The "holy-shit" moment

Toward the end of November 2025, I finally gave it a real shot. I took a complex data model for a music-education tool my bass instructor and I wanted to build and gave a simple paragraph prompt: "Create a SQLAlchemy 2.0 data model with a FastAPI REST wrapper; follow SQLAlchemy best practices." Three hours later, I had a working, tested Python framework --- something that would have taken me a week by hand.

That was the first epiphany: it could write the whole thing. The second epiphany was iteration. I defined standards (PEP compliance, etc.), asked it to make the code compliant, and it did. I decided to go all-in. I set a rule: don't write code by hand. Don't touch the files. Let the tool do it. I wanted to see how far I could take it.

What changed in my day-to-day

The mechanical act of coding --- which is probably two-thirds of my skillset --- is now obsolete for me. Most of us have implemented the same algorithm a hundred times. We can do it quickly. But we can't do it better than the AI does. There's also a lot of boilerplate in any software product that doesn't require higher-order thought. AI is the next evolution of the tooling we've always used to offload grunt work --- like search-and-replace or code templates.

Speed is one win. Iteration cost is the bigger win. When iteration is cheap, your standards can go higher.

Developer persona: why it works (barely)

Here's the key problem: LLMs don't adapt in-session the way humans do. Humans internalize feedback; AI can repeat the same mistake unless external gates force correction. Improvements happen at the model-version level, not through the kind of learning a human does on the job.[12][13]

That means documentation alone isn't enough. The real enforcement mechanism is hard CI gates. If the AI writes code that fails linting, tests, or coverage, it has to fix it. It will do that, but it often needs multiple passes. A human usually gets it right the first time because they've learned the rules.

A silly example: every time the agent tried to submit a PR with the GitHub CLI, it used a here-doc syntax that doesn't work in bash. It repeated the mistake every single time. I had to add a bespoke guideline to stop it. That kind of edge case is frustrating, but solvable --- by hard gates and strict instructions.

The developer persona works because the output is code and the code is gated. That's the core reason it's stable.

Beyond developer: operations and QA personas

Infrastructure engineering isn't just writing code. It's ops and QA. These roles have different priorities and trade-offs. A developer's job is to implement the spec. QA's job is to block it if it doesn't meet requirements. Operations' job is to keep the system alive and restore stability when things fail.

If I want to use AI beyond the developer role, I need those personas to behave differently. An ops persona should not stop to write code when there's an outage. Its job is to restore service quickly, under constraints, while capturing enough diagnostics for RCA later. QA should focus on why testing failed to catch an issue. These roles complement each other. Without them, you end up throwing garbage into production.

The unsolved problem: validating AI behavior outside code

Here's the hard part. For developers, CI gates are the validator. What's the equivalent for an ops persona? How do you stress-test an operational agent? How do you validate behavior when LLMs are non-deterministic and can shift as providers update models --- meaning you can't pin them as tightly as a dependency?[13][14]

With software, I pin versions to keep behavior stable during a development cycle. I can't do that with the LLM behind the tool. So we need a new kind of testing framework for agent behavior. I don't have that yet. I haven't even seen a good answer.

I don't want an autonomous ops agent touching production without predictability. That's one of the biggest open problems I see.

Why I became a convert anyway

By mid-January 2026, I stepped back and looked at what I had built without writing more than a handful of lines by hand: a set of repositories for my bass-instructor/student tool, and the supporting Python libraries. They're not finished, but the foundations are months ahead of where I'd be if I'd done it manually.

Iteration became absurdly cheap. A newer type checker (ty) appeared, and across six repos I had the AI implement it alongside mypy with specific failure rules --- in about 30 minutes. By hand, that would have been one to two days. I can ask the agent to redo an entire API from scratch if I don't like it. I would never ask a human to do that without expecting them to revolt.

I can also raise the bar. I set a policy: 100% line and branch coverage. That's the holy grail. Humans avoid it because it's tedious and the ROI is low. With AI, the ROI changes. The investment is tiny. The standards can be higher.

One more reason I became a convert: it was surprisingly fun. I know the AI isn't a human and doesn't think the way we do, but the conversational loop is a genuine cognitive amplifier --- and it maps well to my personality. I move very fast, I'm intense, and I don't have much patience for slow, defensive feedback cycles. With people, you have to be careful when pointing out mistakes; the first response is often defensive, and you have to protect the relationship to keep the conversation constructive. With the AI, if something is wrong, I can stop and ask, "Why did you screw up?" and do a root-cause analysis without any emotion or social friction. That's refreshing. It's also a confession: I'm not naturally great at working with people who don't share my speed or depth, and that's a flaw I'm aware of. The AI removes that impedance mismatch and lets me iterate at full speed.

The long-term risk

The flip side, especially in technology-consuming organizations: incentives skew short-term. They want profit now, not next year. Some will use AI to slow hiring or reduce headcount rather than build expertise.[2][3][15] That's dangerous.

You can't staff an entire company with the 10 best engineers on the planet. As you scale, you run into scarce top-tier talent and persistent skills gaps, which makes it hard to maintain uniformly elite teams.[15] The question is: how do you build and preserve institutional knowledge? When systems become opaque, maintenance turns into software archaeology --- and the cost of change balloons.

I've spent my career walking into systems that started strong and ended bankrupt. I'm worried we'll accelerate that with AI. We should be using these tools to scale humans, not replace them. And we should be using AI to educate ourselves, not to avoid learning. That takes discipline.

Final thought (and the capstone)

AI shows general-purpose-technology characteristics and, in my view, is industrial-revolution scale --- with the timing and magnitude moving faster than humanity is prepared to absorb.[16][17] It will only work if we invest in people, process, and rigor --- not just speed.

In a follow-up article, I will share my thoughts about AI impact on society and civilization, but there's one last point I'd like to close with.

I fear that Stephen Hawking was right. AI will result in the extermination of humanity and the end of civilization as we know it, but I would merely add the caveat that this is largely true only if we don't change the way we're becoming dependent on it. I am not confident that we as a species have the ability to work together collectively to solve a problem like this (cf. climate change), but I hope to be proven wrong.

Here's the quote:

"With the rapidly approaching inception
Of artificial intelligence
Humanity may well set the stage
For its own demise
Once the exponential rate of intelligence
Reaches critical mass
There will be no turning back
And all of mankind will be exterminated
The human race, who, for centuries
Have looked to the stars for answers
Have always questioned
Whether or not God exists
He does now"

And for your entertainment value, I actually discovered this not as a student of science, but rather as a Death Metal fan when I stumbled across this song:

https://www.youtube.com/watch?v=xixI-q5wFpg&list=RDxixI-q5wFpg&start_radio=1

Listen at your own risk...

References

1. Stack Overflow. "Stack Overflow's 2024 Developer Survey Shows the Gap Between AI Use and Trust in its Output Continues to Widen Among Coders." Stack Overflow Press Release. 2024-07-24. https://stackoverflow.co/company/press/archive/stack-overflow-2024-developer-survey-gap-between-ai-use-trust/
2. CNBC (Hayden Field). "Shopify CEO: Prove AI can't do the job before asking for more headcount." CNBC. 2025-04-07. https://www.cnbc.com/2025/04/07/shopify-ceo-prove-ai-cant-do-the-job-before-asking-for-more-headcount.html
3. Reuters (via Investing.com). "IBM to pause hiring in plan to replace 7,800 jobs with AI." 2023-05-01. https://www.investing.com/news/stock-market-news/ibm-to-pause-hiring-for-roles-that-ai-can-replace-bloomberg-news-3072836
4. Circana. "Most Consumers Are Aware of AI, but One-Third Don't Want It in Their Devices." Circana Press Release. 2026-01-27. https://www.circana.com/intelligence/press-releases/2026/most-consumers-are-aware-of-ai-but-one-third-dont-want-it-in-their-devices/
5. Pew Research Center. "How Americans View AI and Its Impact on People and Society." 2025-09-17. https://www.pewresearch.org/science/2025/09/17/how-americans-view-ai-and-its-impact-on-people-and-society/
6. YouGov. "Americans' top feeling about AI: caution." 2024-04-09. https://today.yougov.com/technology/articles/49099-americans-2024-poll-ai-top-feeling-caution
7. YouGov. "Do Americans think AI will have a positive or negative impact on society?" 2025-01-15. https://today.yougov.com/technology/articles/51368-do-americans-think-ai-will-have-positive-or-negative-impact-society-artificial-intelligence-poll
8. NIST. "AI Risk Management Framework: Generative AI Profile." 2024. https://www.nist.gov/itl/ai-risk-management-framework/ai-rmf-generative-ai-profile
9. Noy, Shakked, and Whitney Zhang. "Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence." Science. 2023-09-08. https://www.science.org/doi/10.1126/science.adh2586
10. Brynjolfsson, Erik, Danielle Li, and Lindsey Raymond. "Generative AI at Work." NBER Working Paper 31161. 2023-04. https://www.nber.org/papers/w31161
11. Sajjadi, Seyedmohsen, et al. "Explainability Does Not Mitigate the Negative Impact of Incorrect AI Advice on Human Decision-Making." Scientific Reports. 2024-01-18. https://www.nature.com/articles/s41598-023-50507-5
12. OpenAI. "How your data is used to improve model performance." 2025-04-28. https://openai.com/policies/how-your-data-is-used-to-improve-model-performance/
13. OpenAI. "Model deprecations." 2025-06-10. https://platform.openai.com/docs/deprecations
14. OpenAI. "How to make your completions outputs consistent with the new seed parameter." OpenAI Cookbook. https://cookbook.openai.com/examples/reproducible_outputs_with_the_seed_parameter
15. World Economic Forum. "The Future of Jobs Report 2025." 2025-01. https://www.weforum.org/publications/the-future-of-jobs-report-2025/
16. OECD. "Is generative AI a general purpose technology?" 2025-01. https://www.oecd.org/innovation/is-generative-ai-a-general-purpose-technology.htm
17. IMF (Kristalina Georgieva). "AI Will Transform the Global Economy. Let's Make Sure It Benefits Humanity." IMF Blog. 2024-01-14. https://www.imf.org/en/Blogs/Articles/2024/01/14/ai-will-transform-the-global-economy-lets-make-sure-it-benefits-humanity