Humans have evolved over billions of years to perceive and interact with the physical world, while AI systems are trained on static datasets using huge numbers/ clusters of processors. 

Already today, many AIs and robots are better than humans in many tasks. They can classify objects in images better than we can. They beat us at many games – like chess and Go, – and they achieved the highest rank in the computer game StarCraft II.

Human cognition, though, involves complex factors like emotions, consciousness, and self-awareness. This is difficult to replicate in machines.

Today’s AIs might appear as intelligent as humans, but they lack a true understanding of context and cause-effect relationships. Current AI architectures – the way the technologies are built – don’t allow to replicate human-like cognition. In fact, in order to fully replicate something, humans must understand the underlying mechanisms – biochemistry, relationships between different parts of the human brain and the body – and decode human consciousness. We dedicate several episodes of our podcast, “AI Snacks With Romy & Roby”, to studying human consciousness. 

Most of today’s AIs and robots are based on deep learning models, which have made impressive advances, but still have major limitations when it comes to general intelligence capabilities like reasoning, common sense understanding, and adaptability across diverse domains. Their performance on tasks requiring zero-shot generalisation to novel situations is still quite poor compared to humans. Recent research has shown that even infinitely increasing model size – adding more data into a model – cannot increase zero-shot performance. Realistically, the only way to improve performance is to invent an architecture that is so radically new that it has only a few things in common with any current approach.

Developing human-like or artificial general intelligence (AGI) systems will require massive amounts of computing power beyond what’s currently available. Advances in specialised AI hardware and innovation in AI chipset design – and potentially quantum computing – may be needed.

Surveys of AI researchers show a wide range of estimates for when AGI might be achieved. Median predictions often fall around 2040-2060. However, there’s significant uncertainty and disagreement among AI researchers and developers.

Claims of AGI powers of large language models from companies like OpenAI are based on their attempts to justify high valuation rather than reflect the true stand of their technology. This does not diminish their achievements; we are only providing a realistic and healthy view of what AIs are and aren’t today.

In our view, rather than relying solely on deep learning, future AGI systems may combine deep learning with other AI techniques like symbolic reasoning, reinforcement learning, and cognitive architectures inspired by neuroscience. Building such technologies requires a lot of money. Unfortunately, governments and commercial institutions don’t invest as they try to capitalize on what is already there – powerful deep learning technologies – which are routed in mathematics from the 1940s, 1950s, and 1960s.

While deep learning as we know it today didn’t fully emerge until the 2000s and 2010s, its foundations were indeed laid in the mathematical and computational work of the mid-20^th century. The resurgence and success of deep learning in recent years have been built upon these early mathematical foundations, combined with increased computational power and the availability of large datasets.