The ‘I’ in AI Isn’t Really Intelligence
Call a Large Language Model ‘intelligent’ and you’ll win a headline; call it a pattern matcher and you’ll win clarity.
The promise of Artificial Intelligence rings on the digital ether in a gold rush of manic proportions – tech giants such as Google, Amazon, Microsoft and Meta have pledged a whopping $750 billion on the construction of data centers to fuel their AI models, with global spending expected to sail past over $3 trillion on AI development by 2029.
Yet the reality behind these trillions demand much more careful consideration.
The Large Language Models (LLMs) of today are essentially probability models that work by predicting the next token given a large amount of text. While useful and sometimes even practical, it is still nowhere close to what humans refer to as ‘intelligence’. Recent studies indicate that LLMs routinely fail tasks that demand sustained, algorithmic, or step-by-step planning; their solutions may be plausible, but still lack nuance or just be plain wrong (read: hallucinations).
Consider LLMs as a world-class chess commentator who can recount thousands of games, patterns and tactics from memory – even advise a brilliant move – but cannot reliably calculate an entire tree of all future outcomes. Or, as a high-roller poker player who has premeditatedly memorized all the tells of his or her rivals, in addition to all the best predetermined statistical plays. Impressive, yes. However, alter the rules of the game or introduce a variant, and their carefully constructed pattern-matching superstructure falls to pieces.
The current power of AI is thus not in real reasoning or understanding, but rather based on gargantuan training, data and computing resources. This distinction is key to a CFO considering whether to rely on a model or a product head promoting an AI ‘assistant’ to their customer.
The Illusion of Intelligence
It’s high time we abolish the romantic notions amplified by popular culture. AI is nowhere close to the usually-described futuristic force capable of independent reasoning or decision-making. What we have instead are a tremendous, but severely restricted, category of extremely advanced pattern-recognizing software.
Such systems excel at limited, well-defined tasks – predicting broad market trends, detecting fraud or sifting through vast datasets – often outperforming human speed in these narrow domains. However, this is not intelligence – it is a task-specific and data-dependent product of human design. Herein lies the nub of the issue: if the capabilities of current AI, particularly the much-hyped LLMs, revolve largely around the ability to match patterns in linguistic data, then what actually is intelligence?
Artificial General Intelligence (AGI) is often proclaimed as more of a theoretical idea, a machine defined by the capacity to reason, learn and apply knowledge across fields, as a human can. This is also exactly the chasm that current LLMs, despite their linguistic dexterity, are unable to bridge. They are masters of statistical correlation and generation, not of causal inference or cross-domain problem-solving. They mimic reason, without truly reasoning things themselves.
What businesses need, then, are not just more advanced LLMs, but a new class of models altogether.
The current research frontier is indeed moving towards just that – beyond raw next-token prediction to the sort of goal-driven architectures and training procedures that can be used to support reasoning: Large Reasoning Models (LRMs).
From Parrots to Conceptualizers
The present class of LRMs simulates the external appearance of reasoning: chaining arguments, employing tools and generating sensible explanations, since they have learned statistical regularities for what reasoning looks like in their training data. That mimesis is strong and frequently beneficial but it is not identical to having internal, causal models of the world.
True reasoning entails building and testing hypotheses, maintaining stable abstractions across modalities, running counterfactuals and revising or updating internal beliefs in the face of new evidence inconsistent with previous assumptions. In contrast, current LRMs represent the world using surface patterns encoded in weights and activations; their “chains of thought” emergent artifacts of pattern recognition.
Practical consequences follow quickly. LRMs are brittle to distributional shifts, prone to confident-but-wrong answers and poor on genuinely novel problem-solving where principles must be transferred across domains. They do not have durable goals, long-term memory with verifiable provenance or an autonomous drive to validate information – all aspects central to human-style deliberation.
For organisations, this implies that LRMs are most effectively viewed as advanced heuristics and enhancers of human judgement rather than strategists in their own right. To extract maximum value, companies should twin such models with rigorous verification and human checks and architecture that externalize planning and responsibility instead of presupposing these models are doing the ‘thinking’ for themselves.
Hybrid models combining longer internal ‘thought’ trajectories, explicit tool use (calculators, search, databases) and reinforced learning-to-reason mechanisms in a way they can plan, verify and chain steps is clearly the next evolution we are gearing towards. Initial surveys and technical reports explain this as an already-concerted swing: stricter and more structured objectives, increased scaffolding and more externalized computation. These would not simply produce plausible text out of taught patterns, they would struggle with figuring out the underlying logic, discover causality and transfer knowledge between domains.
Consider a situation in which an LRM system resembles a mature corporate strategist: studying competitive behavior in one market, extracting strategic lessons and then applying those innovatively and fruitfully in a completely different sort of market. This goes beyond fluency with statistics, it is true strategizing. LLMs are parrots extraordinaire; LRMs, when they mature, will be conceptualizers.
— The path to this deeper realization of intelligence requires a deep recalibration of our expectations and hence, our investments. It implies going beyond the present-day arms race for generative AI infrastructure to a more deliberate focus on systems that can actually reason, learn and apply knowledge.