Microsoft chief announced that it would integrate artificial intelligence into its Bing search engine and Edge browser. Microsoft would also integrate ChatGPT into Microsoft products setting off an AI battle.
The AI development in the last year has been very interesting and I have curiously followed the developments.
I want to write about AI as philosophy and technology in two parts. This essay covers the philosophical part. In the second essay, I will talk about the AI industry today and the trends I believe will shape our future.
I believe philosophy is the key that unlocks artificial intelligence.
Table of Contents
Philosophy and AI
Let me take you for a walk in the Darwin era. My simplest interpretation of Darwin’s philosophy on evolution is that both natural and forced (artificial) selection result in evolution. Based on the conditions (landscape, food, or weather) evolution may differ and produce different outcomes which may alter the course of the evolution. The base premise of evolution as presented by Darwin was that species adapt to the conditions or environment they live in till the extent they can manage the conditions of that environment. Darwin was a naturalist who believed in natural selection as the means to survive and grow. He was a believer that natural selection resulted in genes that ousted the forced selection in the race for survival. With this base theory in mind, the question that arises is, “Is AI a natural selection or a forced selection?” I want to explore this question and understand the right fit for Artificial Intelligence.
We are today teaching robots to evolve autonomously so that they can adapt to changes around them and possibly live alone (natural selection). However, there is a human “oversight” (artificial selection) that controls the evolution of robots. This oversight in my opinion does two things – Control AI-based robots to live and grow in a controlled environment and also fast track the growth based on advancements in technology. Where natural or biological evolution takes place over millions of years, Artificial evolution can happen in hours or minutes (with the exponential development in computing power and technology).
There is another strong argument that separates natural evolution from artificial evolution. The distinction of a cognitive understanding. Let’s take an example – the central processing unit (CPU) of a computer doesn’t really know what addition or subtraction is, it knows the command to perform an addition and does it extremely well with consistency. Computers chess program does not understand that its king is in jeopardy but it kind of understands this in terms of zeros and ones.
Where Darwin tells us the importance of cognitive intelligence, robots tell us the importance of an “operator” who kind of understands the input and output but not the real cognitive importance of human perception and evolutionary process.
How do you build an AI operator based on Darwin’s principles?
The concept of evolution states that if organisms’ mutations allow them to survive and reproduce, that mutation is then passed along and served. If it does not, then mutation dies with its organism. In the algorithm world, it is called neuroevolution. EvoJAX is a Google algorithm that uses Neuroevolution to solve problems. Google published papers on neuroevolution principles of image recognition.
Mixing technology with human principles can be confusing the already complex research. Think of algorithms as Horses. Horses learn throughout their lifetime but are only evaluated on a few traits and metrics, like how fast can they run. But during its lifecycle, a horse grows its muscles, intellect, etc. This is the complexity that mirrors the growth of an algorithm. We may judge an algorithm based on its output of image recognition as good or bad.
An ideal AI operator would allow you to do anything to everything that has ever been imagined by the human mind. However, building such a machine would be impossible not just because of cost but because of the current computing power and availability of technology.
The rate of change in AI is directly proportional to optimization power, i.e. the amount of design effort that is being put in to increase systems intelligence, and inversely proportional to responsiveness to a given amount of optimization.
However, we can mathematically get closer using a perfect bayesian agent. A perfect Bayesian agent is one that makes probabilistically optimal use of available information. Two schools of thoughts on building AI
1. Input — Process — Output: Logical reasoning or self-learning Algorithms
Self-learning algorithm to create AI is not far from the principles of evolution. A process from baby to child to adult. In this process, we have built our knowledge and the process to gain and store it. A process built on pattern recognition and biological evolution. Biological evolution is not practical for a machine but pattern recognition is. Pattern recognition in a machine leads to logical reasoning which is used for self-training the models.
The algorithm follows instructions, it needs to be told step by step on how to bake a cake. Whereas, self-learning algorithms perform a learning process by interacting with the environment for a limited time to solve the problem. It’s like life, which works on self-learning algorithms where we learn from our mistakes, improve bad behaviors, and adapt to our surroundings.
In short, we are building functions that take inputs from the external environment and produce behavior (actions) on the basis of these percepts. This is a time taking process.
2. Top-down approach
But why start from scratch when we already have a ton of data? A second school of thought, the philosophical school of thought on AI believes in a top-down approach. Meaning designing and implementing algorithms that capture cognition rather than building and training on cognition. The challenge to the top-down thesis in AI is the limitation we have with computation power. This can be understood at the level of Turing’s machine. the first machine capable of hypercomputations of trials and errors.
There is even a saying that a function (f) is only computable via a machine if and only if it can be computed by a Turing machine.
Going back to Darwin and assuming you were able to construct a true AI using the above two methods or any other way, what then?. Darwin’s principle of cause and effect. One change in nature drives change in the environment and species. A similar change in machines leads to AI and ML-trained robots like ChatGPT.
But remember, the decision ChatGPT makes is reason based rational decision, which is what humans aspire to as integral decisions. Emotional bias is under check. I am sure we are aware of it, it’s about what are you looking for from that decision. Is it advice on how to handle your kid or where my newspaper is for the day?
In the next essay, I will bring the non-philosophical context and talk about the AI industry today.