The term "artificial intelligence" has been recycled so many times—from data mining to machine learning to deep learning—that it's lost all precision. Ivo Bernardo, co-founder of consultancy DareData Engineering, cuts through the noise by grounding the concept in what it actually is: systems that take decisions autonomously, not because they were explicitly programmed, but because they learned patterns from data. That distinction matters. A Netflix recommendation algorithm, credit-risk scoring systems, and ChatGPT are all AI, but few of us think of them that way because they've already disappeared into the infrastructure.
The real inflection point came with neural networks discovering they could learn feature interactions automatically—something traditional machine learning couldn't do without humans manually engineering variables. When images arrived as training data, convolutional neural networks (CNNs) suddenly had the capacity to see. When language followed, the 2012 AlexNet breakthrough on ImageNet set a template: throw enough GPU compute at the problem, feed it enough data, and the network figures out what matters. This pattern held for a decade. Then in 2017, the attention mechanism arrived—a mathematical trick allowing the network to focus on different parts of a sentence depending on context, solving the ambiguity problem that plagued earlier text models. By the time transformer-based models like GPT hit scale in 2022, the shift wasn't technical novelty; it was compute and data abundance finally colliding.
The uncomfortable truth Bernardo hints at is that we don't fully understand why scaling works. Increasing parameters, training data, and compute should hit diminishing returns—it did for every other approach in machine learning. Instead, GPT models kept improving. That's the magic we're actually chasing: why does "more steroids" keep working? The answer might not be mystical—it could simply be that language, as a training signal, contains enough statistical structure to reward scale up to some ceiling we haven't found yet. But it also means the next limitation is likely waiting somewhere we can't see until we hit it.
On the future of work, Bernardo's instinct is sound but rarely stated plainly: repetitive tasks disappear first—call centre work, formulaic writing, routine data entry. But what survives is craft. Writers who write well, engineers who care about elegant solutions, strategists who can judge context—these see demand increase precisely because AI floods the market with commodity output. The constraint shifts from "can this be done?" to "can this be done with taste?" That's not comforting for everyone, but it's more honest than either boomer optimism or doomerism.