“Universal AI” — Reconciling the Debate between Narrow AI and Artificial General Intelligence

By Jacob Renn, Ph.D. (Chief Technologist at AI Squared) — April 3, 2024

Introduction

In my capacity as Chief Technologist and head of research and data science at AI Squared, I am constantly immersed in the latest technological advancements. This involves not only the evolution of our products and technologies but also a broader contemplation of the trajectory of technology itself. With this post, I aim to shed light on some recent developments within the expansive realm of artificial intelligence (AI) and machine learning (ML). Specifically, I will delve into the transformative impact of generative AI models like ChatGPT and GPT-4, which have surged onto the technological scene. Furthermore, I will propose a theory aimed at reconciling the ongoing debate between narrow AI (NAI) and artificial general intelligence (AGI), offering what I call a “universal AI” (UAI) framework.

Terminology

Before diving into my thoughts on the subject, it’s important that I outline some of the terminology I’ll be using, as personal definitions can vary. Below are some of the key terms which will be utilized in this work:

1. Artificial intelligence (AI): Artificial intelligence is a cross-domain field encompassing aspects of computer science, mathematics, biology, psychology, and other fields. It seeks to build computer systems that perform tasks which otherwise require human intelligence.
2. Machine learning (ML): Machine learning is a subset of AI that focuses on the specific algorithms and methods to teach computational systems to perform tasks using data.
3. Narrow AI (NAI): Narrow AI systems are AI systems designed for specific (or narrow) tasks, but do not generalize their knowledge across a set of domains. Examples include image recognition systems, recommendation algorithms, and even autonomous vehicles. NAI is sometimes also referred to as “weak” AI.
4. Artificial general intelligence (AGI): Artificial general intelligence is often considered the “holy grail” of AI. In contrast to NAI, an AGI system would meet or exceed not just human-level intelligence, but also human-level abilities to self-teach and apply its knowledge of existing tasks across other tasks, possibly leading it to perform any intellectual task a human could perform.
5. Multimodal AI: Multimodal AI refers to AI systems that can process and understand data from multiple modalities, such as text, audio, visual, and other input methods.
6. Large language models (LLMs): Large language models are AI models that typically employ deep learning techniques and are trained on vast amounts of text data to understand and generate human-like language.

A Shift in Thinking: Universal AI

Examining the terminology defined above, one might wonder how long it will take to develop AGI, and the most skeptical reader may wonder if AGI will ever be accomplished. Others, perhaps rightfully so, would also ask whether AGI should ever be developed. These are extremely poignant and powerful questions with equally important answers and consequences. Despite these questions, I have come to believe that by focusing on NAI vs. AGI, we have overlooked a third categorization of AI systems, which I’m calling “universal AI” (UAI).

Ultimately, AI systems are designed to be tools. Rather than focusing on theories of intelligence and how an AI system does what it does (which is extremely important — this is not something that I am in any way indicating that we abandon), we focus on what it does and what it can be used for. For example, let’s consider ChatGPT and other LLMs versus a recommender system that suggests movies based on what someone has previously watched and rated. In terms of utility, ChatGPT far outpaces the recommender system in terms of the pure number of problems one can apply it to. With the correct input, ChatGPT may even be able to recommend movies (though perhaps not as well as the purpose-built recommendation engine). Not only that, but it can also help write wide ranges of documents, summarize text, translate text from one language to another, explain complex topics for people of different age groups, and query other data sources such as search engines and databases in search of answers to user-provided questions.

Despite the utility of LLMs across a variety of tasks, even the most advanced models we have today are still considered NAI systems, since they lack the ability to truly generalize their intelligent capabilities to human levels across modalities and tasks. But even without true intelligence, there is almost no debate that these systems have signified a paradigm shift and reach beyond what other traditional NAI systems do. It is for this reason that I propose thinking about these systems not from a purely intelligence-theory perspective, but from a utilitarian and pragmatic perspective. So, what is UAI, then? My proposed definition is:

UAI is AI which, when posed with any task consisting of direct input and any relevant contextual input, provides a response that either directly solves or improves the solvability of the requested task.

Let’s break this definition down a bit. First, let’s look at the inputs to the AI system as defined: direct input and any relevant contextual input. Think of these as the combination of relevant information that’s needed to approach a problem. Some questions (such as “What is the first-order derivative of f(x) = x2?”) can be solved without any context, but others (such as “What are the highest-rated restaurants within walking distance to me?”) need additional context and cannot be answered without it.

A second, critical part of this definition is the notion of “any task.” In order to truly be universally intelligent, an AI system would have to be able to truly help with any task. This includes purely intellectual tasks such as performing mathematical calculations or designing scientific experiments, but it also includes multimodal tasks such as self-driving cars, translating spoken language, and other similar types of tasks.

Finally, the last part of the definition is that the UAI system “directly solves or improves the solvability of the requested task.” This part of the definition highlights the utility of the concept of UAI. Not all tasks are directly solvable given the current state of information, but a UAI system should be able to understand a given task and applicable context to a degree that improves the solvability of the task.

Let’s think about the concept of UAI from a broader perspective. If a UAI system were to exist, then any time that system is connected into another system in a meaningful way, the connected system would see improvements in its capabilities driven by the UAI system. This is where the utilitarian perspective of UAI comes in, as it focuses on the ability of the AI system to help solve problems instead of how it does so.

Current Systems and the Concept of NUAI

Just as there is skepticism as to whether AGI will ever be created, I expect there to be healthy debate over whether UAI will ever be created. Indeed, it may be possible that a UAI system may exist in the future without ever achieving AGI (or vice versa). Maybe UAI and AGI are inextricably linked, and we cannot have one without the other. While there are still many questions that necessarily surround UAI, the concept of UAI as opposed to AGI introduces a much more measurable scale in which we can test AI systems and the advancements they’re making.

The advancement from NAI to AGI is seen as a monumental leap — a technical chasm of sorts — where a proverbial switch is pulled, and an AI system gains a level of consciousness never before seen. In essence, an AI system is either an AGI or it is not. With the concept of UAI, however, the utility of a system for various tasks is a measuring stick with which one can measure the strength of an AI system. This is extremely beneficial when looking at today’s state-of-the-art AI systems compared to more traditional AI systems. Take, for example, the LLM versus recommender system comparison outlined earlier: The LLM is clearly closer to a UAI than the recommender system.

Some LLMs and AI systems are already so powerful that they may be considered a near universal AI (or NUAI). While these current systems cannot truly help at all tasks (I wouldn’t recommend pulling up ChatGPT and asking it to take the wheel for you while you’re driving, for example), it’s astonishing just how powerful these current systems are at performing nearly any natural language task they’re asked to do. While just how near these current systems are to true UAI, even using just one modality, is a question which could be debated endlessly, the fact remains that these systems are the nearest we have ever been to UAI.

What’s Next?

As NAI systems grow stronger and more powerful, there are a few trends I believe we’re going to see emerge. In particular, I am hopeful of the following:

1. Future NUAI systems will have greater abilities to reason about the world, the environment they operate in, and the context they operate in.
2. NUAI systems will grow to become truly multimodal systems, able to understand textual, photographic, videographic, and auditory data (and possibly more).
3. Future NUAI systems will grow to be more useful across even wider arrays of tasks, and they will necessarily be more truth-seeking and fact-based than current systems as we get closer to truly UAI systems.

All these shifts make me excited to see the future of AI, but it also makes me excited for the future of AI Squared as well. As an AI integration company, we’re focused on bringing the results of AI and other advanced analytics to end user workflows regardless of the web application that’s being used. Traditionally, integrating the results of custom-built models and analytics can be a lengthy process within enterprises, as it requires either altering existing, possibly legacy, tools and applications or building entirely new applications just for a single model. With AI Squared, however, integrating new models couldn’t be easier, as we circumvent that application development process and instead focus on overlaying results directly within the browser window, allowing integration not into just one business application, but all applications a user visits in the browser.

As AI systems become more universal, so will the need to break them out of individual applications. Instead, future AI systems will be able to truly work within human-machine teams, and AI Squared is ready to empower that new way we will work.