The Navy revealed the embryo of an electronic computer today that it expects will be able to walk, talk, see, write, reproduce itself and be conscious of its existence. …The service said it would …build the first of its Perceptron thinking machines that will be able to read and write. It is expected to be finished in about a year at a cost of $100,000.

– New York Times [1958]

Predicting the future is always perilous. Brooks [2018] gives more recent informed predictions, and updates the progress every year.

Over the last decade there has been an explosion of applications that rely on large datasets and immense computation power, fueled by online datasets and the use of vector processing units (including GPUs), especially by large corporations that can afford huge computation centers. However, in the science of AI, integrating more of the dimensions is happening much more slowly.

For the technology, there are some predictions that seem safe, given the current state of the art in Figure 19.1.

For low-stakes decisions where there is abundant homogeneous data, such as vision, text, video, and big-data science, it is likely that more data and more compute power, together with improved algorithms that are being developed, will lead to better predictions in these cases. Universal function approximators, functions that can approximate any functions on bounded domains, such as neural networks, have been shown to be very effective when provided with abundant data and computation power.

Generating images, video, text, code, and novel designs for drugs and other chemicals, in what is known as generative AI, will get more sophisticated. When used for high-stakes decisions, unless they are generated to be provably correct, the predictions will need to undergo critical evaluation to ensure they can be used as reliable components for the decision task.

Improvements in predictive technology are likely to have beneficial outcomes because better predictions lead to better decisions. However, they can also have harmful outcomes for people when the values embedded in decisions do not coincide with the wellbeing of those people. For example, the use of generative AI to produce text, images, video, and music is likely to explode. These techniques could be used in the workflow to create art, possibly enhancing human creativity, or to create deep fakes, designed to mislead. There is an arms race to build and detect these fakes, but note that adversarial networks (including GANs) explicitly build models to counter efforts to detect them.

One medium-term development we can expect is for cases when the world in deployment is different than the world the training data is from, or in transfer learning, when using data from one domain in another. Observational data alone is not sufficient to predict the effect of actions; causality and expert knowledge needs to be taken into account. We expect more interactions between subareas.

There are many cases where there is no abundant data. For example, the SNOMED CT medical ontology has about 100,000 terms for diseases and other underlying causes that humans can have. The long tail of the probability of diseases means that for most diseases there are very few people with the disease. For nearly all of the pairs of diseases, no one in the world has both. We cannot learn the interactions of these diseases from data alone, as there are not enough people to cover all of the interactions. For these cases, more sophisticated models need to be considered, and expert knowledge cannot be ignored. We expect to see more integration of data that has rich metadata and expert knowledge to build hypotheses that match both the data and prior expectation.

For decision making in the world, an agent cannot learn from passive observation, such as text or video alone. Text only contains what someone thought was interesting, and does not contain mundane details that everyone knows. Video does not specify the actions of the agent, and if it did, it does not specify what the agent would have done in other circumstances, which is counterfactual reasoning. Except in artificial domains, an agent is never in the same state twice. An agent that is embodied interacts with the environment. An embodied agent in the real world needs common sense [Brachman and Levesque, 2022b]. This includes being ready to react to the unexpected; something never experienced before.

For high-stakes decisions, preference elicitation from affected stakeholders needs to be taken into account. In general, preferences are not obtainable by learning from data. While maybe it is possible to hypothesize someone’s preferences from their actions, in what is called inverse reinforcement learning, it is difficult to know what they would have done in other circumstances, or they might regret their actions, or their preferences may have changed. The various stakeholders may have very different preferences, and somehow these need to be combined in a fair and transparent way for an agent that affects multiple stakeholders.

Some have suggested that computers will become more intelligent than people, reaching a singularity, in which case people might not be needed. There is much speculation about when computers will be more intelligent than humans. This is difficult to define, as it is not clear what “humans” or “computers” mean. Humans could refer to a random person answering questions about a random culture, an educated person answering questions about their own culture, a world expert, or a person who has access to the Internet and others who are connected. A computer could refer to a standalone phone or laptop, a phone or laptop connected to the Internet, the whole World Wide Web without the humans in the loop, or the World Wide Web with humans in the loop. At the most general level of these it isn’t clear what the difference is.

Some have predicted that AI will lead to a dystopian future ruled by machines, but others see a nirvana – an ideal or idyllic place – where good decisions are made and people are looked after. Which of these arises depends on the people who build the AI and those who regulate the AI.

Turing [1950] concluded his seminal paper with “We can only see a short distance ahead, but we can see plenty there that needs to be done.” This is still true today, even though there have been considerable advances since then.