The Illusion of Intelligence
Are LLMs really intelligent, or just extraordinarily powerful pattern matchers? A critical look at what these systems actually are—and what they're not.
I have been thinking about this for a while. Some will disagree, and that is fine.
Many people genuinely believe that ChatGPT, Claude, and other LLMs are intelligent. Not just useful. Intelligent.
But are they really?
Why We Think They’re Intelligent
The reason we attribute intelligence to these systems is simple: they communicate in human language. And when something speaks our language fluently, our brains instinctively infer a mind behind the words. These tools exploit that beautifully.
But if you look past the language and examine the fundamentals, they are extraordinarily powerful pattern matchers with exceptional memory.
A 2025 EMNLP study showed that even the most capable modern LLMs show consistent performance drops when irrelevant information is embedded into a problem. A human would instinctively filter the noise. The model struggles to do so reliably. It is finding patterns in context, not reasoning the way we reason.
Will We Reach AGI?
This matters when we ask: will we reach AGI, a system that can think and learn across any domain the way a human can? My view is that it is possible, but not on the current trajectory. There needs to be a paradigm shift.
Yann LeCun, Turing Award recipient and founding father of deep learning, argues that true intelligence requires a world model, the ability to simulate consequences and plan, not just predict the next word. That architectural gap is also why hallucinations are not a bug to be patched. They are a symptom.
A Personal Confession
Now here is where I will get personal.
I work daily with LLMs: context & retrieval engineering, fine-tuning, evaluation pipelines, system design, behavioural steering, architectural constraints. I build systems that, at times, genuinely surprise me with what they produce.
And sometimes I catch myself thinking: did I just build something that actually understands?
The answer is no. I am, in the most literal sense, a Master Illusionist. I have learned to construct the conditions under which intelligence appears to emerge. The scaffolding, the guardrails, the carefully engineered context create an illusion so convincing that I occasionally fool myself. And if I can fool myself working inside the machinery every day, it is no wonder the rest of the world does too.
The Real Value
None of this means these tools are not valuable. The automation, code generation, knowledge democratisation, and productivity gains are real and significant. LLMs possess a form of machine intelligence that, when understood correctly, is extraordinary.
The key phrase is when understood correctly.
At Tomorrow Built, this sits at the core of how we work. We know what these systems actually are, where they genuinely help, and where human intelligence needs to lead. The goal was never to sell the illusion. It is to build something real around it.