2000 miles, and the case for human-centred AI infrastructure

Recently, during a long-haul flight somewhere over central Europe, we read a short column that made us think. It landed like a splash of cold water in a world too often drunk on the AI hype. It wasn’t another glossy forecast about artificial general intelligence, or another “golden” piece on the billion-parameter model arms race. It was a grounded, quietly radical article by Professor Gancho Todorov Ganchev, published in @Bloomberg Businessweek BG, titled “За границите на изкуствения интелект” (“On the limits of artificial intelligence”).

In a world increasingly dominated by the promise and the pressure of smart machines, Ganchev’s message was refreshing. Artificial intelligence, he reminded, isn’t some inevitable mighty force. It is a human-built, digital approximation of reality, limited not just by hardware and software but by logic, economics, and ethics. And that’s not a reason to panic, but a motivation to build better.

What resonated most was the article’s refusal to indulge in either dystopia or blind optimism. The article addresses ideas that feel abstract and brings them down to earth. AI cannot and will not ever fully “know” reality. It can simulate, match patterns, and predict, but it does so through layers of approximated logic and ever-growing energy consumption. The digital world we’ve built, impressive as it is, is not infinite, and to pretend it is can lead to costly mistakes.

At Utilink we work towards the sole aim of infrastructure for machine learning. Every day, we think about how to make model training more efficient, how to bring down costs, and how to let small teams do what only Big Tech could afford. And every day we ask the same question: how do we scale intelligence without losing sight of the limits that make it human-compatible?

The answer (in part) is to accept those limits more as architectural foundations rather than obstacles. Accepting that machine learning is an approximation would lead us to think more critically about accuracy, calibration, and context. A model that’s 94% accurate in the lab might fail spectacularly when it meets messy, real-world inputs. That’s why we invest in tools that don’t just pop out answers but expose uncertainty, make decisions testable, and let humans stay in the loop.

Ganchev’s piece also raises an economic point that deserves mentioning. AI, as it is applied today, can disrupt traditional ideas of fairness and efficiency. The technology is often sold as a way to boost productivity, but its benefits don’t always come evenly. We’ve already seen how pricing algorithms or automated loan systems can favour those with the most data, the most resources, or the most power. The application of AI can quietly tilt entire markets, breaking long-held assumptions about free access to information and equal competition. And if you think about it, there is a word for that dynamic – “asymmetry”. And it’s one of the reasons why infrastructure really matters. The platforms we build (…and who gets to use them, how much they cost, whether they are open or not) have an impact on who wins and who is left behind. That’s why at Utilink, we are not trying to build the “biggest” and “smartest” AI tool in the world. We’re simply trying to make it possible for more people to build the model they need, on their terms, without burning through capital or compute budgets just to test an idea.

Ethics isn’t just a layer you slap on once the tech is done. It’s something you embed at the system level. It’s how you set defaults, how you visualise outcomes, and how you let users tune or challenge the results. There are real stakes here, if we stay ignorant, we risk the AI (machine learning) systems growing inequality, concentrating power, and creating processes that will benefit only some (read the “chosen ones”).

But that’s not inevitable, and the antidote by all means is not to stop progress but to design progress more intentionally inclusive. A future shaped by AI should be one dictated by people (plural, diverse, empowered). That means investing not only in bigger training runs, but in better frameworks for accessibility and accountability. For example, it means thinking about energy efficiency as a feature, not a footnote. It means building AI in a way that makes it easy to use and interpret, not only because regulators demand it but because users deserve it.

…We don’t have all the answers (or questions), no one does. But what this article reminded us, high above the clouds that day, is that asking the right things matters more than ever. What are we optimising for? Who benefits from this success? Is there something that gets lost in translation from data to decision?…

At Utilink , we’re committed to keeping those questions at the heart of our work. Not because it’s trendy, but because we believe this is what makes AI sustainable, equitable, and ultimately useful to everyone.