Adapting to the age of AI

Working with AI

I recently met up with an engineer friend of mine to catch up, and once again a good chunk of the conversation was about AI; more specifically the state of shock we're both in over how quickly it's been progressing. It was not that long ago that the output quality was too mediocre to be seriously considered as a substitute for manually writing the code, and many engineers were delaying, or even resisting, adoption. But we are now well and truly past that point, and most engineering teams are leveraging one or more of Claude Code, Codex, Copilot, and countless other flavors.

This is not to say that we're quite at the define-it-and-forget-it stage of code generation; Gergely Orosz from The Pragmatic Programmer has a really good piece on how different companies and teams are managing the higher influx of pull requests, for example (paywalled past the third approach, but the free preview is still insightful). I realize that some of you who have been all-in with AI from the start may disagree with the idea of not fully committing to an agentic workflow, but even OpenAI and Anthropic use a "triage by blast radius" approach to code reviews. I'm not saying that there's only one way to do things, but if two of the leading companies in that space believe it's important to keep humans in the loop for the truly important stuff, it might be worth taking some notes.

Truthfully, for personal projects, tools like Claude Code are a gift. I know so many software developers, including myself, who have had plenty of ideas for side projects for years, some of which we've even contributed a lot of time and effort to. But there was always one more module to add, one more thing to test, one more feature to build. Even if we got as far as a proof of concept or a working prototype, life got in the way, work responsibilities took precedence, and the entire project would get shelved.

Now, you don't even have to be a software engineer, really — unless you want to make the side project into a legitimate business. No matter how good AI gets (and it's getting very good, very fast), I believe the devil will always be in the details. When building software, it's important to take many things into consideration: security, scalability, networking latency, data integrity, performance optimization, cloud/compute/storage/AI costs, search, caching, and so on. That's not even taking under account the demon that is scope creep, one that torments us all and can only be reined in (to some degree) with a lot of experience. If anything, unchecked AI will empower the scope creep demon to feast merrily.

Tools like the /grill-me skill can help alleviate some of these concerns, and good prompt training for non-technical folks will allow many to build software mostly independently; certainly for internal or personal projects. I'm also hopeful that AI will attract new people who were not inclined to write code before to be much more willing to dive in now. But I believe we'll most likely need humans who know and understand the SDLC, and who have experience writing software, for a good while to come. This is something that is well outlined by Sean Goedecke in his LLMs Reward Expertise piece.

But I also worry, as does the aforementioned engineer friend of mine, that the seemingly impregnable moat of finding work as any variation of software engineer is suddenly looking a lot more vulnerable (Sean Goedecke again wrote another great piece on this topic, aptly entitled Software engineering may no longer be a lifetime career). Teams that might have required front and back end engineers, data engineers, devops, QA, and so forth, may now just require a couple of experienced senior people who can make sure that they implement a sound workflow for AI to be able to operate with a high level of trust, even if it comes with some human gating for the most sensitive work. This is probably especially true for smaller and medium-sized companies that are still trying to find their North Star and need to pivot quickly, and is much less applicable to organizations that require a high level of scrutiny, oversight, and compliance. But with time, it's essentially impossible to imagine things going in the opposite direction or remaining stagnant for too long.

I occasionally think about knocker-uppers:

A knocker-up or knocker-upper had the job of rousing sleeping people so they could get to work on time. The role started during the Industrial Revolution, in Britain, the Netherlands, Ireland and some other countries, when alarm clocks were neither cheap nor reliable. By the 1940s and 1950s, the job had more or less entirely died out, although it still continued in some pockets of industrial England until the early 1970s.

Though being a knocker-upper certainly didn't require experience or knowledge, and was surprisingly resilient to be able to last into the 1970s, ultimately automation made this particular job redundant. Even original business models that were able to take advantage of a relatively new technology can quickly become obsolete. A good example of that is internet cafés, which are much less prominent than they used to be now that most everyone has a pocket computer with easy access to free or affordable wifi accessible everywhere but the most remote places.

Automation replacing jobs is not unprecedented; generally speaking it's made life for most people more comfortable and affordable, and allowed for more accessible goods and services than one could ever have hoped for in the past. Materials like aluminum are so ubiquitous in kitchens that it's difficult to imagine a time when Napoleon was compelled to serve his most esteemed guests with aluminum cutlery and plates (the less important folks had to make do with the gold pieces, imagine the shame). So what makes AI so different, and (perhaps understandably) distrusted and vilified?

Brian Merchant at Blood In The Machine has written a few articles on the topic, but the below quote, from a Glassdoor report, in his piece entitled Why office workers are turning against AI is germane to the question I just asked:

Many workers are concerned about AI replacement (20%), while others resent being force-fed AI tools (14%) or think their company is focused on AI at the expense of their core business (13%). Other criticisms include using AI for worker communications or surveillance (10%), moving too slowly on AI (9%), having unrealistic expectations about AI gains (8%), and being in a line of business that is likely to be disrupted or replaced by AI (7%).

These very legitimate concerns are not even taking under account non-work related unease around clean water usage, supply chain challenges and dwindling critical minerals resources, and, uh, "a 10% chance of it killing all humans" — something that feels like it should only belong in the Futurama universe.

Clearly, there is a myriad of strong arguments for office workers to be upset by AI. Perhaps I'm providing an oversimplified comparison, but it feels to white-collar jobs what happened to blue-collar jobs with the advent of machinery and robotics. And if we no longer need blue or white collar jobs — this is not exactly accurate, nor is it necessarily guaranteed; but it is starting to feel like the endgame — then that means, as a society, we will have to completely change our understanding of what it means to earn a living.

Learning with AI

Then there's education, something that every good engineer I've ever worked with cares about deeply, as engineering is the type of career that always requires continued learning to grow and succeed. But arguing with young people to learn different topics that they may never need, and can quickly look up with a number of devices that are easily accessible to them, is only going to prove not just more challenging but to put it bluntly, it's a losing fight. Who can blame so many of them for choosing to become influencers when it's becoming a more compelling argument for generating sustainable income, especially when they can leverage AI to hand-hold them through the process?

The real concern is that they are never going to develop essential skills, much like a 7 or 8 year old me arguing with my mom that I didn't need to learn multiplication tables because I could always use an electronic calculator:

When I ask AI to evaluate a claim, I can check the output against my own judgment. I notice when it oversimplifies. I catch when it omits a competing interpretation. I understand when the confidence of the language exceeds the strength of the evidence. This is auditing the output.

A child usually won't be able to do this—not because children are less intelligent, but because auditing requires the exact domain knowledge that the child is supposed to be developing. You cannot check an AI's analysis of heredity if you don't yet understand what heredity is. You cannot evaluate an AI's interpretation of the French Revolution if you've never read conflicting accounts of it yourself.

The author of the above article, Timothy Cook, then highlights an even more worrying concern, homogenization:

When every student in a class processes information through the same language model, they are learning to reason through the same system. This introduces a new threat vector on the developing mind.

The model's statistical biases become the student's default framing. The model's reasoning structure becomes the student's reasoning structure. LLMs homogenize not just language but also perspective and reasoning strategies. The convergence tracks toward Western, educated, mainstream norms because that's what dominates training data and gets reinforced through alignment (Sourati et al., 2026).

Immediate gratification is understandably more appealing than anything requiring effort; an innumerable amount of businesses have made a fortune from this knowledge. Think about Amazon continually trying to reduce the amount of time to delivery, companies like Grubhub and Postmates convincing some that cooking is for suckers, and how modern video games are designed to hijack your brain:

... Playing video games can satisfy different basic psychological needs, probably also dependent on the specific video game and its genre. Especially fulfillment of psychological needs like competence (sense of self-efficacy and acquisition of new skills), autonomy (personal goal-directed behavior in novel fictive environments), and relatedness (social interactions and comparisons) were associated with video gaming (Przybylski et al., 2010). Specifically, satisfaction of psychological needs might be mainly related to the various feedback mechanisms provided to the player by the game. This elaborate reinforcement and reward schedule has the potential to maximize motivation (Green and Bavelier, 2012).

This means that we'll have to be far more intentional as a society to encourage and reward learning (perhaps taking some notes from video games?), as well as to make education more accessible to those who seek it. In many ways, it already is: universities like MIT and Stanford offer free courses online, and many more free or affordable classes can be found on YouTube, Coursera, and countless other platforms. But as wonderful as those initiatives and platforms are, not all of them are necessarily accessible — say, for financial or censorship reasons. Even without those barriers, sometimes, people just don't want to learn. After all, it's so much easier to stream a new show on Netflix than to try to figure out what the hell an eigenvector is.

And honestly, we should use AI for learning, though we need to be intentional and specific about what we mean by that. While I don't rely too often on AI for learning new topics, it can be a good starting point (especially with RAG), not overly different from using a search engine to seek out sources but allowing you to be more specific with your query to get a customized response. It can also be useful as a teaching assistant by creating recurring study reminders, generating quizzes from reading material (assuming it has access to that material), and helping you put together small projects or games to help put newfound knowledge in application. When you are comfortable with your own abilities, it gives you more opportunities to be creative with the learning process.

Takeaways

I'm worried that this is going to sound like an "AI is evil" article, which is not my intent at all. AI is a tool, it's the next and most logical step towards generalized automation, and it's now ubiquitous: everything we've done to advance society has lead to this moment, and it will continue to move forward, good and bad parts included. Knowing how to use the tool, and choosing to use it appropriately, is frankly hugely beneficial in that it's a great way to save money, time, and effort. It opens new doors that were previously unavailable, out of reach, or even just a bit too distant to be worth it.

There is a lot of uncertainty and anxiety around it, but my takeaways are this:

• Learn how to use AI well and learn when to use it appropriately, but also do what you can to learn as much about the underlying technology and its impact. You can be upset and distrustful at the companies and individuals who are pushing it onto every aspect of our lives, but the technology itself is quite impressive.
• Even if you fully embrace AI, don't give up on undertaking challenging technical problems. They are always frustrating and time consuming, but it will always be rewarding to acquire knowledge and gain experience through effort and dedication. It will make you more confident in your prompting as well.
• Keep reading, keep learning: AI may be able to generate output faster than you, but you are the best line of defense from when it's "confidently wrong." The only way to catch its mistakes is if you intimately know the topic you're working with. It will make you more specific in your prompting as well, which will reduce the amount of mistakes and misguided or unintended outputs.
• There will be times when AI is unavailable, broken, or too expensive. You will likely still need to do some stuff manually. The best way to be prepared is to stay prepared. The last thing you want to see is that 529 error when there is a production issue.

I would love to hear your thoughts, and if you have some takeaways you'd like to share! Also please let me know if this was useful, and if you have any feedback for me. I appreciate you taking the time to read through this. 🫶

— Yacine