The responses all seem a little harsh. If you poke through the user's history, this kind of thing isn't new for them, going back to pre-LLM-could-do-this days.
I think the docs are light because the primary goal for them is not to provide everyone with a well documented linux on esp-32-s31 guide, but rather this that they shared:
"I'm currently working on a hackable music player and I used to prototype with the OG esp32, and tbh if I wasn't for Bluetooth audio I'd move on to S3 already" ( from the links in this comment https://news.ycombinator.com/item?id=49134987 )
> If you poke through the user's history, this kind of thing isn't new for them, going back to pre-LLM-could-do-this days.
Is there another set of comments or history I should be looking at? The only pre-LLM comment is about doing a prototype for something with the old ESP32. That was a common hobby microcontroller introductory project, but there’s a world of difference between playing with an OG ESP32 and porting MMU Linux to a new platform.
I think we should be honest about what this is: Someone spent their tokens letting an agent attempt bring up of Linux on the platform and it got something to work. I’m appreciative that it was shared. However, given the lack of useful documentation (the key MMU doc is basically empty) and the lack of other explanations, I don’t think we should be reading more into this than as a pure LLM agent proof of concept.
There isn't much interesting about a rocket engine other than it's rocket science and it's awesome if it works. A basic mockup of one is in PrusaSlicer sample files.
It takes no effort to play LLM or not, you don't need to know anything about the topic or put anytime into reading and understanding the submission, just find that em-dash and you can turn every discussion into meta slop.
I also think an LLM isn't going to just spit this all out and viola, it boots. I imagine there were many rounds of jtag debug, copy/paste into the LLM prompt, with enough human knowledge/context to say the right thing, suggest some existing implementation snippet, etc.
> I imagine there were many rounds of jtag debug, copy/paste into the LLM prompt
You may not be familiar with modern LLM tools.
You don’t need to copy/paste anything. You can easily instruct your agent to access the serial port and JTAG debugger and it will easily handle it through subsessions.
The tools have moved rapidly this year. Something like this is entirely doable by attaching the right cables, telling the agent where to access everything, and then keeping it fed with enough tokens to keep going.
There are some amazing projects doing reverse engineering or porting to micros fully automated. The results are still full of typical LLM output problems, but it’s amazing what can be brute forced with enough tokens in an LLM loop.
When I’m evaluating MCU platforms for new projects I’ll some times set an LLM loose on each to get a proof of concept running so I can do benchmarks or testing. It can save a lot of time finding showstoppers or roadblocks before I waste a lot of time on doing the problem correctly.
Obviously vibecoded but interesting nevertheless. The agent left everything marked as untested in the README but the output snippets toward the bottom imply that it got something working enough to log in and run some commands.
Nobody is going to mistake this for a carefully crafted port of Linux but it at least serves as a proof of concept.
The real downside of the vibecoding is that we don’t get any helpful information about what it took to get it done with thoughtful analysis from a human. Just a chunk of code in a GitHub repo with some half-coherent README. There’s a docs folder, but the documentation about the MMU part just says that there are two MMUs across a couple lines of notes. Okay, great.
Given the context of a vibe-coded project, would it be reasonable to assume the example outputs are outright hallucinations, without seeing indisputable evidence to the contrary?
Sad that’s where my mind goes, but this is what the world has been training me to believe. And these doubts now eclipse the skepticism that I developed toward things that humans posted on the internet.
>Espressif's radio firmware blobs are closed source, and must run within ESP-IDF's FreeRTOS framework. It's near impossible to reverse-engineer them (not to mention legal risks.)
It's not nearly impossible, it has already been done :)
The responses to agent-aided dev seem a little over the top to me.
The derisive "vibe-coded" label is one thing, but the least compelling argument or complaint of them all is the truly worn out one about barriers to entry that you and I and so many others worked soooo hard to climb over are now being obliterated, allowing any casual Moe to stroll into your domain. How many professionals -- djs, photographers ... hell, SEO experts -- pretty much anyone technical and possibly creative have watched technology make work and the accumulation of competitors almost too easy.
If you just remember that no one in this time-space or anyother AFAICT is forcing you to consume or even read about projects like this, you'll find life goes on and you're actually still an expert who people who need experts will value materially (which should be your thing, otherwise wth are you complaining about?)
In fact, if you keep your mind open you might -- not definitely but _just_ _might_ -- find a piece of something useful amidst the slop. One fellow's trash ...
What I remember was they said it could come back if people needed it but it was broken for looking (months, years) at a time. So even if they didn't remove it, 6.12 might be the last working version with it anyway
Thanks for the links. This doesn't seem to be a true RISC-V MMU (according to the Sv32 specification) integrated into the CPU core itself, but just a peripheral designed for memory mapped SPI flash and PSRAM. So as far as I understand there is no true process isolation with page faults and dynamic paging.
Ok, I see. The S3’s "MMU" is just an external-memory mapper, not a virtual-memory MMU. The S31 apparently has both, that mapper plus an architectural CPU-side Sv32 MMU; that offers indeed a lot of interesting possibilities. Even sel4 would run on this machine as it seems. The ARM world has no microcontroller with a true MMU as far as I know. Risc-V now has at least two (here is the other one: https://www.bunniestudios.com/blog/2026/baochip-1x-a-mostly-...).
The Microchip PIC32MZ MCU has an MMU as well. But not with wireless options in a 8x8 QFN80 package like this ESP32-S31. That's pretty small. No flash though.
I can't find it now, but someone ported nommu linux to those cortex-m3 "bluepill" development boards that could be bought for ~$2 (they were built wrong and liquidated for a massive discount).
Seeing the title of this submission reminded me of it because I messed around with it a bit a few years ago.
Yeah, nommu Linux basically cannot run normal ELF since there is no virtual memory which is needed for relative addressing and relocations. You are mostly left with classical formats like AT&T a.out only
You wouldn’t be able to guarantee execution of arbitrary ELF, but couldn’t you intentionally build non-PIC-compiled ELF executables where the section base addresses as defined in the header must match the MMU region “slots” the host provides?
I think the docs are light because the primary goal for them is not to provide everyone with a well documented linux on esp-32-s31 guide, but rather this that they shared:
"I'm currently working on a hackable music player and I used to prototype with the OG esp32, and tbh if I wasn't for Bluetooth audio I'd move on to S3 already" ( from the links in this comment https://news.ycombinator.com/item?id=49134987 )
reply