On the home page we pictured the three terms as three floors of one house. Here we take a closer look at what each floor holds, why experts disagree about exactly where the first floor ends, and how to read news that someone has “achieved AGI” accordingly.
AI is more than a chatbot
Artificial intelligence has been with us longer than it seems. The filter that sorts your spam, the satnav that finds the fastest route, the recommendation for your next film, the program that flags a suspicious spot on an X-ray — all of that is AI, and none of it chats with you. The field covers very different methods. Some learn from data (machine learning, which today's models are built on); others follow rules written by people or systematically search through possibilities, like a chess program.
Beware of similar-sounding words, too. Generative AI produces content: text, images, music. General AI handles a wide range of different tasks. A chatbot generates text and still need not be general; a system that generates nothing at all can be highly capable across many tasks.
What AGI means
AGI, artificial general intelligence, is the term for general-purpose AI. The trouble is that everyone sees something slightly different in it — the researchers Shane Legg and Marcus Hutter collected over seventy definitions of intelligence alone. Some mean the ability to learn and reason in any domain the way a person can. Others, such as OpenAI in its charter, speak of systems that outperform humans at most economically valuable work. Others again stress that the system must be able to learn a new task it was never prepared for.
This guide uses AGI for a system that handles a wide range of tasks at roughly human level and can learn new ones. That is a working definition, not a standard. And that is exactly why the question “is model X AGI yet?” makes no sense until we say which tasks, compared with whom and under what conditions. Assessing breadth, level of performance and autonomy separately is what Levels of AGI, a proposal from researchers at Google DeepMind, suggests; it too is a proposal for discussion, not a binding scale.
What ASI means
ASI, artificial superintelligence, is superintelligence: clear superiority over the best humans in practically all significant cognitive domains. Calculating faster than a mathematician or winning one test is not enough — a calculator and a chess engine do that today. Bostrom's detailed definition is in his paper.
The definition speaks only of capability. A superintelligence would need no human body, human personality or consciousness. Nor would it be infallible: even the smartest system needs data, which are sometimes missing, and the future stays uncertain for it too.
The difference between AGI and ASI is therefore mainly one of level of broadly applicable performance. And the arrival of general AI would set no date for superintelligence.
Four questions instead of one label
When you want to judge a particular system, it helps to split it into four questions:
- Breadth: how many different kinds of task can it handle?
- Level: how well does it do them compared with people who do them professionally?
- Reliability: does it succeed repeatedly and notice its own mistakes?
- Autonomy: how much work does it do without human direction?
A model can be an excellent mathematician and a poor organiser at the same time. Another may reliably clear a pile of administration without reaching the top in science. A single “AGI yes/no” label hides these differences; the four questions show them.
Autonomy is also not authority. A system may be able to plan a purchase and still have no right to make it — that depends on the permissions people give it. More on the AI agents page.
What not to confuse with AGI
- Consciousness is subjective experience. The ability to solve tasks says nothing about it, in either direction.
- Agency is action directed at a goal. A narrowly focused agent can be highly autonomous and not general at all.
- General-purpose model is a technical term and, in the EU, a legal one for a model with broad uses. It does not mean that anything has legally “reached AGI”.
How to read news of “AGI achieved”
When you see a headline that someone has achieved AGI, look for three things: which definition they used, what concrete evidence they presented and who verified it independently. You will often find people arguing about the word while agreeing on what the system can actually do. How to tell real progress from hype is covered on When and how it might arrive.