Immortality? Are we ever going to die?
Recently I have learned more about AI and LLMs (from my new favourite teacher Andrej Karpathy). I learned about how they work and what they can bring to our society. In these process of learning this new technology, I realised that we might never die or more accurately we will die physically but our minds, our thougths, our characteristics will stay in this world forever.
These LLMs are trained based on this process (from my own understanding, might be wrong as I'm new to this, you can also find the steps here): First they collect data from the internet, and really THE WHOLE INTERNET, and convert them into one single zip file, this is zip file is a neural network model with all the paramaters inside or the "weights" (this is crazy to me, as how large the internet is and they can do this, just wow) this step is called pre-tranning step, this step allows us to have something that can generate the "next" words very good. For example, they can guess the next word of this sentence: Knock! Knock" ..... , the model will likely (this is all base on probability or the parameters inside them) produce these words "Who's there?". But so how does it useful? Well it is not just yet and thus we need the second step.
Second, from the base model, we can create conversations to train these models, and these conversations are actually human created, this step is called fine-tunning step or post-trainning. The models will learn these conversations and be able to talk just like how these converations are. For example, I have a conversation with my friend like this:
me: hi how are you?
friend: good, what about you?
me: good, how is your bitcoin going? (I just thought of this as I write this :))
friend (who is an expert in cryptocurrency): yeah bits is going ok (brah brah, and then he goes on and mention something that only expert know and I don't)
The model will learn from this conversation (not a very good example, but hopefully you get the idea), and act/talk as my friend (an expert in cryptocurrency), and so when I ask this model a question about cryptocurrency they will generate answer just like him.
But the models will only be as good as my friend (the expert) and can't beat him since they were trained in his data. Therefore, to make them smarter, we introduce the next step. Third, reinforcement learning, in this step the model will go and do their own thinking, think of it like this, your math teacher shows you how to solve a math problem, but in order to be good at math you also need to go and do the practice problems in your textbook and learn how to solve math questions for yourself, and the concept is the same for these models, they go learn and improve by themselves. (currently as I'm writing this post, people are trying to improve this step even further).
That is the whole process of these LLMs, and so why do I think we will live forever? Well because think of our lives as dataset, for step one, creating a base model (that big zip file), we can use datas from our parents, our grandparents, great-grandfather to create this zip file your own (although it is right now expensive to do that for individual, but here it is the idea). For step two, we can fine-tunning the model base on our own data, so your own conversations everyday with friends, family, co-workers, etc. Already in this step we have our own AI clone of yourself, think like you, speak like you, write like you. How cool and kind of scary is this? Of course we can go further and make this clone of yourself do its own homework or reinforcement learning, and this model can be a better version of your own self (something every mom want hahah). And what cool about this is the model can stay alive as long as it has people to talk with.
But is this fun? do we as humman want to live forever? maybe it is true to think that it is great for people who miss their late mother can be able to talk to them again? but then what is the meaning of death anymore? isn't it a much more meaningful relationship when we had good goodbye?