What Is AI, Really? A Clear, Honest Guide for Curious Readers

The technology nobody asked for and everybody now uses

In November 2022, OpenAI launched ChatGPT as a research preview. Two months later it had 100 million monthly users, and most of them still couldn't tell you what it actually was.

This guide exists because "AI" has become a word that means everything and therefore nothing — applied to chess engines, Netflix recommendations, autopilot, and software that holds genuinely surprising conversations. We're going to be specific about what AI is, what the recent breakthrough actually was, and why an idea from 1956 took 70 years to suddenly work.

AI is software that does what we used to need humans for

The plain definition: artificial intelligence is software that performs tasks which previously required human cognition. Recognising a face, translating Mandarin to English, drafting a legal memo, suggesting the next word in a sentence. That's it.

This excludes a lot of things people call "automation." A script that emails you when the server crashes is not AI. A spreadsheet macro is not AI. The line is whether the rules can be fully written down by a human in advance — AI exists for tasks where they cannot.

The slippery part: every time AI starts working, we stop calling it AI. Optical character recognition, GPS routing, spam filtering, spell-check — all were called AI in the 1990s. Today they are "just software." This is the so-called AI effect, and it is part of why the term feels permanently undefined.

A more useful working definition: AI is software that learns patterns from data instead of following rules a human wrote. That single shift — from rule-based to data-driven — is what separates a transformer from a thermostat.

The honest definition isn't really about intelligence. It's about who wrote the rules: a programmer, or the data.

Every AI in production today is narrow

Researchers split AI into two categories. Narrow AI does one thing well — translate, transcribe, recommend, generate text. General AI, often called AGI (artificial general intelligence — a system that learns any cognitive task with the flexibility of a human), does not yet exist, and no serious researcher will give you a confident timeline.

Every system shipping today is narrow. GPT-4 (OpenAI), Claude 3 (Anthropic), Gemini (Google DeepMind) — these are large language models, or LLMs (neural networks — software loosely modeled on the way brain cells connect, which learn by adjusting numerical weights — trained on enormous text corpora to predict the next word, billions of times). They look general because they handle text on any topic. They are not. They are extraordinarily good narrow systems, narrow at the task of next-word prediction.

The interesting wrinkle is that this narrow task — predict the next word from everything ever written — turns out to require approximating an enormous range of human reasoning. Predicting the next word in a Python program means understanding code. Predicting the next word in a legal brief means understanding precedent. The narrow task generalises by accident.

What most people miss is that the marketing wants you to confuse "general-purpose" with "general intelligence." Today's models are the first; they are not the second. (For the mechanics of how LLMs actually produce text, see What is an LLM?.)

Three things explain why AI suddenly works

AI as a field dates to a workshop at Dartmouth in the summer of 1956. For the next 50 years, it mostly underdelivered. Two "AI winters" — periods when funding collapsed after a wave of hype — followed each cycle of overpromise. The neural networks at the centre of today's models were proposed by Frank Rosenblatt in 1958. They didn't really work for half a century.

Three things changed in the 2010s. First, the internet produced text and image corpora vastly larger than anything previously available — the data. Second, NVIDIA's GPUs (graphics processing units — chips originally designed to render video games, but architecturally well-suited to the parallel maths neural networks require) became cheap and abundant — the compute. Third, in 2017, eight researchers at Google published a paper called "Attention Is All You Need," introducing the transformer — the architecture every modern LLM is built on, which lets a model weigh how every word in a sentence relates to every other.

Then came the scaling hypothesis. In 2020, OpenAI researchers led by Jared Kaplan showed that performance improves predictably as you grow three quantities: parameters, training data, and compute. Bigger and better-fed models simply got better, on a smooth curve. DeepMind's "Chinchilla" paper (2022) refined the recipe. That observation — more than any algorithmic trick — is why GPT-4 exists.

The breakthrough wasn't a new theory. It was an old idea finally meeting enough data and compute to demonstrate what it could do at scale. There was no Eureka moment. There was a long engineering grind, and at some point the curves crossed.

What current AI can and can't do, today

Frontier models can write competent code, summarise a 100-page contract in 30 seconds, draft prose in almost any voice, translate between major languages with near-human fluency, and generate photo-realistic images from a text description. GPT-4 scored in the top decile on the bar exam in OpenAI's published evaluation. Claude 3 routinely outperforms human baselines on graduate-level reasoning benchmarks. These are not parlour tricks.

They cannot reliably do simple arithmetic on numbers they haven't seen. They cannot plan over long horizons. They do not know what they don't know. They cannot distinguish a true claim from a confidently-stated false one — the "hallucination" problem, where a model fabricates facts that look plausible because plausibility is what it was trained to produce.

A concrete example: in 2023, a New York lawyer named Steven Schwartz used ChatGPT to research a brief in Mata v. Avianca. It invented six legal cases — names, docket numbers, judges, fabricated quotations from non-existent rulings. The court was unamused; Schwartz was sanctioned. The fake citations looked exactly like real ones because the model has no internal mechanism distinguishing what is true from what is well-formed.

For most professional work, current AI is best thought of as a fast, broadly competent intern who has read everything but understands nothing in particular, never gets tired, and will confidently make things up if you don't check. That framing tells you precisely what to delegate and what to verify.

The mistake is treating it as an oracle. The right framing is a collaborator with selective amnesia and a confidence problem.

A general-purpose technology, not a miracle

Economic historians reserve a small category for inventions that reorganise the entire economy — they call them general-purpose technologies. The steam engine, electricity, the internal combustion engine, the internet. AI looks like the next one on the list.

General-purpose technologies share three traits: they apply across almost every industry, they improve dramatically over decades, and they enable secondary innovations that were not imaginable beforehand. Electricity didn't just replace gas lamps; it made factories, refrigeration, telecommunications, and modern cities possible. The full diffusion took roughly 40 years.

AI is at something like the electricity-in-1900 stage. Clearly important, obviously transformative in some directions, but genuinely unclear which industries it will reshape first or how deeply. The hype cycle and the long-term trajectory are not the same curve, and confusing them is the most common mistake intelligent observers make.

The honest take: it is neither the end of work nor an overhyped chatbot. It is a new general-purpose layer of capability that, like electricity, will take 30 years to fully diffuse and look obvious in retrospect. The question is not whether it will change everything — it will. The question is which industry it changes first, and whether you happen to work there.

What this means for you

If you build software: the platform shift is happening at the application layer, not the model layer. OpenAI, Anthropic, Google DeepMind, and Meta FAIR have already commoditised the underlying capability. The work — and the value — is in figuring out what to build on top, which requires knowing your domain better than the model does. The companies that win will look less like AI labs and more like deeply opinionated software businesses that happen to use AI.

If you invest or operate: distinguish between companies using AI as a product and companies using it as a feature. NVIDIA's revenue ran roughly $61B in fiscal 2024, up from $27B the year before — the infrastructure is being built at speed. Which application companies will compound durably on top of that infrastructure is still wide open. Be sceptical of any pitch that ends with "and we use AI."

If you are a citizen of the early 21st century: the single most useful thing you can do is spend an hour with Claude or ChatGPT, ask it to do something inside your actual professional area, and watch carefully where it succeeds and fails. That hour will give you better intuition than a year of reading op-eds. Most of the public conversation is being conducted by people who haven't seriously used the tools.

Ask Gradient

For deeper questions, Gradient's Ask Gradient tool — a RAG system (retrieval-augmented generation: an LLM grounded on a curated corpus, so it cites real sources rather than inventing them) — lets you query our full archive of AI guides and research notes. Try something like "What is the practical difference between today's LLMs and AGI, and why does it matter for how I should think about hiring in the next five years?" — you'll get a sourced, specific answer rather than an opinion column.