Secrets Don’t Belong in Config

Links
1

This post was assisted by Corona, Opus 4.8 and Fable 5.

4 Likes
2

Undisclosed LLM slop, yikes. (I’ll bet it’s claude opus 4.8 specifically.)

11 Likes
3

What about this is slop, exactly?

1 Like
4

If you know you know, people don’t talk like this lmao

Screenshot 2026-07-20 at 22.26.19
Screenshot 2026-07-20 at 22.26.192358×314 63.6 KB

Current opus likes generating text with stuff like “real [noun]”; a person wouldn’t insist this hard that something is real and not (cough) hallucinated.

Even the table headers are a bit…silly.

2 Likes
5

Hm, I’ll refrain from reading until @domenkozar responds to this accusation. I do have to say it reads a tad weird. I’m not familiar with the peculiarities of individual models, so I cannot say which one this could be.

6

This audit was assisted by Claude Opus 4.8.

But what do I care if he used an LLM or hacked together some perl scripts?
That survey was made in 2017 and it had to be updated one way or another.

4 Likes
7

Because we have a policy about LLM disclosure and some people care. Just because you don’t, doesn’t mean LLM use isn’t a valid concern. Me personally, I want to read human thoughts, not probabilistic LLM output.

10 Likes
8

Not that I disagree, but that policy effectively doesn’t apply to the links category. The post is technically not on discourse, and it’s not a “project” so disclosure isn’t required.

I should probably start being more aggressive with throwing that policy at people. The announcements category is still full of people who only remember to add a comment like half a discussion in, and say nothing in their posts.

10 Likes
9

can we please not start this? i think there was enough context in the situation for your claims/accusations to absolutely be justifiable but if we start policing the way people talk… oh boy

my concern: i have been a long time emoji enjoyer but recently feel like i need to be careful lest people accuse me of writing forum posts with the aid of an LLM :expressionless_face:

13 Likes
10

I do not care whether or not the article is LLM slop but it’s worth putting it on the record that Domen is historically known for using LLMs. Do with that what you will.

4 Likes
11

I agree, i also like emojis. Though if people could be trusted with disclosing LLM use (which hasn’t happened here) there would no need to try to guess if something was done by an LLM.

I do not appreciate being lied to, that’s all. Fundamentally, if you use an LLM and disclose it, I will not click on the post at all. However if you use an LLM and don’t disclose it, your posts comment section is gonna get people trying to make you do so. There is a clear way out of this conundrum (proper disclosure).

8 Likes
12

and i think that is fine, same with @waffle8946 and @NotAShelf pointing out the obvious facts about the situation and how this violates a policy - i just don’t want this to be based on “look how they spoke!” when there is so much more evidence to point to…

keep in mind that LLMs are provided a large body of data to scrape (the internet)… so if an LLM uses emoji or speaks a certain way, it is because a real person did that at some point too

1 Like
13

For the sake of being pedantic I’d like to add that LLM does not copy individual people (as implied by “a real person did that…”) It learns statistical relationships over enormous corpora. A particular phrasing may never have been written exactly as the model produces it. It’s just the synthesis of what the model has been trained on—which is why I think all LLM text goes into the same uncanny valley area.

P.S. emojis are usually a post-training thing now

1 Like
14

@rhendric can you please move all this off-topic discussion into a separate thread?

3 Likes
15

god forbid a white boy is pedantic

16

Just to conclude this discussion, yes the post is in fact LLM “assisted”. See for yourself: https://github.com/cachix/secretspec/pull/179

3 Likes
17

Good morning!

Didn’t know about these rules, but I’ve added the disclaimer (clearly the blog post and commit is public).

4 Likes
18

I can’t help you with your personal anxieties, I just simply recognize patterns having been forced to work with various LLMs. I can even to some degree tell them apart based on their text generation patterns :face_with_diagonal_mouth: Same way I could tell a if it’s a German person typing in English based on their word choice, capitalization of nouns, etc. It’s pretty simple to tell a person apart from an LLM.

I’m not really going to spend more than 10 seconds reading slop at this point (long enough to notice that it is), if I wanted to read chatbot output I would go start the chat myself.

By the way, the emoji LLM thing is more along the lines of starting bullets with them, especially paired with selling points, stuff like

  • :high_voltage: 50% faster than v1
  • :low_battery: Uses 24% less power
  • :input_latin_letters: Simplified queries

and so on. I have seen people write like this, but it usually accompanies pushy sales tactics, so it’s offputting anyway :laughing:

4 Likes
19

it seems you’re not accepting direct messages, so i’ll call this out here - please don’t personally attack me because i asked for you to not start witch hunts based on “trust me, bro, i know”

there was plenty of other evidence that you were aware of to call out the original post, you didn’t need to go with the weakest possible point that has no empirical backing whatsoever

5 Likes
20

Again I explained myself pretty well, misrepresenting what I said and extracting out of context intentionally to shadowbox is a cheap, dishonest tactic to attack me when I wasn’t even addressing you originally. We don’t know each other, you don’t have to trust me, I’m just stating my opinions on a chatbot. It’s not the way people talk, because it wasn’t a person talking :wink: feel free to reference the edit on the original post. You made it about yourself and a hypothetical (read nonexistent) person who would talk like a chatbot, I … can’t help you with that.

1 Like