I hadn't heard the term library operating system before, but what is described at that link is nothing like your project.
The "kernel" referred to at that link is a full actual kernel, with networking, proceed scheduling, device drivers and hardware access, etc. It means the full implementation of those features. The reason it can be linked into the application's process as a library is because it's taking about a situation where there's only a single application running on the whole machine (presumably a virtual machine). The kernel and application then run together in a single memory space, with the kernel code (maybe the whole thing?) running in kernel mode.
You have implemented the user mode API that makes system calls to the (Windows) kernel. That is not a kernel in any sense. It is confusing and wrong to call it a kernel.
User mode Linux does actually run something like a real kernel in user space. OP is just providing a POSIX API with no kernel implementation whatsoever. It just forwards those calls to the Window's API. The library OS concept linked to above is different again – that is an actual kernel running in kernel mode but as part of a unikernel (kernel code and application code linked together into a single process).
Still pretty cool but overshadowed by the claim that they've written a whole kernel.
Good point, using wine as the interpreter for PE executables is another cool part of the WINE project.
But it is the case that you can build a Windows application from source on Linux and link it with libwine, in the same way you can build a Unix program from source on Windows and link it with cygwin.dll -- I seem to remember some time ago CorelDRAW (or at least Corel Photo-Paint) was made available for Linux that way, officially supported by Corel.
> we uncover that RL-trained models excel at low k (e.g., pass@1) but are consistently outperformed by base models at high k (e.g., pass@256).
This is a weak argument. I think I get what we are trying to say, but let's take this to the extreme, say pass@10^10^100. Just like a group of monkeys could write Shakespeare if given enough time, a complete random model could probably outperform an RL-trained model at pass@10^10^100. Would we then say the random model can reason too?
Of course the correct reasoning trace will be in the base model's distribution, just like any other well-formed, coherent paragraph. Kind of makes me think, maybe sampling efficiency _is_ intelligence?
When the article starts with an AI generated image that adds nothing to the explanation, it tends to make me suspicious if the article itself was written by an AI as well...
