AGI

AGI is the abbreviation for “Artificial General Intelligence,” roughly translated as: general artificial intelligence. It refers to a computer system that solves arbitrary intellectual tasks as well as a human. Today’s programs can each do only certain things: write texts, recognize images, play chess. A general system like that, by contrast, would learn new things without anyone having to specifically prepare it for that. So far there is no program to which this applies. AGI is therefore a goal, not a product you can buy.

Why the term matters

A great deal of money revolves around AGI. Companies like OpenAI, Google DeepMind, or Anthropic explicitly name it as their corporate goal. Investors are pouring billions into this promise. Anyone who wants to understand the news about AI therefore needs to know what is meant.

The term also has a political dimension. If a machine really took over all knowledge work, that would have consequences for labor markets and safety. This is exactly what people who call for stricter AI regulation point to. Others consider the debate premature and good marketing.

One problem remains: there is no recognized definition. Some understand it to mean a system that can handle any office job. Others set the bar at scientific discoveries. Because the yardstick is missing, any company can claim to be close to it.

How one approaches the goal

Today the most important path leads through large language models. These are programs that have learned from vast amounts of text which word is likely to come next. They answer questions, write program code, and translate. This range surprises even many researchers themselves.

The widespread hope is called scaling: more data, more computing power, larger models. In recent years this has indeed reliably led to better results. Whether this path will ever end in a general intelligence, nobody knows. There could also be a limit beyond which additional size brings hardly any benefit.

Because there is a difference between capability and understanding. A language model recognizes patterns in texts very well. But it has no lasting memory across conversations and no goal of its own. It sometimes fails at tasks a child can solve, such as simple spatial reasoning. Many experts therefore consider additional building blocks necessary: reliable memory, independent planning, learning from experience during operation. A comparison helps: a student who knows all the formulas by heart is not yet a mathematician. What is missing is the ability to apply them sensibly in an unfamiliar situation.

Where you encounter the term

In business news, AGI frequently appears in reports about funding rounds. When a startup raises billions, the goal is often summed up in this word. It also appears in contracts: the dispute between OpenAI and Microsoft revolved, among other things, around who decides when AGI has been reached.

Anyone who reads sentences like “our model is a step toward AGI” should remain skeptical. In product marketing the term is elastic. It only becomes useful when it is accompanied by a statement of which task the system can concretely solve. Something similar applies to time estimates: some company executives promise a few years, independent researchers often name decades or hold back entirely.