From The Past to The Future of AI: A Trajectory

AI looks sudden, but it stands on decades of hardware and research. The first microchip was built in 1958, and today’s chips hold billions of transistors. The first microprocessor, the part that works as the “brain” of a computer, followed in 1971.

These inventions sparked a wave of continuous improvements in computing power, setting the stage for the AI tools we see today. And yet, visionaries like Nikola Tesla, Alan Turing, and Gordon Moore saw it coming long before the rest of us. Tesla imagined a world of wireless communication and automation, Turing laid the foundation for machine intelligence, and Moore predicted the relentless march of computing power that shaped the tech-driven reality we live in today.

I won’t go into the technical details, but it’s humbling to think that just over 100 years ago, letters were still the norm for long-distance communication. Today? We can talk to anyone, anywhere, in real time.

Bringing it back to the present, it’s safe to say AI is starting to be everywhere. If you’re reading this, chances are you’re wondering: Will AI become sentient and take over the world, or is it just another tech bubble waiting to pop?

We’ve seen hype cycles before. Quantum computing is advancing, but nowhere near the pace once predicted. The metaverse was sold as the future of digital interaction, yet it failed to gain mass adoption. Crypto disrupted finance, but it didn’t replace banks like some believed it would.

AI has more hype than evidence in some areas and measurable change in others.

To understand the true potential of artificial intelligence, we need to look back 500 million years, to when the first neural networks emerged in living organisms. Biological intelligence has been evolving ever since, leading to the complex human minds we have today. In contrast, machine intelligence was formally introduced by Alan Turing in 1950 and coined as “Artificial Intelligence” in 1955.

Mammals (including us) have undergone a half-billion-year process of unassisted learning, shaped by natural selection. Our brains don’t operate in isolation – they’re integrated with the rest of the body – so evolutionary fine-tuning has made our intelligence efficient. By comparison, AI is still in its infancy. Yet it’s advancing at a rate that’s astonishing in human terms.

The first mathematical model of an artificial neuron appeared in 1943 – a binary system where neurons either fire (1) or stay inactive (0). This crude approach laid the groundwork for artificial neural networks (ANNs). But as our biological research advanced, scientists realized a single real neuron is far more complex – possibly equivalent to 1,000 artificial neurons. The human brain, with its ~86 billion neurons, remains far beyond our full understanding.

How smart is AI? It depends on the model, on its training data, the training methods, the number of parameters, and the computational power used during training. It’s not straightforward to define “intelligence” in a system that mainly predicts words or patterns from data. Yet we find ourselves drawn to it, and that raises questions about us:

• Why are we so fascinated by talking to a computer?
• What does it mean to think, to reason?
• How do we judge if a system is behaving “intelligently”?

Even though nobody can fully predict the new, often surprising behaviors a large language model might exhibit, many people rely on AI because it gets things done. But our willingness to use AI based on practical benefits alone doesn’t fully capture the sense of wonder (and occasional unease) around these systems. In some ways, it points back at our own curiosity and our tendency to see human-like qualities in machines, even if we recognize they’re different from us.

Modern AI, especially Large Language Models (LLMs), predicts how words tend to follow one another – often described as learning “probability distributions over text.” Critics call them “stochastic parrots,” saying they merely repeat patterns from their training data. However, neural scaling laws and recent research suggest that once LLMs reach a certain size and complexity, they start combining learned “skills” in unexpected ways. They can solve math problems, write poetry, reason about cause and effect, and even show flashes of creativity.

Researchers have studied how new abilities can “emerge” when language models grow large. For instance, they use a “skills mix” test – a bipartite graph that randomly connects language skills to different text chunks – to see if the model can combine skills it hasn’t been directly trained on together. The success of these tests suggests that, beyond simple pattern-matching, models might be forming new kinds of knowledge.

Language isn’t spread out evenly. We use some words – like “the,” “is,” and “and” – very often, while thousands of other words appear only rarely. This pattern is described by Zipf’s Law, which explains why a handful of words dominate our speech while many others show up far less frequently.

Because LLMs learn from this natural distribution, they can quickly master the common words but struggle more with unusual or specialized ones. This means it gets harder to see big leaps in performance once they reach a certain size – almost like an “efficiency ceiling” created by language itself.

Meanwhile, our own brain hums along at about 20 watts of energy, juggling a massive range of tasks in real time. By contrast, training a frontier AI model can use millions of watts for months.

Large Language Models (LLMs) are typically trained to predict the next word in a sentence, which has led to large gains in writing and coding. Whether they reason or only make sophisticated guesses is still an open question.

That’s where Reinforcement Learning (RL) comes in. RL provides clear rewards for correct or useful outputs, helping AI refine its approach. Take Tesla’s vision system, which originally included “human in the loop” instructions during training, engineers manually labeled or wrote rules for the AI to follow.

This manual layer eventually limited the model’s ability to discover new strategies on its own. By moving to a more autonomous RL-based approach, the system began learning directly from real-world data and automated feedback, improving its performance faster.

A similar principle applies to coding AI. OpenAI’s research (arXiv:2502.06807v1) shows that relying solely on human annotations can hold models back, whereas verifiable rewards (like checking if code runs or passes tests) let the AI self-correct with less human guidance. When combined with chain-of-thought prompting, models outline their reasoning step-by-step and learn to fix errors or explore alternative solutions.

This shift from purely predictive text to reasoning-driven AI marks a “third wave,” potentially moving us closer to Artificial General Intelligence (AGI) by enabling machines to learn, adapt, and solve unfamiliar problems on their own.

So, could these methods take us from narrow AI (great at specific jobs) to Artificial General Intelligence (AGI), a system with the flexible thinking skills of a human? And if we get AGI, do we quickly move on to Artificial Superintelligence (ASI)?

There’s no consensus. An AI might excel at coding or math puzzles but fail at social responsibility or at handling unpredictable “real-world” scenarios. The step from narrow AI to a system that can handle virtually any task remains huge.

Still, techniques like RL and chain-of-thought might help AI advance rapidly, especially if future models connect to real-world data through vision, robotics, or advanced memory. As AI improves its understanding of the world through experience, we may see the line between “machine” and “biological” intelligence blur even more.

Artificial Intelligence stands at a crossroads between wonder and worry. It can create tremendous prosperity or deepen existing inequalities. It can give us new forms of creativity or replace entire industries. It can help us understand intelligence or hide it behind processes we barely grasp.

AI learns differently from humans, mostly by processing huge datasets to predict what comes next, and it is already useful. The open question is how we guide it.

1. Will we keep humans “in the loop,” or allow AI to develop with minimal oversight?
2. Will we plan for major shifts in jobs, or let them hit us by surprise?
3. Will AI remain “just another tool,” or become something that forces us to rethink the nature of life and mind?

We don’t have all the answers yet. But we do know AI is more than just hype and a real step forward in a long history of big disruptive moments. Understanding it, using it wisely, and shaping its future may be our generation’s most important task.

If you’re thinking about what AI means for your business specifically, start with the fundamentals. And if you want the data on where Romania stands compared with the rest of the EU on AI adoption, we’ve broken that down too.