engineering area, debates about whether or not utilizing LLMs in coding is an effective or a nasty thought are raging. On the extremes, there are some individuals who suppose that any use of LLMs in coding is indefensible, and different persons are able to throw out the entire area of software program engineering as too costly and pointless. I’m fascinated by this dialogue, significantly from my sociologist’s perspective, as a result of a lot of the dialog appears to be about whether an LLM is useful. Is it so fatally flawed that it’s a complete waste of time, or is it so unimaginably highly effective that it’s a idiot’s errand to keep away from it? As an alternative of desirous about it so narrowly, I truly actually wish to speak concerning the broader context of software program engineering within the context of LLM know-how. (I contemplate software program engineering solely adjoining to my area of machine studying engineering, however I’ve labored alongside many extraordinarily proficient SWEs in my profession, at tech firms of assorted sizes, and I’ve had the prospect to look at the function fairly intently. A few of the evaluation beneath additionally applies to the MLE area, however not all of it.)
Context
Throughout my very own profession I’ve seen rather a lot change within the American tech economic system, and in the best way the job of “coding” has been perceived and described. It’s a little bit of a trope but it surely was definitely true for a time period that “studying to code” was touted because the quickest technique to a white collar profession with a gradual paycheck and future-proof prospects, each for younger folks and for career-changers looking for a spot within the shifting economic system. Coding bootcamps came and (mostly) went, for a time giving folks a comparatively fast technique to acquire the foundational abilities of the job.
In lots of fields, nonetheless, shortages of certified staff can swiftly flip right into a glut of expertise with nowhere to go (ask legal professionals), particularly when excessive or at the very least snug salaries are on supply. Whereas some folks suppose that is taking place in software program engineering, it’s not clear if that’s essentially the case but, as a result of there are different elements in play too. Within the mid-2010s, many tech firms and startups had been scrambling for hyper-growth and hiring as quick as attainable, utilizing low cost enterprise capital to fund their expansions with out regard for profitability, however that’s definitively over now. Tech layoffs, together with many software program engineers, exploded within the years instantly following the beginning of the COVID pandemic, largely as a response to the unsustainable hiring of previous years and the sudden jolt to the economic system of lockdowns. After that interval of layoffs, hiring for SWE expertise slowed in lots of sectors. Developer skillsets are diversifying and specializing, so there should still be sub-fields with robust hiring, however plenty of younger SWEs particularly are struggling to interrupt in.
Junior vs Senior
This brings us round to now. Some folks argue that there aren’t too many software program engineers, however there are too many inexperienced or junior software program engineers, and that employers are nonetheless determined for knowledgeable hires. What does that basically imply? Is it true? With out doing tutorial analysis to actually examine (if anybody has carried out good analysis on this query, please ship it to me!), we are able to nonetheless check out the problems concerned. First, let’s outline phrases.
I’m not utilizing anyone’s in-house leveling phrases, or speaking about job titles right here, however I wish to make clear what the variations is perhaps in somebody’s function relying on the expertise and abilities they bring about.
A junior SWE is somebody who wants fairly a little bit of assist, teaching, and managing to supply good high quality work. They’re not but at a spot of working independently or mentoring others. They’re nonetheless studying how the job is completed. This doesn’t imply they don’t produce worth! Junior engineers do write code and get work carried out, and so they can typically study in a short time and change into extra productive. Nevertheless, a senior SWE is somebody who will be anticipated to work largely independently, with minimal supervision required for good work to be carried out and objectives to be met, and they are often relied on to assist these junior people alongside. They begin to have the ability to remedy harder issues, write elegant code that is top of the range, and even start being a part of technique conversations about the best way to construct your complete codebase for the lengthy haul.
This is kind of the sample for many expert professions — you’ll study some quantity of what that you must know from education or coaching, however you actually learn to be a member of the career and do the job from expertise and observe. There’s no set rule about how a lot time it takes so that you can go from junior to senior on this framework, as a result of it’s all about what you are able to do and the capabilities you will have. Your perceived worth as a employee (and your pay) go up as you get higher on the job over time.
LLMs Enter Stage Left
With this, we are able to begin to inquire about how that shifts, and the way the existence of LLMs and LLM primarily based coding instruments truly might (to make use of a cliche) disrupt the context of the career now.
As a person, should you’re a senior SWE, you may discover an LLM device like Github Copilot or Cursor helpful for doing scut work, resembling getting unit assessments working, writing fundamental docs, or sketching out the beginning of a challenge that you simply’ll then fill in primarily based by yourself information. Nevertheless, your individual information is absolutely important for this to be efficient. As we in machine studying will let you know, LLM output is probabilistic at its core. If the coaching knowledge has many examples of code which are related to what you’re doing, it’s going to be fairly prone to produce code for you that works. However eventually (in all probability sooner) it’s going to generate one thing that received’t work. That is perhaps failing loudly or failing silently, it is perhaps opening up a safety gap, or it would simply be a suboptimal method of doing a process. The factor that issues at that time is what you do in response.
Assuming you’re a senior SWE with important expertise writing all these items by hand your self, you will have the abilities to identify the errors, and work out the best way to repair them easily so the top product works nicely. The LLM in all probability makes you extra productive in the long run, since you might summary away a number of the work that isn’t fascinating and doesn’t require a lot thought. In case you’re not a senior particular person, nicely, then it’s possible you’ll find yourself going again to the LLM to ask it to repair its personal errors. And it would try this, however there’s all the time a danger of one other mistake being launched, and on down the highway. There’s actually no telling how lengthy this might take, and it’s arduous to match the time with how lengthy it might have taken you to study the underlying code and simply remedy it your self.
Nevertheless, numerous folks will argue that the LLM repeating workflow is ok. Individuals might imagine that simply being there to click on the hover button to inform Copilot to “Repair” over and over is ample, even when it’s not a very satisfying job. Nevertheless, there are critical arguments to be made that the output can be significantly flawed, whether or not in readability, maintainability, safety holes, or different points.
“Certain, Gen AI supercharges growth, but it surely additionally supercharges danger. Two engineers can now churn out the identical quantity of insecure, unmaintainable code as 50 engineers.” — https://www.zdnet.com/article/10-professional-developers-on-vibe-codings-true-promise-and-peril/
This query has been addressed by many writers lately. Nevertheless, I’m much less at this second in whether or not Copilot can get to the specified end result finally than I’m in what this course of and method of working does to us.
Broaden the Aperture
Keep in mind that we’re right here to speak concerning the impact of the LLM on the career. Within the pre-AI period, a junior engineer would begin engaged on a ticket, and make some errors, introduce a bug or two, and so they’d mainly be armed with StackOverflow and the docs, plus friends to collaborate with for assist, to determine the best way to get it proper. As soon as they muddled by means of and obtained the code working, they’d make a PR and get suggestions from extra skilled friends, who would supply suggestions and corrections. As soon as these corrections had been made, a PR might be authorized and merged.
I feel the “muddling by means of” a part of this course of is definitely pedagogically and professionally fairly vital. When you’ve handed by means of the educational preparation for the software program engineering function, the sensible preparation is what that you must develop the capabilities to develop. I’m a powerful believer in the concept making errors and fixing them is a tremendously beneficial technique to study. As soon as the junior dev engaged on that little bit of code has labored arduous, fastened their very own bugs, written the assessments and docs, and gone by means of PR evaluate, they know that little bit of code higher than some other particular person round. That course of is the productive cognitive labor that brings them a step nearer to being an engineer who can work independently, and tackle extra senior duties efficiently.
On Being Environment friendly
Nevertheless, from the surface, this course of seems to be gradual. It seems to be like they’d to spend so much of time flailing round to simply get to the top results of code that works, and absolutely it might be extra environment friendly to both simply make an LLM do it, or make a senior dev do it?
If effectivity is your solely standards, then positive, possibly that is true. However I wish to push again strongly on the premise. We do numerous issues with out desirous about broader context or influence, particularly within the tech business, and prioritizing effectivity immediately in any respect prices is a recipe for future catastrophe. Let’s spin out the narrative a bit — if effectivity is the precedence, both we now have an LLM do the majority of the work, or a senior dev does it, or possibly a little bit of each, like an LLM does the primary stab and a senior dev critiques it. Setting apart any questions of high quality (about which there ought to be questions), what’s the function of a junior SWE on this area?
Does a junior SWE watch the LLM do a lot of the coding? Do they only exist as an observer? Do they click on “okay” once in a while because the LLM goes alongside, exemplifying “vibe coding”? Or do they not exist in any respect on this state of affairs? Loud voices in the AI space certainly are arguing that they disappear altogether. In that case, within the speedy second, employers in tech save a big sum of money on salaries, so inventory worth goes up.
The thought of vibe coding is utilizing an LLM primarily based device to write down the code in your challenge, with minimal precise hand-writing of code concerned. The identify comes from the precept that you simply behind the keyboard have a normal sense of the “vibe” of what you need your end result to be, however you both don’t know the best way to do it your self or don’t wish to do it your self, so that you’re outsourcing the cognitive work of coding to a pc. You’ll simply describe the broad strokes of what you need, in human language, and let the LLM work out a technique to obtain that by means of code.
Hollowing
Nevertheless, we all know {that a} wholesome career must be bringing in new expertise as senior employees progress and finally retire. Senior engineers are nonetheless being wanted, as a result of they will remedy the issues and write the complicated code that neither junior devs nor LLMs can produce. Their expertise and abilities are undeniably beneficial and essential to good software program being produced. However the place do we predict senior engineers come from? Identical to senior docs, or senior attorneys, they begin out as junior folks. Their capabilities have to be acquired by means of expertise.
If we construct a working atmosphere the place junior software program engineers now not exist, we’re hollowing out the pipeline of expertise the career wants. As an alternative of getting a funnel of expertise from tutorial examine into the sphere, younger folks don’t have any actual technique to get to the extent of expertise that’s in highest demand.
Conclusions
I respect readers coming together with me on this dialogue, as a result of it’s vital to consider the know-how we make ubiquitous not simply from our personal particular person views however from a broader view. How we make use of LLM know-how within the area of software program engineering may have a critical influence on the way forward for the career and what it seems to be like — and, I’d argue, that is true for a lot of different fields as nicely the place incorporating LLMs is changing into modern. I’m not arguing for zero LLM utilization in coding, by any means — it may be extremely helpful when you have the expertise and abilities to profit from it. (There are a lot of vital conversations to have about the way it might influence our ability ranges and the upkeep of our abilities, however that’s a special subject for an additional day.)
However in any career, senior practitioners don’t simply seem — they’re created, and that strategy of creation takes work and time, and will not be probably the most instantly “environment friendly” factor to do. Effectivity within the speedy second simply isn’t crucial factor — it will possibly’t be, if we wish the longer term to be vivid.
I’m talking at two occasions in July 2025, together with a free virtual conference called “Agents in Production” — please join should you’re !
Learn extra of my work at www.stephaniekirmer.com.
Additional Studying
AI Killed My Job: Tech workers
Tech workers at TikTok, Google, and across the industry share stories about how AI is changing, ruining, or replacing…www.bloodinthemachine.com
10 professional developers on vibe coding’s true promise and peril
Is vibe coding the future of software or a security nightmare in disguise? Here’s what experienced developers think…www.zdnet.com
Frontiers | Some Evidence on the Cognitive Benefits of Learning to Code
Please see the submitted files.www.frontiersin.org
Trusting your own judgement on ‘AI’ is a huge risk
Web dev at the end of the world, from Hveragerði, Icelandwww.baldurbjarnason.com
Jones, Capers. 2014. The Technical and Social History of Software Engineering.
Layoffs.fyi — Tech Layoff Tracker and DOGE Layoff Tracker
LIVE] Tracking all tech startup layoffs — and lists of employees laid off — since COVID-19 was declared a pandemic…layoffs.fyi
