No one born today will ever live in a world where they are smarter than AI. But what does that really mean? Is there a way for human beings to understand how smart AI is? Is IQ a fair test? Is it even reasonable to apply ideas about human intelligence to machines? If IQ tests aren’t right, how will we know when an AI model is smarter than we are?

Furthermore, can we separate the benchmarks of intelligence from benchmarks of practical capabilities? After all, information is not knowledge – knowing how to do something is quite different from the ability to actually do it.

With this in mind, let’s explore some ways to think about machine intelligence in the context of human capabilities.

Exploring the Concept of Rating AI Intelligence

OpenAI (the creators of ChatGPT) have proposed a five-level scale for AI intelligence. It offers a structured approach to categorizing the progression of AI systems, providing a framework for assessing their impact on society. Importantly, this scale is not yet fully defined and OpenAI clearly says it is under development, but here are the aggregated descriptions of the proposed five levels:

• Level 1: Narrow AI (ANI – Artificial Narrow Intelligence) – AI at this level is designed to perform specific tasks with proficiency but lacks the ability to generalize across different domains. It operates within predefined parameters and excels at tasks such as natural language processing, image recognition, and recommendation systems. Examples include virtual assistants like Siri and Alexa, chatbots like ChatGPT, and recommendation algorithms used by Netflix and Amazon.

   

  Level 2: Problem-Solving AI – This level of AI can solve basic problems at a level comparable to a person with a PhD. It can handle more complex tasks within specific domains and exhibits a higher degree of cognitive ability than Narrow AI, although it still operates under domain constraints. These systems would be used in specialized research areas such as medical diagnosis, legal document analysis, and financial modeling.

  Level 3: Autonomous AI Agents – AI agents at this level are capable of taking autonomous actions on behalf of users. They can perform tasks that require a sequence of decisions, adapt to new situations, and make choices without human intervention, but their autonomy is still domain-specific. Achieving Level 3 gets you AVs (autonomous vehicles), autonomous customer service agents, and advanced robotics in manufacturing.

  Level 4: Innovative AI – AI at this stage can create new innovations, demonstrating a high level of creativity and advanced problem-solving capabilities. It can generate new ideas, designs, and solutions, surpassing human capabilities in many creative and technical fields. Level 4 AI could be used for drug discovery, automated research assistants, and to create new artistic works or scientific theories.

  Level 5: General AI (AGI – Artificial General Intelligence) – AGI can perform tasks that entire organizations of people currently handle. This level represents highly autonomous systems that surpass human intelligence in most economically valuable tasks, integrating and generalizing knowledge across various domains. These hypothetical systems could manage entire companies, develop comprehensive strategies across multiple industries, and perform complex interdisciplinary research.

Is There A Level 6?

There is no Level 6 proposed by OpenAI, but their scale stops at AGI. The next level would be ASI (Artificial Superintelligence). No one knows what that is, BTW. But the hypothesis is that it would be a system so smart and capable that humans would not understand how it worked or what it was doing. At the moment, this is science fiction – but that’s only for the moment. ASI may never be achieved, but if generative AI has taught us anything in the past few years, it’s taught us to never say never.

Understanding Human IQ and Its Measurement

Human IQ, or Intelligence Quotient, is a measure designed to assess cognitive abilities relative to others. It was never meant to apply to machines. Traditional IQ tests evaluate various mental faculties, including logical reasoning, mathematical ability, spatial visualization, and language skills. These tests are calibrated to ensure an average score of 100, with scores above 130 indicating exceptional intelligence. Marilyn vos Savant, with an IQ of 228, holds the record for the highest IQ ever recorded.

However, IQ tests and their results can be controversial and are not universally accepted as definitive measures of intelligence. They can be culturally biased and may not capture the full spectrum of human intelligence, such as emotional or creative intelligence. Understanding the limitations and applications of IQ tests is crucial in the broader context of measuring cognitive abilities. That said, since we are exploring ways to hypothetically measure AI against human intelligence—an area with no defined metrics—let’s use IQ for this hypothetical thought experiment.

The Broader Perspective on Measuring Intelligence

Before we do our calculations, it’s important to consider that intelligence is a multifaceted concept that extends beyond traditional IQ. Howard Gardner’s theory of multiple intelligences suggests that intelligence is not a single general ability but a combination of various distinct capacities, such as linguistic, logical-mathematical, musical, and interpersonal intelligences. This broader perspective highlights that cognitive abilities can manifest in diverse ways, and a comprehensive understanding of intelligence must consider these multiple dimensions. In the context of AI, this means that evaluating AI capabilities should go beyond simple metrics and consider the range of tasks and contexts in which AI can operate effectively. Currently, there are no standardized measurement methods for comparing human and AI intelligence. So, let’s invent one.

Hypothetical Correlation Between AI Parameters and IQ

GPT-3.5 is said to have been trained on about 175 billion parameters, OpenAI considers the number of parameters used to train GPT-4 a trade secret. Informed speculation suggests that GPT-4 could be trained on as many as 1 trillion parameters. In our scenario let’s hypothesize a direct relationship between the number of training parameters in AI models and their IQ scores.

Assuming GPT-4, with an estimated 1 trillion parameters, has an IQ of 155 (a speculative figure suggested by some enthusiasts), what would happen if we double or quadruple the number of parameters?

We’ll use two different approaches to calculate the potential hypothetical IQ of subsequent AI models: linear scaling (overly simplistic and assumes a direct relationship between parameters and IQ), logarithmic scaling (more realistic, as it highlights that merely doubling parameters doesn’t yield proportional gains). I won’t bore you with the math, but these two approaches will provide a directional understanding.

If GPT-Next is trained on 2 trillion parameters:
– Linear Scaling IQ: 310
– Logarithmic Scaling IQ: 218

If GPT-Next is trained on 4 trillion parameters:
– Linear Scaling IQ: 620
– Logarithmic Scaling IQ: 310

In our hypothetical, just doubling the number of parameters used to train GPT-Next will create an AI with an IQ near the top of tested human IQ (218) or that is beyond human comprehension (310). Using these hypothetical calculations, doubling the training parameters to 4 trillion creates an AI that would likely operate outside of human understanding which brings us to the science fiction-based concept of ASI.

The Implications of ASI

As AI models approach these hypothetical high IQ levels, their cognitive abilities would surpass human intelligence to an extent that their thought processes and decision-making would become opaque and unfathomable. We might still understand what an AI with an IQ of 310 was thinking, but an ASI with an IQ approaching 620 would operate on a level beyond human comprehension, posing significant challenges in terms of control, alignment, and ethical considerations. If it ever happens, the development of ASI will necessitate robust safety measures and governance frameworks to ensure that its actions align with human values and interests.
 

We Need A Better Way To Think About This

While the notion of assigning IQ scores to AI is purely hypothetical (and almost definitely wrong), it underscores the potential trajectory of AI development and the transformative impact it could have on society. Maintaining a focus on ethical and responsible AI development is more important than it has ever been. And, being honest about the limitations and implications of measuring AI capabilities will help us understand how to share the world with an intelligence so fundamentally different from our own – we don’t yet know how to define or measure it.

More By This Author: