The Edge of AI: Exploring the Emergence of AGI through Self-Organised Criticality in Large…

Artificial intelligence (AI) research has rapidly evolved over the last decade, producing a series of increasingly sophisticated large…

Artificial intelligence (AI) research has rapidly evolved over the last decade, producing a series of increasingly sophisticated large language models (LLMs). Yet, we remain at the edge of a monumental leap: the emergence of general artificial intelligence (AGI).

I have been sucked deep down the rabbit hole of Generative AI over the last 6 months and this article brings together all the learning and research I have been pursuing. I am proposing, what I hope is a novel perspective on a roadmap for AGI, investigating the idea that this might occur at a critical point of self-organised criticality (SOC), a concept originating from the fields of mathematics, theoretical physics and complex systems.

The Interplay of Large Language Models, Reinforcement Learning, and Self-Organised Criticality

LLMs, like GPT-4 from OpenAI, have already shown promising capabilities, exhibiting emergent behavior as they scale up. The addition of reinforcement learning (RL) — learning from feedback and adjusting responses accordingly — further enhances their potential. Yet, I believe that the most fascinating conjecture stems from the application of the SOC concept to this domain.

SOC refers to the tendency of large systems to self-organize into a critical state, where a minor disturbance can cause large-scale effects — a phenomenon seen in a myriad of natural systems. Applying this to LLMs, I propose that as they grow and learn through RL, these Generative AI systems might self-organize and reach a critical point of complexity and learning dynamics. This critical point could be where AGI — an AI as capable as a human across the full range of cognitive tasks — might emerge.

Designing a dynamic LLM to evolve and self-organise

I propose that an LLM, combined with RL and an optimally designed reward function and access to previous outputs, could reach a state of SOC. By continually refining its behavior and responses, the model could grow and become more adaptable, displaying more general cognitive capabilities over time.

To achieve this, the model could use self-reflection as its RL reward function. This would need to be carefully tuned so that the model could self evaluate on metrics such as coherence, grammatical correctness and adherence to the input prompt.

Using self-reflection could reduce the amount of feedback data needed as the model is essentially generating its own feedback. To avoid the potential for overfitting, the model could evaluate its outputs against previous outputs and assess the diversity of its responses over time and adjusting its behaviour accordingly. This would involve the model developing a form of meta-learning or self-awareness about its own performance (very human!).

Exploring the Critical Point

Identifying the “critical point” where an LLM could reach a state of SOC is challenging due to the high-dimensional and non-linear nature of these models, but critical to ensure that we have appropriate safe guards in place and are prepared for the advent of AGI.

However, it should be possible to identify the approach of AGI as the LLM will begin to exhibit an increasing number of certain characteristics such as power-law distributions and long-range correlations as it approaches the “critical point”. These are common characteristics in mathematical and theoretical physics systems that exhibit SOC, so should be expected in an LLM that is demonstrating emergent behaviour as it increases in size.

For LLMs specifically, this could be detected in the frequency of word use exhibiting Zipf’s law, which is a type of power-law distribution. Another indicator might be seen in the activation patterns of neurons in the neural network of an LLM exhibiting long-range correlations.

Challenges and Safety Concerns

Safety and ethical considerations are paramount. As AI systems become more powerful, ensuring they align with human values and avoid harmful consequences is vital. This points to the importance of AI safety research, especially when discussing the potential emergence of AGI.

Moreover, AI alignment remains a critical challenge. As AI systems grow in complexity, ensuring their actions align with intended human objectives becomes increasingly difficult.

This becomes even more of an imperative in the design of an LLM that is built to be dynamic and habe the ability to evolve and self-organise through RL.

Conclusion

I believe this hypothesis presents a novel approach to the future design of LLMs and sets out a potential roadmap to reaching AGI through increasing emergent behaviour. I am no AI expert or researcher, but have become fascinated by the potential of LLMs and I believe the idea of self-organised criticality offers a fresh and intriguing perspective.

The road ahead is filled with exciting possibilities, critical challenges, and vital ethical considerations. As we continue to probe the depths of AI, concepts like self-organised criticality can provide valuable new lenses to view and understand the complexity of this journey.

And maybe, just maybe, in understanding this complexity, we might stumble upon the path that finally leads us from the edge of AI, into the realm of AGI.

---

This article was researched and written with help from ChatGPT, but was lovingly reviewed, edited and fine-tuned by a human.