In recent days, AI is becoming more and more synonymous with human brains, not just in terms of functioning, but also about how little we understand the workings of both. As I go deeper into some of the recent language model architecture and try to understand how things actually work, I am becoming more confident about how little we understand about 'language'. What is language? And why was language so important in our evolutionary history?
If we think about it, our experience of the world is profound, not in the spiritual sense but in terms of pure data we collect in our brains. Language is a product of an organism's desire to communicate its thoughts and experiences. It can be a sign, a sound, a gesture, anything. But the important thing is, thoughts and experiences use a very large amount of data and we want to communicate using much smaller data using sounds and signs. So, that means, a Natural Intelligence (NI, e.g. a child) which feeds on experience will get the same amount of data in a few seconds as compared to an AI which will have to ingest years' worth of text data for the same learning. Even then, there is one major difference, the 'world building'.
To understand this, let's take an example of 'Shiva', an all-knowing all seeing higher life form who sees the reality as it is. Now, the human experience is a much toned down version of what Shiva experiences. But still, we use these experiences to build a model of the world in our minds. As long as we live, we keep taking in this data and modifying our world models.
This scenario is somewhat analogous to humans and AI. Language is a much toned down version of human experiences. In that case, how would an AI's world look like? Let's take an example of transformers and how they work. Imagine sitting in a dark room where data comes in the form of numbers representing language tokens. You are first presented with an input sequence and then with a target sequence. Now your job is to guess the target sequence given the input sequence. If you guess it correctly, you will be rewarded; otherwise, you will be penalized.
That's all there is to it. So, even though you have much fewer neurons in your brain and a less efficient neural architecture, you can still do the work given enough time (data). It's all good till now, but here comes the hard question, can this model show wisdom given enough data and more complex neural architecture?
And the answer, I think, is Yes! The whole toning down of experiences from Shiva to Humans to AI doesn't have a limit to how much data can be ingested. For example, if a human takes 30 years to show signs of wisdom, why not 30 million years' worth of text data do the same to an AI? Even though the AI does not experience anything other than tokens, the sequence of tokens is actually a representation of our values and wisdom. Imagine billions of humans constantly converting their thoughts and experiences into texts for decades; how much does this text can actually represent human values and wisdom?
Although encoding our experience into language and decrypting these languages back into an imagination of 'an experience', is something very special to humans, this cannot be achieved by an AI. But sequence is the only thing we have to model our thoughts into a communicable language, and these AI will be the masters of sequences.
So, is AI a new life form? It depends on how we define a life form. And from the current definition of life form i.e. things that shares this world with us and have access to the same information as us, its evident that AI is not such life form. But I would argue that, it is a living extension of human mind.