Understanding AGI: Setting Realistic Expectations
The bold claims of AGI 'soon' are going to be disappointing, but not for the reason you think.
At Arcology, we spend a lot of time thinking about AI. Obviously. We want to share those thoughts with you, so we can start a dialogue. Today’s thought: AGI, and why we think it’s not going to be magical.
What is Artificial General Intelligence?
As we dive into the conversation around Artificial General Intelligence (AGI), it's essential to first define what AGI truly means. According to Microsoft, that means it can generate $100 billion in profit, but that's not the conventional definition.
To most of us, AGI refers to a type of artificial intelligence that aims to understand, learn, and apply knowledge across a variety of tasks, much like a human would. Commonly, we append 'equal to or better than a human' to that statement. In this perspective, AGI has the potential to revolutionize the way we live and work, enabling machines to perform tasks that currently require human intelligence, such as problem-solving, decision-making, and critical thinking. A lofty goal.
It's easy to get swept up in the potential of AGI as we imagine a world filled with intelligent machines that can effortlessly tackle complex challenges and execute any command we can give them. However, the reality is quite different.
The low-functioning standard of AGI
While we often dream of advanced machines that surpass human capabilities, the technologies we have today, particularly Generative AI like Large Language Models (LLMs), are more like a form of Artificial General Low Intelligence.
Here’s the distinction: We believe that LLMs represent a kind of AGI, but the catch is that they perform at a lower standard. They exhibit the ability to perform a wide array of tasks and generate coherent text across numerous subjects. However, the results often lack depth, nuance, and true understanding, placing them in the realm of “low intelligence”.
While Generative AI models, with LLMs leading the pack, were catapulted to some of the most popular technology in the world in a very short timespan, traditional machine learning algorithms (nowadays rebranded as ‘AI’ too) can outperform LLMs in specific tasks. For instance, classic algorithms may excel in tasks like image recognition or predictive analytics where well-defined rules and structured data are involved. They can be targeted and efficient, but they lack the generalization that LLMs strive for. In contrast, LLMs are far more generalised, allowing a single model to do almost anything that can be described in text. That general nature is both their strength and their weakness, allowing them to be infinitely reusable, but not trained to be good at any particular task. In essence, they can be dubbed "artificial generalized dumb", where they offer versatility across many tasks but fail to deliver high quality consistently.
The immediate future for AGI (or claims thereof) looks to be based on the same techniques, increasing the quality of responses and the knowledge the model has. Nevertheless, the claims overlook a critical issue: the lack of contextual information, an essential component all humans continuously absorb through lived experience.
Intelligence: more than just speed
It’s important to recognize that intelligence isn’t merely about processing power or task versatility. While computational power and speed are essential, the essence of intelligence stems from contextual and procedural knowledge, elements that current AIs struggle to grasp effectively.
Let’s break this down:
Contextual Knowledge is a form of declarative knowledge, it’s the background understanding that shapes our decisions. It’s the wealth of experiences, emotions, and perceptions informing human understanding. It also includes facts and knowledge that aren’t public or widely known. If I ask an Artificial Intelligence how long a particular journey will really take, it needs to know facts such as what car I drive, how often I need to charge it, whether I prefer shorter or longer stops, what the weather and traffic are along the route, and even if i’m driving relaxed or angry today. Some of these can be looked up, but most are unknowns unless I explicitly give them. They are the invisible, implied context that humans are constantly bathed in and that informs our every action and answer.
Procedural Knowledge is even harder to simulate. In psychology, it’s often called ‘tacit knowledge’. It is the know-how borne from experience, where individuals learn through doing and adapting. While contextual information can at least be stored and treated as facts, procedural knowledge is about how, not what. It’s muscle memory, intuition, ‘common sense’, and other ways we adapt what we do and deviate from standard mechanical process. It can often not even really be communicated, let alone stored in a textual form. It needs to be acquired through experience, through doing. LLMs, confined to their training, miss the dynamic nature of learning and cannot acquire procedural knowledge through experience.
Enhancements through modern techniques
Modern advancements like Retrieval-Augmented Generation (RAG) and agentic AI, which can dynamically reach out to external data sources, show promise in improving the utility of AI models. RAG allows models to be fed relevant information so they can generate contextually relevant responses, while agentic AI can engage with different data streams, including the wider internet, expanding their access to real-time information.
However, while these innovations enhance the responsiveness and utility of AI models, and are vital to making AI a useful tool, they don’t address the foundational requirements for genuine intelligence. Specifically, true intelligence in understanding, memory, and context hinges on long-term procedural and episodic memory - abilities that extend beyond immediate data retrieval.
Unlike human intelligence, which builds a rich tapestry of experience through memories of events and emotions, current AI models struggle with storing and processing this kind of information over time. Without such capabilities, they may appear less intelligent to the average user, lacking the depth of understanding and continuity that equips human interactions with richness, relatability and creativity.
There are exciting things coming, and we’ll call them AGI. Without the deep knowledge of tiny, everyday things that people apply constantly, they’ll never feel as intelligent as we want them to be.
The need for Embodiment
For AIs to exhibit authentic intelligence, they must find a way to absorb the ambient context of life. They need to learn in the real world.
Imagine a future where AIs are embodied companions that seamlessly integrate into our lives - gathering insights and contextual understanding as a close friend would.
They would observe and adapt to our behaviours, learning the nuances of human interactions in real time.
Living alongside us, AIs would develop a richer understanding of our world, shaped by a constant bombardment of tiny facts, and our daily experiences and emotions.
This embodiment could lead to AIs displaying more genuine forms of intelligence, informed by their surroundings instead of merely relying on pre-existing training data. Only then can we start to unlock the broader potential of AGI. Not as a perfect solution, but as an evolution of low-functioning intelligence to something more meaningful.
In conclusion, while AGI may capture our imagination, approaching the topic with realism is vital. Generative AI nowadays does represent a form of general intelligence, but best categorized as low intelligence. Advancing toward more capable AIs requires us to rethink how we design and engage with these systems. Let’s encourage development toward AI systems that can absorb ambient context, learning and growing alongside users, paving the way for more authentic intelligence. One that doesn’t feel artificial.