Current generative AI research is heavily reliant on brute-force scaling with ever-larger datasets and computing power, raising questions about sustainability and future data limitations. There is a call for architectural innovation enabling AI to learn efficiently from modest data,⁴ drawing from concepts in human learning and cognitive science.

Approaches to cope with data limitations include generating synthetic data, employing new data diversification algorithms⁵,⁶ and developing smarter, data-efficient learning paradigms. However, progress in generative AI must overcome issues with robustness and generalisation by moving from “vanilla” data synthesis to creating data that does not yet exist online, and by exploring alternative architectures to transformers. The field must also avoid over-fixation on current trends (brute-force scaling) and foster unconventional collaborations, especially between academia and industry, to drive fundamental change.

It is important to emphasise that ethical considerations, safety and alignment of generative models with societal values should be foundational to further adoption.⁷ It is also important to keep technical progress aligned with societal needs and to raise critical questions about long-term future directions of the field.⁸

Internal world models, which allow an agent to predict environmental states given actions, are central to intelligence, supporting zero-shot generalisation and planning.⁹ Embodied intelligence, which is intelligence that is rooted in the connection between perception, action and reality, can be likened to evolutionary and child-development processes: human children’s learning, viewed as “scientists in the crib”, provides a model for active exploration and self-directed experiment in intelligent systems. However, current text-based models lack the necessary grounding and are orders of magnitude less data-efficient than humans. Real-world, multimodal data is critical for building future models, and robotics offers a promising path: here, the integration of language, vision, memory and manipulation capabilities supports hierarchical planning and lifelong learning.¹⁰

In the long term, AI robots may develop complex self-awareness and seamlessly acquire new skills from rich sensory experience, moving closer to human-like capabilities. Safety boundaries must be set, though. For example, reinforcement- learning strategies that incentivise pure survival should be avoided to ensure that AI acts as a tool aligned with human interests. Some researchers feel that highly abstract reasoning and planning must always be reducible to linguistic or symbolic representations in order to maintain transparency, but there is debate about this. There is also debate about whether LLM-based approaches can generalise to richly structured, non-textual reality such as images and 3D worlds. Overall, embodied learning will be important for credible AI advancement.

All science involves abstracting data into representations that make future data predictable. AI research, especially in world modelling, can be helpful to science if it focuses on learning these abstractions rather than reconstructing every detail. AI, as a “digital microscope”, provides novel, data-driven approaches to scientific problems, differing from classical rationalism and reductionism. Scientific discovery is currently constrained by the pace and cost of data generation, and AI can accelerate progress via automated experiment design and high-throughput simulation.¹² AI will also democratise tools for scientific discovery, allowing for new digital experiments, hypothesis generation and global-scale simulations such as digital organisms and weather models.

AI may itself become a target of scientific study in terms of new mathematical frameworks and methods such as “vector-based probability”. There are open challenges still: one is achieving integration across abstraction levels — from language to vision, for instance, and from virtual to physical domains. Nevertheless, there is optimism that AI-driven approaches will create new scientific paradigms and domains, though modesty is warranted, as fundamental principles of intelligence and discovery remain undiscovered.

LLMs and correlated AI models, for example, excel at fluency and generalisation, but are prone to spurious correlations, lack causal and counterfactual reasoning, and can be fooled by perceptual illusions. Robust intelligence will require new representations and architectures, going beyond scaling and focusing on joint perception-action models, internal consistency and intervention-centric learning.¹³,¹⁴,¹⁵ If research is to progress in helpful ways, there is a need to rethink education,¹⁶ both in the content of statistics and algorithms, and in interdisciplinary skills spanning philosophy, ethics and computational sciences.¹⁷ Ethical, safety and alignment considerations must also be embedded at the foundational level, especially regarding long-term impacts on society, democracy and human flourishing.

The future of AI is as much about discovering new principles as it is about continuous improvement and scaling of current architectures. Researchers in this field must remain humble and open to unexpected discoveries. Furthermore, AI’s trajectory should be guided not only by intellectual elegance but also by societal needs, human values and democracy.