AI vs AGI: The Line Got Blurry When Nobody Was Looking

The textbook distinction is clean. AI is narrow: it does one thing — recognise faces, rank videos, hold a lane on the motorway. AGI is general: it does anything a human can, reasons across domains, and maybe one day wakes up with opinions. Neat, teachable, and increasingly wrong at the edges.

Because the thing we actually shipped doesn't sit cleanly on either side of that line.

The framing that stopped working

For years, "narrow AI" meant a model that did exactly one task and fell apart the moment you asked for anything else. Fair description, at the time. Then large language models arrived and quietly broke the taxonomy. A single model that writes code, drafts an email, explains a contract clause, and translates Latin is not "narrow" in any sense that matches how we used the word. It's also plainly not general intelligence — it will invent a citation with total confidence and then miscount the letters in "strawberry."

So what is it? The honest answer is that it's a new category the two-bucket model never had a slot for: broad but shallow. General in surface area, unreliable in depth. Calling it "narrow AI" is technically defensible and practically misleading, and that gap is where most of the confused arguments come from.

"AGI" is a word doing too much work

Here's the part these explainers skip: nobody agrees what AGI means. Ask ten researchers and you'll get a benchmark score, an economic threshold ("can do most remote work"), a philosophical bar ("understands, doesn't just predict"), and a couple who think the term is mostly marketing. These aren't small differences — they produce completely different answers to "are we close."

That's why the goalposts appear to move. It isn't always dishonesty. It's that we keep reaching capabilities we once called AGI-complete — passing the bar exam, holding a coherent conversation — and then deciding, on reflection, those weren't quite what we meant. The definition is downstream of whatever machines can't do yet.

The timeline question, honestly

You'll see predictions ranging from "a few years" to "never." Both ends should make you suspicious, and for the same reason: a confident prediction requires a definition, and we just established there isn't a shared one. Add that many of the loudest forecasts come from people raising money or defending a reputation, and the honest posture is discomfort. I don't know when, or whether. Anyone who tells you they know is selling something — a product, a fear, or a worldview.

What I'm fairly sure of is that it won't feel like a single arrival. It'll be the same thing we've already watched happen: capabilities landing unevenly, some domains falling fast, others staying stubbornly hard, and the label "AGI" getting applied in hindsight once we've stopped being impressed.

Why the distinction still matters when you're building

For an engineer, the useful move is to set the consciousness question aside. It's a fascinating philosophy seminar and a terrible design input. The systems in front of you today are capable-but-unreliable tools, and that's the whole spec you need: powerful enough to be worth using, wrong often enough that you must design for it. Guardrails, evaluation, a human on the decisions that matter. You build the same way whether or not there's a mind in there, because from the outside the failure modes are identical.

The AI-versus-AGI debate is genuinely interesting. It's just mostly irrelevant to shipping something that works on Tuesday.

What I actually think

We built something stranger than either bucket predicted — not the narrow tool, not the general mind, but a broad, shallow, confidently-wrong assistant that's more useful than "narrow AI" implies and less trustworthy than "intelligence" suggests. The interesting question was never "is it AGI." It's "what is this specific thing good and bad at" — and that one you can actually answer.