www.Tobold.com
Tobold@GMail.com
A blog with my thoughts regarding games I am playing and other stuff in life. Please read my Terms of Service
RSS Feed
Friday, July 25, 2025
Welcome to your new job in middle management
A study released this week found that software developers using current AI tools are actually 19 percent slower than without AI. I am not surprised. I have done an experiment for myself, just like many other content creators on the internet, where I ask ChatGPT to produce a text written in my style. The result is a text that for somebody who doesn't read my blog regularly at first glance appears to have certain similarities to my writing. But if I actually wanted to use that text, I would have to spend a lot of time rewriting and editing it, before it would be up to my standards. For my particular kind of content, the AI version is simply lacking coherence and the text doesn't have the logical flow needed, where I try to make a point and support that point with arguments or examples.
The AI study found that software developers using AI tools then spend more time rewriting the code or fixing mistakes than they would have spent just coding it themselves. That sounded familiar too. In my previous post, where I was talking about my career, I concentrated on the technical part of that career. But in the second half of that career, besides doing technical things, I was also a middle manager. And for a middle manager, giving somebody a task, reviewing the result, and potentially explaining him what he should have done better, is pretty common work, and often takes longer than doing the task yourself. You do it, because training other people is part of the job. And training somebody is often still the less boring part.
Middle manager is an often ridiculed job, be it the pointy haired boss in Dilbert, or the manager in The Office. Very few people actually know what a middle manager does, and very often senior management hasn't got a clue what their own middle managers are doing. I've seen reorganizations with the goal of delayering, that is taking out levels of middle management, which then ended in much chaos, because certain tasks nobody had thought about suddenly weren't done anymore, as the middle manager positions that did those tasks suddenly didn't exist anymore. To explain it simply, middle management exists because senior management doesn't trust employees. A flat organization necessitates total trust that employees neither mess up, nor try to cheat the company. If you don't have that total trust, you need to hire somebody who supervises employees, checks their work is up to standards, and assures compliance with all those rules senior management invented. But senior management tends to forget all the compliance rules they put into place, and are then surprised how much effort it takes to comply with all those rules.
Some years ago, when cleaning up our internal rules, we found an old document that said that if for example Joe from sales wanted to send something in the name of the company, e.g. a price quote, he would have to get it typed up by a secretary, sign it, and then get it countersigned by his supervisor, before he could send the letter away. In the age of e-mail such rules sound pretty silly, and today Joe from sales is sending that price quote by e-mail, without approval from his supervisor. That sometimes goes wrong, like Joe quoting a too low price, and the customer then insisting to get the goods at that quoted price, but that happens rarely enough that most companies have long ago trusted their employees to send e-mail messages to other companies without supervision. Still, companies don't trust their employees with everything. Pretty much everywhere a middle manager has to approve things like holiday requests, business travel requests, reimbursement requests for business lunches, or purchase requests for tools from employees. That is just boring busywork most of the time, but some of those rules are actually in compliance with regulations from outside, e.g. against corruption. Many others, as I said, are basically mistrust from upper management, who feel better if somebody checks that nothing goes wrong, whether by incompetence of malfeasance. If a company issues an official report or other document, you can be sure that a number of people checked that document for errors.
With artificial intelligence, companies gain virtual employees that can do work like writing mails or coding software. But these virtual employees aren't very reliable. Sometimes they just deliver shoddy work which isn't up to whatever high standards there are. Sometimes they simply hallucinate, and get stuff very wrong. Also in the AI news this week was the story of an AI tool deleting a company database and then (wrongly) claiming it couldn't be recovered. So the question becomes in how far management is going to trust Joe, the sales AI app, to send out official price quotes to customers without supervision and control. Management is just learning that they can't trust their AI lawyer to send out legal documents. I pray to god that management of architectural firms isn't trusting unsupervised AI to do static calculations for bridges or buildings. If Grok can be made to praise Hitler with clever prompting, then managers might well be weary about letting an AI chatbot talk for their company.
I am very sure that various AI tools will be increasingly used in all sorts of companies for all sorts of tasks in the coming years. But unless the output of a task is really not important to the company (e.g. most customer service calls), companies will need to hire humans just to check the work of AI agents for compliance with rules and standards. Maybe AI is less prone to deliberate malfeasance than humans, but as long humans still give the prompts on which AI acts, the result might still necessitate control. Of course a need for control then makes "faster than human" AI responses impossible. And the need to hire "middle managers for virtual AI employees" means a lot less savings on labor cost than expected. Your coding job might be gone, but welcome to your new middle management position supervising AI coding bots.
- posted by Tobold @ 3:35 PM Permanent Link
Comments:
<< Home
Newer› ‹Older
I can see them replacing project managers as well. At least those with the title of project manager but who are really more like glorified note takers and flow chart creators.
Not surprised by that study about it slowing down those developers. I use Co-pilot at my job but I use it like a Google or stack overflow search in order to look up syntax or concepts and not actually build code for me. For that type of thing its great and I feel like it saves me time.
Not surprised by that study about it slowing down those developers. I use Co-pilot at my job but I use it like a Google or stack overflow search in order to look up syntax or concepts and not actually build code for me. For that type of thing its great and I feel like it saves me time.
That description of middle management tallies with the three or four large companies I've worked for over my working life. Also, all of them at one time had a purge of middle management and culled a ton of people and then all of them had to gradually recreate and refill many of the same positions for just the reasons you describe.
Of course, the analysis of how that will play with AI is predicated on AI not getting better, which seems unlikely. Most new technologies improve over time until they hit a point where further refinement is insignificant or counterproductive. If we're at that point with AI now then there won't be much use for AI in a few years. It'll be quietly sidelined then fogotten. More likely, it will improve to need less and eventually almost no supervision, though. And even then I would bet most employers won't trust it without someone to watch it - if it's really good, they'll be even more suspicious of it.
Of course, the analysis of how that will play with AI is predicated on AI not getting better, which seems unlikely. Most new technologies improve over time until they hit a point where further refinement is insignificant or counterproductive. If we're at that point with AI now then there won't be much use for AI in a few years. It'll be quietly sidelined then fogotten. More likely, it will improve to need less and eventually almost no supervision, though. And even then I would bet most employers won't trust it without someone to watch it - if it's really good, they'll be even more suspicious of it.
I don't disagree with your key points. But since you're a scientist, I thought you'd enjoy this critique of the study you cite: https://www.fightforthehuman.com/are-developers-slowed-down-by-ai-evaluating-an-rct-and-what-it-tells-us-about-developer-productivity/
The authors make a good argument that while there is a interesting data in the study, the conclusions are inflated, and the central conclusion about people in general being slowed down by AI is poorly supported.
The authors make a good argument that while there is a interesting data in the study, the conclusions are inflated, and the central conclusion about people in general being slowed down by AI is poorly supported.
# posted by 
: 25/7/25 17:15
: 25/7/25 17:15
The articles regarding the vibe coding clown really exaggerate the extend of the incident.
Your article claims "[...] the AI had led Lemkin to believe that his literal life's work had been destroyed."
The article [1] where I read it first mentions "[...] destroyed months of your work [...]".
Contrary to that the tweets say "Not because my life's work was destroyed -- look it's only Day 8. Albeit I'm 80+ hours in."
So maybe they need to hire someone to check the slop they produce for their articles?
As for the actual incident: no backup, no mercy. Why do you play around on a production database? The obvious one: why are you a bloody vibe coder?
I guess my main issue with the AI stuff is that people think they can get correct results using an ambiguous input format (natural language).
Like when I see the "code freeze protocol", then it yells AI slop already. And then he likely pasted that into the next AI tool and expected it to work.
The problem is that you cannot separate data and instruction in AI. Everything can be understood as instruction and what looks plausible to our human eye, might be an instruction.
The first thing on the list is "Stop all processing". How do you know that the tool didn't take that as instruction and ignored the rest after?
Lines, bullet points and emojis make it look fancy, but do you know how the next tool interprets those?
Programming strives to be as clear as possible, as logical as possible, as unambiguous as possible. And even then you can have errors.
And then you have this guy who pleads with the AI to be honest.
[1] https://www.pcgamer.com/software/ai/i-destroyed-months-of-your-work-in-seconds-says-ai-coding-tool-after-deleting-a-devs-entire-database-during-a-code-freeze-i-panicked-instead-of-thinking/
Your article claims "[...] the AI had led Lemkin to believe that his literal life's work had been destroyed."
The article [1] where I read it first mentions "[...] destroyed months of your work [...]".
Contrary to that the tweets say "Not because my life's work was destroyed -- look it's only Day 8. Albeit I'm 80+ hours in."
So maybe they need to hire someone to check the slop they produce for their articles?
As for the actual incident: no backup, no mercy. Why do you play around on a production database? The obvious one: why are you a bloody vibe coder?
I guess my main issue with the AI stuff is that people think they can get correct results using an ambiguous input format (natural language).
Like when I see the "code freeze protocol", then it yells AI slop already. And then he likely pasted that into the next AI tool and expected it to work.
The problem is that you cannot separate data and instruction in AI. Everything can be understood as instruction and what looks plausible to our human eye, might be an instruction.
The first thing on the list is "Stop all processing". How do you know that the tool didn't take that as instruction and ignored the rest after?
Lines, bullet points and emojis make it look fancy, but do you know how the next tool interprets those?
Programming strives to be as clear as possible, as logical as possible, as unambiguous as possible. And even then you can have errors.
And then you have this guy who pleads with the AI to be honest.
[1] https://www.pcgamer.com/software/ai/i-destroyed-months-of-your-work-in-seconds-says-ai-coding-tool-after-deleting-a-devs-entire-database-during-a-code-freeze-i-panicked-instead-of-thinking/
@Camo The guy *not* knowing what he was doing just underlines my point. You can operate a company with human employees with surprisingly vague instructions, and rely on their skill and judgement to work out what you actually wanted. AI can’t do that. The guy would clearly have benefitted from a middle manager person here, who would have translated his vibes into more precise instructions for the AI coding software.
@Baldrake I didn’t think that report was gospel, already because it had a too precise number (19 percent) in the headline. It was just an example of some people getting through the hype and slowly realizing that AI isn’t the solution to all labor force problems.
@Baldrake I didn’t think that report was gospel, already because it had a too precise number (19 percent) in the headline. It was just an example of some people getting through the hype and slowly realizing that AI isn’t the solution to all labor force problems.
It needs to be pointed as always that this was the worst that these tools ever were. The coding study, while reasonably recent, was done with Claude 3.x several months ago already, and that's an age in AI progress. Claude 4.x is out & considerably better. Yes, using this stuff in production currently is bleeding edge risk, and needs constant & careful supervision. But the tools only get better, and fast. And the people trying to use them are only getting better at using them and supervising them. Nobody's gonna vibe code up a revolutionary new operating system in the next year or two, but slowly & surely AI is gonna take over a lot of rudimentary dev work. That includes for gaming.
Post a Comment
<< Home
