Let’s play a fast recreation: I’ll present a graph and attempt to guess what it’s about.
No, it isn’t a crypto coin crashing a couple of hours after being minted. And never, it is usually not an oscillatory/wavy graph made with pure CSS, however a harsher reality.
I already gave it away with the title, however it nonetheless hits like a ton of bricks to know it’s the steep decline within the variety of questions requested on Stack Overflow. You possibly can see its peak round 2014 with greater than 200,000 questions requested in a single month. However now in 2026, it’s struggling to even hit 3,000 questions a month.
We don’t need to be consultants within the subject to search out out the wrongdoer. You guessed, it’s AI… largely.
Whereas AI is painted because the Stack Overflow killer, the reality is Stack Overflow’s downfall began lengthy earlier than ChatGPT’s launch in late 2022.
By group accounts and in addition from private expertise, moderation since its peak in 2014 has been (and nonetheless is) one of many main causes for the dearth of questions.
As the location grew, Stack Overflow wanted a greater technique to average the tons of of hundreds of questions requested each month: the inevitable wall that forum-based communities hit after they scale past a sure level. There are a number of methods to attempt to clear up this, however the route Stack Overflow took may not have been the most effective:
On Stack Overflow, we shut or delete questions that may’t be answered right away – it’s not very sociable, however it scales splendidly.
It’s clear Stack Overflow wasn’t specializing in the amount of the questions however somewhat on the standard of them, whereas avoiding duplicates as a lot as potential. This sample was in favor of Google searches for questions that have been already answered and, therefore, residing on pre-answered questions as a substitute of on customers making new or duplicate ones.
It wasn’t useful both how the group appeared to shut upon itself, making it tougher for freshmen to even ask a query. And in the event you’re like me, you most likely need to inquire with out being informed you’re silly, as if getting punished for desirous to study.
Generative AI was the ultimate nail within the coffin. I can’t complain about this, as AI seemingly gives the identical solutions with out judgment (in truth, perhaps an excessive amount of encouragement) nor delay, so I can see why individuals would possibly favor asking an LLM as a substitute.
Nevertheless, as I dug deeper into this, my concern was not about simply Stack Overflow, however the tech ecosystem at giant. Questions like, are we nonetheless asking questions? Are we nonetheless searching for to be higher? Or can we all depend on LLMs, and solely on LLMs, for recommendation? That saved ringing in my thoughts as I continued my analysis.
I consider that, past the autumn of Stack Overflow, these questions linger greater than ever. How AI has usually impacted our workflow, how we are able to use it in problem-solving, and what we are able to do about this as builders.
Drawback-Fixing and AI
Is AI a greater programmer than you? What makes a programmer higher than others is as subjective because it will get, however some are desirous to say that AI can write code higher than you. In accordance with that analysis:
AlphaCode achieves human-level downside fixing abilities and code writing capability as proven by efficiency in programming competitions.
A minimum of that’s when it was examined in opposition to Codeforce’s (a web based code competitors website) issues, the place I admit it will probably and can carry out higher than your common programmer. However most builders don’t care about Contest issues past a technical interview; they know being a software program developer is a lot greater than that.
AI writing high quality code is a particularly nuanced subject and lacks a decisive conclusion. Nevertheless, in the event you take the time to analysis, you’ll discover that AI-generated code has numerous flagrant variations. In accordance with the analysis from Cornell:
AI-generated code is mostly easier and extra repetitive, but extra vulnerable to unused constructs and hardcoded debugging, whereas human-written code displays better structural complexity and the next focus of maintainability points.
Okay, so it will probably generate easy code, however can it write good code? Even clear up issues higher than a software program engineer would?
In accordance with MIT analysis, AI can write good code, however it can not probably suppose and make choices like a software program engineer. AI can not compete on that degree but, at the least with out operating into loads of bugs.
Drawing on each first-hand expertise and suggestions, if all you do is copy-and-paste AI-generated code with out cautious consideration, you’re sure to hit severe bugs and probably even vulnerabilities. The truth is, VeraCode revealed an article stating that “[…] 45% of AI-generated code incorporates safety flaws,” after testing for safety vulnerabilities in 100 AI fashions. That’s a big share of code that’s flawed security-wise and would have price implications for any consumer who desires to “vibe-code” with out doing thorough checks.
Enjoyable reality: GitHub launched the outcomes of its AI in software program improvement survey in August 2024, and over 97% of its respondents have used AI outdoors or inside their work. That’s even apart from the businesses implementing using AI in your present code workflow. It’s actually in every single place; there’s nearly no escaping its utilization
However, does that imply it’s all unhealthy? The reply to that, in my view, is not any. In accordance with analysis performed by Harvard Enterprise Evaluate, AI is efficient for serving to clear up issues (let’s not additionally ignore the trade-off from the research that AI workflows lead to much less motivation). In essence, it’s maybe greatest used to reinforce problem-solving effectiveness.
Which means that, as AI is taking on industries and being integrated into our each day work, it nonetheless gained’t substitute your creativity and problem-solving strategy, which you’d must sort out distinctive on a regular basis challenges. It’s troublesome to copy.
Like each different software, AI has its limits, and with out human craftsmanship behind it, the software is sort of ineffective. craftsman makes use of all of the instruments at his disposal to realize his objectives, AI being simply one among them.
“The effectiveness of the software is decided by the ability of the craftsman who created it and the ingenuity with which he makes use of it.”
Craig D. Lounsbrough
The large hazard is not only safety vulnerabilities, however over-dependence on the software, which I consider will result in an eventual decline within the variety of code craftsmen within the coming era. How ought to newer and skilled builders go about this?
Some Recommendation
Here’s a record of questions I ask myself when choosing up AI in my improvement work:
- Am I asking the LLM smaller, particular questions? This fashion, I can confirm every course of step-by-step somewhat than eyeballing the entire system code as an entire. I’m nonetheless a developer within the sense that I’m not leaving the LLM to do all the work.
- Am I evaluating the output when it’s completed? In different phrases, do I perceive what it did? Would I be comfy modifying the generated code if I do know a greater strategy, or when I’ve to keep up it sooner or later?
- Am I checking the software’s references? This can be extra geared in direction of analysis as a substitute of straight code output. The place precisely are its solutions coming from? Are these good sources? Are there others? It’s vital to know the software isn’t citing a fictional supply, however somewhat, developing with trendy and tried-and-true approaches.
- Have I examined the work? Did the software perceive the duty and contemplate all edge instances? That is maybe crucial query as a result of understanding how individuals use your software is one thing a machine is much less inclined to know than a human.
What occurs once we cease asking?
Take into consideration this: if we cease asking questions, how will AI be skilled sooner or later? Applied sciences change and enhance over time. What’s up to date now will quickly develop into old style. Take CSS, for instance. With the current CSS updates (nesting, view transitions, container queries, and many others.), we’re writing CSS vastly completely different than even a couple of brief years in the past. You wouldn’t need to be caught with an outdated and clumsy answer skilled from code written many years in the past. If we cease asking questions and answering them, don’t you suppose that will make the LLMs lag behind? That’s simply me speculating, however I believe it’s straightforward to think about that being the case.
We can not deny Stack Overflow’s service over time. It acquired us asking. It acquired us answering. It acquired us pondering. The query we should always all ask ourselves is,Will LLMs do the identical?
I’ll go away you with this quote from Stack Overflow co-founder Jeff Atwood:
Stack Overflow is you. That is the scary half, the nice leap of religion that Stack Overflow is based on: trusting your fellow programmers. The programmers who select to take part in Stack Overflow are the “secret sauce” that makes it work.
