From Intelligence to Superintelligence: Navigating Change and Shaping the Future with Causal AI
Professor Javen Qinfeng Shi
The rapid evolution of artificial intelligence (AI) is reshaping our understanding of intelligence and expanding the boundaries toward superintelligence. At the heart of this discussion is causal AI, a powerful tool for understanding and navigating the complexities of our changing world.
By uncovering root causes, identifying latent variables, and resisting spurious correlations, causal AI supports generalisation across diverse domains and predicts the outcomes of interventions. More crucially, it addresses a pivotal question: What is the optimal sequence of interventions needed to achieve desired outcomes under real-world constraints? Drawing on past successes and future projections, this article illustrates AI’s societal impact and offers insights into how artificial general intelligence (AGI) and superintelligence could shape the future of work and life.
Creativity as the spark of intelligence
Intelligence extends beyond mere knowledge; it encompasses creativity — a crucial element often absent from traditional data storage systems like books or hard drives, that cannot be considered intelligent regardless of the volume of information they contain. Surprisingly, even simple neural networks, like Multi-Layer Perceptrons (MLPs), exhibit a form of creativity. Through their vector embeddings, MLPs can accurately predict unseen instances, sometimes surpassing the limitations imposed by their training data.
For instance, an MLP trained on handwritten digits data with 50% incorrect labels can achieve a mere 5% testing error, effectively outsmarting its “teacher” by a significant margin. Without supervision, model-based reinforcement learning (RL) agents demonstrate creativity by constructing a “world model” and developing effective strategies or policies to achieve specific goals. Large language models (LLMs), grounded in transformer architecture and reinforcement learning, showcase remarkable creativity, occasionally producing results that diverge from human expectations, a phenomenon sometimes referred to as “hallucination.”
Narrow AI and the race to AGI
Despite the creativity and impressive performance of deep neural networks, RL, and LLMs, they remain confined within the realm of narrow AI, which requires developing or training a separate model for each specific task. What the world truly desires is AGI: a single model capable of effectively addressing all tasks.
Are LLMs the only path to AGI? No! However, technology giants like Microsoft, Meta, and Tesla/xAI are aggressively investing billions of dollars in LLMs and graphics processing units (GPUs), viewing these models as the most promising and expedient route to AGI, even if it may cost billions or trillions to build. Aspirations and attention are now shifting toward creating superintelligence (SI), or AI systems that could exceed the collective intelligence of humanity. Recently, Safe Superintelligence (SSI), co-founded by Canadian computer scientist Ilya Sutskever, raised $1 billion with only 10 employees, underscoring the urgency and immense potential of this field.
Causal AI as the way of change
The essence of change is causality, manifesting through the dynamics of cause and effect. Causality provides three core powers that significantly enhance understanding and decision-making.
First, it offers a deeper truth that goes beyond mere correlation. Causality identifies root causes, discovers latent variables, builds immunity against spurious correlations, and improves generalisation, or the ability of AI systems to apply or extrapolate their knowledge to new data which might differ from the original training data, across diverse domains and distribution shifts.
Secondly, causality empowers proactive change by modeling the consequences of interventions and answering ‘what if’ counterfactual questions. More importantly, causal AI holds the key to determining the ideal sequence of interventions, given available resources or budgets, to optimise future outcomes.
Lastly, causality provides a holistic view by integrating diverse data, tasks, and domains at a fundamental causal level, rather than at a superficial level of effects. This unified model aids in advancing toward AGI and SI.
The interplay between compression and causality: benefits for AGI
Compression is widely acknowledged as fundamental to the success of LLMs, as it distills and refines knowledge, enhancing both creativity and intelligence. However, it is less commonly understood that optimal compression inevitably captures causality. The most effective compression, when applied across diverse data and tasks with sufficient variability, naturally leverages the mechanisms of change. If these mechanisms are not incorporated, it suggests that further compression could still be achieved, indicating that the original compression was not truly optimal. While the exact methods for achieving the best compression remain unclear and may vary, one certainty is that any future AGI or SI that achieves optimal compression will necessarily incorporate causality as a fundamental component.
AI’s impact on society and the future of work
As AI, robotics, and automation continue to advance, many are concerned that machines will eventually replace human jobs. However, this concern is not solely about the potential loss of jobs; it is rooted in the fear of losing income and, with it, the stability and security that come from work.
Historically, human labour has been indispensable to production, with income tightly linked to one’s role in generating value. Animals and machines have enhanced our capabilities, but they could never fully replace us. This necessity gave rise to economic structures that incentivised human productivity and secured livelihoods.
In a world where AI and robotics can handle much of production, we may find that this traditional bond between work and income no longer serves the same purpose. In this envisioned future, the focus shifts from individuals working out of necessity to a system where wealth generated by AI and robotics could be distributed to all. If we were to reach a stage where everyone’s job was replaced simultaneously, the challenge would not be in providing income but in designing an equitable way to share these resources.
However, this transition period is crucial. As automation gradually transforms industries, many people may face temporary or permanent job loss, creating economic and social challenges. A measured, collective response is needed to ensure that individuals are supported during this period of change. The transitional phase calls for policies that bridge the shift from work-based income to a new economic model that values human life beyond traditional labour.
The longer-term picture, however, is hopeful. By moving beyond the need for humans as tools of production, we open a future where everyone can pursue fulfillment without the constraints of financial survival. Freed from routine labour, people could focus on personal growth, creativity, relationships, and contributions that enrich society in ways that AI cannot. Imagine a world where everyone has the freedom to explore their true passions, deepen their understanding of life, and find purpose in their unique talents.
Causal AI can play a crucial role in navigating this transition responsibly. By analysing and anticipating the effects of each decision, causal AI can help policymakers and technologists design systems that support society as it adjusts to these profound changes. And by cultivating values such as kindness, wisdom, and compassion, we can ensure that the future we build is inclusive and humane. Together, we can guide AI’s power to benefit everyone, fostering a world where humans are free to pursue deeper meaning and live in harmony with one another and the planet.
Professor Javen Shi is the Director for Causal AI Group at AIML.
