{"id":4879,"date":"2026-11-20T18:00:00","date_gmt":"2026-11-20T17:00:00","guid":{"rendered":"https:\/\/msalamon.pl\/?p=4879"},"modified":"2026-11-13T18:15:22","modified_gmt":"2026-11-13T17:15:22","slug":"embedded-developer-in-the-age-of-ai","status":"publish","type":"post","link":"https:\/\/msalamon.pl\/en\/embedded-developer-in-the-age-of-ai\/","title":{"rendered":"Embedded developer in the age of AI: which job is yours?"},"content":{"rendered":"\n<h1 class=\"wp-block-heading\">Embedded developer in the age of AI: which job is yours?<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">In just two years, more than a quarter of the jobs in one programming occupation in the US disappeared. Exactly 27.5%. Meanwhile, in an occupation with almost the same name, in the same country and over the same period, the drop was 0.3%. So practically nothing.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Some internet AI experts say you don&#8217;t need to learn anymore, because AI will do everything for us. They say it especially when they&#8217;re selling their own AI tools, which is quite telling, isn&#8217;t it? \ud83d\ude09 I don&#8217;t fully agree. I still think you need to know your stuff in whatever you want to do with AI. <strong>AI should be your booster, even a turbo booster, not a complete replacement.<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">This article is the text version of the video that opens my &#8220;AI w Embedded&#8221; series. If you&#8217;d rather watch, it&#8217;s all here (in Polish):<\/p>\n\n\n\n<figure class=\"wp-block-embed is-type-video is-provider-youtube wp-block-embed-youtube wp-embed-aspect-16-9 wp-has-aspect-ratio\"><div class=\"wp-block-embed__wrapper\">\n<iframe loading=\"lazy\" title=\"Kt\u00f3ry zaw\u00f3d zast\u0105pi AI? Sprawdzi\u0142em 4 badania\" width=\"750\" height=\"422\" src=\"https:\/\/www.youtube.com\/embed\/Uq_tnWHn1PQ?feature=oembed\" frameborder=\"0\" allow=\"accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share\" referrerpolicy=\"strict-origin-when-cross-origin\" allowfullscreen><\/iframe>\n<\/div><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1600\" height=\"900\" src=\"https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/01-bls-dwa-zawody.jpg\" alt=\"Two jobs in the US: computer programmer \u221227.5% of jobs, software developer \u22120.3% (BLS, 2023\u20132025)\" class=\"wp-image-4865\" srcset=\"https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/01-bls-dwa-zawody.jpg 1600w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/01-bls-dwa-zawody-300x169.jpg 300w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/01-bls-dwa-zawody-1024x576.jpg 1024w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/01-bls-dwa-zawody-768x432.jpg 768w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/01-bls-dwa-zawody-1536x864.jpg 1536w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\" \/><figcaption class=\"wp-element-caption\">Two jobs in the US: computer programmer \u221227.5% of jobs, software developer \u22120.3% (BLS, 2023\u20132025)<\/figcaption><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">AI posters and my apps that &#8220;suck&#8221;<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Design is a great example. People with zero design knowledge jumped on making posters with AI. And now, when I fly literally all over Europe, I see exactly the same posters everywhere. I bet you can spot them at first glance too.<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"900\" height=\"1338\" src=\"https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/01b-mem-plakat-ai.jpg\" alt=\"An AI poster you&#x27;ve probably seen too ;)\" class=\"wp-image-4877\" srcset=\"https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/01b-mem-plakat-ai.jpg 900w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/01b-mem-plakat-ai-202x300.jpg 202w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/01b-mem-plakat-ai-689x1024.jpg 689w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/01b-mem-plakat-ai-768x1142.jpg 768w\" sizes=\"auto, (max-width: 900px) 100vw, 900px\" \/><figcaption class=\"wp-element-caption\">An AI poster you&#x27;ve probably seen too \ud83d\ude09<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">I&#8217;m an example of this myself. I write my own web apps for my company with AI. And sure, AI can write them. Not in half an hour, because I&#8217;ve been writing one of them for over three months now, but it does write them. Except they just look like crap. I don&#8217;t know anything about design, and I don&#8217;t even know how to steer the agent so that my app looks usable and modern. Things I know nothing about end up at the bottom, together with the mud.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">But my opinion isn&#8217;t proof yet. And since I&#8217;ve recently become a so-called internet researcher \ud83d\ude09 let&#8217;s look into it. Relax, I won&#8217;t be doing my own research, because I don&#8217;t know the first thing about that either. We&#8217;ll go through four studies and reports from the last year and a half, plus one big survey. Ones where we know who made them and on what data.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Programmer or developer?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The numbers from the intro come from the US Bureau of Labor Statistics (BLS), reported in March 2025. The bureau distinguishes two things:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Computer Programmer<\/strong>, someone who writes and tests code based on a task and a specification they got from someone else. Let&#8217;s say, a regular code monkey,<\/li>\n\n\n\n<li><strong>Software Developer<\/strong>, someone who designs the solution, defines the requirements and is responsible for the whole thing.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">The first one gets the task. The second one figures out how the task should be done in the first place.<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1600\" height=\"900\" src=\"https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/02-definicje.jpg\" alt=\"A programmer writes code from a specification, a developer designs the solution and owns the whole thing\" class=\"wp-image-4866\" srcset=\"https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/02-definicje.jpg 1600w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/02-definicje-300x169.jpg 300w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/02-definicje-1024x576.jpg 1024w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/02-definicje-768x432.jpg 768w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/02-definicje-1536x864.jpg 1536w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\" \/><figcaption class=\"wp-element-caption\">A programmer writes code from a specification, a developer designs the solution and owns the whole thing<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">For a long time these two jobs were merged, and many people did both the solutions and the code typing. Lately that has been changing, and in embedded you can see it very clearly. There&#8217;s someone who takes an example from HAL or from the internet, copies it, changes the pin number and is happy that the LED blinks. And on the other side, there&#8217;s someone who knows:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>why I2C works on your desk but hangs on the target board or doesn&#8217;t work at all,<\/li>\n\n\n\n<li>why the code works in the debug build but suddenly stops in release,<\/li>\n\n\n\n<li>why the device turns into a brick after an update in the field, at the customer&#8217;s site.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">The first kind of work, typing code from a specification, AI already does today. Faster and cheaper. <strong>Code is cheap today, and there&#8217;s plenty of it.<\/strong> By itself it&#8217;s no longer a big value. AI doesn&#8217;t do the second kind of work yet, but who knows if that&#8217;s just a matter of time.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">To be fair, one thing has to be said. The fact that the employment drop coincided with the release of ChatGPT is <strong>a correlation, not proof<\/strong>. The programming gold rush was ending long before the AI revolution, and at the same time the industry was simply cooling down after the pandemic. But the direction is, I&#8217;d say, pretty clear. I&#8217;ve even heard of programmers retraining from C to C+E \ud83d\ude09<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1600\" height=\"900\" src=\"https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/03-korelacja.jpg\" alt=\"The employment drop coincided with the release of ChatGPT, but it&#x27;s a correlation, not proof\" class=\"wp-image-4867\" srcset=\"https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/03-korelacja.jpg 1600w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/03-korelacja-300x169.jpg 300w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/03-korelacja-1024x576.jpg 1024w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/03-korelacja-768x432.jpg 768w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/03-korelacja-1536x864.jpg 1536w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\" \/><figcaption class=\"wp-element-caption\">The employment drop coincided with the release of ChatGPT, but it&#x27;s a correlation, not proof<\/figcaption><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">Stanford: canaries in the coal mine<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The second study is Stanford, the August 2026 version. The researchers took data from ADP, a company that runs payroll for millions of workers in the US. Month by month, occupation by occupation, broken down by age. And they compared occupations where AI can do a big part of the work with ones where AI can do almost nothing.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The result? <strong>People aged 22\u201325 in the occupations most exposed to AI have 19% lower employment<\/strong> than they would have if it had grown the same way as in other occupations. Among older, experienced workers there&#8217;s no such gap.<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1600\" height=\"900\" src=\"https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/04-stanford-juniorzy.jpg\" alt=\"Stanford: employment of people aged 22\u201325 in AI-exposed occupations is 19% lower\" class=\"wp-image-4868\" srcset=\"https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/04-stanford-juniorzy.jpg 1600w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/04-stanford-juniorzy-300x169.jpg 300w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/04-stanford-juniorzy-1024x576.jpg 1024w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/04-stanford-juniorzy-768x432.jpg 768w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/04-stanford-juniorzy-1536x864.jpg 1536w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\" \/><figcaption class=\"wp-element-caption\">Stanford: employment of people aged 22\u201325 in AI-exposed occupations is 19% lower<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">How it happens matters. Nobody is laying people off en masse. Companies simply hire fewer young people. The door isn&#8217;t closing on those who are already inside. It&#8217;s closing a little on those who are only trying to get in.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">But the most interesting part is something else. The researchers split occupations by how people actually use AI in them. <strong>The drops are where AI replaces a person in a task. Where AI supports a person, employment stays flat or grows<\/strong>, especially among experienced workers.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The authors called the study canaries in the coal mine. Miners used to take a canary underground, because the bird sensed gas first. If the canary died, you knew you had to run. Young people are that canary: changes in the job market show up in their group first. The thing is, this study shows what is happening, not why. The researchers ruled out a few other explanations, like interest rates or remote work, but you can&#8217;t prove from this data, black on white, that it&#8217;s AI&#8217;s fault specifically.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">What I see in my own company<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">And honestly, I see a bit of this myself. For several months my whole team has been working with AI agents. We learn from each other, run internal workshops and use them really intensively.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Before, we couldn&#8217;t keep up with all the tasks we came up with. Today I have the feeling that more things are under control. The agents took over the boring, repetitive work that burns us out the most as people: clicking through systems, moving data from one place to another. Our focus shifted to building systems and more creative things.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The effect is exactly like in that study. For now I&#8217;m not thinking about hiring anyone else. But if I do look for someone, and sooner or later I probably will, I&#8217;ll expect that person to be good with AI, ideally with agentic work. That&#8217;s just my observation, not a study. But I see a very strong shift towards these tools, also when I talk to other business owners and programmers.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">So what? Is AI killing us?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">According to this data, no. There&#8217;s no wave of layoffs that was so loudly announced. But where AI can do a person&#8217;s job, companies stop hiring people at the entry level. Juniors are the ones with the problem. The thing is, you&#8217;re a junior so that you can learn and one day become a senior, right?<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">And here&#8217;s the second catch.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Anthropic: what you learn when AI helps you<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Anthropic, the company behind Claude, ran an experiment in January 2026. They took 52 programmers, mostly beginners. Not a huge number, but at least we have something. Everyone wrote Python, but nobody knew the library they were given. They had 35 minutes to add two new functions to it. Half of them had AI to help, which could see their code and, on request, simply gave them a ready solution. The other half wrote everything completely on their own.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Right after that, everyone took a test, without AI. Pure knowledge from their heads. The test was about what they had just used: reading code, finding bugs and understanding how it works.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The AI group finished the task a mere two minutes faster, and scored <strong>50%<\/strong> on average on the test. Those who wrote on their own scored <strong>67%<\/strong>. The AI group did worst at finding bugs, that&#8217;s where the gap was biggest.<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1600\" height=\"900\" src=\"https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/05-anthropic-nauka.jpg\" alt=\"Anthropic: test score of 50% in the AI group and 67% without AI\" class=\"wp-image-4869\" srcset=\"https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/05-anthropic-nauka.jpg 1600w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/05-anthropic-nauka-300x169.jpg 300w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/05-anthropic-nauka-1024x576.jpg 1024w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/05-anthropic-nauka-768x432.jpg 768w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/05-anthropic-nauka-1536x864.jpg 1536w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\" \/><figcaption class=\"wp-element-caption\">Anthropic: test score of 50% in the AI group and 67% without AI<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">It gets more interesting inside the group that used AI:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>those who handed all the writing to AI finished fastest and scored <strong>39%<\/strong>,<\/li>\n\n\n\n<li>those who had AI fix their bugs scored <strong>24%<\/strong>, even less,<\/li>\n\n\n\n<li>and those who let AI write the code and then kept asking why it was done this way scored <strong>86%<\/strong>. Better than those who wrote without AI!<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">What follows from this? When you&#8217;re learning something new and hand it over to AI, the task will get done, but you won&#8217;t learn anything. And you lose most of what you need most when working with AI: <strong>the ability to find a bug<\/strong>. If you&#8217;re at the beginning of your path, the way you use AI decides whether you&#8217;ll one day become that second kind of job, the one that solves complex problems, and not just a regular code monkey.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">In this experiment, AI only sat in a side panel, as a chat. The authors themselves write that with agentic tools like Claude Code the effect will probably be even bigger. The sample was small, and some subgroups had just two or three people, which the authors stress themselves. But I&#8217;ve known this from online courses for years: <strong>if you only copy, you don&#8217;t understand.<\/strong> And today you can copy faster than ever.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Duke: AI agents on real boards<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Now the study that matters most to us, because it&#8217;s the only one I&#8217;ve found that tested AI agents strictly in embedded. And on real hardware, not just on whether the code compiles.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Duke University, March 2026, so pretty fresh. 42 peripheral tasks on three platforms: Arduino, ESP-IDF and Zephyr on nRF. 378 experiments, each checked on a board. The agent (Claude Sonnet 4.5) got the tasks in three variants:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>on its own,<\/li>\n\n\n\n<li>with skills, meaning instructions written for it by another AI,<\/li>\n\n\n\n<li>with skills written by a real, living, flesh-and-blood embedded expert.<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1600\" height=\"900\" src=\"https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/06c-duke-zephyr.jpg\" alt=\"Duke: the agent on Zephyr, 57% without skills, 45% with skills from AI, 93% with skills from an expert\" class=\"wp-image-4870\" srcset=\"https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/06c-duke-zephyr.jpg 1600w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/06c-duke-zephyr-300x169.jpg 300w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/06c-duke-zephyr-1024x576.jpg 1024w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/06c-duke-zephyr-768x432.jpg 768w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/06c-duke-zephyr-1536x864.jpg 1536w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\" \/><figcaption class=\"wp-element-caption\">Duke: the agent on Zephyr, 57% without skills, 45% with skills from AI, 93% with skills from an expert<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">In the video I showed Zephyr, here&#8217;s the full table. The number is the share of tasks that worked on the board on the first try:<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>Platform<\/th><th>No skills<\/th><th>Skills from AI<\/th><th>Skills from an expert<\/th><\/tr><\/thead><tbody><tr><td>Arduino<\/td><td>88%<\/td><td>93%<\/td><td>98%<\/td><\/tr><tr><td>ESP-IDF<\/td><td>62%<\/td><td>62%<\/td><td><strong>95%<\/strong><\/td><\/tr><tr><td>Zephyr<\/td><td>57%<\/td><td><strong>45%<\/strong><\/td><td><strong>93%<\/strong><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">On Arduino the agent manages without help. The more serious the ecosystem, the more it falls apart. On Zephyr it delivered 57% of the tasks on its own. With skills from another AI, 45%, so worse than with nothing! And with skills from an expert who added their experience, 93%. <strong>57, 45, 93.<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">What was in those expert skills? Nothing magical. Pin polarity in devicetree. Explicit pull-ups instead of pins floating in the air. Things you know once you&#8217;ve already fallen flat on them.<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1600\" height=\"900\" src=\"https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/06d-devicetree-ekspert.jpg\" alt=\"What was in the expert skills: pin polarity in devicetree and explicit pull-ups\" class=\"wp-image-4871\" srcset=\"https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/06d-devicetree-ekspert.jpg 1600w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/06d-devicetree-ekspert-300x169.jpg 300w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/06d-devicetree-ekspert-1024x576.jpg 1024w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/06d-devicetree-ekspert-768x432.jpg 768w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/06d-devicetree-ekspert-1536x864.jpg 1536w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\" \/><figcaption class=\"wp-element-caption\">What was in the expert skills: pin polarity in devicetree and explicit pull-ups<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">And for me that&#8217;s the heart of the matter. <strong>Your knowledge doesn&#8217;t lose its value.<\/strong> It&#8217;s what makes the difference between an agent that works half the time and one that works almost always. And knowledge made up by another AI can actually hurt.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Registers, HAL, AI<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">I often hear that AI isn&#8217;t real programming. But in embedded we&#8217;ve been through this before. On STM32 we first wrote on registers, so we had CMSIS. Then came SPL, then HAL, and you also heard that it&#8217;s not real programming. That too much happens on its own, and you need to write everything by hand and keep it under control.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Today hardly anyone writes everything on registers. Of course it happens. I even have a course on <a href=\"https:\/\/stm32narejestrach.pl\" target=\"_blank\" rel=\"noopener\">STM32 on registers<\/a> (in Polish), because people want it and need it: you understand the hardware better that way. But those who understand registers also use HAL better.<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1600\" height=\"900\" src=\"https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/07-rejestry-hal-ai.jpg\" alt=\"The same task on three levels: registers, HAL and AI. HAL does the same thing every time, AI doesn&#x27;t\" class=\"wp-image-4872\" srcset=\"https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/07-rejestry-hal-ai.jpg 1600w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/07-rejestry-hal-ai-300x169.jpg 300w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/07-rejestry-hal-ai-1024x576.jpg 1024w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/07-rejestry-hal-ai-768x432.jpg 768w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/07-rejestry-hal-ai-1536x864.jpg 1536w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\" \/><figcaption class=\"wp-element-caption\">The same task on three levels: registers, HAL and AI. HAL does the same thing every time, AI doesn&#x27;t<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">In my opinion it&#8217;s similar with AI. With one important difference: <strong>HAL does exactly the same thing every time. AI doesn&#8217;t.<\/strong> The same prompt can produce completely different code. With AI we lose determinism, and in some embedded sectors that can be a serious problem.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">The bottleneck: verification<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">That&#8217;s why our main job stops being writing code. It becomes checking. Does it actually work on the board, is there a race between tasks, does the chip even come back up after a reset.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Developer surveys from this year say exactly the same. In the Sonar survey (over 1,100 developers), it turned out there&#8217;s a lot of AI code, and <strong>verification has become the bottleneck<\/strong>.<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1600\" height=\"900\" src=\"https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/08-weryfikacja.jpg\" alt=\"Verification is the bottleneck of working with AI\" class=\"wp-image-4873\" srcset=\"https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/08-weryfikacja.jpg 1600w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/08-weryfikacja-300x169.jpg 300w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/08-weryfikacja-1024x576.jpg 1024w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/08-weryfikacja-768x432.jpg 768w, https:\/\/msalamon.pl\/wp-content\/uploads\/2026\/11\/08-weryfikacja-1536x864.jpg 1536w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\" \/><figcaption class=\"wp-element-caption\">Verification is the bottleneck of working with AI<\/figcaption><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">Which job is yours?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Back to the question from the start.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">If your work is translating specifications into code, AI already does that today. And it will do it better and better.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">If your work is knowing what should be built, why exactly this way, and being able to check that it works, you&#8217;re in a very good place. And AI can take all that code typing off your plate.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">That&#8217;s exactly how I&#8217;ve been working for the last few months. The agents write code for me, and my job is to make sure they know what I want, to check it in the end and to steer them towards the solution as their senior.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">What&#8217;s next?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">In the next video I show what this looks like for me tool-wise: what I have installed, what it runs on and how I start working with an agent on STM32. You&#8217;ll find it here: <a href=\"https:\/\/youtu.be\/deAV4LyvqvI\" target=\"_blank\" rel=\"noopener\">My setup for working with an AI agent<\/a> (in Polish).<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">On the &#8220;AI w Embedded&#8221; list I write about what I&#8217;ve tested with AI agents on STM32 and where I fell flat on my face (in Polish). Sign up: <a href=\"https:\/\/aiwembedded.pl\" target=\"_blank\" rel=\"noopener\">aiwembedded.pl<\/a>. We can also chat on the <a href=\"https:\/\/linki.msalamon.pl\/ai_embedded\" target=\"_blank\" rel=\"noopener\">&#8220;AI w Embedded&#8221; Discord<\/a> (in Polish).<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>And you? Which job is yours?<\/strong> Let me know in the comments.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Sources<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li>US job market, programmer vs developer (Washington Post, 03.2025): <a href=\"https:\/\/www.ibj.com\/articles\/computer-programming-ranks-among-hardest-hit-occupations\" target=\"_blank\" rel=\"noopener\">summary<\/a><\/li>\n\n\n\n<li>BLS, Computer Programmer occupation: <a href=\"https:\/\/www.bls.gov\/ooh\/computer-and-information-technology\/computer-programmers.htm\" target=\"_blank\" rel=\"noopener\">bls.gov<\/a><\/li>\n\n\n\n<li>Stanford, &#8220;Canaries in the Coal Mine&#8221;: <a href=\"https:\/\/digitaleconomy.stanford.edu\/publications\/canaries-in-the-coal-mine\/\" target=\"_blank\" rel=\"noopener\">digitaleconomy.stanford.edu<\/a><\/li>\n\n\n\n<li>Anthropic, &#8220;AI assistance and coding skills&#8221; (01.2026): <a href=\"https:\/\/www.anthropic.com\/research\/AI-assistance-coding-skills\" target=\"_blank\" rel=\"noopener\">anthropic.com<\/a><\/li>\n\n\n\n<li>Duke University, AI agents in embedded and IoT (arXiv, 03.2026): <a href=\"https:\/\/arxiv.org\/abs\/2603.19583\" target=\"_blank\" rel=\"noopener\">arxiv.org\/abs\/2603.19583<\/a><\/li>\n\n\n\n<li>Sonar, State of Code Developer Survey 2026: <a href=\"https:\/\/www.sonarsource.com\/state-of-code-developer-survey-report.pdf\" target=\"_blank\" rel=\"noopener\">PDF<\/a><\/li>\n<\/ul>\n\n\n<div class=\"kk-star-ratings kksr-auto kksr-align-left kksr-valign-bottom\"\n    data-payload='{&quot;align&quot;:&quot;left&quot;,&quot;id&quot;:&quot;4879&quot;,&quot;slug&quot;:&quot;default&quot;,&quot;valign&quot;:&quot;bottom&quot;,&quot;ignore&quot;:&quot;&quot;,&quot;reference&quot;:&quot;auto&quot;,&quot;class&quot;:&quot;&quot;,&quot;count&quot;:&quot;0&quot;,&quot;legendonly&quot;:&quot;&quot;,&quot;readonly&quot;:&quot;&quot;,&quot;score&quot;:&quot;0&quot;,&quot;starsonly&quot;:&quot;&quot;,&quot;best&quot;:&quot;5&quot;,&quot;gap&quot;:&quot;0&quot;,&quot;greet&quot;:&quot;&quot;,&quot;legend&quot;:&quot;0\\\/5 - (0 votes)&quot;,&quot;size&quot;:&quot;24&quot;,&quot;title&quot;:&quot;Embedded developer in the age of AI: which job is yours?&quot;,&quot;width&quot;:&quot;0&quot;,&quot;_legend&quot;:&quot;{score}\\\/{best} - ({count} {votes})&quot;,&quot;font_factor&quot;:&quot;1.25&quot;}'>\n            \n<div class=\"kksr-stars\">\n    \n<div class=\"kksr-stars-inactive\">\n            <div class=\"kksr-star\" data-star=\"1\" style=\"padding-right: 0px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" data-star=\"2\" style=\"padding-right: 0px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" data-star=\"3\" style=\"padding-right: 0px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" data-star=\"4\" style=\"padding-right: 0px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" data-star=\"5\" style=\"padding-right: 0px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n    <\/div>\n    \n<div class=\"kksr-stars-active\" style=\"width: 0px;\">\n            <div class=\"kksr-star\" style=\"padding-right: 0px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" style=\"padding-right: 0px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" style=\"padding-right: 0px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" style=\"padding-right: 0px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" style=\"padding-right: 0px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n    <\/div>\n<\/div>\n                \n\n<div class=\"kksr-legend\" style=\"font-size: 19.2px;\">\n            <span class=\"kksr-muted\"><\/span>\n    <\/div>\n    <\/div>\n","protected":false},"excerpt":{"rendered":"<p>One programming job in the US lost 27.5% of positions, another 0.3%. I go through studies on AI and the job market and check what your knowledge is worth when an agent writes the code.<\/p>\n","protected":false},"author":1,"featured_media":4874,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":"","_links_to":"","_links_to_target":""},"categories":[200],"tags":[],"class_list":["post-4879","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai-in-embedded"],"_links":{"self":[{"href":"https:\/\/msalamon.pl\/en\/wp-json\/wp\/v2\/posts\/4879","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/msalamon.pl\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/msalamon.pl\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/msalamon.pl\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/msalamon.pl\/en\/wp-json\/wp\/v2\/comments?post=4879"}],"version-history":[{"count":1,"href":"https:\/\/msalamon.pl\/en\/wp-json\/wp\/v2\/posts\/4879\/revisions"}],"predecessor-version":[{"id":4880,"href":"https:\/\/msalamon.pl\/en\/wp-json\/wp\/v2\/posts\/4879\/revisions\/4880"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/msalamon.pl\/en\/wp-json\/wp\/v2\/media\/4874"}],"wp:attachment":[{"href":"https:\/\/msalamon.pl\/en\/wp-json\/wp\/v2\/media?parent=4879"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/msalamon.pl\/en\/wp-json\/wp\/v2\/categories?post=4879"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/msalamon.pl\/en\/wp-json\/wp\/v2\/tags?post=4879"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}