GPT-5.6 Sol Ultra: $2,000 in tokens, 437,000 lines of AI-written code in one hour.

zester

zester

This is basically the current state of AI coding when you put it in the hands of an actual software engineer. I had it build a display server and compositor around a minimal Wayland-like design, but using a ring-style protocol. It benchmarks around 10%+ faster across the board, with some areas doing considerably better, and uses Vulkan instead of GLES.

From there, it built a 2D graphics engine, GUI toolkit, Flexbox layout engine, font and PNG rendering, a 3D graphics engine, and an AV1 decoder. Most of the rendering is Vulkan-based, with CPU rendering used where it makes more sense. 437,000 LOC of C++ 1hr time.

streamable.com

Watch output | Streamable

Watch "output" on Streamable.
streamable.com streamable.com

This is how I make sure it actually understands what it’s building. I first develop the concept or algorithm, help it implement it in C++, and then have it write a short document explaining how the algorithm works. In this case, I take that documentation, give it to a fresh AI, and have it implement the same algorithm in a different programming language. 300+ LOC of Crystal Lang 2min time.

streamable.com

Watch Screencast From 2026-08-08 17-24-54 | Streamable

Watch "Screencast From 2026-08-08 17-24-54" on Streamable.
streamable.com streamable.com

Neither the design for the compositor nor the smart background erase brush exists in the wild. In fact, the background erase brush only exists in Photoshop and Affinity Photo. I couldn’t find any other image editor that has the same feature.


The 437K LOC only had 5 or 6 compile errors, no warnings, and the code is perfectly understandable and well commented. The 300+ LOC had no compile errors or warnings at all. In my opinion, the code quality of both is at the highest level.
 
myuserid said:
It was one hour of fun and you can be proud of your AI agent.
Now, do this yourself, it will be months of fun and you can be proud of yourself.
Click to expand...
It has nothing to do with being proud of AI. I code professionally for a living, and these are the tools we use. I stress its capabilities because any legitimate software development firm carries liability insurance. Once an algorithm can produce the same code with fewer errors, you’re going to have a problem because now you’re looking at a situation where it may become more expensive to insure code written by humans. And honestly it's a better coder than problably 90% of the human population, I couldn't write the code above and I have been coding for 30 years.

If your company isn’t doing the same and preparing to position your people to take on engineering roles instead of being code monkeys, then you’re not thinking about their futures. And no human is realistically writing almost half a million lines of code in a month, and definitely not for $2,000.

Some people might not like AI, but this is our new reality. I would suggest people find peace with it and figure out where they fit. I don't have a problem with it because I am not threatened by the existence of clever math algorythms.
 
6502 said:
I guess AI can generate thousands lines of code only if such type of software already exist. If it is something new, it will generate garbage.
Click to expand...

That’s kind of true in a way. It can produce new things from ideas that already exist. But producing something completely new independently is a task even humans struggle with.

Most things in our reality that were created by humans weren’t entirely new ideas. They were variations or improvements on ideas that already existed. Programming languages are a good example. Every language that exists wasn’t a completely new invention. They are variations, extensions, or evolutions of ideas that came before them.
 
It is interesting what will happen if hundreds of lines generated by AI turns out to be identical to part of GPL software. AI has to recognize licenses.
 
I asked Codex to check and optimize 20 lines of assembly code, explicitly stating that the code implemented the avrgcc ABI. It removed registers that were no longer in use—which was fine—but it also removed the copying of the stack pointer into Y register (r29:r28). That was a mistake, because any C function might need to access data in stack via Y register if there aren't enough registers available.

I assume AIs are trained much more on C++ than on low-level AVR code.
 
6502 said:
It is interesting what will happen if hundreds of lines generated by AI turns out to be identical to part of GPL software. AI has to recognize licenses.
Click to expand...
The Free Software Foundation has already published technical analysis explaining that code-generating AI learns from training data and attempts to generate new code and does not functioning as a database that simply retrieves and copy-pastes source files.

 
doul said:
I asked Codex to check and optimize 20 lines of assembly code, explicitly stating that the code implemented the avrgcc ABI. It removed registers that were no longer in use—which was fine—but it also removed the copying of the stack pointer into Y register (r29:r28). That was a mistake, because any C function might need to access data in stack via Y register if there aren't enough registers available.

I assume AIs are trained much more on C++ than on low-level AVR code.
Click to expand...
Were you using a paid account? What model(5.6 sol, terra, luna, ..) and reasoning level(low, medium, high, extra high, max or ultra) were you using?
 
So the takeaway is "feed correct/defined requirements into AI and tell it to generate code and it gives back good code"
How is that different from "project managers define correct requirements and humans code to those requirements"?

Programming has always been about solid defined requirements. Crappy software is often traced back to poorly defined requirements, regardless of who implements them.
 
zester said:
AI slop does not mean everything produced with AI is slop.
Click to expand...
There's a problem with the term AI. If a technical achievement is made possible with "AI", the actual computer tech somehow remains out of sight. This doesn't happen accidentally, and it can't have a reason except a commercial one.
 
I remember the 90s, when a single computer (IBM PC/XT 8086) at a transportation company with a program for determining the optimal route and calculating the cost of transporting various types of cargo using various trucks (trucks with semi-trailers, vans, dump trucks, tankers, refrigerated trucks), could replace 10 people, and they were fired, leaving just one computer operator. The task was defined by a single transportation company specialist, and the software implementation was handled by a single programmer. Now, a single transportation company specialist is enough. A programmer is no longer needed....
 
balanga said:
Can AI create a FreeBSD device driver for unsupported WiFi or Bluetooth devices?
Click to expand...
I have no idea. Here is what I do know: it has been used to make drivers for unsupported USB devices on Linux. I created a driver for the Native Instruments Maschine Mikro MK3 based on a spec released by the company. Someone else online, made a driver for a gamepad. The Linux kernel maintainers are using AI, and Linus Torvalds says it's useful.

Right now, I am using two AI agents in a clean room environment where one has access to BlueZ and the other has access to Android's Fluoride Bluetooth stack. The coding agent with access to the BlueZ source code looks at the source and steers the second agent. The second agent will then either write new code or pull code from the Fluoride Bluetooth stack. I'm not sure how well it's going to work. At the moment, I'm about three days in, and we are able to detect and list Bluetooth devices. I imagen freebsd could do the same type of thing and start converting linux drivers to freebsd.
 
ZioMario

bhyve Post in thread 'Enabled the nVidia GPU passthrough inside a Linux virtual machine via QEMU accelerated with BHYVE.'

monwarez said:
So basically you won't take the time to learn how to code, and expect the developer to waste time and read at your bullshit AI generated code which will then taint the mind of such developer reading it thus preventing a clean room re-implementation.

Was that necessary to also pollute the freebsd-virtualization mailing list ?
Click to expand...

With this kind of (rude/inadequate) language there can't be any (polite) answer, at least from me. I want respect, I'm not your brother.

Edit: not so much the last few messages in that thread, more that it fizzled out unresolved (so far as I could tell).
 
zester said:
Were you using a paid account? What model(5.6 sol, terra, luna, ..) and reasoning level(low, medium, high, extra high, max or ultra) were you using?
Click to expand...
Following your question, I took a look because I'm a bit lost—the world of AI is evolving and changing so fast! :-o
So, I have a paid "Plus" subscription that gives me access to:
- ChatGPT in "chat" mode with GPT-5.6 Sol (thinking effort: low, medium, high)
- Codex Cloud without model choice (more or less the same as the chat version)—connected to my GitHub repository
- ChatGPT in "work" mode (this is new): options to choose the model (5.6 Sol, Terra, Luna) and adjust thinking effort and speed. It offers GitHub connectivity, but it’s not a coding agent; it’s a workspace agent.

It's not a professional environment like the one you use.
 
doul said:
Following your question, I took a look because I'm a bit lost—the world of AI is evolving and changing so fast! :-o
So, I have a paid "Plus" subscription that gives me access to:
- ChatGPT in "chat" mode with GPT-5.6 Sol (thinking effort: low, medium, high)
- Codex Cloud without model choice (more or less the same as the chat version)—connected to my GitHub repository
- ChatGPT in "work" mode (this is new): options to choose the model (5.6 Sol, Terra, Luna) and adjust thinking effort and speed. It offers GitHub connectivity, but it’s not a coding agent; it’s a workspace agent.

It's not a professional environment like the one you use.
Click to expand...

In the Codex CLI, you can use the /model command, then go down to More reasoning to increase the reasoning effort.

1787060530361.png


1787060598449.png


1787060641516.png


It’s better at some languages than others. C, C++, and Python are top tier.

One of the first things I used AI for was learning about compression algorithms, and I quickly realized how insanely difficult they are to develop. I’m not talking about using something like zstd. I mean actually creating a new compression algorithm.

Creating a new image or video format is easy. Getting a 60-minute video down from something like 160 GB without destroying it is the hard part. This type of math is going to be way outside the capability of most software developers.

Go to diagrams.net https://app.diagrams.net/ and start designing whatever you want to build as a flowchart. The more detailed, the better. Think of yourself as the engineer and AI as the code monkey. I don’t mean just telling it, “Make me a program.” You need to engineer the entire thing from the ground up.

Break the code into small areas of concern. That saves tokens, makes debugging easier, and lets you actually understand what it’s writing instead of dumping thousands of lines on you at once. You can have it explain concepts you don’t understand and write examples.

Some days AI will feel brilliant, and other days you’ll be sitting there thinking, “WTF, why are you so stupid?” Model behavior and performance can vary, so don’t rely on it to do the engineering for you. You design the system. AI helps you build the pieces.

Slop comes from laziness. With experance and effort it will become a force multiplyer.

I use it extensively to write C bindings for Ruby. I love programming in Ruby, but its ecosystem has a lot of abandoned packages. I also don’t trust third-party librarys anymore for any language do to the massive volume of malware showing up in open-source ecosystems.
 
You must log in or register to reply here.
Back
Top