Microsoft execs worry AI will eat entry level coding jobs Microsoft Azure CTO Mark Russinovich and VP of Developer Community Scott Hanselman have written a paper arguing that senior software engineers must mentor junior developers to prevent AI coding agents from hollowing out the profession's future skills base. The paper, Redefining the Engineering Profession for AI, is based on …
Business leaders are only looking towards their stock price in the current quarter. There's no long term thinking at all. Technology is constantly changing but many corporations went to contractors so they could swap technologies without having to pay for training. Short term thinking.
They will do the same here. Avoiding hiring EIC developers will cut labor costs now. And as you say, expect advances in AI to reduce demand for senior coders as they retire. The faster the better.
Crapilot will put it tegether for you.
You can dispense with stableboys and strappers if you "bolt the horses to the stable door"
The ensuing bloody shambles resemble much of what emerges from Microsoft's stables nowadays.
The only reason I know LLMs generate bad code is because I have <mumbles> years of experience generating it myself the old fashioned way.
The entire value that LLM sellers claim is "cut headcount" buy at the moment the only headcount it could possibly remove it the graduate level codes at the the likes of Infi, Accenture etc.
Without having expertise in doing something yourself how can you be sure someone or something did a good job
[/rant]
I think it's a bit like using a code generator or a compiler. If in these programs there's a bug you have to look at the output and see where's the error. And you have to know the target language or the assembly. With LLM you have the added problem that they aren't deterministic and you can't reproduce the bug and worse it could generate different bugs at every iteraction.
You can use a feedback loop and tweak the model until it generates a better result, but you can do this only if you know the correct answer anyway.
When an AI, even the almighty Opus 4.6, encounters a problem that is not well documented on the internet and therefore was not part of its training data, it will run circles around the developer without ever actually implementing what the developer asks. It will hack, tweak, and adjust irrelevant things, but never truly solve the problem.
This is because the model has no real understanding of what the developer is doing. It only predicts the next token.
The whole idea of prompting is that you must phrase things in a way that nudges the model towards predicting what you want. You cannot simply say “draw me an owl” if it has never seen an owl before. But if it has seen feathers, you can slowly guide it from there. By the time you finally get the owl, you could probably have drawn it yourself.
Yes, and even the "good" output is very rarely up to my standards of production-quality. AI is still a very helpful assistant in many cases, but it's just that: an assistant.
It's not ready to be a worker yet, and I don't think we're anywhere near as close as a lot of people want / assume us to be. I think some fundamental problems need to be overcome first, and the time spans for solving problems of that magnitude tend to be measured in decades. That doesn't mean it won't happen tomorrow; it just means that it's lottery-odds unlikely.
This is not a complicated problem to understand. Some of the basic threads.
- LLM training is dependent on pre-existing knowledge. No people, no new knowledge.
- As people retire and are replaced with AI, this means, again, no new knowledge. No new knowledge means that you are creating a knowledge ceiling.
- AI isn't "actual intelligence". It's an output of known data that can handle the imprecise input of human beings. It's only as good as both the input and the output.
If it gets implemented as another productivity tool then it has a realistic future. Otherwise, it's just a plan to fail.
"AI is increasing productivity only for senior developers while reducing it for juniors – is a "hot topic in all our customer engagements... they all say they see it at their companies.""
This is exactly what has been reported in the various stories in 'El Reg' over the last 6 months at least.
'AI' can help the experienced devs because they know what is 'good' & 'Bad' code based on real knowledge.
New devs using 'AI' tools will not know when their level of knowledge/experience is actually a 'stick to beat their own backs'.
'AI' is very good at 'plausible' because the 'interWebs' is full of 'plausible' but bad code.
It is not deliberately there trying to cause issues, it is simply some people will not see the issues with their code yet publish for all to see and for 'AI' to 'learn' from.
Analogy:
No matter how hard I try, I will not be able to teach someone to speak good conversational French IF I cannot speak conversational French myself.
The 'AI' is making the 'best efforts' to do what it is asked BUT IF I cannot works at a higher level I will not see 'ALL' the issues with the answers.
This is exactly what is happening now ... lots of people with limited skills trying to 'bootstrap' their skills without the ability to see when they are learning bad ideas or NOT seeing the flaws because the 'AI' code is beyond their own abilities.
'AI' looks like magic BUT it is actually digging a deeper and deeper hole as we are deskilling the new devs as the older more highly skilled devs are priced out of the market OR see where the wind is blowing and are jumping ship before they are pushed.!!!
:)
Even if a junior dev manages to figure out the LLM is generating crap and navigate through why the LLM is generating crap, their own skills are completely neutered. They're just the maid for the LLM and have no idea how any of this crap works. I see it every day. We have one guy who had the brilliant idea to do a config generation app for [internal stuff I can't say, but it's annoying and complicated]. If some of the network guys had this tool it would save a lot of time. So he decided to make one with Claude.
It's 6 months later and it still breaks on normal use cases every single time he has a release candidate. He has no idea why so he goes back to Claude with 'no no this isn't working please fix this' and Claude regenerates the code. And breaks two other things. Of course he has no idea how to write unit tests. He asked Claude to do some and it wrote some hilariously trivial stuff that doesn't really test anything - his broken code passes the unit tests just fine but always breaks when anyone tries to do anything with it. And he doesn't have the skills to go in and see 'oh it did this wrong, maybe I can tweak this...' He's allowed to completely waste his time like this because one of the VPs is 'oooo, AI!' FWIW I think I could write the app, go through a couple rounds of testing and adjustments and confidently document it and deploy a 1.0 in about a week of work (two weeks actual time because I wouldn't be doing it full time, and waiting for tester feedback).
So this kid isn't learning shit about software engineering. He's just a sloperator poking at a black box. And that's our new generation of junior software engineers.
This is the same "engineer" mindset of some of the electrical installation students I encountered.
"So, join all the red wires together?"
No, the incoming live connects to the outgoing live, the returning switched live connects to...."
"So, (desparate voice) join all the red wires together......?"
If you don't understand what you're actually doing, your tools are useless.
1990 - remember IBM and AD/Cycle?
Now....that attempt was dismissed by many as "marketing"!!!
Plus ca change....plus c'est la meme chose!
Link: https://www.georgeschussel.com/wp-content/uploads/articles/XD1120050411_promise%20of%20ad-cycle.pdf
AI is so expensive AI companies can probably stay afloat only if they take a large cut of peoples' salaries - which means people will be fired or not hired, so companies can increase profits (and in turn, being locked-in by AI providers, but they can't see that yet).
Companies won't pay expensive AIs in addition to the costs of wetware, even if productivity increases, because increased poductivity may not only mean higher profits. Even if you can release new software faster, your customers won't pay more, nor is assured you can find enough more customers. The plan is to cut costs adopting AI - after all that's been the main procedure to increase profits easily - witout good new products - and someone working at MS should know it.
But in that time, I've witnessed 'developers' slowly morphing from a bunch of self-confessed nerds who got into their profession out of sheer interest in anything to do with computers to a group largely made of people who took some coding courses because they were told it would lead to a good career.
There are some good people in the latter group (and the first group still exist, albeit in fewer numbers), but it's becoming increasingly rare to find the few who actually understand what their code is doing below the level of the language, or know their way around a shell prompt, or understand the hardware. The few that succeed do so because they form an interest in the quality of their work and learn from their own mistakes.
The worrying thing is that the LLMs will stop this cycle. They aren't really capable of self-improvement. The output they make that actually works does so because it's basically copying some training material that was created in the first place by somebody competent. If people coming through the ranks in junior positions start to lean heavily on the LLMs, they'll never develop the understanding needed to become competent senior devs themselves. If, long term, the population of competent senior developers starts to dwindle as a result, this will mean fewer 'good' examples of code for the LLMs to consume. Leaving the cycle to its ultimate conclusion basically guarantees enshittified code everywhere that nobody understands.
In that similar time I've seen passionate talented enthusists who have been steered into those "do coding, get a good career" courses come out of the other end into "IT labouring" office IT jobs, resetting passwords and changing printers - it's a computer, that's IT, this is IT!!!!, stop complaining - blinking bewlidered, 50 grand in debt, wondering why their skills have been ignored, then going off to stack supermarket shelves or wash dishes, doing their talented coding in their own time. Complete. Utter. Waste.
Two publicans in my favorite pub were working for an IT consulting company. One of them has a PhD by the way. When layoffs came, they decided to invest the severance and open a pub. It worked for them, and they are happy with the new job. Maybe it's easier to deal with drunk customers rather than project managers.
What could happen with AI displacing juniors at least in the mind of the managers, it that people, ene who is passionate about computers, will stop going to study computer science and IT related stuff, but will pursue other careers, keeping using computers as an hobby. Finding competent people will be difficult.
The last item in that sodden list of agentic bad code brought to mind a story from my early days as a nascent programmer -- in fact the very first assignment in my very first programming class in college in the middle 1960s.
In it we were supposed to do some very simple operations on the machine's registers1 and for some reason I struggled with producing the expected result from a shift register.
Running out of time, I simply coded the tiny program to write out the expected answer, fully expecting that I would take my lumps and get zero points for that question.
When the assignment papers2 were returned I was somewhat shocked to find that the student assistant who did the grunt work grading of assignments for the professor had given me full marks, as the Brits are wont to say, having obviously only looked at the answer and not the rather slothful cheat in the code.
At that moment I became enlightened.
_________________
1 In those days, introductory courses often meant learning about the bare metal of the machine, eschewing such frippery as assemblers, languages, be they high and low level, and (who needs `em?) operating systems.
2 No fancy electronic submissions back then.
"In it we were supposed to do some very simple operations on the machine's registers and for some reason I struggled with producing the expected result from a shift register."
Do new coders/devs still start at the very begining like this ???
Knowing what the hardware does is useful.
Coding in assembler teaches you 'what really happens'.
It teaches you the difference between 'sloppy' code and well thought out design.
:)
Let's check the list...
1. Significant bugs
2. Implementing inefficient algorithms
3. Duplicating common code throughout the codebase
4. Dismissing crashes and hangs as not relevant
5. Leaving debug code behind
6. Making code work for specific tests but not generally
Jesus Christ...we've crossed the "Indian Outsourced" line...
why?
Why bother doing 3 years of a programming degree when AI is taking all the jobs anyway, and companies likel microslop* are cutting entry level jobs in order to save money(and then splurge it on AI)
Whats the point of years of student loan debt you'll never pay back on your min wage job cleaning toilets** while never using the stuff you were trained with?
Then you get companies complaining that theres no skilled workers out there........... well you wont help train the entry level folks so duuuurhh
Best advice I have for school leavers, get into grave digging/undertakers, or accountancy*** or tax inspector.
*microslop...... I like it,and it kinda fits
**nowt wrong with cleaning toilets tbh... just wish some people learned to flush.. jeez
*** oh our accountant said there AI coming for his job too... so thats out
> Why bother doing 3 years of a programming degree when AI is taking all the jobs anyway, and companies likel microslop* are cutting entry level jobs in order to save money(and then splurge it on AI)
Because having a degree still lets someone claim they are a better person than someone who doesn't have one. Humans are all about hierarchies and putting others down.
AI will make toxic credentialism worse. If an LLM can replace all of us, then it becomes even more important to have that credential to be "better than" the competition.
