Despite progress in deep learning and foundational models, recreating the kind of reasoning-driven intelligence that humans have developed may involve coupling it with other approaches. There are question marks over whether the statistical patterns these algorithms learn can ever equate to true understanding.³⁵ They also require vastly more data and energy to learn than humans, are notoriously bad at transferring knowledge between domains and are incapable of continuous learning. Improving the reasoning capabilities of the models and their extrapolation ability becomes a central challenge.

Future Horizons:

×××

5-yearhorizon

An entirely unpredictable era for AI achievement

At this timescale predictions about AI are inherently unreliable. If scaling laws hold, then Alternative AI approaches are likely to become intellectual backwaters. However, if fundamental barriers to scaling arise, then the most advanced AI systems are likely to be complex architectures incorporating a wide range of modules running very different kinds of algorithms. Incorporating the best of all of these approaches will make it possible to create generally intelligent AI systems that exhibit adaptability, generalisability, common sense and causal reasoning.

10-yearhorizon

Emerging scaling laws determine the future of deep learning

Even at this timescale, the horizons are unclear. The scaling laws that have governed progress in deep learning could begin to peter out, renewing interest in alternative AI approaches, especially given ethical and other concerns over unintended consequences of deploying deep learning models. This could result in a shift to hybrid systems that make use of many different AI techniques to boost explainability, efficiency and flexibility. Alternatively, further scaling of model size could lead to the spontaneous emergence of these missing ingredients in deep learning models, causing interest in alternative approaches to wane.

25-yearhorizon

Deep learning continues to dominate

The continued success of deep learning overshadows alternative approaches and diverts resources away from them. Nonetheless, theoretical advances continue and progress is made on building the software tools required to scale these approaches up. Advances in cognitive science provide more clues about the missing ingredients behind general intelligence.

Alternative approaches could solve some of these issues. Neurosymbolic AI, for instance, combines statistical learning with structured representations of prior knowledge, often inspired by cognitive science. It appears more capable of common-sense reasoning, and decisions are more readily interpretable by humans.³⁶ Google DeepMind recently used the approach to solve complex geometry problems at the level of human experts.³⁷ While deep learning focuses on finding correlations in data, Causal AI seeks to give machines a deeper understanding of cause and effect that could be crucial for general intelligence.³⁸ And Bayesian AI approaches use the statistics of probability to help AI better understand the uncertainty of the real world while achieving continuous learning.³⁹

So far, these approaches have not achieved the same success as deep learning. Many researchers also believe that similar capabilities will emerge spontaneously as deep-learning models become larger and more sophisticated. But our theoretical understanding of deep learning lags far behind the practice, making it difficult to spot potential roadblocks to progress. It is possible that some combination of these approaches may be necessary if we want to recreate the kind of general intelligence seen in humans.

Alternative AI - 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.