AI vs AGI

A distinction that decides what you promise

You are building on systems that exist. The debate about general intelligence is about systems that do not. Keeping those apart is not philosophy; it is what stops you promising a stakeholder something no model does.

The short version: today's models are extremely capable at tasks resembling their training, within a bounded context, with no persistence you did not build. General intelligence, as usually defined, means transferring to novel problems without task-specific setup. Nothing you can call over an API does the second thing.

Four properties of what you can ship

Naming the limits is more useful than naming the goal, because each one has a workaround you are responsible for building.

No memory between requests. A model starts every call knowing nothing about the last one. Everything that looks like memory is you resending a transcript or fetching notes, which is what external memory is for.

A bounded window. There is a hard limit, and quality degrades well before it. Context failure modes documents degradation from around 32k on one model and past 100k on another, against advertised windows far larger.

A knowledge cutoff. The model's world stopped on a date. Retrieval is how you move past it, and what is RAG covers the mechanism.

No verification of its own output. A model producing a wrong answer and a right one is doing the same thing with the same confidence, which is why evaluation metrics and a human in the loop exist.

An AGI, by most definitions, would have none of these limits. Your architecture is largely a response to all four.

The claim to avoid making

The practical harm from blurring this is a promise you cannot keep.

"It understands our business" is the sentence to watch. What is true is that it produces plausible text about your business given the context you supplied. Those sound similar and fail differently: one implies it will notice when something is wrong, and the other does not.

Two others in the same family. "It learns from our users" is false unless you built a feedback loop and retrained something. "It will get better over time" is false for a pinned model, and Anthropic Claude notes that pinned snapshots are now the norm, so improvement is an upgrade you perform rather than something that happens.

Capability is jagged

The useful mental model is not a single level of intelligence that keeps rising. It is an uneven surface.

A model can write a competent essay on a subject and fail to count objects in a photograph, which is among the documented vision weaknesses in multimodal AI. It can produce working code and be wrong about arithmetic. Neither is surprising once you stop expecting a single axis.

The consequence for building: you cannot infer from one capability to a neighbouring one. A model that handled your extraction task tells you nothing reliable about your classification task. That is why choosing the right model argues for twenty of your own cases rather than a leaderboard, and why the argument holds however capable the models get.

What to do with the debate

Ignore it, mostly, and keep two habits.

Build so the model is replaceable. Whatever arrives next, a thin interface and an evaluation set of your own means adopting it is a test rather than a rewrite. That discipline is worth having regardless of what the next release does.

Describe the system honestly, internally and to users. What it is for, what it refuses, and what it is unreliable at, which is the scope work in know your customers and use cases and the disclosure argument in AI safety and ethics.

Neither habit depends on who is right about AGI, which is the point of separating them.

Further reading

• What is an AI engineer: the work that exists because of these limits.
• Context failure modes: the measured shape of the bounded window.
• Choosing the right model: testing capability rather than inferring it.
• AI safety and ethics: describing the system honestly.