We found an undocumented bug in the Apollo 11 guidance computer code
juxt.pro
juxt.pro
First: this is indeed a real bug in the AGC software. However, it did not go unnoticed for the whole program. It was discovered during level 3 testing of SATANCHE, and late development branch of the Command Module software COMANCHE. It was assigned anomaly number L-1D-02, and was fixed between Apollo 14 and 15. There are two known surviving copies of the L-1D-02 anomaly report:
* https://www.ibiblio.org/apollo/Documents/contents_of_luminar...
* https://www.ibiblio.org/apollo/Documents/contents_of_luminar...
The fix described in the article is partially complete, but as noted in the anomaly report there's a little bit more to it. Rather than just adding the two instructions to zero LGYRO, they restructured the code a bit and also cause it to wake up pending jobs. You can compare the relevant sections of the Apollo 14 and Apollo 15 LM software here:
* Apollo 14: https://github.com/virtualagc/virtualagc/blob/master/Luminar...
* Apollo 15: https://github.com/virtualagc/virtualagc/blob/master/Luminar...
The bug would not manifest silently in the way described in the article. For starters, LGYRO is also zeroed in STARTSB2, which is executed via GOPROG2 on any major program change: https://github.com/virtualagc/virtualagc/blob/master/Luminar...
This means that changing from any program to any other program would immediately resolve the issue. This is almost certainly a large part of why it took them so long to notice. Hitting BADEND while actively pulse-torquing is quite rare, and avoided by normal procedure. The scenario presented in the article can't happen since the act of starting P52 will zero LGYRO.
Moreover, in the very specific scenarios in which the bug can be triggered and remain, it results in multiple jobs stacking up attempting to torque the gyros. Eventually the computer runs out of space for new jobs -- similar to what happened on 11 -- and a 31202 (the Apollo 12+ equivalent of 1202) is triggered.
Since the issue was found before the flight of Apollo 14, a further description of how it might occur and what the recovery procedure should be was added to the Apollo 14 Program Notes: https://www.ibiblio.org/apollo/Documents/LUM159_text.pdf#pag...
Some other notes:
> Ken Shirriff has analysed it down to individual gates
I've done the bulk of the gate-level analysis. :)
> the Virtual AGC project runs the software in emulation, having confirmed the recovered source byte-for-byte against the original core rope dumps.
We've only been able to do that in very specific circumstances and only for subsections of assorted programs, but never for a full program. Most AGC software either comes from a program listing, from a core rope dump, or from reconstruction using changelogs and known memory bank checksums. We've disassembled all of the rope dumps into source files that assemble back into the same binary, but the comments and labels will be different from what was in the original listing. And to be extra clear: I've never had the opportunity to dump a module containing Apollo 11 software for either vehicle. Our sole source for both programs is a pair of printouts in the MIT Museum's collection.
> Margaret Hamilton (as “rope mother” for LUMINARY) approved the final flight programs before they were woven into core rope memory.
Jim Kernan was the rope mother for Luminary at least up through Apollo 11. Margaret was the rope mother for Comanche, the CM software, and was later promoted to lead the software division. Their positions at the time of 11 can be seen on this org chart: https://www.ibiblio.org/apollo/Documents/ApolloOrg-1969-02.p...
> Their priority scheduling saved the Apollo 11 landing when the 1202 alarms fired during descent, shedding low-priority tasks under load exactly as designed.
This is a huge topic on its own, but the AGC software was not designed to shed low-priority jobs. Ironically, the lowest priority job during the landing was the landing guidance itself, with high-priority jobs being reserved for things that needed quick response like antenna movements or display updates. If the computer were to shed the lowest-priority jobs, it would shed the landing guidance. This memo contains a list of all jobs active during the landing and their priorities: https://www.ibiblio.org/apollo/Documents/CherryApollo11Exege...
> For example, the ICD for the rendezvous radar specified that two 800 Hz power supplies would be frequency-locked but said nothing about phase synchronisation. The resulting phase drift made the antenna appear to dither, generating roughly 6,400 spurious interrupts per second per angle and consuming roughly 13% of the computer’s capacity during Apollo 11’s descent. This was the underlying cause of the 1202 alarms.
The frequency-lock prevents phase drift, so the phase is essentially fixed once the power supplies are up. Ironically, however, the bigger issue is that one reference was 28V while the other was 15V. Initial testing on actual Apollo hardware suggests that at least for Apollo 11, this voltage difference was the key contributor rather than the phase difference: https://www.youtube.com/watch?v=dT33c70EIYk
(I mention this so more people can know the list exists, and hopefully email us more nominations)
The anomaly report pdf is quite long. The relevant pages for the bug report are 51 and 52: https://www.ibiblio.org/apollo/Documents/contents_of_luminar...
Your experimental work on the voltage difference is fascinating, I appreciate you sharing the link to the demo. There's something about seeing results come off real hardware that you just don't get from computer simulation. Watching your setup brought back some of the excitement I felt driving my Makerbot via a stylus input 15 years ago (probably the last time I seriously engaged with hardware) [1]. Thanks!
One of the more interesting things they have been working on, is a potential re-interpretation of the infamous 1202 alarm. It is, as of current writing, popularly described as something related to nonsensical readings of a sensor which could (and were) safely ignored in the actual moon landing. However, if I remember correctly, some of their investigation revealed that actually there were many conditions which would cause that error to have been extremely critical and would've likely doomed the astronauts. It is super fascinating.
I don't think the numbers you quoted are outliers, though. The F-100 lost ~900 out of 2,300. The F-106 lost ~120/342. That's a pretty big list of planes with a 1/5-1/3 loss rate.
There was a fuel tank mounted between the engine and cockpit so if it took enough of a hit to puncture right through (not hard, in practice) the failure mode was that the cockpit was now full of a 350mph jet of burning petrol.
Still, it did the job.
Death being a layer of aluminum away changes your mind.
For complex reasons, available CPU time during landing was lower than expected (it was stolen by radar pointing peripheral). This caused regularly scheduled job to spawn before previous instance finished. As such, this caused two effects: job instances were suspended before finishing by new instances in the middle of the routine, and that pilling up of the old instances eventually exhausted resources and caused kernel to panic and reboot. Rebooting during landing sounds scary, but that actually was fine: such critical tasks were specifically designed to automatically restart from previously saved checkpoint data in the memory.
What was more dangerous, was the suspended tasks before restarts occured. First, it meant routine wasn't executing to the end, which in actual flight caused blanked displays (as updating the display was the last thing routine was doing). Any more CPU time stolen, and it could be interrupted even earlier, eg. before it sends the engine commands.
Another issue is that in case of fluctuating load, new instances could actually begin running to the end, and then previously suspended job instance could be resumed, potentially sending the stale data to the displays and engine.
And finally, while each job instance had it own core and VAC set properly managed by the kernel (think of it as modern kernel switching between task stacks), that particular routine wasn't designed to be reentrant. So it was using various global variables ("erasables") for its own purpose, that when interrupted in unluckly place might have caused very bad behavior.
How likely all of above is to occur, depends on the exact profile of fluctuating load caused by the confused radar peripheral. I guess that's why Mike Stewart is trying to replicate these issues with real CDU.
TC BANKCALL # TEMPORARY, I HOPE HOPE HOPE
CADR STOPRATE # TEMPORARY, I HOPE HOPE HOPE
TC DOWNFLAG # PERMIT X-AXIS OVERRIDE
CADR is an AGC assembly directive defining a "complete address" including a memory bank, in this case a subroutine to be called by the preceding BANKCALL (TC = transfer control, i.e., store return address and jump to subroutine), which switches to the memory bank specified in the CADR before jumping to the address specified in the CADR.
For a brief explanation of AGC subroutine calls, see [1].
CAR and CDR in Lisp come from the original implementation on the IBM 704, where pointers to the two components of a cons cell were stored as the (C)ontents of the (A)ddress and (D)ecrement fields of a (R)egister (memory word).
(CADR x) is just shorthand for (CAR (CDR x)), i.e., a function that returns the second element of a list (assuming x is a well-formed list).
[1] https://epizodsspace.airbase.ru/bibl/inostr-yazyki/American_...
“I tried to keep to Shelley’s unusual (and non-standard) rhyme scheme for the sonnet, but I departed from it in the second-to-last line for poetic reasons. For a language which excels in stealing words from other cultures, English has an appalling lack of rhymes.”
Perhaps with deeper analysis, and a few choice new words this issue could be remedied.
Although that’s a paradoxically tedious engineering solution to improve a languages beauty.
From another angle how comes other languages are more poetic, are they older and have had more time to evolve to be more poetic? Or were the speakers who wrought the language just more poetic.
One of AI’s strengths is definitely exploration, f.e. in finding bugs, but it still has a high false positive rate. Depending on context that matters or it wont.
Also one has to be aware that there are a lot of bugs that AI won’t find but humans would
I don’t have the expertise to verify this bug actually happened, but I’m curious.
But, I do think their explanation of the lock acquisition and the failure scenario is quite clear and compelling.
The intro says “We used Claude and Allium”. Allium looks like a tool they’ve built for Claude.
So the article is about how they used their AI tooling and workflow to find the bug.
They used their AI tool to extract the rules for the Apollo guidance system based on the source code.
Then they used Claude to check if all paths followed those rules.
Could the "AI native language" they used be Apache Drools? The "when" syntax reminded me of it...
https://kie.apache.org/docs/10.0.x/drools/drools/language-re...
(Apache Drools is an open source rule language and interpreter to declaratively formulate and execute rule-based specifications; it easily integrates with Java code.)
It's not even clear you read the article
> We found this defect by distilling a behavioural specification of the IMU subsystem using Allium, an AI-native behavioural specification language.
A.k.a. as fabricating. No wonder they chose to use "AI".
2nd paragraph starts with: "We used Claude and Allium"
And later on: "With that obligation written down, Claude traced every path that runs after gyros_busy is set to true"
Please, keep your offensive comments to yourself when a clarifying comment might have sufficed.
> We found this defect by distilling a behavioural specification of the IMU subsystem using Allium, an AI-native behavioural specification language.
https://github.com/juxt/Apollo-11/tree/master/specs
have many thousands of lines of code in it.
Anyways, it seems it would take a dedicated professional serious work to understand if this bug is real. And considering this looks like an Ad for their business, I would be skeptical.
A guarded switch, no less.
But personally I'm trying to be more generous about this sort of thing: it is very very difficult to explain subtle bugs like this to non-technical people. If you don't give them a story for how it can actually happen, they tend to just assume it's not real. But then when you tell a nice story, all us dry aged curmudgeons tut tut about how irreverent and over the top it is :)
Finding the middle ground between a dry technical analysis and dramatization can be really hard when your audience is the entire internet.
The repro runs on my computer, that's positive.
However, Phase 5 (deadlock demonstration) is entirely faked. The script just prints what it _thinks_ would happen. It doesn't actually use the emulator to prove that its thinking is right. Classic Claude being lazy (and the vibe coder not verifying).
I've vibe coded a fix so that the demonstration is actually done properly on the emulator. And also added verification that the 2 line patch actually fixes the bug: https://github.com/juxt/agc-lgyro-lock-leak-bug/pull/1
I see this a lot in AI slop, which I mostly get exposed to in the form of shitty pull requests.
You know when you're trying to explain Test-Driven Development to people and you want to explain how you write the simplest thing that passes the test and then improve the test, right? So you say "I want a routine that adds VAT onto a price, so I write a test that says £20+VAT is £24, and the simplest thing that can pass that test is just returning 24". Now you know and I know that the routine and its test will break if you feed it any value except £20, but we've proved we can write a routine and its test, and now we can make it more general.
Or maybe we don't care and we slap a big TODO: make this actually work on there because we don't need it to work properly now, we've got other things to do first, and every price coming up as £20+VAT is a useful indicator that we still have to make other bits work. It doesn't matter.
The problem is that AI slop code "generators" will just stop at that point and go "THERE LOOK IT'S DONE AND IT'S PERFECT!" and the people who believe in the usefulness of AI will just ship it.
If anything, if you try to cram a ton of complexity into a few kb of memory, the likelihood of introducing bugs becomes very high.
Rust specifically does not forbid deadlocks, including deadlocks caused by resource leaks. There are many ways in safe Rust to deliberately leak memory - either by creating reference count cycles, or the explicit .leak() methods on various memory-allocating structures in std. It's also not entirely useless to do this - if you want an &'static from heap memory, Box.leak() does exactly that.
Now, that being said, actually writing code to hold a LockGuard forever is difficult, but that's mainly because the Rust type system is incomplete in ways that primarily inconvenience programmers but don't compromise the safety or meaning of programs. The borrow checker runs separately from type checking, so there's no way to represent a type that both owns and holds a lock at the same time. Only stacks and async types, both generated by compiler magic, can own a LockGuard. You would have to spawn a thread and have it hold the lock and loop indefinitely[0].
[0] Panicking in the thread does not deadlock the lock. Rust's std locks are designed to mark themselves as poisoned if a LockGuard is unwound by a panic, and any attempt to lock them will yield an error instead of deadlocking. You can, of course, clear the poison condition in safe Rust if you are willing to recover from potentially inconsistent data half-written by a panicked thread. Most people just unwrap the lock error, though.
Oh dear. I strongly suggest this author look specification up in a dictionary.
Nor did they make any mistakes when they described how they produced a specification, (and indeed, that it is a specification) despite your insinuation otherwise, for a similar reason.
Maybe instead of pointing towards dictionaries, stop pretending that you lack reading comprehension, and get off of your high horse please.
[1] In the repo, the "reproduce" is just a bunch of print statements about what would happen, the bug isn't actually triggered: https://github.com/juxt/agc-lgyro-lock-leak-bug/blob/c378438...
It seems the difference between this and conventional specification languages is that Allium's specs are in natural language, and enforcement is by LLM. This places it in a middle ground between unstructured plan files, and formal specification languages. I can see this as a low friction way to improve code quality.
>The Apollo Guidance Computer (AGC) is one of the most scrutinised codebases in history.
What? AGC programs were developed by relatively small team and pretty much left alone since then. Architecture is rather quirky when viewed with modern sensibilities. There's not much people that are familiar with it. Compare it to widely used software like libcurl or sqlite. Or perhaps to Super Mario Bros, which was extensively analyzed for competitive speedruns reasons. Surely that dwarfs amount of knowledge about Apollo code.
>2K of erasable RAM and a 1MHz clock. The AGC’s programs were stored in 74KB of core rope
How about picking a unit and staying with it? AGC has 2K words of RAM, where each word has 15 bits of usable data (physically it's 16 bits, but one bit is used for parity). Maximum amount of ROM that could be installed is 36K words. (but they switch to KB, which is not only inconsistent with previous sentence but the number is also wrong! It's 72 KiB, 73.728 KB or 67.5 KiB, 69.12 KB depending whether you include parity or not) (maximum of 64K ROM words could be addressed by architecture design, but isn't available in any real hardware)
And yes, there is 1.024 MHz clock in the system, which is revelant for peripherals, but you probably want to know how fast it executed instructions. One memory cycle takes 11.71875 μs (85 1/3 kHz), and most instructions take 2 such cycles (one for operation, second for fetching next instruction) (each memory cycle is long enough for read from ROM, or read and write to RAM. ROM speed was the limiting factor, by standard of core memories it wasn't particularly fast. AGS backup computer used core for both RAM and ROM and had memory cycle time of 5 μs) (in case you are confused, "core memory" and "core rope memory" refers to quite different things!).
If you think I'm nitpicking, try writing an emulator and wondering why you have to sift through all that slop. You could give the correct numbers, you know?
>“My secret terror for the last six months has been leaving them on the Moon and returning to Earth alone”, Collins later wrote of the rendezvous. A dead gyro system behind the Moon, with Armstrong and Aldrin on the surface waiting for a rendezvous burn that depends on a platform he can no longer align, is exactly that scenario. A hard reset would have cleared it. But the 1202 alarms during the lunar descent had been stressful enough with Mission Control on the line and Steve Bales making a snap abort-or-continue call. Behind the Moon, alone, with a computer that was accepting commands and doing nothing, Collins would have had to make that call by himself.
You know what an orbit is? That it goes around? That you could just wait for a while and speak with Mission Control? What even is this scenario? That your guidance system failed, and you for some inexplicable reason are considiering immediately leaving back for Earth right now leaving your pals behind? (with a manual burn, I guess, since guidance is dead?) You just wait for contact with Houston and tell them what happened. They pore over the program listings and find the bug. They radio you back appropiate VERB and NOUN commands for poking right values into memory. The End. And besides, spacecraft can be tracked and orbit determined from Earth, so even if the PGNCS did fail completely LM would just get necessary orbit information from Mission Control. (also in case guidance fails in either LM or CM, either one can have active role during rendezvous. And LM have extra backup system, the previously mentioned AGS)
The whole thing of "we found a minor deadlock bug in AGC program, what a shock!" is bizzare. It's not a small program. If you have any experience with software, of course you know it has bugs! They iterated on the software, releasing new software for most missions, adding new features, and, fixing bugs they found. What a concept!
Ughhhh… I know this is probably legit here, but reading these words make me lose interest sooo fast these days…
therefor decided not gonne use any llm for blogging again and even tho it takes alot more time without (im not a very motivated writer) i prefer to release something that i did rather some llm stuff that i wouldnt read myself.
Another one: "Two instructions are missing: [...] Four bytes."
One more: "The defensive coding hid the problem, but it didn’t eliminate it."
This insistence that certain stylistics patterns are "tell-tale" signs that an article was written by AI makes no sense, particularly when you consider that whatever stylistic ticks an LLM may possess are a result of it being trained on human writing.
For what it’s worth, Pangram reports that Marcus’ article is 100% LLM-written: https://www.pangram.com/history/640288b9-e16b-4f76-a730-8000...
73% judged GPT 4.5 (edit: had incorrectly said 4o before)to be the human.
https://arxiv.org/abs/2503.23674
Not only are people bad at judging this, but are directionally wrong.
> Our experiments show that annotators who frequently use LLMs for writing tasks excel at detecting AI-generated text, even without any specialized training or feedback. In fact, the majority vote among five such “expert” annotators misclassifies only 1 of 300 articles, significantly outperforming most commercial and open-source detectors we evaluated even in the presence of evasion tactics like paraphrasing and humanization.
Even though they are perfect for usage in writing down thoughts and notes.
“An em dash… they’re a witch!”… “it’s not just X, it’s Y… they’re a witch!”
that's a strawman alright; all the comments complaining how they can't use their writing style without being ganged up on are positive karma from my angle, so I'm not sure the "positive social reactions" are really aligned with your imagination. Or does it only count when it aligns with your persecution complex?
My hunch that this is substantially LLM-generated is based on more than that.
In my head it's like a Bayesian classifier, you look at all the sentences and judge whether each is more or less likely to be LLM vs human generated. Then you add prior information like that the author did the research using Claude - which increases the likelihood that they also use Claude for writing.
Maybe your detector just isn't so sensitive (yet) or maybe I'm wrong but I have pretty high confidence at least 10% of sentences were LLM-generated.
Yes, the stylistic patterns exist in human speech but RLHF has increased their frequency. Also, LLM writing has a certain monotonicity that human writing often lacks. Which is not surprising: the machine generates more or less the most likely text in an algorithmic manner. Humans don't. They wrote a few sentences, then get a coffee, sleep, write a few more. That creates more variety than an LLM can.
Fun exercise: https://en.wikipedia.org/wiki/Wikipedia:AI_or_not_quiz
Someone probably expended a lot of time and effort planning, thinking about, and writing an interesting article, and then you stroll by and casually accuse them of being a bone idle cheat, with no supporting evidence other than your "sensitive detector" and a bunch of hand-wavy nonsense that adds up to naught.
I also hate this style of plastic, pre-digested prose. Its soulless and uninteresting. Maybe I've just read too much AI slop. I associate this writing style with low quality, uninteresting junk.
If there is constant vigilance on the part of the reader as to how it was created, meaning and value become secondary, a sure path to the death of reading as a joy.
More importantly, it's an article about using Claude from a company about using Claude. I think on the balance it's very likely that they would use Claude to write their technical blog posts.
Your job doesn't require you to think or expend effort?
AI writing also tends to show these indicators over and over, consistently, over passages of text. It is very hard for humans, even if they are really familiar with AI writing, to be that consistent, and almost impossible for them to be that consistent for more than a sentence or two. Writing a long blog post by hand that is believably "AI-written" takes the amount of purposeful skill you'd need to forge an entire painting or ancient document.
The problem is that people either look for the wrong things, look for obsolete things ("delve" is dead and modern LLMs have killed it), or extrapolate things from indicators that are extremely narrow and specific.
In fact, the latter is the opposite of terseness. LLMs love to tell you what things are not way more than people do.
See https://www.blakestockton.com/dont-write-like-ai-1-101-negat...
(The irony that I started with "it's not just" isn't lost on me)
But an LLM wouldn't write "It's not just X, it's the Y and Z". No disrespect to your writing intended, but adding that extra clause adds just the slightest bit of natural slack to the flow of the sentence, whereas everything LLMs generate comes out like marketing copy that's trying to be as punchy and cloying as possible at all times.
Don't understand how these tools exist.
They found that Pangram suffers from false positives in non-prose contexts like bibliographies, outlines, formatting, etc. The article does not touch on Pangram’s false negatives.
I personally think it’s an intractable problem, but I do feel pangram gives some useful signal, albeit not reliably.
What's making it even more difficult to tell now is people who use AI a lot seem to be actively picking up some of its vocab and writing style quirks.
Not sure how I feel about the whole "LLMs learned from human texts, so now the people who helped write human texts are suddenly accused of plagiarizing LLMs" thing yet, but seems backwards so far and like a low quality criticism.
> The specification forces this question on every path through the IMU mode-switching code. A reviewer examining BADEND would see correct, complete cleanup for every resource BADEND was designed to handle.
> The specification approaches from the other direction: starting from LGYRO and asking whether any paths fail to clear it.
> *Tests verify the code as written; a behavioural specification asks what the code is for.*
However this is a blog post about using Claude for XYZ, from an AI company whose tagline is
"AI-assisted engineering that unlocks your organization's potential"
Do you really think they spent the time required to actually write a good article by hand? My guess is that they are unlocking their own organizations potential by having Claude writes the posts.
Given I'm familiar with Juxt since before, used plenty of their Clojure libraries in the past and hanged out with people from Juxt even before LLMs were a thing, yes, I do think they could have spent the time required to both research and write articles like these. Again, won't claim for sure I know how they wrote this specific article, but I'm familiar with Juxt enough to feel relatively confident they could write it.
Juxt is more of a consultancy shop than "AI company", not sure where you got that from, guess their landing page isn't 100% clear what they actually does, but they're at least prominent in the Clojure ecosystem and has been for a decade if not more.
It seems to look at sections of ~300 words. And for one section at least it has low confidence.
I tested it by getting ChatGPT to add a paragraph to one of my sister comments. Result is "100% human" when in fact it's only 75% human.
Pangram test result: https://www.pangram.com/history/1ee3ce96-6ae5-4de7-9d91-5846...
ChatGPT session where it added a paragraph that Pangram misses: https://chatgpt.com/share/69d4faff-1e18-8329-84fa-6c86fc8258...
It’s becoming a problem in schools as teachers start accusing students of cheating based on these detectors or ignore obvious signs of AI use because the detectors don’t trigger on it.
I don’t mind that they let an LLM write the text, but they should at least have edited it.
It is:
- sneering
- a shallow dismissal (please address the content)
- curmudgeonly
- a tangential annoyance
All things explicitly discouraged in the site guidelines. [1]
Downvoting is the tool for items that you think don't belong on the front page. We don't need the same comment on every single article.
You can’t downvote submissions. That’s literally not a feature of the site. You can only flag submissions, if you have more that 31 karma.
Optimistically, I guess I can call myself some sort of live-and-let-live person.
Consider that by submitting AI generated content for humans to read, the statement you're making is "I did not consider this worth my time to write, but I believe it's worth your time to read, because your time is worth less than mine". It's an inherently arrogant and unbalanced exchange.
Note: the guidelines are a living document that contain references to current AI tools.
> Consider that by submitting AI generated content for humans to read, the statement you're making is "I did not consider this worth my time to write, but I believe it's worth your time to read, because your time is worth less than mine". It's an inherently arrogant and unbalanced exchange.
This is something worth saying about a pure slop content. But the "charge" against the current item is that a reader encountered a feeling that an LLM was involved in the production of interesting content.
With enough eyeballs, all prose contains LLM tells.
We don't need to be told every time someone's personal AI detection algorithm flags. It's a cookie-banner comment: no new information for the reader, but a frustratingly predictable obstacle to scroll through.
But they won't do that, because deep down they feel shameful about it (as they should).
> Don't post generated comments or AI-edited comments. HN is for conversation between humans.
The same principle applies to submissions. If you couldn't be bothered to write it, don't ask me to read it. HN is for humans.
The short sentence construction is the most suspicious, but I actually don't see anything glaring. It normally jumps out and hits me in the face.
1. Use Short Sentences
It seems like almost every discussion has at least someone complaining about "AI slop" in either the original post or the comments.
There is some real content in the haystack, but we almost need some kind of curator to find and display it rather than a vote system where most people vote on the title alone.
There might be a market for your alternative though. Should be easy enough to build with Claude Code.
By asking AI to write the article for you, you're asserting that the subject matter is not interesting enough to be worth your time to write, so why would it be worth my time to read?
Sure, let me have a look.
He wrote 8 similarly lengthy blog posts in just 2 months: https://www.juxt.pro/blog/from-specification-to-stress-test/ https://www.juxt.pro/blog/three-paradoxes/ https://www.juxt.pro/blog/what-outlasts-the-code/ https://www.juxt.pro/blog/composition-at-a-distance/ https://www.juxt.pro/blog/new-vocabulary-for-an-old-problem/ https://www.juxt.pro/blog/softwares-second-heroic-age/ https://www.juxt.pro/blog/capability-hyperinflation/
They contain a lot of classic LLMisms:
"Implementation is the shrinking currency. Not because it’s worthless, but because supply is exploding."
His past writing was much, much less wordy: https://henrygarner.com/
Seeing comments warning about the AI content of a link is helpful to let others know what they’re getting into when they click the link.
For this article the accusations are not about slop (which will waste your time) but about tell-tell signs of AI tone. The content is interesting but you know someone has been doing heavy AI polishing, which gives articles a laborious tone and has a tendency to produce a lot of words around a smaller amount of content (in other words, you’re reading an AI expansion of someone’s smaller prompt, which contained the original info you’re interested in)
Being able to share this information is important when discussing links. I find it much more helpful than the comments that appear criticizing color schemes, font choices, or that the page doesn’t work with JavaScript disabled.
This got me thinking: what if LLMs are used to do the opposite? To condense a long prompt into a short article? That takes more work but might make the outcome more enjoyable as it contains more information.
You're fighting an uphill battle against the inherent tendency to produce more and longer text. There's also the regression to the mean problem, so you get less information (and more generic) even though the text is shorter.
Basically, it doesn't work
> Please don't post shallow dismissals, especially of other people's work. A good critical comment teaches us something.
> Please don't complain about tangential annoyances—e.g. article or website formats, name collisions, or back-button breakage. They're too common to be interesting.
>> Please don't post shallow dismissals, especially of other people's work. A good critical comment teaches us something.
>> Please don't complain about tangential annoyances—e.g. article or website formats, name collisions, or back-button breakage. They're too common to be interesting.
They don't. people. tangential.
Speaking of the HN guidelines, they also say this:
> Don't post generated comments or AI-edited comments. HN is for conversation between humans.
Who gives a crap if it was written by an LLM. Read it or don’t read it. Your choice.
If it conveys the idea and your learn something new, then it’s mission accomplished.
A Note on the Process
To be clear about what happened here: Claude wrote this article.
What a horrible world we live in where the author of great writing like this has to sit and be accused of "being AI slop" simply because they use grammar and rhetoric well.
If an LLM wrote that, then I no longer oppose LLM art.