Amazon Web Services developer advocate Darko Mesaroš has argued that curiosity, technical judgement and a deeper understanding of software engineering will become more important than coding skills as artificial intelligence (AI) transforms the profession.

Mesaroš, a Distinguished Developer Advocate at Amazon Web Services (AWS), showcased his work during an online video call with international media, including the Cyprus Mail, demonstrating how he had connected a 1984 IBM PC XT to a modern artificial intelligence model.

The demonstration, which also involved a vintage teletype machine, was designed to illustrate how technology has evolved while highlighting the importance of skills that remain relevant regardless of the tools engineers use.

In a garage near Seattle, Mesaroš has connected the decades-old computer to an AI model, allowing it to communicate with technology that would otherwise be far beyond its capabilities.

“What if it could talk to a large language model (LLM)?” he asked, demonstrating the setup.

The computer operates as a basic terminal, sending and receiving characters through a serial connection at just 300 baud. A serial-to-Wi-Fi modem emulator converts the connection into network traffic, which reaches an application running in a container.

The application uses Strands, an open-source software development kit released by AWS for building AI agents, to pass conversations to a model running on Amazon Bedrock.

The same approach works with a Commodore 64 and a 1982 DEC teletype machine, which prints AI-generated responses on paper.

“The real hero here is standards. In the past we agreed on the modem protocol, on serial communication, on how terminals move information. Today we’ve agreed on how agents call tools, and on protocols for talking to agents. Blessed be the standards we all agree on. They make it easier for everyone to build,” Mesaroš said.

He acknowledged that the experiments were not necessarily practical, but argued that they encouraged a more deliberate approach to technology.

“Is this useful? Maybe not. Is it fun? Absolutely,” he said, adding that older computers forced him to concentrate because they could only perform one task at a time.

“Old computers can only do one thing at a time. That forces me to focus and be intentional about what I’m doing,” he said.

Mesaroš’s central argument is that software engineering has never been limited to writing code, despite coding traditionally consuming much of developers’ time.

“AI has taken over the coding part of software engineering,” he said, arguing that the profession also involves understanding customers, scalability, operations and observability.

He recalled that earlier in his career he worked as a systems administrator, alongside specialists responsible for networks, storage and databases.

The arrival of cloud computing changed those roles, bringing them together under the broader discipline of platform engineering.

“The same thing is happening to software engineering now. From people who purely write code to people who do much more and are much closer to their customers,” he said.

Mesaroš believes autonomy rather than speed will determine which developers succeed in the AI era.

“A lot of people focus on how fast you can produce software. I don’t think fast is what’s winning. Autonomy is. When a developer has the autonomy and the capacity to build things for their customers, that’s what brings value,” he said.

He advised developers whose main ability is writing code to broaden their knowledge and understand more than simply how to write functions.

Curiosity and attention, he argued, will distinguish those who use AI effectively from those who simply accept whatever it produces.

“Anybody can prompt an LLM, say ‘build me a thing’ and hit accept. But then you’re just one of a hundred thousand people doing the same thing. If you’re curious, if you’re willing to go deep, tinker, and find the interesting path rather than just the efficient one, you’ll actually build something,” he said.

He warned against allowing AI tools to make every decision, describing the danger of becoming a passive approver of automatically generated code.

“Don’t fall into the hole of letting the tool generate all your code and delegating every decision to the LLM. You’ll quickly realise you had zero opinion in the thing you built,” he said.

His own experiments provide opportunities to develop those skills. While troubleshooting a faulty serial card in the IBM computer, he used AWS’s Kiro AI coding tool to write x86 assembly language, exploring a type of low-level programming that has limited practical application in modern software development.

“I strongly believe we have no reason to write bad software anymore. We have the tools, the LLMs, the agents. We can troubleshoot, debug and optimise to our heart’s content,” he said.

On development methods, Mesaroš described vibe coding, in which developers repeatedly prompt AI until an application works, as a useful and fast approach.

However, for more serious projects, he recommended spec-driven development, which he described as the more mature version of the technique.

The approach, supported by Kiro, involves working with AI to establish requirements, a technical design and a list of tasks before writing any code.

“It’s a safer, more correct way to work,” he said, adding that the approach had existed long before AI but was tedious to carry out manually.

He considers it particularly important for software exposed to the internet.

Mesaroš also addressed concerns about the growing cost of AI computing and the resources required to operate increasingly powerful models.

He argued that the future would not necessarily involve sending every request to the largest available model, pointing instead to token economics and smaller, faster models capable of determining which system is best suited to a particular task.

Such models could operate locally and direct straightforward requests to cheaper systems while reserving more powerful models for complex problems.

He also argued that software developers had been wasteful with computing resources long before the arrival of AI, with many applications designed around the assumption that resources were effectively unlimited.

“With the right intent and the right developer mindset, we can be far more efficient. I’m a toxic optimist. Developers who stay true to their craft will keep tinkering and making more efficient things,” he said.

The changing nature of software development also raises questions about how junior engineers will gain experience if AI agents take over routine coding tasks.

Mesaroš acknowledged that expectations had risen considerably, with considerably more required of junior developers than five years ago.

However, he warned that relying entirely on AI could prevent newcomers from developing the troubleshooting skills needed to advance.

“Experience comes from troubleshooting hard problems. The danger is letting a junior hand everything to the LLM. Then they never get that experience,” he said.

He believes that as development teams take on more demanding projects involving performance and scalability, junior engineers will have opportunities to tackle harder problems and progress towards senior positions.

His advice to students was to use AI extensively as a learning tool, but also to experiment and deliberately make mistakes.

“Use as much AI as you can to learn, scratch your curiosity, build things, and, more importantly, break things,” he said.

“One of the questions I ask when I interview people is, tell me the last thing you broke. Fixing what you broke is where you actually learn.”

He added that Amazon continued to invest in early-career employees, hiring large numbers of junior developers and interns because it believed they could achieve a great deal when given opportunities and access to modern tools.

Despite advocating AI adoption, Mesaroš was sceptical about claims that current models could become self-aware.

“In the architecture of a transformer model, I don’t think so. It’s a token-producing machine. You put something in, and it guesses the next word. It’s very good at guessing, but that’s what it does,” he said.

He argued that natural-language interfaces made AI systems appear more human than they actually were, while dismissing fears of runaway technology with a practical observation.

“I’m the guy with a power switch next to me. I can always flip off a GPU,” he said.

Ultimately, Mesaroš said his vintage computer experiments were intended to reinforce the human qualities that would remain essential as AI took over more technical tasks.

“Curiosity and intentionality are what will separate the developers of the future, the good ones, from the ones who merely use AI,” he said.