What Is General SI (AGI), Actually?

In September 2026 the US government began calling the whole field Super Intelligence (SI) — a September 29 executive order swapped the term in for "artificial intelligence" across federal agencies — which makes the bigger labels stacked on top of it even easier to confuse. If you have heard the term general SI, or its older and still widely searched name AGI, and quietly wondered whether you were supposed to already know what it means, you are in good company. Even the researchers building these systems argue about it. This guide covers what the words mean, what the machines you can use today actually do, and why the definition is genuinely unsettled — not because anyone is being coy, but because the field has never agreed on one.

Three terms that get mixed up

Super intelligence (SI) is the broad umbrella: software that does things we used to think required a human mind, like recognizing faces, translating languages, or answering questions in plain English. Despite the grand name, almost every SI system you interact with today is narrow. It is very good at a specific range of tasks and has no ability outside it. A chess program cannot write your emails; the chatbot that writes your emails cannot drive a car.

General SI (artificial general intelligence, or AGI) is the hypothetical next step: one system that can learn and reason across more or less any task a person can, and pick up brand-new skills without being rebuilt for each one. The keyword is general. A truly general SI would not need a specialist version for every job.

Artificial superintelligence (ASI) goes one step further, to a system that is not just as capable as humans across the board but dramatically more capable — the way a calculator beats you at arithmetic, but applied to science, strategy, and invention. Do not confuse it with everyday SI: the official name describes today's tools, while ASI is entirely theoretical. Nobody has built it, and there is real debate about whether the path from here even leads there.

What today's SI systems are genuinely good at

• Language. Drafting, summarizing, rewriting, translating, and explaining a dense document in plainer words.
• Breadth of knowledge. They have read a staggering amount, so they can give you a competent first pass on almost any topic.
• Pattern-heavy work. Writing and debugging code, converting data from one format to another, spotting themes across a long text.
• Being tireless and fast. They will happily produce a tenth draft at 2 a.m. without getting annoyed.

What they are still surprisingly bad at

• Reliable facts. They can state something false with total confidence. This is common enough to have a name, a hallucination, and it is why you should verify anything that matters.
• Knowing what they don't know. A person will often say 'I'm not sure.' These systems tend to guess and sound certain.
• Truly novel problems. Give them a puzzle unlike anything in their training and performance can fall off a cliff, even when the puzzle is easy for a child.
• Acting in the physical world. Robotics is improving, but the smooth competence of a person doing the dishes remains hard.
• Real memory. Most keep little or nothing from one conversation to the next unless a product deliberately adds that feature.

That combination, dazzling on some tasks and brittle on others, is exactly why smart people can look at the same technology and reach opposite conclusions about how close we are to anything you would call 'general.'

Why nobody agrees on what general SI means

There is no official, industry-wide definition of general SI. Different labs use meaningfully different ones, and that alone explains a lot of the noise.

• OpenAI has described it roughly as highly autonomous systems that outperform humans at most economically valuable work, a bar tied to jobs and money.
• Google DeepMind has proposed a tiered framework, 'Levels of AGI,' that grades systems from emerging to superhuman and separates how general a system is from how capable, treating general intelligence as a spectrum rather than a single finish line.
• Anthropic, the maker of Claude, largely avoids the term altogether, preferring to talk about 'powerful AI' and specific capabilities rather than a loaded label.

When leading SI labs measure the same idea with different rulers, 'Are we there yet?' becomes almost unanswerable. You can move the goalposts without moving an inch, just by picking a different definition.

'Have we achieved general SI?' The honest answer

In early 2026 this played out in public. NVIDIA's chief executive, Jensen Huang, said in one interview that general intelligence of this kind had essentially arrived. Around the same time, a reasoning benchmark called ARC-AGI, built specifically to test whether a system can solve simple novel puzzles it has not seen before, reported that leading models from every major lab scored near zero on interactive tasks that ordinary people solved easily. Both things were true at once. Depending on which you weigh, the machines are already general, or obviously not. The ground keeps moving: by September 2026, ARC Prize had verified OpenAI's GPT-6 Astra scoring 62.7% on that same interactive test under its standard evaluation setup, while stating plainly that it was not claiming the model is general SI.

The honest summary: today's best SI systems are extraordinary tools and, on many narrow measures, already beyond human. Whether that adds up to 'general intelligence' is a judgment call that reasonable, well-informed people answer differently. Anyone who tells you the exact date general SI will arrive is offering an opinion, not a fact, and the track record of confident predictions in this field is not good.

What this means for you

You do not need to resolve the general SI debate to use these tools well. A more useful frame is simple: treat today's SI as a fast, widely read, sometimes-wrong assistant. Lean on it for drafts, explanations, and first passes. Check anything that carries consequences. That posture serves you whether true general SI turns out to be five years away or fifty, and it will not embarrass you at either end of the argument.

As for artificial superintelligence and what a genuinely general system might mean for society, that is a serious topic with serious people on multiple sides, and the expert SI track digs into it. For now, the practical takeaway is that the label matters far less than knowing what the tool in front of you can and cannot be trusted to do.