The language and image recognition capabilities of artificial intelligence (AI) systems have developed rapidly.

This chart zooms into the last two decades of AI development. The plotted data stems from several tests in which human and AI performance were evaluated in five domains: handwriting recognition, speech recognition, image recognition, reading comprehension, and language understanding.

Within each domain, the initial performance of the AI is set to –100. Human performance is used as a baseline, set to zero. When the AI’s performance crosses the zero line, it scored more points than humans.

Just 10 years ago, no machine could reliably provide language or image recognition at a human level. However, AI systems have become much more capable and are now beating humans in these domains, at least in some tests.

This series of nine images shows how these have developed over just the last nine years. None of the people in these images exist; all were generated by an AI system.

This is one of the critical evolutions of AI systems in recent years: not only do they perform well on recognition tasks, but they can also generate new images and text with remarkable proficiency.

Even more importantly, since 2021, the highest-performing AI systems – such as DALL·E or MidJourney – can generate high-quality, faithful images based on complex textual descriptions.

The ninth image in the bottom right shows that even the most challenging prompts – such as “A Pomeranian is sitting on the King’s throne wearing a crown. Two tiger soldiers are standing next to the throne” – are turned into photorealistic images within seconds.

A key takeaway from this overview is the speed at which this change happened. The first image is just eight years older than the last.

In the coming years, AI systems’ ability to easily generate vast amounts of high-quality text and images could be great – if it helps us write emails faster or create beautiful illustrations – or harmful – if it enables phishing and misinformation, and sparks incidents and controversies.

Current AI systems result from decades of steady advances in this technology.

Each small circle on this chart represents one AI system. The circle’s position on the horizontal axis indicates when the AI system was made public, and its position on the vertical axis shows the amount of computation used to train it. It’s shown on a logarithmic scale.

Training computation is measured in total floating point operations, or “FLOP” for short. One FLOP is equivalent to one addition, subtraction, multiplication, or division of two decimal numbers.

All AI systems shown on this chart rely on machine learning to be trained, and in these systems, training computation is one of the three fundamental factors that drive the system's capabilities. Other critical factors are the algorithms, the input data, and the parameters used during training.

The chart shows that over the last decade, the amount of computation used to train the largest AI systems has increased exponentially. More recently, the pace of this change has increased. We discuss this data in more detail in our article on the history of artificial intelligence.

The recent evolution of AI, particularly large language models, is closely tied to the surge in computational power. Each dot on this chart represents a distinct language model. The horizontal axis shows the training computation used (on a logarithmic scale), measured in total floating point operations (“FLOP”). The vertical axis indicates the model's performance on the Massive Multitask Language Understanding (MMLU) benchmark, an extensive knowledge test composed of thousands of multiple-choice questions across 57 diverse subjects, from science to history.

As training computation has risen, so has performance on these knowledge tests.

OpenAI's GPT-4, released in 2023, achieved an 86% accuracy on the MMLU benchmark. This far exceeds the 34.5% accuracy achieved by non-expert humans, and comes close to the 89.8% accuracy estimated for hypothetical human experts¹ who excel across all 57 subjects covered in the test.²

AI technology has become much more powerful over the past few decades. In recent years, it has found applications in many different domains.

A lot of this was achieved with only small investments. But this has increased dramatically in recent years. Investments in 2021 were about 30 times larger than a decade earlier.

Given how rapidly AI developed in the past – despite its limited resources – we might expect AI technology to become much more powerful in the coming decades, now that the resources dedicated to its development have increased so substantially.

The machines that power AI systems rely heavily on specific hardware. These include central processing units (CPUs) and graphics processing units (GPUs), which allow them to analyze and process vast amounts of information.

More than 90% of these chips are designed and assembled in only a handful of countries: the United States, Taiwan, China, South Korea, and Japan.

While reporting on AI tends to focus on software and algorithmic improvements, a few countries could, therefore, dictate the direction and evolution of AI technologies through their influence on hardware.