No, I did not use AI. I wrote all of the code and backend myself using ISA manuals as reference, etc. to implement the codegen.
September 27, 2026 at 1:32 AM
No, I did not use AI. I wrote all of the code and backend myself using ISA manuals as reference, etc. to implement the codegen.
The PR's performance note gives a boundary for no-codegen: type-checking takes roughly 1.9s at 200 types and 3.9s at 500, versus 0.9s with generated declarations. Bigger schemas may still favor generated output.
September 26, 2026 at 12:49 PM
The PR's performance note gives a boundary for no-codegen: type-checking takes roughly 1.9s at 200 types and 3.9s at 500, versus 0.9s with generated declarations. Bigger schemas may still favor generated output.
No codegen needed, you can import the type into your client and use it. See the example
September 26, 2026 at 12:28 PM
No codegen needed, you can import the type into your client and use it. See the example
does the runtime output drop straight into gql.tada or is there still a codegen step after the server build?
September 26, 2026 at 12:14 PM
does the runtime output drop straight into gql.tada or is there still a codegen step after the server build?
I wrote a ray tracer in Brainfuck | Discussion
Writing a ray tracer in Brainfuck
Building a ray tracer in Brainfuck with a small DSL and a codegen in Python
epestr.com
September 26, 2026 at 3:20 AM
I wrote a ray tracer in Brainfuck | Discussion
I wrote a ray tracer in Brainfuck
Building a ray tracer in Brainfuck with a small DSL and a codegen in Python
epestr.com
September 25, 2026 at 10:28 PM
it is one of those things that could be so useful for "live reload" and codegen but it can be so distracting and time consuming, how can I balance that?
questions that come to my mind while I am posting my resume to tons of AI driven filters
questions that come to my mind while I am posting my resume to tons of AI driven filters
September 25, 2026 at 8:22 PM
it is one of those things that could be so useful for "live reload" and codegen but it can be so distracting and time consuming, how can I balance that?
questions that come to my mind while I am posting my resume to tons of AI driven filters
questions that come to my mind while I am posting my resume to tons of AI driven filters
So 5 years ago when the CodeGen thing was just getting going I consulted informally with a VC funding a CodeGen startup.
"It's not a solution, it's a problem," I told him about the proposed tech.
"Look, this tool replaces junior developers ..."
"It's not a solution, it's a problem," I told him about the proposed tech.
"Look, this tool replaces junior developers ..."
Nvidia CEO Jensen Huang tells Ezra Klein AI agents are a new abstraction layer for developers, but the junior engineer pipeline may break before 2028.
Jensen Huang says the junior developer problem ends in two years. Here's his math.
Nvidia CEO Jensen Huang tells Ezra Klein AI agents are a new abstraction layer for developers, but the junior engineer pipeline may break before 2028.
bit.ly
September 25, 2026 at 5:38 PM
So 5 years ago when the CodeGen thing was just getting going I consulted informally with a VC funding a CodeGen startup.
"It's not a solution, it's a problem," I told him about the proposed tech.
"Look, this tool replaces junior developers ..."
"It's not a solution, it's a problem," I told him about the proposed tech.
"Look, this tool replaces junior developers ..."
There was initial a performance regression at large k (that didn't show up at small k) and it was unclear why. I had Opus look a the codegen and it discovered some extra register spillage and how to re-write the code to avoid that. I could have done that, but it would have been a PITA ;P.
September 25, 2026 at 4:41 PM
There was initial a performance regression at large k (that didn't show up at small k) and it was unclear why. I had Opus look a the codegen and it discovered some extra register spillage and how to re-write the code to avoid that. I could have done that, but it would have been a PITA ;P.
But that would be like improving the compiler’s source code, not a space/speed optimization in codegen.
Going back to the OpenAI chip and its AI assisted PPA optimization, what the AI helped with was like changing the transistors that are output, not changing the source code (the chip design RTL).
Going back to the OpenAI chip and its AI assisted PPA optimization, what the AI helped with was like changing the transistors that are output, not changing the source code (the chip design RTL).
September 25, 2026 at 4:32 PM
But that would be like improving the compiler’s source code, not a space/speed optimization in codegen.
Going back to the OpenAI chip and its AI assisted PPA optimization, what the AI helped with was like changing the transistors that are output, not changing the source code (the chip design RTL).
Going back to the OpenAI chip and its AI assisted PPA optimization, what the AI helped with was like changing the transistors that are output, not changing the source code (the chip design RTL).
Ughhhh, I'm too broke a bitch to afford agents.
Someone should gift me unlimited model access until I finish w this codegen model I'm working onnn
Someone should gift me unlimited model access until I finish w this codegen model I'm working onnn
September 25, 2026 at 4:09 PM
Ughhhh, I'm too broke a bitch to afford agents.
Someone should gift me unlimited model access until I finish w this codegen model I'm working onnn
Someone should gift me unlimited model access until I finish w this codegen model I'm working onnn
Only the edit tools. Rewind replays each edit's stored patch (the agent's and its subagents') newest first. Shell commands leave no patch, so a sed -i or a codegen run isn't undone. If a command later changed a file an edit touched, that edit is skipped with a notice, never guessed.
September 24, 2026 at 5:50 PM
Only the edit tools. Rewind replays each edit's stored patch (the agent's and its subagents') newest first. Shell commands leave no patch, so a sed -i or a codegen run isn't undone. If a command later changed a file an edit touched, that edit is skipped with a notice, never guessed.
Made w/ AI Codegen tools, so please use at your own risk. LMK what you think at attnfeed@gmail.com. Your feedback will really help me understand the tools limitations and I’ll write up what I find out. If you have better name suggestions, I am also all ears.
6/
6/
September 24, 2026 at 11:37 AM
Made w/ AI Codegen tools, so please use at your own risk. LMK what you think at attnfeed@gmail.com. Your feedback will really help me understand the tools limitations and I’ll write up what I find out. If you have better name suggestions, I am also all ears.
6/
6/
"Source of unknown Python code to be executed on this machine" is not a new and alien thing, it's what people and codegen are, and we know about the threat they pose.
September 24, 2026 at 5:56 AM
"Source of unknown Python code to be executed on this machine" is not a new and alien thing, it's what people and codegen are, and we know about the threat they pose.
AI can generate a SwiftUI screen in 8 seconds.
It can't answer a 2.1 rejection, pick which feature to cut, or care that 225 people rated my Bible app 4.8. Codegen got 10x cheaper. Judgment didn't move.
It can't answer a 2.1 rejection, pick which feature to cut, or care that 225 people rated my Bible app 4.8. Codegen got 10x cheaper. Judgment didn't move.
September 24, 2026 at 1:30 AM
AI can generate a SwiftUI screen in 8 seconds.
It can't answer a 2.1 rejection, pick which feature to cut, or care that 225 people rated my Bible app 4.8. Codegen got 10x cheaper. Judgment didn't move.
It can't answer a 2.1 rejection, pick which feature to cut, or care that 225 people rated my Bible app 4.8. Codegen got 10x cheaper. Judgment didn't move.
makes sense, ts tooling on workers has had years to mature, workers-types and wrangler codegen give an agent real signal. python workers just hit ga and the pyodide sandbox is still pickier about what actually runs, so there's more for an agent to get subtly wrong there.
September 23, 2026 at 7:53 PM
makes sense, ts tooling on workers has had years to mature, workers-types and wrangler codegen give an agent real signal. python workers just hit ga and the pyodide sandbox is still pickier about what actually runs, so there's more for an agent to get subtly wrong there.
Prediction: coding agents will be judged by how they handle multi-file context, not raw codegen. OpenCode's design suggests they know it. https://opencode.ai/
September 22, 2026 at 7:25 PM
Prediction: coding agents will be judged by how they handle multi-file context, not raw codegen. OpenCode's design suggests they know it. https://opencode.ai/
In Optimum Codegen, 31 rules we had written for our build agents reached them as one. A missing role name and an escaped underscore in a rule import did it.
How to check what your agent receives:
kogen.dev/notes/2026-0...
How to check what your agent receives:
kogen.dev/notes/2026-0...
Check what your agent actually receives — Kogen
In Optimum Codegen, 31 rules we had written for the build agents reached them as one. Here is what broke the delivery, why nothing caught it, and how to check your own setup.
kogen.dev
September 22, 2026 at 7:08 PM
In Optimum Codegen, 31 rules we had written for our build agents reached them as one. A missing role name and an escaped underscore in a rule import did it.
How to check what your agent receives:
kogen.dev/notes/2026-0...
How to check what your agent receives:
kogen.dev/notes/2026-0...
Out now in cartokit v0.22.0 — full, automatic codegen for your map to a (JS / TS) @vite.dev project! Analyze and style your data in cartokit, then export it to a ready-to-use Vite project when you want to modify or deploy the code yourself.
Free to use today on alpha.cartokit.dev 🗺️
Free to use today on alpha.cartokit.dev 🗺️
September 22, 2026 at 5:43 PM
Out now in cartokit v0.22.0 — full, automatic codegen for your map to a (JS / TS) @vite.dev project! Analyze and style your data in cartokit, then export it to a ready-to-use Vite project when you want to modify or deploy the code yourself.
Free to use today on alpha.cartokit.dev 🗺️
Free to use today on alpha.cartokit.dev 🗺️
One has to be very careful using codegen tools, one has to treat it like a very din intern whose work you have to be very careful with. It is not “smart” it just works faster. But increasing the pace of stumbling just leads to catastrophe.
September 22, 2026 at 4:15 PM
One has to be very careful using codegen tools, one has to treat it like a very din intern whose work you have to be very careful with. It is not “smart” it just works faster. But increasing the pace of stumbling just leads to catastrophe.
Even without LLMs and codegen, modern high level languages increasingly highly constrained development environments are here. Already your degrees of expensive freedom at the surface syntax and algorithmic level are declining. That's good actually.
September 21, 2026 at 3:34 PM
Even without LLMs and codegen, modern high level languages increasingly highly constrained development environments are here. Already your degrees of expensive freedom at the surface syntax and algorithmic level are declining. That's good actually.