I Dropped Git for Speed (and AI Let Me Get Away With It)
I recently did something I preach never to do: I dropped Git for speed. On an already-built project that only needed quick fixes, I skipped my sacred git → CI → GitHub flow and deployed straight to production, losing all tracking. The interesting part isn't that I did it — it's that AI made the shortcut not just possible but rational. And coming from someone who considers Git non-negotiable (I wrote about that here), that's the thing keeping me up at night.
TL;DR
- Under pressure I used an instant deploy instead of Git+CI: blazing fast, but blind.
- AI shifts the math: often delegating everything to the model costs less than inspecting by hand.
- The numbers tell of a changing era: Stack Overflow is collapsing, AI-assisted code is exploding.
- Security: deploying with no tracking exposes you to every kind of attack. It's only acceptable if you know what you're doing.
- But it's a double-edged sword: AI raises quality or multiplies chaos. It depends who's driving.
The confession: instant deploy instead of Git
Development was already done. It only needed debugging and tweaks, fast, without even time to breathe after months of fighting with code, tests and bugs. So, for the first time in years, I set aside the workflow I defend tooth and nail and switched to an instant-deploy tool: write, save, it's live. Hyper-fast, zero friction.
The problem is you lose all tracking. No history, no "go back to yesterday's version", no record of who-changed-what. I felt lost, like I'd gone back twenty years. And it's exactly that "back" that bothered me the most.
The nightmares I can't shake
Because that instant deploy dredged up memories I thought were buried. The mega-servers hosting hundreds of sites over FTP, the whole team with access, everyone editing files live, in production. Project trees on the server that never matched the ones on local machines. Syncing done with rsync — efficient yes, elegant no, but still better than pure chaos.
Git came along and brought order to all of that. Going back to deploying "blind", even with modern, lightning-fast tools, tastes like putting a foot back in those years. The difference, today, is that I have an ally I didn't have back then.
The numbers: a changing era
It isn't just a feeling. Two trends, opposite and connected, describe the shift underway.
| Signal | What the data says | Period |
|---|---|---|
| Stack Overflow questions | Sharply down: in April 2025 posts were roughly 64% below the year before, and over 90% below the 2020 peak | post-ChatGPT (Nov 2022 →) |
| AI adoption among developers | 80% of new GitHub developers use Copilot in their first week | 2024 |
| AI-written code (big tech) | Google: from 25% (Oct 2024) to over 30%; Microsoft: 20–30% "written by software" | 2024–2025 |
| Code quality | Duplicated ("cloned") code blocks rose from 8.3% to 12.3% between 2021 and 2024 | 2021–2024 |
The message is twofold and uncomfortable. On one side, the place where for fifteen years we went looking for answers is emptying out: AI answers first, and in private. On the other, that same industrially-generated code is showing the first signs of quality erosion — more copy-paste, less refactoring. For the first time, in 2024, copied code overtook refactored code.
The anti-pattern that's saving me
Here comes the part that makes me feel like a heretic. With AI's "superpowers", paradoxically, it's often faster and cheaper to delegate everything to the model — hand it the problem, let it fix, iterate, retry — than to manage it by hand: read the logs, inspect the variables, reason line by line. Getting it wrong costs almost nothing, in time and money; far less than an hour of a person's work. So you iterate, and iterate, until it works.
I know it sounds like an anti-pattern. In a collaborative environment I'd never do it: no tracking, no review, opaque decisions. But on an already-built project that only needs debugging speed, right now it's saving me. The trick is knowing exactly which of the two worlds you're in.
| When the shortcut makes sense | When nobody should touch it |
|---|---|
| Already-built project, fixes-only phase | A system to build or to grow |
| Working solo, full context in your head | Teamwork, many hands on the same code |
| Reversible, low-cost mistakes | Critical production, data or payments |
| You just need speed, today | You need maintainability, for years |
The hidden price: security
Let's be blunt: deploying live, with no tracking, is awful for security. It exposes you to basically every kind of attack. No review to catch a secret committed in the wrong place, no history to understand when and how a hole got in, no clean restore point. It's the exact opposite of "prevent rather than cure".
And yet, in an early, contained phase, it really can be a gift from the sky — on one condition only: that whoever does it knows exactly what they're walking into, knows what to do and why they're doing it. It's not a tool for someone coding at random. Speed without competence, here, isn't a shortcut: it's an accident waiting to happen.
The double-edged sword
The truth is AI cuts both ways. It can raise quality — depending on who uses it, on experience, on the rules you give it, on how well you know the tools and the "vibe coding". But it can also generate far more chaos, genuinely much more, precisely because it lowers the cost of producing code without lowering the cost of maintaining it. Same tool: what changes is who holds the handle.
Where I land
I don't have a clean thesis to sell you, and be wary of anyone who does. My honest position is this: Git stays sacred for collaborative work and for any system meant to last. The shortcut I took is a conscious exception, on one project and in a specific context, not the new rule. AI hasn't made Git useless: it's made it easier to pretend it is. The discipline still matters — it's just that now it has to be chosen, because the tool offers you, in one click, the most convenient way to throw it away.
FAQ
Does AI make Git obsolete?
No. AI makes writing and fixing code faster, but it doesn't replace traceability, collaboration and rollback. If anything: the more code you produce quickly, the more you need a system to know what changed and to go back when something breaks.
Why is Stack Overflow declining?
Because many developers ask ChatGPT, Perplexity or editor-integrated assistants first, getting immediate, contextual answers. Stack Overflow's own data shows a collapse in question volume that accelerated after ChatGPT launched in late 2022.
Is delegating everything to AI a good idea?
It depends on context. On an already-built project that only needs fast, reversible fixes, it can be rational. On a system to build, to maintain, or shared across a team, no: there you need understanding, tracking and discipline, otherwise you pile up technical debt at record speed.
If AI is making you move fast but you fear you're piling up chaos, that's exactly the problem I help keep under control: let's talk or see what I do as a Fractional CTO. And if you missed it, here's the flip side: why I use Git as a litmus test when I judge a team.
Sources: Eric Holscher — Stack Overflow's decline, The Pragmatic Engineer, GitHub Octoverse 2024, Fortune — Google AI-generated code, GitClear — AI Copilot Code Quality.
Related articles
- Why I Use Git as a Litmus Test When I Judge a TeamVersioning discipline is still the fastest way to tell whether a dev team is serious. Here is why — and where Git actually comes from.
- How I put the Italian Constitution and codes into a knowledge graphThe Constitution and seven codes (criminal, civil, procedure, road, consumer, insurance): ~6,445 articles ingested one by one, with ELI identifiers and 86,672 authoritative amendment relations from Normattiva. Here’s the method, the numbers and the real limits.
- I downloaded EU law and added it to the Italian-law graphFrom the EU Publications Office bulk dump (56,292 acts) I extracted and linked thousands of European statutes — regulations, directives, decisions — to the Italian-law graph, with CELEX and ELI. Here’s the method, the real numbers and what’s still missing.