There's a paradox at the center of the AI boom. The systems attracting hundreds of billions of dollars in investment are, by their own designers' accounts, not "intelligent" in the way the word implies — and they still fail at things a child would not. Making sense of that gap is the key to understanding both why the money is flowing and where the technology heads next. This is analysis, not prophecy.

What today's AI actually does

Most of the tools people mean by "AI" today are large language models (LLMs) — systems trained on enormous amounts of text to predict the next word in a sequence, as IBM's technical overview describes. Trained at vast scale, that simple objective produces something remarkable: software that can draft an email, summarize a report, write working code and hold a fluent conversation.

But the mechanism matters. An LLM is, at heart, a very sophisticated pattern-matcher over language. It has no model of the world, no senses, no goals of its own. That's why it can produce fluent, confident text that is simply wrong — the failure mode known as "hallucination" — and why it can stumble on arithmetic or logic that seems trivial. It's generating plausible language, not reasoning about reality. When critics say AI is "not smart," this is what they mean: competence without comprehension.

Why the boom is still real

None of that makes the technology a mirage. Even as a pattern-matcher, AI is economically useful right now — automating drafting, coding, translation, customer support and analysis at a scale that changes how businesses operate. Measures of adoption and capability have climbed steeply in recent years, as Stanford's annual AI Index has documented. The investment case doesn't rest on the software being conscious; it rests on it being good enough at valuable tasks to justify the spending on chips, data centers and energy that has rippled through markets.

That's the tension driving the debate: real, monetizable utility on one hand; genuine limits on the other.

Where it's trying to go next

The frontier is a set of attempts to push past pure next-word prediction. Several directions stand out — described here as the industry's stated aims, not guarantees:

  • Reasoning models. Newer systems are trained to "think" in steps — working through a problem before answering — which improves performance on math, logic and coding, though it doesn't eliminate errors.
  • Agentic AI. Rather than just answering, "agents" are given tools and permission to take actions — booking, buying, running software, completing multi-step tasks. This is where much commercial hope now sits, and where reliability matters most.
  • Multimodal and "world" models. Systems that combine text with images, audio and video — and, researchers hope, build richer internal representations of how the physical world behaves — aim to address the "no model of reality" gap directly.

Each is an attempt to convert fluent language into dependable action. Whether they arrive at something deserving the word "intelligence" is exactly what the field disagrees about.

Why it matters

For the economy, the distinction is not academic. If AI's value comes from automating concrete tasks, the returns depend on reliability — an agent that books the wrong flight or a model that invents a legal citation destroys the savings it promised. That is why the next phase is less about bigger chatbots and more about trustworthy systems that can be safely handed real work.

For investors and companies, the honest framing is the useful one: today's AI is a powerful, imperfect tool, not a mind. The money chasing it is betting that the imperfections shrink faster than the costs — and that the leap from "sounds right" to "is right" is one the technology can actually make. Boursel gives no investment advice; the takeaway is to judge AI by what it reliably does, not by how human it sounds while doing it.