Exploring the human factors that make software engineering so unique, so difficult, so important and all consuming. Learning to work with the systems, not against them.
The Human Software 285 - When does AI help, and when does it hinder?
|
Last couple of weeks, I've been rebuilding some Windows base images in order to comply with corporate patching policies. The new images are CIS hardened which means they follows guidelines set out by the Center for Internet Security benchmark. This ultimately means that the images are restricted in what they can do, what they can access, what is installed upon them by default. These security measures work in opposition to the automation we already have in place for our customers. This is the natural tension of security that we see in any form of software. Users need software to be useful, but adding unnecessary features and complexity increases the likelihood of exploitation. As an experiment, I asked Github Copilot to do the work for me to fix the problems with the Windows images. Copilot itself uses a variety of different LLM models (ChatGPT, Claude and others). None of the models could really help fix any of the weirdness that we saw happening. Essentially, the code suggestions it made cost me more time and effort in the long run and it disappeared down rabbit holes of code complexity which were actually dangerous in themselves. Eventually, I took to LinkedIn to vent my frustration. Full post below but here's the closing remark. Config management is where you need to be forensically accurate and focus on repeatability and a complete lack of surprise. Unsurprisingly this is where LLMs really show their limitations. Despite LLMs sucking for certain tasks, it definitely excels at quickly and effortlessly spotting patterns and explaining things to you that should be obvious. For example, as we near the end of the year I've been organising my thoughts. I asked ChatGPT to analyse my archive of blog posts and provide me with an an overview of what I like to talk about. It not only did that and it also drew me a picture to illustrate the themes I like to write about. This is not bad. It's not perfect, but I do like to talk about a lot of these things. Alongside what I write about, I also wanted to know more about how I could more effectively reach a wider audience. I'm experimenting with social media - for example I post quite a lot on LinkedIn and increasingly on BlueSky and now Instagram too. So likewise when I asked ChatGPT what kind of a newsletter this is it told me a few home truths about the kind of writer I am and what I should do more of. You are not a “5 tips to improve your sprint planning” type of writer. You’re a reflective, human-centered, philosophical, narrative-driven writer. So trying to force yourself into “value delivery content” is like asking a poet to publish “10 Productivity Hacks.” ChatGPT understands me. Should I be worried? Until next time. p.s. to celebrate Black Friday, I’ve introduced a special limited-time price across all retailers. If you've not picked up your copy of "Human Software" yet, be sure to grab a bargain! When to Say Stop to Copilot’s Unhelpful HallucinationsPublished on November 14, 2025 It’s quite a regular requirement in my day job that we update the Linux and Windows images associated with our EC2 instances in AWS. There’s a platform team dedicated to rolling out company-specific versions of our base images and recently it was, once again, time to update. So I’ve spent a week or so using… Read More »When to Say Stop to Copilot’s Unhelpful Hallucinations
|
The Human Engineer
Exploring the human factors that make software engineering so unique, so difficult, so important and all consuming. Learning to work with the systems, not against them.
January is over. The longest month! I managed to get a break in Chamonix, nominally to do some skiing, but for the most part it was marvelling at the beauty of it all. The world around us. So, I'm in the alps taking a break from the uncertainty of the present and facing my own mortality on the gentlest slopes near Mont Blanc. It's wonderful being up there with friends, I feel very lucky, and I also don't feel any real need to push myself like I would have done in the older days. I went out...
A few weeks before Christmas I asked ChatGPT a series of questions along the lines of "ok, so what next?" I was out of ideas. I was tired. My freelance contract was coming to an end so I was already looking for a new one. Launching Human Software had been exhilarating but exhausting. I'd burned the candle at both ends on social media plus done some podcasts (a few of which are yet to see the light of day) and also put myself in front of bookshops and chased reviews and talked talked talked...
REBRAND ALERT!! So it's been a while since I renamed this newsletter but I feel it's due a slight sidestep following the launch of my book. So welcome to episode 286 overall, but episode #1 of The Human Engineer. Despite me constantly rename this newsletter, over these years the subject has never really varied too much. I talk about software systems and how they relate to human systems. I find my work increasingly focusses on the human side of the this divide - because it is a divide right?...