Blogging with an LLM assistant
20 points by vbernat
20 points by vbernat
I honestly don't mind if writing is somewhat idiosyncratic. It can help build a sense of voice, and especially the sorts of edits that were illustrated in this post didn't, in my opinion, really fix a problem, so much as make the text tend toward the average.
so much as make the text tend toward the average
This makes it seem like there's no merit in tidying up your grammar, which I disagree with, though the merit is small and perhaps more a matter of self-pride. To me, "making text tend toward the average" would be more like washing your blog post in the slop machine, rather than making (to use its words) surgical edits to fix incorrect grammar or over-complicated prose.
I would be more inclined to saying that these edits indeed fixed a problem, but one that isn't very bothersome.
I also don’t feel like hiring an editor for something I see as an hobby.
I find this really interesting. You almost certainly are still paying, but I suppose this removes the friction of hiring/waiting for revisions/reduces the cost. On the other hand, LLMs are trained on misappropriated materials of editors and authors before you. I am not asking to be instigative, but of genuine curiosity: does this not bother you? Were other automated grammar tools not sufficient in some way? What caused you to write this article differently (w.r.t. the final note)? (edit: saw your comment :D)
I already have a subscription. Another nice thing we an LLM compared to something like DeepL or Grammarly is it's ability to handle markdown directly.
For the ethic issues, I acknowledge them and it's difficult to reconcile with my usage.
LLMs are trained on misappropriated materials of editors and authors before you
Doesn't this apply to human editors and authors too?
There's just no way you're making that argument in good faith. A person reading a book that they (bought, borrowed from a library, etc.) or a blog (meant for human consumption) is trivially not comparable to mass-scraped text, often without compensation.
So, is it about scale? I'm struggling to understand your argument, because scale was not part of your original comment. If you read hundreds of books, you should feel bad but if you read only 20 in your life it's ok? What about reading in public libraries, should it make me feel guilty?
It's kind of boring having the exact same debates every 10-20 years, when the new thing appears.
You overlooked 'misappropriated' in your initial reply, which is a big part of the issue. Those aspects have been widely covered elsewhere, so I won't repeat them here; but they may help you understand the wider argument.
I am struggling to understand what part bit you don't understand that made you think someone is suggesting that a person reading hundreds of books would be a problem, or where public libraries come into it.
But to engage on a wider point in good faith- yes, scale can absolutely be a problem. A person cutting down an old tree in their back yard might not be a problem, but mass deforestation is. A person making a campfire is not a problem, but vast wildfires are. A human reading a blog is not a problem; armies of scrapers affecting sites availability and resource usage are. A human enjoying or being inspired by a piece of art is not a problem, but an automated plagiarism machine is.
Honestly, it's the same discussion, again and again: Software, music, movies, books, and so on. And the argument is at the core the same: "intellectual property" rights. A stupid and extractive concept, that protects incumbents (from successful artists to trillion dollar companies) against new challengers.
I want information to be free. I want to be able to create using AI, without stupid restrictions imposed by copyright holders (because this is where this discussion eventually ends up).
The complaint here is plagiarism, which is more related to fraud than intellectual property.
If I read a book in the library and it inspires me to write, I can see:
LLMs can't, or at least don't in practice. If they were better about this, I suspect everyone would be a lot less mad.
I think the other complaints are fundamentally the same too. Scraping, spam, automation, resource use etc. are bothersome because they entail a lack of respect and reciprocity for everyone around them, much more so than that they are fundamentally bad
Information should be free. But this mantra should not be used as an excuse for the pillaging of the commons, while society is unable to support the unpaid creation of information. Companies with the resources to do mass scraping also have the means to support the people creating the content they so depend on. They don't, because they know they don't have to, and this imbalance is an economic source of my concern.
On the other hand, writing is a creative process. It is an emotionally and intellectually vulnerable thing to do. We do this because to write is to share oneself with the community. A machine is not a member of that community, and a company which trains models off of that vulnerability is consuming the very soul of the writer, churning it into floating points, and emitting a homogenized and disingenuous simulation of the writers it consumed. That, more than anything, offends me most, and perfectly encapsulates the problem I have most with the tech industry: the commoditization of the self. So, no, it is nothing to do with "copyright" and everything to do with the abuse of material that humans wrote for humans, not for companies to commoditize or the machines they commoditize them with.
That argument can be made with closed proprietary models, but I don't see how it applies to open models.
Open models are still created without the consent of or support for the authors in pretty much all cases afaik. The technology is still largely dependent on senseless scavenging. The authors' voice is still used by others in contexts they did not consent to.
I don't really believe in copyrights and I think making information freely available is a net positive for humanity. It's worth noting that authors themselves don't exist in a vacuum and they too learned from others. The complaint here is largely that automation makes the process more widely accessible. The idea of senseless scavenging assumes that human learning is somehow more noble or intentional when it is really just as opportunistic. A painter does not ask every artist who influenced them for permission, and a writer doesn't seek consent from every author whose sentence structures they absorbed. The difference with LLMs is only scale and speed rather than the underlying principle. And if an author's voice is used in a context they did not consent to, the same thing happens every day when people quote them out of context or remix their work in fan creations, or adopt a similar style of their own. Our entire culture and civilization are built on reusing and repurposing existing ideas.
I still think this is missing the point. Humans draw from other humans to find their own voice to express themselves. LLMs draw from humans to produce the average of its input, and don't experience the vulnerability of creation. Maybe that doesn't matter to you, and that's fine! But it's a big part of why I partake in others' work.
Consider Adam Neely's content on generative AI in music for a better formulated expression of why the "artists learn from artists too" argument doesn't hold.
Except that LLMs are a tool humans use. The intent ultimately comes from the human using the tool. The whole premise that people just push a button and content magically falls out is completely false, and it's intentionally used to misdirect from how these tools are used in practice.
What matters to me is that whatever content I engage with stimulates my thinking or emotions in a way that I find meaningful. How that content is produced is not really something I find terribly relevant. What matters to me is whether I learned something interesting, whether I got some new ideas out of it, or if something I saw or heard triggered some emotions within me.
And regarding the video, I feel like the whole human connection argument completely misunderstands how aesthetic experience actually works. A human creator is not required for us to feel emotion, just look at how people react to a beautiful sunset or a majestic mountain range. Not only that, but when we listen to music or look at a painting we rarely have any real insight into what the artist was actually feeling or trying to convey anyway. Typically, what happens is that we projecting our own internal state and personal baggage onto art. If an LLM can organize audio in a way that acts as a mirror for our own emotions, it is still creating a meaningful experience because the meaning comes from within ourselves just as it does with art produced entirely by human hand.
If that's your view I would gently suggest there are better ways to articulate it than "ya gotta problem with people reading books in libraries, huh??". As Addison pointed out that comes across as rather bad faith. I suggest this as I am not entirely unsympathetic to your worldview; I don't think picking fights in comment threads is the best way to win hearts and minds.
You added color that was not there in my comments.
From my point of view: Someone wrote a sincere article explaining how as a non-native English speaker finds value using AI to improve his writing, and comments here ask him more or less, "but how does it feel using the ultimate plagiarism machine that steals from creators"?
As a non-native English speaker myself, maybe I should have filtered my comments through ChatGPT first, to avoid misunderstandings.
Είμαι πολύ περίεργος να μάθω τι κάνει κάποιος που δεν μιλάει Ελληνικά όταν βλέπει ένα σχόλιο σε μία άλλη γλώσσα. Χρησιμοποιεί αυτόματη μετάφραση που έχει εκπαιδευτεί πάνω στον κόπο χιλιάδων δημιουργών ή όχι;
Someone wrote a sincere article explaining how as a non-native English speaker finds value using AI to improve his writing, and comments here ask him more or less, "but how does it feel using the ultimate plagiarism machine that steals from creators"?
And the question was sincere! I was genuinely curious as to how the author reconciled with this, and they responded. I'm not judging the author for that because I have no real understanding of their experience and have no reason to extrapolate a belief about their morality based on their decision to use these tools. I do want to understand these decisions, though, because I don't see the world the same way, and I want to understand what considerations people make when using these tools.
Personally, I use them extensively. I have always published my articles under Creative Commons licenses, and my code under Open Source ones, and throughout the years, the licenses I use are even more permissive (where I started with GPL, I'm now with MIT for example).
I'm grateful to everyone else who has done the same and has helped me lean, improve and create using their work. I'm also happy that search engines can scrape everything I've written and make it indexable for the 5 people per year that may need it, I'm happy that small and larger companies have taken advantage of my code contributions to solve their problems or become more productive.
Every time I enter a domain where intellectual property rights are prominent, I realize how lucky developers are. There is so much free and open code out there to read, copy and learn from. Try to ask ChatGPT to give you the music score of something Beethoven wrote 200 years ago (100% public domain), to see what protecting intellectual property leads to. Example: https://farcaster.xyz/vrypan.eth/0x078d2394
I want llm models trained on everything they can get their hands on, and I don't want legislation and armies of lawyers deciding what can and cannot be used and how much it costs. Now, if you think that these companies make excessive profits, or that they have the obligation to return back to the community (I agree), just tax them accordingly and use the money to invest in research, education, libraries, arts.
I had a paragraph about that in an earlier version of the article, but it was unconvincing and I didn't want to pretend I have an answer about everything, so I just removed it.
Είμαι πολύ περίεργος να μάθω τι κάνει κάποιος που δεν μιλάει Ελληνικά όταν βλέπει ένα σχόλιο σε μία άλλη γλώσσα. Χρησιμοποιεί αυτόματη μετάφραση που έχει εκπαιδευτεί πάνω στον κόπο χιλιάδων δημιουργών ή όχι;
I used both my browser's translate selection function, and translate.google.com. I don't know if they use AI (translate.google.com might not have a few years ago, it's been around for years). Both where nearly identical, but the browser translation used "it" ("Does it use automatic translation ...") where Google used "he" ("Does he use automatic translation ..."). I wish I had a better answer to this.
Thanks for being more moderate than me. :-)
They definitely use some kind of ML, trained on existing texts.
It's "he". In a way, this is the difference between vibe-coding (or slop content), and AI-assisted generation: Is the human in position to evaluate the result? If he knows the language (Rust, Go or Greek in this case), they will adjust and guide accordingly, if not, they just take what's given and hope for the best.
In cases similar to the ones described in the article, I know English well enough (vbernat too, I guess, but I'll speak for myself) to evaluate if the translation maintains the original sentiment and style, and I can steer it in the direction I want. I'll do it out of self-interest (distribution), but readers benefit too (by getting a carefully curated/edited translation instead of an automated one).
I think that if you see the whole process, motives and benefits through this lens, it's far from "AI slop". Imho, this also applies to coding, and I find it infuriating when someone classifies a project as "vibe-coded slop" because "no human can push so many commits in one day", without checking the actual architecture design and code quality.
Sorry, I know I'm digressing, but this is a persistent theme in lobste.rs lately, and I wanted to share all these for too long, it just happened you triggered it.
The original Google Translate (2006) was a statistical model trained on UN and EU Parliament documents, which are required to be translated to multiple languages. They have long since moved to ML models trained on (among many other things) web scraping.
You added color that was not there in my comments.
I embellished the tone to illustrate the belligerence inherent in a bad-faith line of reasoning. If we were instead discussing the problems with overuse of cars, and someone instead replied with, "What about driving to pick up groceries for my sick grandma, should it make me feel guilty?" I would also call that out. I wanted to illustrate that words landed with "I want to start an argument", not "I want to discuss how we can get to a point of universal information use".
From my point of view: Someone wrote a sincere article explaining how as a non-native English speaker finds value using AI to improve his writing, and comments here ask him more or less, "but how does it feel using the ultimate plagiarism machine that steals from creators"?
The author has responded to that specifically to acknowledge the difficulty they have reconciling that, which leads me to believe it is a worthwhile thing to bring up.
As a non-native English speaker myself, maybe I should have filtered my comments through ChatGPT first, to avoid misunderstandings.
I'm not sure what you mean by that; and alas my understanding is extremely limited so I will have to apologize by saying that your final paragraph is all ελληνικά to me.
Edit to add: I've seen your reply to a sibling post- your other points are made reasonably! If I can be so bold as to make another humble suggestion: start with that next time :)
You almost certainly are still paying
Assuming there is a subscription already being used for other reasons, copyediting (which does not need the most expensive tier) is unlikely to exceed the quota for the time periods when it happens, so immediate marginal payment is probably zero.
(If there is no subscription, rather than paying per-call it makes more sense to run a sufficient-for-copyediting model on CPU or, for some hardware, iGPU; marginal expense is not zero but probably lower than token prices for lower-end proprietary hosted models)
I'm in the same shoes. I'm not a native speaker. I still found it a worthwhile challenge to learn to write well; I think it helped me immensely in my career and hobbies. It's also not that complicated. For me, it boiled down to three observations:
You're not writing for yourself. Put yourself in the shoes of your reader. Understand what they know and don't know, what drives them, and how much they actually want to learn about your craft. In particular, don't assume they've read every other thing you've ever posted, want to click on every outgoing link in your article, or look up every bit of jargon you use.
The hardest thing for an author is to delete text. But a good part of what an editor does is removing unnecessary words, weird tangents, and other chaff. So, get in the habit of deleting.
Finish the article, take a lunch breach or go for a walk, and then re-read the whole thing before you publish. You'll be surprised how many issues you can catch just by spacing these tasks apart.
I'd also recommend flipping through The Economist Style Guide. You don't need to follow it religiously, but it's a good read and it will make you go "huh" here and there.
That said, I understand why not everyone sees the benefits. In the era of LLMs, your coworker / rival blogger can now effortlessly produce professionally-looking text, so maybe it's all for naught. So, if you want to use an LLM as a critic / editor, I think that's fine. Just be aware of the limitations. In my experiments, their sycophancy gets in the way: both Gemini and ChatGPT always praise me for objectively crappy explanations or analogies while suggesting cosmetic edits that rob me of my voice... convergence toward the bland.
The problem you might run into is guilt by association: most people who use LLMs for blogging just have the tool auto-generate a slop-post from a short prompt. So if someone gets a whiff that you're relying on an LLM, they might assume you're one of the slop-mongers who are ruining the web.
These are interesting suggestions. 1. is good advice but it's easy to lean too far into it and come out with something that reads as a dry presentation -- there was someone in IRC looking for feedback on an article about Story Points that I think illustrates the issue. The article is not bad, but I think the author has perhaps done too much exposition. 2. Antoine de Saint-Exupéry agrees :) 3. I would perhaps add the suggestion to temporarily change the font when reviewing; all the better if it's one you never use commonly yourself- it also helps jog the brain out of overfamiliarity with the text.
For more advice, I defer to the sage wisdom of Alex / mangopdf :)
Yeah, for #1, it's not that you shouldn't be yourself, just that you should think about the audience. I work with infosec and found this particularly important in professional settings. When we write bug reports or request improvements, the reader doesn't want to learn security. They're just a developer or a PM desperately trying to decipher what you want from them and why it matters.
For hobby blogging, I think it applies in a different way. You get like-minded readers, but if they come to you from search or social media, you can't expect them to know all your blog-lore. If it's a tenth post in a series, it can't be a prerequisite that they read all the previous installments first.
As for your last link... yeah ;-)
This article is my reaction to the story about banning LLM-generated submissions. Some users were against any use of LLMs, including as an helper for copyediting. I know this may be difficult to change people minds on such a polarized subject, but it is my tentative to show LLMs can be the evolution of a spell and grammar checkers.
As an experiment, I didn't use an LLM to copyedit this specific post (nor translate it to French).
For what it’s worth: I like your personal voice better. It has a kind of „Ok, let’s do this“ energy to it, a kind of slightly clipped cadence.
I read two other posts and the tone is not worse but this one has a recognizable rhythm.
(Caveat: My native language is German, so my perception might be different from what you are going for.)
I can't stress enough just how unhealthy these LLM witch hunts are. These tools exist and they have legitimate uses, as the author of the article clearly demonstrates. Going after people for using these tools is harassment and all it does is create a toxic atmosphere here.
In my view, the focus should always be on the actual quality of the content. Was the article interesting, did you learn something by reading it, did it articulate a new idea or a concept you weren't aware of. Did it make you think and give you a useful perspective on the subject. These are the questions people should be asking themselves when engaging with submissions.
Focus on substance over style, stop perseverating over whether an LLM might've been used, and stop harassing people. Focus on the quality of the content and discuss the actual subject of the submission. If the content is low quality then mark it as spam or explain why you think it's low quality.
The fact that these basic things have to be stated is incredibly depressing.
I wonder if Clojure users are more likely to be in favor of LLMs, or at least their use, than those who use other languages, since Clojure developers by definition already seek a mindset of getting more stuff done now, faster, in a pragmatic way.
That's possible, I'd also imagine this would be more true for people using high level languages, especially FP style where we are already used to abstracting implementation details and focus on writing declarative code. Going from that to making specifications that LLMs fill in doesn't feel like a huge jump, while it might be jarring for somebody who is used to focusing on the minutiae of the implementation details.
AI is useful outside of the narrative process.
My approach is to use LLMs when I'm researching an idea (it's simply more efficient than Google and links to sources).
Once I'm confident enough to speak to the topic, I do so literally, using Whisper to transcribe.
Then I have the LLM insert that into my blog, with explicit instructions to not alter any of the transcribed text, because preserving my own voice is mission critical.
Then I editorialize the prose manually, making any verbal-only phrasing make sense in a written context.
Finally, I prompt several LLMs to function as a panel, trying to surface anything objectionable or outright false in my post (I'm human so I make mistakes and want to be aware of them). If anything salient is in the notes, I'll manually adjust my post.
Point is LLMs are used fortuitously, and you don't need to let their prose bleed into the text.
I can see how it could be useful as a language translation layer, especially since romantic and germanic languages are so idiomatic.
You should use a different title for this post. People are immediately triggered by it and will stop reading.
I thought that adding "assistant" would make it clear, but yes, it may not be the best strategy.
Something more explicit like “Blogging with LLMs as a non-native speaker” would be more effective I think because it would state your motivating use case upfront. But parent might even drop the “with LLMs” part.
I tried something like this myself, especially on the translation front, and concluded that LLMs are woefully not up to the task. Perhaps French is easier for LLMs to handle and produce satisfactory results than Polish, but the blog posts that I translated using Claude Opus from English to Polish are just... terrible, honestly.
I suspect part of it is in the nature of how autoregressive LLMs operate. In Polish sometimes what you put at the end of the sentence can change the declension of what comes before, which is not the case in English. So if the model, in the process of generating a sentence, doesn't "know" whether it'll have a masculine or feminine subject at the end of the sentence, it will put incorrect endings on preceding words.
So at the end of the day I only have a few articles translated to Polish, because the price of a competent translator is not really justifiable for an independent blog, and I am hopeless at context switching so I can't do it myself properly without expending a lot of effort that could be better spent on original content.
I promise you that in French you also sometimes need to know the grammatical gender of a word yet to be written. I have observed pretty OK results translating between English and French with local LLMs.
I have now tried translating from English to Russian (which should be grammatically not completely dissimilar to Polish, although of course there are details), including some pages from your blog and sentences specifically constructed to require reasonable but larger lookahead. This is still a local LLM (although a 35B-A3B, so on the iGPU side of consumer-local). I'd say that style and word choice could use some improvement but it needs mostly isolated fixes, not rewriting — at least if I go paragraph by paragraph. If I paste an entire long post at once, the results are … not so good. Not so good, as in I have caught it putting Latin letters into the output where a thing was clearly translatable. Of course a local LLM won't charge me extra for using API and even injecting false history to prime the translation style.
I think some of this is a training data quality/quantity difference between languages, but for Polish, unless someone thought that understanding Lem too well is a safety risk (this would not be untrue…), it could probably still be doable with some slicing and example-style injection into the history.
There are many issues where translating with small-ish chunks per request seems to improve things; when I mass-translated things, I wanted to auto-slice anyway just to avoid translating things that definitely needed to be copied verbatim (this is both a performance issue and a «very annoying to check for hallucinated changes» issue!) — so I have a more positive impression than people dropping a txt file from Project Gutenberg into attachments and asking for a translation.
By the way, speaking of an older kind of LLMs that were not called LLMs, how would you rate DeepL on English-to-Polish translation of a single paragraph?
Isn't the quality good enough for you to do a second pass to fix the issues? The LLM is here to go faster: it translates and keep the markdown markup. I am still reading aloud the text and fix the remaining issues. But in my case, the LLM produces grammatically correct text. It just doesn't sound like I want, so, yes, French and Polish may be different.