With US President Donald Trump recently pushing the use of “superintelligence” (SI) instead of artificial intelligence (AI) in national discourse, the acronym has gained unprecedented attention. It is the right acronym but the wrong meaning. In my opinion, SI should stand for “self-intelligence”.

It probably explains why academics, critics and even major AI companies are calling for a temporary pause in AI development. They must have realised the technology’s current trajectory is headed down the wrong path.

Producing an intelligent answer is vastly different from independently acquiring knowledge, deciding what to do and acting appropriately in an unpredictable world. I believe the next great challenge for AI is moving from machine learning towards achieving machine self-intelligence.

This evolution is best understood by three broad historical stages: AI 1.0, AI 2.0 and AI 3.0. In AI 1.0 – the first era – we asked whether we could make a machine to think.

Humans supplied the knowledge and logical rules for computers to solve specific problems. While expert systems successfully automated tasks requiring deep human expertise, their cognitive boundaries remained strictly confined to what humans programmed into them. We had simply computerised select aspects of human thought.

AI 2.0 is machines that learn. The transition occurred when we stopped manually programming every rule. Instead, machines began learning patterns directly from data. This rise of machine and deep learning expanded what computers could accomplish. Armed with massive data sets and vast computing power, we built today’s generative AI.

We shifted the question from “How do we make a machine to solve this?” to “How can a machine learn to solve this?”The results have been extraordinary.

Google DeepMind’s Gemini Robotics 2, for example, combines vision, language and action to control robotic bodies and perform complex physical tasks. Nvidia’s Cosmos 3 is designed to help physical AI systems reason about environments, simulate possible future states and generate actions. These advances suggest AI research is increasingly confronting a problem I studied for more than 20 years: how to connect perception to action through intelligence.

AI 3.0 is machines that acquire intelligence for themselves. I define intelligence as the self-ability that links perception and action to achieve intended outcomes. The key concept here is self-ability.

By machine self-intelligence, I do not mean deep consciousness, sentience or a machine developing a humanlike “self”. It is a machine’s capability of perceiving its environment, acquiring relevant knowledge, understanding its meaning, making decisions and taking appropriate actions. This is what I call AI 3.0.

In AI 1.0, humans provided machines with rules for thinking. In AI 2.0, machines learn patterns from data. In AI 3.0, machines must be capable of acquiring and using knowledge through their own interactions with the world.

Consider a cup on a table. A large language model trained on millions of text descriptions can perfectly explain what a cup is, its uses and the physics of ceramic breakage. A robot confronting a physical cup faces a dynamic problem. It must locate the cup, evaluate its shape and material, calculate grip force and continuously adjust its motor controls. If the robot detects the cup is filled with boiling coffee, this real-time signal must immediately alter its handling behaviour to avoid a spill.

While scaling compute and data sets has driven massive progress, static data is not the same as live knowledge. An autonomous machine does not just need more outdated data sets; it needs to understand what real-time information means in its immediate environment. This is why we must pivot towards knowledge models.

My work on KnowNet explores large knowledge models and how machines can progressively construct, organise and use knowledge rather than rely on detecting patterns in data sets that might be outdated. The critical question is no longer how much data a machine can learn from but how much useful knowledge it can acquire on its own.

Humans are not born with all the knowledge they will ever need. We perceive, receive guidance and constantly update our mental models. Autonomous machines will similarly encounter circumstances their designers cannot anticipate. Therefore, intelligent machines must be teachable.

A humanlike teachable mind in machines is able to translate raw environmental signals into structured concepts. Ultimately, a teachable machine must continue to learn from humans and its environment, recognise when its knowledge is lacking and take steps to acquire what it needs.

As machines gain autonomy, the responsibility grows. If a digital assistant generates an incorrect sentence, the harm is usually minimal. If an autonomous vehicle or industrial robot makes an incorrect physical decision, the outcome can be catastrophic.

Therefore, AI 3.0 must prioritise reliable autonomy. This requires robust safeguards, clear objective functions, constant human oversight and the capacity for machines to explain and correct their own errors.

The next major leap in AI will come from machines capable of recognising what they do not know, acquiring that missing knowledge and deciding how to act safely. That shift would mark the transition from machine learning to machine self-intelligence, and that is the real SI.