> Are you really saying that you can keep more of the code in your head than AI with a one million token context window or that you can read and understand various markdown files in a repo faster?
Of course. A million tokens is 50,000 lines of code. That's nothing.
It's not that you memorize a million lines of code. It's that you develop a solid mental model of how those lines of code are organized, how they work together, what the important gotchas are, what hurt performance in the past. You know the system well. Then when you are interacting with something, you can generally make predictions about the best way to make a change to that system, and you know in which modules to look when diagnosing a problem, so you can work your way around the codebase efficiently.
Most large mature projects are millions of lines of code. Chrome, Photoshop, databases like oracle or postgres.
But the LLM with that million token context window has that 50,000 line memory and then its general training data. That's it.
So your "value" as a developer who is using AI is to bring that mental model with you so you can supervise the AI and guide it to do surgery on that million line codebase with it's puny 50,000 line knowledge of your codebase.
But when you stop coding, you begin to lose your mental model. So, yes, there is a short term burst of productivity as your mental model fades but you are still good enough to guide the LLM. But after a while -- and depending on how deeply that code has been burned into your mind, it could be 6 months or a couple of years -- you will no longer be able to effectively supervise the LLM. You'll be like that noob that wanders into a big codebase they don't understand anymore. Then you wont be able to supervise the LLM.
You have to keep coding in a codebase to maintain your understanding of it. It really is just like having the LLM do your homework for you. Use the LLM to help you learn, to help you do your work, but keep doing the exercises yourself, so that your mental model remains fresh.
> It's not that you memorize a million lines of code. It's that you develop a solid mental model of how those lines of code are organized
As you do when you are leading a team or are more concerned with the overall system - people, business processes, architecture, etc.
I don’t need to write every single if statement to do that
> But the LLM with that million token context window has that 50,000 line memory and then its general training data. That's it.
Absolutely no person us keeping the entire 50K lines of code in memory at the same time - they have a mental model of how the components interact. You don’t lose that mental model when you use LLMs.
> So your "value" as a developer who is using AI is to bring that mental model with you so you can supervise the AI and guide it to do surgery on that million line codebase with it's puny 50,000 line knowledge of your codebase.
Isn’t that what I’ve been saying the entire time?
> But when you stop coding, you begin to lose your mental model. So, yes, there is a short term burst of productivity as your mental model fades
Is you scope of responsibility only what you code yourself? Before LLMs and before I got into consulting, I was responsible for the architecture of a startup with three teams and I did my own MVPs to prove out concepts so I wouldn’t be an “architect astronaught”. I went months with out a line of code. But I did understand the system.
Absolutely no one is paying me to write for loops or understand every line of code. I get paid …decently…for understanding systems. The business, the personalities, conflicting priorities, teasing out the level of technical maturity of the customer, designing the architecture, addressing cross cutting concerns, the modularity of the code etc. the code is no more important that I do it by hand today than it is for me to be writing assembly (or even c) instead of using a compiler.
It’s also not important I know how AWS procures hardware when I submit a YAML file like it was when I was partially responsible for an on prem server room with a whopping 3TB SAN in 2004.
My “mental model” is based on again the processes I’ve perfected to go from requirements of a large implementation -> happy paying customer from an empty git repo and an empty AWS.
The “code” is a minuscule part of any large project. In most enterprise companies that can be outsourced to lower paid people in non US countries. There is no moat around “I know how the for loops work”. For now there is a moat to being able to fly out to a customer site and work with “the business” or on a Zoom call and being able to understand systems. The people who pay you to help you make money or save them money don’t care
But I better be able to explain to the PMO where the project stands, to legal whether we are meeting our contractual obligations, to sales/account management whether the customer is happy and will pay us.
On the customer side, I need to be able to intelligently answer questions that come my way from the security folks, the DevOps department, the bean counters and the development department who have to maintain the project.
Do I need to understand the why and how the system interacts and the behavior of the system? Of course, do I need to know how every for loop was written instead of a while loop? No, I don’t know that if I have a team under me or if I’m responsible for validating the Salesforce integration we outsourced (when I was at a startup).
Coding - in enterprise dev - has been an undifferentiated commodity where the comp hasn’t kept up with inflation in a decade. I saw this happening and was the reason I aggressively moved up the stack as far as scope and responsibility. With LLMs I can take on much larger scoped projects by myself and get them done in 40 hours
> You have to keep coding in a codebase to maintain your understanding of it. It really is just like having the LLM do your homework for you. Use the LLM to help you learn, to help you do your work, but keep doing the exercises yourself, so that your mental model remains fresh.
I graduated from college in 1996. I’ve been assigning other people to do a large part of my “homework” for over a decade. That could be other in house developers or another company to do things like integrations with Salesforce, Workday etc. now it’s Claude and Codex.
Of course. A million tokens is 50,000 lines of code. That's nothing.
It's not that you memorize a million lines of code. It's that you develop a solid mental model of how those lines of code are organized, how they work together, what the important gotchas are, what hurt performance in the past. You know the system well. Then when you are interacting with something, you can generally make predictions about the best way to make a change to that system, and you know in which modules to look when diagnosing a problem, so you can work your way around the codebase efficiently.
Most large mature projects are millions of lines of code. Chrome, Photoshop, databases like oracle or postgres.
But the LLM with that million token context window has that 50,000 line memory and then its general training data. That's it.
So your "value" as a developer who is using AI is to bring that mental model with you so you can supervise the AI and guide it to do surgery on that million line codebase with it's puny 50,000 line knowledge of your codebase.
But when you stop coding, you begin to lose your mental model. So, yes, there is a short term burst of productivity as your mental model fades but you are still good enough to guide the LLM. But after a while -- and depending on how deeply that code has been burned into your mind, it could be 6 months or a couple of years -- you will no longer be able to effectively supervise the LLM. You'll be like that noob that wanders into a big codebase they don't understand anymore. Then you wont be able to supervise the LLM.
You have to keep coding in a codebase to maintain your understanding of it. It really is just like having the LLM do your homework for you. Use the LLM to help you learn, to help you do your work, but keep doing the exercises yourself, so that your mental model remains fresh.