This is a linkpost for https://canaryinstitute.substack.com/p/send-me-the-prompt
New Comment
truly are someone who speaks English as a second language, and you wrote the post in your native tongue, only using Claude to translate it.
Would this result in Claudeisms? In my experience, you can simply prompt an LLM to translate only while lifting the original style across languages.
I've read that Claudish persists across languages in normal discussion (e.g. on Xitter someone claimed the tics are all the same nature in German). I'm not sure if it would be noticeable to a human for translation only.
And while I'm fluent in Spanish (CEFR ~B2 although never formally tested), my grasp of nuance isn't good enough to pick up tics along those lines, so I'll have to defer to those who are at that C1/C2 level. By all means elaborate on what you've observed, though!
However my understanding is that per the EU reg, the translation would still be watermarked, and thus detectable in many of the scenarios I think matter (e.g. grant funding).
I think you could test this by having one Claude instance translate an English post to a different language (e.g. German), then have a second Claude instance translate the German post back to English.
In the spirit of Empiricism, I did that with this post: I asked Claude to translate it to German, then I asked a separate Claude instance to translate it back to English.
German version
Schick mir den Prompt
Wenn du willst, dass die Leute das, was du postest, tatsächlich lesen, solltest du das kostspielige Signal senden, das zeigt, dass es kein Slop ist.
Crosspost von canaryinstitute.ai/blog/send-me-the-prompt.
Ich habe gerade einen Xeet auf Xitter [1] gesehen, der den effizientesten Kompressionsalgorithmus erklärt, den ich je gesehen habe:
Wie vieles an gutem Humor enthält er ein Körnchen Wahrheit, nämlich: du hast Claude, und ich habe Claude, also sind die einzigen neuen Bits die in dem Prompt, den du benutzt hast. Außerdem haben KIs den (wohlverdienten) Ruf, schreckliche Schreiber mit einer miserablen Theory of Mind zu sein, und es kostet große Mühe, sich durch all die Berge von Slop zu schürfen und am Ende auch nur ein Nugget an Erkenntnis in der Hand zu halten. Und wenn es dir nicht wichtig genug war, diese Erkenntnis so aufzupolieren, dass sie schön ist, warum sollte es mir wichtig genug sein, mich durch diesen Slop zu wühlen?
Um das klarzustellen: Das ist keine moralische Behauptung. Ich sage nicht „du hast eine KI für dich schreiben lassen, also bist du ein schlechter Mensch“ [2] . Was ich tatsächlich sage, ist: „Jetzt, wo es 20-mal einfacher ist, etwas zu schreiben, schreibt JEDER 20-mal mehr … aber ich habe immer noch genauso viel Aufmerksamkeit wie vorher.“ Ich werde darauf keine tausend Wörter verschwenden [3] , wenn der Plot doch alles sagt:
Nur zu, benutz KI zum Entwerfen, Erkunden und Korrekturlesen. Ich kann dir gar nicht sagen, wie sehr es meine Schreibblockade löst, „Claude, mach einen Entwurf“ zu sagen … und sobald ich ihn sehe, verstehe ich sofort, was mein Essay nicht sagen soll, und kann dann schreiben, was er sagen soll [4] . Aber wenn dein Text bei Pangram als „100 % KI“ angezeigt wird, dann ist mein Prior, dass du zehn Sekunden gebraucht hast, um das zu generieren, und ich Minuten bräuchte, um zu bestätigen, dass es Müll ist.
Das gilt wahrscheinlich doppelt für formularnahe Bewerbungen und Anträge wie Stellenbewerbungen, Förderanträge, Fellowships usw. Jetzt, wo es so viel einfacher ist, eine Bewerbung einzureichen, suchen Organisationen nach einfacheren Wegen, Bewerbungen auszusieben. Das ist zu einem so drängenden Problem geworden, dass mein Team bei einem kürzlichen Hackathon für FAI/IFP ein SlopChecker-Repo [5] erstellt hat (wir haben übrigens „Best In Show“ gewonnen!). Automatisch zu prüfen „wurde das mit KI geschrieben?“ ist ein sehr leicht anzuwendender Test, und während manche Organisationen rechtlich verpflichtet sind, diese Bewerbungen trotzdem zu berücksichtigen, werden viele Organisationen deine Bewerbung einfach sofort wegwerfen.
Nun bist du, geneigter Leser, vielleicht wirklich diese eine Person von zwanzig, die ein langes, nuanciertes Gespräch geführt und alle möglichen Ideen erkundet hat, bevor sie die meisten davon verworfen und nur die wertvollen behalten hat. Aber ich kann das nicht erkennen, ohne die Mühe zu investieren, wofür ich, sorry, keine Zeit habe.
Oder vielleicht bist du, geneigter Leser, wirklich jemand, der Englisch als Zweitsprache spricht, und du hast den Post in deiner Muttersprache geschrieben und Claude nur zum Übersetzen benutzt. Aber wie zuvor kann ich das nicht erkennen, bis ich die Mühe investiert habe, und dafür hat doch kein Mensch Zeit [6] .
Bei Leuten aus diesen beiden Kategorien wäre ich vielleicht geneigt, zu lesen, was ihr gepostet habt, wenn ihr einen Link zum ursprünglichen Gespräch teilt [7] . Aber für den Rest von euch: Spart uns beiden etwas Zeit und schickt mir einfach den Prompt.
Verwandte Posts
- Don't Let LLMs Write For You, von JustisMills. Das befürworte ich, und ich halte es zugleich für eine gute Art, KI zu BENUTZEN, inklusive Verfeinerung durch Widerspruch. Ähnlich wie ich es benutze.
- Posting Some Prompts, von Arjun Panickssery. Arjun hat, glaube ich, wirklich nützliche Prompts? Aber ich kann es trotzdem nicht ertragen, die Berichte zu lesen; sie waren wahrscheinlich gut dafür, dass ER etwas versteht, aber da das, was ER wissen wollte, und das, was ich wissen will, nicht dasselbe sind, ist das Signal-Rausch-Verhältnis nicht gut genug.
- A Thoughtful Defense of AI Writing, von Michael Samoilov. Dem widerspreche ich wegen des Verhältnisses von „produzierten Dingen“ zu „meiner Zeit“.
- And Yet, Defend your Thoughts from AI Writing, von Michael Samoilov. Sein Folgestück befürworte ich.
English version
Send me the Prompt
If you want people to actually read what you post, you should send the costly signal that shows it isn't slop.
Crossposted from canaryinstitute.ai/blog/send-me-the-prompt.
I just saw a Xeet on Xitter [1:1] explaining the most efficient compression algorithm I've ever seen:
Like a lot of good humor, it contains a grain of truth, namely: you have Claude, and I have Claude, so the only new bits are the ones in the prompt you used. On top of that, AIs have a (well-earned) reputation for being terrible writers with a lousy theory of mind, and it takes real effort to pan through all the mountains of slop and come away holding even a single nugget of insight. And if you didn't care enough to polish that insight until it was beautiful, why should I care enough to wade through the slop?
To be clear: this isn't a moral claim. I'm not saying "you had an AI write for you, so you're a bad person" [2:1] . What I'm actually saying is: "Now that it's 20x easier to write something, EVERYONE writes 20x more … but I still have exactly as much attention as before." I'm not going to waste a thousand words on this [3:1] when the plot says it all:
Go ahead and use AI for drafting, exploring, and proofreading. I can't tell you how much saying "Claude, write a draft" dissolves my writer's block … and the moment I see it, I immediately understand what my essay shouldn't say, and can then write what it should [4:1] . But if your text shows up as "100% AI" on Pangram, then my prior is that it took you ten seconds to generate and would take me minutes to confirm it's garbage.
This probably goes double for form-like applications like job applications, grant proposals, fellowships, and so on. Now that it's so much easier to submit an application, organizations are looking for easier ways to screen applications out. This has become such a pressing problem that at a recent FAI/IFP hackathon my team built a SlopChecker repo [5:1] (we won "Best In Show," by the way!). Automatically checking "was this written with AI?" is a very easy test to apply, and while some organizations are legally required to consider these applications anyway, many organizations will just toss your application straight in the bin.
Now, dear reader, maybe you really are that one person in twenty who had a long, nuanced conversation and explored all kinds of ideas before discarding most of them and keeping only the valuable ones. But I can't tell that without putting in the effort, which, sorry, I don't have time for.
Or maybe, dear reader, you really are someone who speaks English as a second language, and you wrote the post in your native language and only used Claude to translate it. But as before, I can't tell until I've put in the effort, and nobody has time for that [6:1] .
For people in these two categories, I might be inclined to read what you posted if you share a link to the original conversation [7:1] . But for the rest of you: save us both some time and just send me the prompt.
Related Posts
- Don't Let LLMs Write For You, by JustisMills. I endorse this, and I also think it's a good way to USE AI, including refinement through disagreement. Similar to how I use it.
- Posting Some Prompts, by Arjun Panickssery. Arjun has genuinely useful prompts, I think? But I still can't stand to read the reports; they were probably good for HIM to understand something, but since what HE wanted to know and what I want to know aren't the same, the signal-to-noise ratio isn't good enough.
- A Thoughtful Defense of AI Writing, by Michael Samoilov. I disagree with this one because of the ratio of "things produced" to "my time."
- And Yet, Defend your Thoughts from AI Writing, by Michael Samoilov. His follow-up I endorse. %%% /llm-output
Interestingly, the round-trip-translated version smells like AI to me but Pangram judges it to be 100% human-written.
For a great deal more than 1,000 words, see the post I stole the inspiration for the figure from, at https://12gramsofcarbon.com/p/meditations-on-ai-and-the-future. ↩︎ ↩︎
Though I'll freely admit that Claude is my errand boy for fact-checks, citations, and often footnotes too. ↩︎ ↩︎
You can try it yourself at https://slop-checker.com; the source code is at https://github.com/nawagner/SlopChecker. ↩︎ ↩︎
Though you really do have my sympathy; I learned a second language as an adult and know how frustrating it is not to be able to be the clever, witty person you are in your native tongue. ↩︎ ↩︎
In the interest of transparency, here's mine: https://claude.ai/share/6a2ffb8b-b06f-498d-873e-0a4a78b78de3. ↩︎ ↩︎
thanks for testing! I notice that the re-un-translated version is almost exactly what I wrote originally, which on the one hand is great. OTOH, I didn't find LLM writing as annoying as everyone else seemed to- but apparently my writing has always been like that. I now have to go explicitly remove dashes just so no one thinks it's slop. :: le sigh ::
Interesting. That's obvious but I didn't think about it. The solution would be to be transparent on the fact that this is a translation and to provide the initial text in appendice.
I like the idea of having "AI usage note: originally written in Spanish, translated by Opus 4.7 to English, see conversation at [LINK]", that's teh kind of thing that could be automated as part of a pipeline system such as SlopChecker.
I do genuinely like the democratization aspect, but some communities I'm in, folks haven't done the extrapolation to "what if everyone does this?", and this post is partly well-intentioned advice for them.
If you want people to actually read what you are posting, you should send the costly signal showing it isn't slop.
Crossposted from canaryinstitute.ai/blog/send-me-the-prompt.
I just saw a Xeet on Xitter [1] that explained the most efficient compression algorithm I've ever seen:
Like a lot of good humor, it contains an element of truth, to wit: you have Claude, and I have Claude, so the only novel bits are those in the prompt you used. In addition, AIs have a (well-deserved) reputation as horrible writers, with abysmal theory of mind, and it requires great effort to pan through all the piles of slop and emerge with any nugget of insight. And if you didn't care enough to clean up that insight so that it's beautiful, why should I care enough to dig through that slop?
To be clear, this isn't a moral claim: I'm not saying "you used AI to write for you, therefore you're a bad person" [2] . I'm actually saying, "Now that it's 20x easier to write something, EVERYONE is writing 20x more... but I still have the same amount of attention". I won't waste a thousand words on this [3] , when the plot says it all:
By all means, use AI for drafting, exploring, and proofreading. I can't tell you how much it reduces my writer's block to say "Claude, make a draft"... and then upon seeing it, I can instantly understand what I don't want my essay to say, and can then write what I do want [4] . But if your piece is showing as "100% AI" on Pangram, then my prior belief is that it took you ten seconds to generate this, and would take me minutes to confirm that it's garbage.
This is probably doubly-true for form-adjacent applications such as job postings, grants, fellowships, etc. Now that it's so much easier to submit an application, orgs are looking for easier ways to filter out applications. This has become such a pressing issue that my team created a SlopChecker [5] repo at a recent hackathon for FAI/IFP (we took "Best In Show", incidentally!). Automatically checking "was this written with AI?" is a very easy test to apply, and while some orgs have legal requirements to consider those applications anyway, plenty of orgs will just throw your application out immediately.
Now maybe you, dear reader, truly are that 1 person in 20 who had a long, nuanced conversation and explored all kinds of ideas before discarding most of them, only keeping the valuable ones. But I won't be able to tell that without investing the effort, which, sorry, I don't have time to do.
Or maybe you, dear reader, truly are someone who speaks English as a second language, and you wrote the post in your native tongue, only using Claude to translate it. But as before, I can't tell until I've put the effort in, and ain't nobody got time for that [6] .
For folks in those two categories, I might be inclined to read what you posted if you share a link to the original conversation [7] . But for the rest of y'all: save us both some time, and just send me the prompt.
Related posts
https://x.com/QuinnyPig/status/2092088373687328791 ↩︎
No, I'm saying you're a bad person for OTHER reasons. ↩︎
For much more than 1000 words, see the post whence I stole inspiration for the figure at https://12gramsofcarbon.com/p/meditations-on-ai-and-the-future. ↩︎
Although I freely admit that Claude is my gofer for fact-checking, citations, and often footnotes. ↩︎
You try it yourself at https://slop-checker.com; the source code is at https://github.com/nawagner/SlopChecker. ↩︎
Although you truly have my sympathy; I learned a second language as an adult and know how frustrating it is when you can't be the clever, witty person you are in your mother tongue. ↩︎
In the spirit of transparency, here's mine: https://claude.ai/share/6a2ffb8b-b06f-498d-873e-0a4a78b78de3. ↩︎