Intention - The Next Layer of Abstraction

Homo Sapiens will be surpassed in raw intelligence very quickly if it hasn't happened already.

What is intelligence? To answer that question, we have to answer what is the utility of it. Every situation that requires an action can be described as a function. The processor gets a set of input, processes it and produces an output. For example, test question, exam time, and passing requirements are the inputs; the student is the processor (where the function sits), and the answer script is the output. At its core, we are agents who inherit a model when we are born which has been pretrained for millions of years in reinforcement learning environments and sits in our genes. We go through more learning in life which is post training captured by the neuroplasticity of the brain. This is the basis of why AI is literally everywhere. Neural nets are universal function approximators.

Intelligence (model/function) is the combination of two things: data and the model architecture.

Data is simply repetition (total epoch) and experience (diverse data). Kobe Bryant, although being one of the greatest basketball players of all time, would throw the ball in the basket for decades for improvement. Training epochs or repetition matters because the model becomes more intelligent (better output) with each learning cycle. Experience on the other hand defines the range (data diversity for generalization). Elon Musk is obviously not the best engineer in his companies but he has simply accumulated data in many crucial domains (diplomacy, fund raising, leadership etc.) which makes him the CEO.

However, given the same amount of data some human or model outsmarts its peers. The differentiator here is the architecture quality. RNN or LSTM lost to the Transformer as the default for modelling the large language models (LLM) not because more data but the information dynamics of transformer is simply superior. One of my favorite anime "Ping Pong the Animation" depicts how an unmotivated prodigy wins against a player who sacrificed everything for winning. "Naturally gifted" people possess architectural advantages (better abstraction capability, freak genetics of Mike Tyson) which helps them produce better results with similar training budget.

Humans have limits to how much data we can store in the brain. Abstraction helps us manage large information, but then our context window is limited. How many ideas we can juggle at the same time is limited by the attention capacity. Furthermore, abstracting out gets rid of details. These phenomena are also applicable to artificial intelligence but the key difference is SCALE; infinite improvement to context window, data storage and compute is only applicable to silicon intelligence.

The competition is uneven because we can't really join the brains of Newton and Leonardo da Vinci, and then that to all human beings. It also sucks that we lost the beautifully trained weights of someone like Gauss, tuned exceptionally for mathematical reasoning. Artificial intelligence is infinitely additive and lasting. Even if we don't want to admit it, in many areas AI already outperforms most human practitioners (coding, summarizing).

Many see AI as a problem as it is going to be more intelligent than us. I think we will be freed from the execution layer to the higher level of abstraction: Intention or goal setting. Previously people used to count manually, calculators freed us from literal counting. Very few people know how a calculator does it but we are 100% confident in the calculator's result. It lets us work on even a higher abstraction level: coding (instructions for data processing and routing). Again, very few coders fully understand how their codes are capable of orchestrating billions of electrons to produce the correct result. We don't need to as we have abstracted that out to the compiler for working on higher levels.

Then why are we afraid that all codes will be written by the machines eventually? What is the point of knowing the syntax of the coding languages? Why is the logical understanding not enough? Furthermore, what is wrong with not knowing exactly what selection sort or insertion sort algorithms are? Isn't just knowing when to sort and AI writing the optimal algorithm enough? Pushing the argument deeper, very few people care about knowing the real inventor. What is wrong with AI making deep theoretical breakthroughs? An AI model disproving an 80 year old conjecture, with mathematicians then interpreting the result for us. The best theoreticians are going to be doing this more from now on.

Intention becomes the new layer of abstraction for human. We will care less about what theory is being used to store ten times more data in a single chip but think what to store. What type of data is worth producing. Bold ideas will emerge. Imagine Michael Jackson developing his own website with the help of the best web developer of all time. Optimization...