Massive AI models running on servers are incredibly powerful, but to make use of them, data has to be streamed back and forth via the cloud. This is not ideal for privacy-sensitive applications such as healthcare and smart homes, or safety-critical ones such as autonomous vehicles or robotics, where network lags could lead to accidents.

Future Horizons:

×××

5-yearhorizon

Simple AI models become ubiquitous

The number of devices capable of running simple AI models increases dramatically. This allows companies to generate far more useful insights from sensors and machinery throughout the supply chain, leading to significant efficiency gains. On-device processing also becomes standard for privacy-sensitive consumer applications like fitness tracking, health monitoring and smart homes.

10-yearhorizon

Autonomous robot deployment expands

Breakthroughs in simulation technology make it possible to train autonomous vehicles and robots far more quickly and cheaply than before, massively expanding their deployment. While state-of-the-art models remain large and centralised, the AI most people interact with on a daily basis becomes highly distributed. Both the training and running of these models is increasingly done on devices at the edge of the network.

25-yearhorizon

Humanity experiences ambient intelligence in the environment

Improvements in both hardware and software mean it is now trivial to add AI to almost any device. Humanity enters the era of “ambient intelligence” in which every element of the human-made environment responds intelligently to us in an intuitive and almost undetectable way. This dramatically simplifies people’s everyday lives, but raises deep questions about distinctions between human and machine agency.

Hence, running models embedded on “intelligent devices” will be imperative for many of the most promising AI applications.²⁸ However, smaller machines with limited power supplies are unable to support the computing and energy requirements of today’s leading models. This is spurring efforts to develop more efficient chips specialised for embedded applications, and optimisation techniques that can make models smaller, more energy-efficient and faster.²⁹³⁰³¹ Analogue AI chips are showing promise as a way to dramatically reduce the power required to run leading models.³²

Training these intelligent devices will also require innovations. Federated learning is a promising approach that distributes training over many smaller devices, reducing bandwidth requirements and improving privacy.³³ Reinforcement learning, in which AI learns to perform a task by repeatedly performing actions to maximise a carefully chosen reward, will keep improving the training of robots and other automated machines. This is slow and costly, though, and collecting enough real-world training data is a challenge. New open robotics datasets can help researchers pool resources, but training models in simulations and then porting them over to devices may work better.³⁴

Intelligent devices - Anticipation Scores

The Anticipation Potential of a research field is determined by the capacity for impactful action in the present, considering possible future transformative breakthroughs in a field over a 25-year outlook. A field with a high Anticipation Potential, therefore, combines the potential range of future transformative possibilities engendered by a research area with a wide field of opportunities for action in the present. We asked researchers in the field to anticipate:

The uncertainty related to future science breakthroughs in the field
The transformative effect anticipated breakthroughs may have on research and society
The scope for action in the present in relation to anticipated breakthroughs.

This chart represents a summary of their responses to each of these elements, which when combined, provide the Anticipation Potential for the topic. See methodology for more information.