Artificial General Intelligence and Large Language Models

AGI vs. Human Intelligence

Surprisingly, there is no unified definition for ‘general intelligence’ or even ‘intelligence’. Historically, the prevailing thought was that only humans possessed the capacity to be classified as intelligent. Thus, animals like dogs, cats, elephants, or even chimpanzees were not considered intelligent. This brings us to the somewhat sad realisation that ‘intelligence’ typically refers exclusively to ‘human intelligence. Why ‘sad’? Because we know that dogs and cats can plan complex hunting strategies, elephants can draw, and chimpanzees can count and learn to understand words (e.g., Koko, who learned up to 1,000 signs). These abilities are driven by the neocortex, a brain organ common to mammals.

There is also a new, more radical view. Postulating that “the very process of development has an intelligence of its own: for example, if you take a tadpole (the precursor to a frog) and manually scramble its facial organs, those facial organs will relocate back to the correct place as the tadpole matures. …. Levin defines “intelligence” as the capacity to achieve the same goal via different means, and over the years, he and others have documented case after case of such adaptability. “[2]

What’s so special about the human brain? The short answer is that the human neocortex is larger (relative to body mass). More detailed answer: it has a greater capacity that allows humans to form language and reasoning. Language facilitates communication and the transfer of knowledge, which has significantly accelerated human evolution over the last 20,000 years. How does the human brain produce language? Most scientists describe it as an “emergent phenomenon.” This phenomenon is aptly described in a recent Quanta Magazine article: “The world is full of such emergent phenomena: large-scale patterns and organisation arising from innumerable interactions between component parts. Yet, there is no agreed scientific theory to explain emergence. Loosely, the behaviour of a complex system might be considered emergent if it can’t be predicted from the properties of the parts alone”

[1]. In simple terms, at some level of brain (neural network) capacity, language can emerge from complex behaviours of subtle processes in the brain – sounds nice. Still, unfortunately, it does not answer the central question: How?

Let’s summarise: Intelligence lacks a clear definition and has historically been considered exclusive to humans. However, animals demonstrate cognitive abilities driven by their neocortex, challenging this view. Additionally, developmental processes, such as tadpoles rearranging facial organs, suggest inherent intelligence. The human brain’s larger neocortex supports language and reasoning, which are key for communication and knowledge transfer. However, the exact mechanisms of how these abilities emerge remain unclear, highlighting gaps in our understanding of intelligence.

Could ChatGPT be Classified as a Junior AGI?

Large Language Models (LLMs) like ChatGPT and Google Gemini are primarily statistical models that rely on mathematical formulas to identify patterns, unlike the human neocortex or neural networks, which learn and adapt from complex patterns. They employ techniques like word2vec to describe the correlation between words, representing them as vectors in high-dimensional space to capture their meanings and relationships. Until recently, no one would have bet that language models would

show any signs of intelligent behaviour. However, starting with GPT-3.5, these models exhibited an “illusion” of human-like conversation capabilities. This sparked a debate about whether to treat such models as primitive AGI. With the latest GPT-4.0, which allows people to pass almost any exam better than the average human, this debate has become even more intense.

To understand the implications of this debate, we can reflect on the ‘Chinese Room’ experiment, the famous thought experiment designed by John Searle. The ‘Chinese Room’ argues that a computer can process symbols to mimic understanding a language without genuinely comprehending it, illustrating the difference between syntactic processing and actual understanding. This concept is crucial in understanding the limitations of current LLMs, which, despite their impressive capabilities, may not truly understand the language they process. Modern LLMs can be considered infant AGIs, sparking vast philosophical and scientific debates and opening immense possibilities for future transformations. Imagine robots like Atlas that speak, look, and walk like humans, working in parking lots, coffee shops, supermarkets, and even as teachers in the near future. This potential future underscores the importance of the ongoing debate about the true nature of intelligence in artificial systems.