Ken Sansom

Artificial General Intelligence, or AGI, is the hypothetical next step in the evolution of artificial intelligence (AI). It refers to systems that exhibit human level cognitive abilities rather than narrow, task specific skills.

Today’s AI systems excel at pattern matching for constrained problems. They are trained on vast amounts of data to perform specific tasks such as image recognition, language translation, content generation, recommendation systems, and question answering. Lately, most of us have used these AI applications (e.g. ChatGPT, Claude, Gemini, etc.). These systems can appear intelligent, but their competence is largely confined to the domains they were trained on. Increasingly, they are multimodal, meaning they can process different types of data input, including images, text, and speech, which increases their perceived comprehension.

Of course, all of us have discovered that they do make mistakes and even after you correct them, they’ll repeat the mistake. I like to remind people that the answer is probabilistic not deterministic…meaning the possibility that your AI tool of choice is wrong always exists.

AGI would be fundamentally different.

An AGI system would be capable of general reasoning, meaning it could apply knowledge across domains it has never encountered before. It would demonstrate transfer learning, acquiring new skills without retraining or reprogramming. An AGI agent could plan independently, adapt to novel situations, and reason abstractly, forming concepts rather than merely recognizing patterns.

At present, AGI remains more theory than reality. However, active research is underway across governments and the technology sector as researchers explore pathways from narrow AI toward more general intelligence. It is a global effort by leading organizations including Amazon, Google, OpenAI, Microsoft, Mistral AI, the Beijing Institute for General Artificial Intelligence, and many others.

Will AGI succeed? No one knows. The leap from today’s AI systems to truly general intelligence is significant. But humans have repeatedly demonstrated an ability to solve problems once thought impossible.

In fact, there is still debate about what characteristics a system would need to have to be regarded as AGI. As mentioned, we would expect it to be able to learn, operate autonomously, demonstrate creativity, plan, and exhibit the ability to apply knowledge across domains it hadn’t been trained on. In 1950, Alan Turing, proposed the Turing Test, which was a measure of a computer’s ability to converse at a human level. It isn’t a valid test of AGI, as it doesn’t assess other areas outside of conversational ability.

If achieved, AGI could deliver profound benefits, including accelerated scientific discovery, autonomous research and engineering, breakthroughs in medicine and climate science, and unprecedented productivity gains.

Yet the risks are equally profound. What happens if human goals and AGI objectives diverge? What if intelligence scales faster than governance, concentrating power at a rate no society can absorb? And if AGI inherits not only our intelligence and capabilities but also our cognitive biases and frailties, how do we prevent it from amplifying our worst tendencies rather than our best?

Like life itself, AGI is unlikely to arrive all at once. It will emerge gradually through incremental improvements in reasoning, autonomy, and transfer learning. The challenge ahead is not merely building more capable systems but ensuring that autonomy and adaptation are guided by meaningful guardrails.

What do you think?