Generative AI
The Machine That Learned to Predict the Next Word — and Accidentally Learned Everything Else
The most powerful AI systems in the world were not designed to reason, write poetry, or pass medical exams — they were designed to guess what word comes next in a sentence.
The Idea
There is something almost embarrassing about the core task at the heart of large language models: predict the next token. Not 'understand language', not 'model human cognition' — just, given everything before this point, what probably comes next? It sounds trivial. It turns out to be anything but. The reason is that to consistently predict the next word across billions of sentences drawn from the full breadth of human writing, a model must implicitly learn an extraordinary amount about the world. It must learn grammar, yes, but also causality, geography, social dynamics, logic, and the conventions of dozens of disciplines. There is no shortcut. You cannot reliably predict what word follows 'the patient was diagnosed with' without absorbing a great deal of medical knowledge. You cannot complete legal arguments, debug code, or finish a Shakespeare sonnet without having internalised the deep structure of those domains. This is why researchers were genuinely surprised when early large models began demonstrating what appeared to be reasoning abilities nobody had explicitly trained for. The capabilities emerged — not from a new architecture or a clever objective, but simply from doing the humble task of next-token prediction at sufficient scale, on sufficiently rich data. Generative AI, then, is less a technology that was engineered to be brilliant and more one that stumbled into brilliance by being relentlessly drilled on a deceptively simple exam.
In the World
In 2020, OpenAI released GPT-3, a model with 175 billion parameters trained on a vast sweep of internet text. The team had not explicitly taught it to translate between languages, write functional code, or answer arithmetic questions. Yet when researchers began probing it, they found it could do all three — imperfectly, but unmistakably. One moment that shook the AI research community came when Jason Wei and colleagues at Google Brain began documenting what they called 'emergent abilities' — capabilities that appeared suddenly in larger models but were essentially absent in smaller ones. It wasn't a smooth curve of improvement; it was more like a threshold effect. Below a certain scale, a model would be hopeless at multi-step arithmetic. Above it, the ability would simply be there. This was unnerving precisely because nobody had a clean explanation for it. The models weren't programmed to acquire these skills. They appeared to be a side-effect of the compression process — of learning to represent enough of the world's knowledge compactly enough to make good predictions. The implication is profound and a little unsettling: we have built systems whose full capabilities we cannot predict in advance, trained on data we cannot fully audit, exhibiting behaviours we cannot entirely explain. The next-token prediction objective is clear. What it produces, at scale, is not.
Why It Matters
Understanding what these systems actually are — prediction engines that generalised far beyond their training task — changes how you should think about their outputs and their limits. When a language model gives you a confident, fluent answer, it is not retrieving a verified fact or running a logical proof. It is producing the most statistically plausible continuation of your prompt, given its training. That is often genuinely useful, sometimes brilliant, and occasionally completely wrong in a way that sounds utterly convincing. The fluency and the accuracy are decoupled. This also reframes the ongoing debate about whether these models 'truly understand' anything. That question may be less interesting than the practical one: what kinds of tasks does the next-token prediction objective equip a model for, and what kinds does it silently fail at? Pattern-rich domains with stable underlying rules — coding, translation, summarisation — tend to go well. Novel reasoning chains, real-time facts, and tasks requiring genuine accountability tend to expose the seams. Knowing this, you become a sharper user of these tools: delegating confidently where they excel, and staying appropriately sceptical where statistical plausibility and factual accuracy are most likely to diverge.
A Question to Ponder
If a system learns to mimic the outputs of human reasoning well enough to be practically indistinguishable from it, at what point does the distinction between mimicry and understanding stop mattering — and to whom?
Get a new one of these every morning.
Start learning with Thinkable