> the Claudeisms
This is hand-waving. Please be more specific.
> made it such a slog
On the flip-side, I didn't find it a slog at all. What if you're wrong?
My gut feeling agrees. The rule of three is one of the stronger signals, can't stamp that out of the AI even if you wanted :-)
I'll push it to GH later, it's nothing fancy but it has been quite good in my experience. Here is highlights which it used
``` tricolon coordinated VERB run: “…break / on them collapse / unlock everything .” (3 members) tricolon coordinated VERB run: “…guard physical addresses / not DRAM coordinates / you rearrange the” (3 members) tricolon coordinated NOUN run: “…handful of data / it to z3 / the translation matrix” (3 members) tricolon coordinated NOUN run: “…view / the elaborate fences / locks / security checks the” (4 members) tricolon coordinated VERB run: “…Read it / the alias map / pipe” (3 members) ```
I'd just like to address this real quick because some people seem to think this is just a "hunch" that has some probability of being false; there is absolutely nothing more certain on planet Earth than the LLM involvement in this writing. It is difficult to come up with things that are certain enough to compare this to to convey the lack of doubt that exists.
I am not going to make fun of you for not being able to tell, although I do find it surprising that people seem to struggle in both directions with telling AI and human writing apart (are our brains really that different?) - I just want it to be clear that some of us can pick up Claudisms within just a couple of sentences with no effort. A Claude-generated sentence, in isolation, may not ring any alarm bells. A few of them in a row, however, that's a load-bearing smoking gun right there.
We can certainly argue to what extent undisclosed LLM involvement is an issue or not, though frankly I don't like reading LLM writeups so I would greatly prefer if people would stop using LLMs for public facing documents. But, it is at least worth making this much clear: we can tell.
Go ahead and throw the comment into your favorite unreliable AI detector. Even though I suspect they're mostly garbage, my writing is just so far away from what AI models do that it doesn't even matter.
edit: I caved into temptation and checked. Big fat zero on GPTZero.
The em dashes are the most obvious stereotypical tell, but that doesn't really matter that much (I actually like them and occasionally used them pre-AI). It's hard to put a finger on, but the most annoying LLMism to me is the overdramatic, staccato, almost "epic" way they talk. It feels like a 2009 lens flare effect over everything, it sounds like a stereotypical hacker in a CSI show.
> the last generation whose datasheets document the DRAM controller's translation registers — and show that they can't be locked
> When your code dereferences *p, it appears to access the DRAM at p. It does not — p is a virtual address
> Physical addresses are really more of a suggestion.
> That's the exploit. All of it.
The worst part is that this stuff is genuinely cool and deserves to be dramatic. And I like stereotypical, campy hacker speak! But LLMs are, IDK... bad at it? Or maybe it just becomes a bore to read the same. Exact. Dramatic. Voice. From literally everyone. After you've heard it enough times.
None of this is against Mr. Domas. He seems like a cool person, with a cool voice, and I want to read his voice, not Claude's.
> What if you're wrong?
I definitely could be! Apologies if I am. But with all the em dashes and such, and having read his previous work, I felt confident enough to mention it. And as the sibling comment says, it really is something you just learn to spot over time.
Likewise, I don't mind the diagrams. Though they do often have the same flaw as other text, being that the LLM throws in EVERYTHING, vs. a handmade one that'd generally have more taste and discretion to it. That can kind of work in its favor here, since the point is just to show the complexity of the stack, but on the other hand the reader lacks confidence that every item in there is "really" a part of the stack (which I would be fully confident in for this author, had he written it by hand) and not just some process related to memory/DRAM that the LLM decided to toss in.
Damn. Now my artist-mode in Emacs skills are useless.
I hate these especially much: It's at the same both both overly dramatic, it's presented as some great reveal that will change everything, while at the same time being completely trivial and only detracts from the explanation. If you have no idea that memory addresses are translated you will understand absolutely nothing from the text or even what this is all about. If you want to explain what an MMU is, just do that instead and don't present it as some great revelation.
But some equally dramatic phrasings could just as well be something that leaves you astonished. You never know. You have to skim the text to find what is useful information and what is just filler. The signal-to-noise is low.
It's called slop for a reason.
Swizzling "randomizes" bank/rank/channel distribution, which makes unlucky access patterns less likely. (Something I'd like to research is microbenchmarking different access patterns to infer the swizzle pattern and defeat physical ASLR)
https://jxself.org/titanic.shtml
He did it well. On "security", the author loves more to own his code/adata than anything. as did the PDP10/ITS hackers.
But why on earth do they have to use AI to write their writeups?!
But I have a new favorite way of demonstrating this:
https://github.com/search?q=owner%3Axoreaxeaxeax+load-bearin...
Guess how many of these are from before 2025.
Ring -1 needs DRAM, so it tells the memory controller to give it some blocks. The memory controller hands back a “physical” address, and promises not to let anything but ring -1 access that address.
The exploit takes advantage of that control register to remap the same DRAM blocks to a different physical address. Since the memory controller only promised to protect the physical address it handed back, that protection is bypassed when using the new address.
There are several theoretical ways to mitigate this exploit, but it remains to be seen if the system is sufficiently field-upgradeable to defend.
The article would be better with just the instructions and audited output. All the LLM added bloat is tiring and distracting; it's like an article from New-Yorker or Wired.
This was a relatively complicated post of the sort that we are lucky to get in any form, AI-assisted or otherwise. Does it meet my personal stylistic standards? No, it's too LLM-ish. Assuming I cared about the presentation at all -- which I don't always, but would here -- I wouldn't be able to stop myself from fixing that in the process of reviewing it. Is it the usual bucket of slop? Emphatically no.
You are literally asking this to people who clearly care...
> In today's present, I wouldn't bother writing the article myself neither besides giving the instructions and auditing the output.
Then why exactly are you even bothering to reply to me instead of having Claude do it and auditing the output? If HN didn't have a rule against it, would you even bother replying yourself?
But really, there's a fair bit more to unpack here than just that. Why wouldn't you bother? Is writing a README.md about some project you worked on really that hard? Even with heavy LLM assistance, I'd wager to guess this project, which clearly involved working on real hardware, was more than just prompting. So clearly there was human effort other than prompting. And I do respect that, but I want people who write things to respect my time. I'm not asking them to disclose every tool they use, I'm asking them to not waste our time with crappy irritating Claude writeups. Whether it's explictly specified or not, we know.
Frankly I struggle to believe that people don't really mind if someone else speaks for them in their own voice, just because they're too fucking lazy to speak for themselves anymore. We've had competent GenAI for like a year or two, at this rate people are going to forget their potty training in another few months.
> Substance is what matters
Substance matters, which is not great for LLMs, because they put out text that has far more fluff than substance. What, however, is far worse for LLMs, is the fact that kick and scream and cry all you want, but: style and presentation matters, too.
It is absolutely true that if you just dropped a very brief blurb that all AMD CPUs from a certain generation can be pwned it would have a decent chance to hit the HN frontpage just out of sheer interestingness. That is not because the style and presentation doesn't matter, it's just that the substance is significant in spite of the bad presentation and style.
And absolutely, we can easily forgive someone for simply not being very good at the presentation and style part, certainly I'm not really an expert at it. But this author has released plenty of great hits before, so I damn well know they can. It's a serious disappointment to see them downgrade to irritating, grating Claude garbage output.
He should also be able to fuse away this access forever, to be fair. But out of the box, when I get a new laptop, I should be able to read and write every byte of DRAM.
> Run `platform_check` first and do not use `SKITTER_FORCE=1` casually. Start with the read-only `dram_state` and `dram_carveouts`, then `dram_dump --dry-run`. Avoid `dram_poke` until maps have been freshly collected and calibrated. Do not bypass fingerprint checks, calibration, fencing, or verification.
Claude's (apparently externally-mandated?) lobotomization continues to be concerning. :-/