Imagine that you're given a business problem to solve. You represent the process of writing the code with a graph - each vertex is a git commit. We consider the space of all possible git commits, so the graph is infinite. All vertices are connected with directional edges, and each edge has a value "cost". If you are in commit A and you want to go to commit B, you have to pay the cost from A to B. Your goal is to find a relatively short path from empty git commit to any vertex which contains code that has some specific observable business properties.
You might notice that not everyone is equally smart, so when giving this task to real people, we'll associate "speed" with each person. The higher the speed, the lower the paid costs when traversing the graph. I'll leave the specifics vaguely undefined.
Since a part of the task is to discover information about the graph, we also need to specify that every person has some kind of heuristic function that evaluates how likely given node is to get you closer towards some vertex that can be considered a goal. Obviously, smarter people have heuristic functions that are more closer to ground truth, while stupid people are more biased towards random noise. This also models the fact that it takes knowledge to recognize what a correct solution is.
This model predicts what we intuitively think - smart specialists will quickly discover connections that take them towards the goal and pay low costs associated with them, while idiots will take the scenic route, but by and large will also eventually get to some vertex that satisfies the business requirements, even if it's a vertex that contains mostly low-quality code, because for idiots the cheap edges that seem good at first glance are the only edges they can realistically traverse.
Obviously, if you have a group of people working on the same task, you'll reach the business goal faster. Therefore, a group of people is equivalent to one person with higher speed, and some better heuristic.
This conclusion suddenly creates a well-known, but interesting situation - each smart specialist can be replaced by a group of idiots. Or, the way I heard it, "the theorem of interns - every senior can be replaced by a finite number of interns".
What AI does is it increases people's speed. Not the heuristic function, but the speed. Importantly, the better the heuristic function, the smaller the speed gains. Makes sense - an idiot who doesn't know shit and copy-pastes things from ChatGPT will have massive speed gains, while a specialist will only modestly benefit from AI.
From business perspective though, by having more idiots write more slop with more AI we traverse the graph significantly faster. Sure, we still take the scenic route, and maybe even with AI we take the really fucking long scenic route, but because the speed is so high, it doesn't matter.
And because AI supercharges idiots more than smart specialists, we have a situation where the skill of working with idiots is more valuable on the job market than the skill of doing your job right. Your goal isn't to find the shortest path, or the prettiest code, your goal is to prompt AI as quickly as possible to get you to any vertex that satisfies the business requirements.
Your graph model lack the aspect of increasing complexity. As you traverse the graph every available node gets increasingly more distant. In some areas of the graph less so than others, a good heuristic function not only identifies a single shortest path, but also dense areas of possible value in the graph.
The question is if blind speed scales quicker then distances grow.
That's true, and I guess the reason why we're building so many datacenters is to answer the question how far exactly will blind speed take us, assuming that we fail to make substantial improvements to AI architecture.
Imagine that you're given a business problem to solve. You represent the process of writing the code with a graph - each vertex is a git commit. We consider the space of all possible git commits, so the graph is infinite. All vertices are connected with directional edges, and each edge has a value "cost". If you are in commit A and you want to go to commit B, you have to pay the cost from A to B. Your goal is to find a relatively short path from empty git commit to any vertex which contains code that has some specific observable business properties.
You might notice that not everyone is equally smart, so when giving this task to real people, we'll associate "speed" with each person. The higher the speed, the lower the paid costs when traversing the graph. I'll leave the specifics vaguely undefined.
Since a part of the task is to discover information about the graph, we also need to specify that every person has some kind of heuristic function that evaluates how likely given node is to get you closer towards some vertex that can be considered a goal. Obviously, smarter people have heuristic functions that are more closer to ground truth, while stupid people are more biased towards random noise. This also models the fact that it takes knowledge to recognize what a correct solution is.
This model predicts what we intuitively think - smart specialists will quickly discover connections that take them towards the goal and pay low costs associated with them, while idiots will take the scenic route, but by and large will also eventually get to some vertex that satisfies the business requirements, even if it's a vertex that contains mostly low-quality code, because for idiots the cheap edges that seem good at first glance are the only edges they can realistically traverse.
Obviously, if you have a group of people working on the same task, you'll reach the business goal faster. Therefore, a group of people is equivalent to one person with higher speed, and some better heuristic.
This conclusion suddenly creates a well-known, but interesting situation - each smart specialist can be replaced by a group of idiots. Or, the way I heard it, "the theorem of interns - every senior can be replaced by a finite number of interns".
What AI does is it increases people's speed. Not the heuristic function, but the speed. Importantly, the better the heuristic function, the smaller the speed gains. Makes sense - an idiot who doesn't know shit and copy-pastes things from ChatGPT will have massive speed gains, while a specialist will only modestly benefit from AI.
From business perspective though, by having more idiots write more slop with more AI we traverse the graph significantly faster. Sure, we still take the scenic route, and maybe even with AI we take the really fucking long scenic route, but because the speed is so high, it doesn't matter.
And because AI supercharges idiots more than smart specialists, we have a situation where the skill of working with idiots is more valuable on the job market than the skill of doing your job right. Your goal isn't to find the shortest path, or the prettiest code, your goal is to prompt AI as quickly as possible to get you to any vertex that satisfies the business requirements.