I think you're right, but rather surprisingly I have some experimental data on this question that is inconclusive at best. My hobby project is symbolasm, which is intended to close exactly the gaps you describe: by naming variables and introducing types, it re-adds much of the context that assembly language erases. I ran a quick'n'dirty experiment to see if it would make the LLM more efficient at writing code than raw assembler.
The task was rewriting an MP3 decoder from Rust into ARM assembly language, Cortex-M33 flavor specifically, optimizing for both speed and code size. The output is approx 3k instructions. The tl;dr is that Opus 5.5 nailed the task, whether writing assembly or symbolasm. It took approximately 2x tokens for the latter, which I think is largely explained by the fact that it was dealing with a novel language nowhere represented in its training set.
The quality of the result was comparable; the raw asm was slightly faster and slightly less dense, which I consider insignificant because I think you can get the same effect just by tuning the prompt. In particular, if I were re-running the experiment, I might set a strict code size budget and then ask it to optimize for speed within that budget, allowing comparison on a single scalar.
The experiment was compelling in another way: the code was 3x denser and 1.5x faster than -O3 output from Rust. So for certain applications, going to assembler will be worth it, especially as the cost of generating the code goes to zero.
If you have an Adafruit Fruit Jam, you can run the demo on that board [1]. It can stream 640x480 video off an SD card at 60fps, as well as display a static image from flash.
Also possibly of interest to HN, the decoder[2] is written in a new low level language called symbolasm[3], which seeks to combine the advantages of assembly language (it's possible to reason about and optimize performance with few limits) with some of the affordances of C, especially named variables and struct fields.
I notice that assembly language is not on this list. I've been working on a modern syntax for assembler, with C-like types[0], something of a successor to qhasm. I've written a race-the-beam renderer[1], which is a good fit because of its incredibly tight performance constraints. I've been doing a combination of coding by hand and Opus 5.5. Anecdotally, I find Claude can code in it just fine, something I'd like to make more quantitative.
Dan Luu has done that kind of evaluation[2], and found that LLMs do well in a range of languages, but do underperform in assembler. My hypothesis is that expressing programmer intent clearly, and having the compiler verify things like consistent placement of variables in registers, will fix the problems that hold LLMs back.
This has been a hobby project, but it feels ready to break containment. I'm open to a research opportunity where I'd be able to work on it for real - get in touch.
My symbolasm project[1] is exactly that: an assembler syntax designed to make programming more mentally tractable. I expect it to help for both human and LLM coders. Apologies for the spare repo, I haven't written good readmes yet.
In this syntax (though I'm not yet targeting x86_64) it would look something like a = round(fma(a, b, c)) where the variables have the appropriate vector type (and have been placed into registers with @zmm1 syntax when first assigned).
The design of the language is entirely by hand, as was the implementation when I started, but lately I have been using AI to speed implementation.
The task was rewriting an MP3 decoder from Rust into ARM assembly language, Cortex-M33 flavor specifically, optimizing for both speed and code size. The output is approx 3k instructions. The tl;dr is that Opus 5.5 nailed the task, whether writing assembly or symbolasm. It took approximately 2x tokens for the latter, which I think is largely explained by the fact that it was dealing with a novel language nowhere represented in its training set.
The quality of the result was comparable; the raw asm was slightly faster and slightly less dense, which I consider insignificant because I think you can get the same effect just by tuning the prompt. In particular, if I were re-running the experiment, I might set a strict code size budget and then ask it to optimize for speed within that budget, allowing comparison on a single scalar.
The experiment was compelling in another way: the code was 3x denser and 1.5x faster than -O3 output from Rust. So for certain applications, going to assembler will be worth it, especially as the cost of generating the code goes to zero.
[1]: https://github.com/raphlinus/symbolasm