As a passionate explorer of the ever-expanding world of artificial intelligence, Ive pondered if we humans have towards the word "artificial", particularly in "Artificial Intelligence." The term "artificial" is sometimes associated with "fake," as in the case of "artificial leather," this suggests that non-biological forms of intelligence are inherently inauthentic. I believe this creates a bias, in which AI is perceived as "not real intelligence." I think it's time to challenge these biases and appreciate intelligence in a more nuanced manner.
When it comes to evaluating intelligence, whether human or artificial, we're often tempted to use a single metric or number, like an IQ score or SAT result. While this approach works well for humans, machines are not humans, and intelligence is a complex construct that defies easy quantification. Currently, GPT-4 achieves the equivalent of about a 140 IQ on the SATs. However, it lacks on some aspects of human intelligence where [almost] any human would pass (example). Recognizing this, I believe that we need propose a more comprehensive, multi-dimensional approach to assessing AI systems. There are dimension where we don't typically measure humans simply because it's not that relevant (unless they are completely socially awkward or sociopats), but AI is not human, and it will lack in unexpected placed while excelling on others.
What I'm saying is that we need an measurement that can be applicable to human and non-human intelligence, but that it measures for more than what we currently measure.
To illustrate, let's think of intelligence as analogous to a car's horsepower. Horsepower is a common metric used to describe a car's capabilities, but it doesn't capture all the nuances of a car's performance, such as acceleration, braking, or aerodynamics. Similarly, a multi-dimensional approach can offer a more complete picture of AI's abilities, moving beyond the limitations of a single number.
Before reducing AI's capabilities to a single value, let's consider five dimensions of intelligence:
- Language understanding
- Logical reasoning
- Creativity and innovation
- Memory and learning
- Decision-making
We can evaluate each of these dimensions independently, normalize them, and eventually combine them into a single metric using the geometric average. This method includes areas not frequently considered in human intelligence assessments but should be compared when evaluating AI (or AGI) to human values.
The geometric average involves multiplying the values and taking the nth root, where 'n' is the number of values. This method is more sensitive to disparities between the values, providing a more intuitive and meaningful result when the values are significantly different from one another.
For example, consider two dimensions with scores of 250 and 0. The arithmetic average would be (250 + 0) / 2 = 125, while the geometric average would be √(250 * 0) = 0. As shown here, the geometric average offers a more accurate representation in this case. An AI with very high Language and Logical reasoning but no Memory would have a lower value than a standard average.
These are just some ideas, but I see the bias towards "artificial" as a limiting factor towards embracing the wave of AI.
As we progress further into the age of AI, it is crucial to create and establish accurate evaluation methods that consider various aspects of intelligence. By refining our understanding of AI and moving away from the "artificial" label, we can foster a more inclusive and comprehensive view of what constitutes intelligence. This shift in perspective will not only enable us to develop better AI systems but also to integrate them more effectively into our society, and detect unwanted outliners.
Moreover, this approach will encourage researchers and developers to focus on creating AI systems that are well-rounded and beneficial to humanity as a whole. By highlighting the importance of social and emotional alignment, we can ensure that AI systems are designed with empathy, ethics, and human values at their core.
In the long run, embracing a more nuanced understanding of intelligence and adopting a multi-dimensional approach to AI evaluation will help us maximize the potential benefits of AI, while minimizing its risks.
Now, let's address social and emotional alignment/intelligence separately. I've intentionally left it out of our list of dimensions because it deserves its own spotlight. This aspect is what separates good from bad AGI and should be handled carefully. Emphasizing this ensures that AI systems are developed (aligned) with ethical, empathetic, and human-centric considerations in mind.
This preliminary work/article suggests that as AI systems become more prevalent, we should seriously consider establishing a benchmark for AI, which could be renamed to Authentic Intelligence. For instance, in the future, we may say, "This system has an AI of 120, or an Authentic Intelligence [quotient] of 120, with an Emotional Alignment of 90." For the general public, 120 would suffice, but a more careful mind would consider the 90 as well.
In conclusion, by adopting a multi-dimensional approach and using the geometric average, we can overcome the bias associated with the term "artificial" and foster a more accurate understanding of AI's capabilities and limitations. By doing so, we can more safely embrace the remarkable potential that AI holds for our future and develop AI systems that are both powerful and attuned.
No comments:
Post a Comment