Is it meaningful to talk about different levels of human and machine intelligence?

Comparing levels of human and machine intelligence is not straightforward. AI systems today excel in specific domains where there is a lot of training data, and outperform humans in many small domains, while there remain tasks where humans perform better than AI.

‘Intelligence’ is also not a clearly defined concept, and is difficult to measure. Illustrative of this is the ‘AI effect’, which is the effect that once a task is accomplished by AI, it has tended to be seen as less representative of intelligence. However, there is a real sense in which AI systems have increasing capabilities over time, and it is useful to discuss different levels of intelligence to understand and frame this.

Changing capabilities

AI systems are becoming more powerful. The best models are making use of more and more compute, contain more trainable parameters, and are making use of more data. At the same time, AI algorithms have become more efficient, requiring less to achieve comparable results.

Compute and data required to achieve the same accuracy, from Epoch