From TFA: "Eric Schwitzgebel writes that . . . There’s a huge cognitive difference between nodding along while reading something and actually productively generating a text. Two reasons: First, once the text is on the page, it’s easy to passively let the approximate word suffice, rather than thinking about word choice in the same effortful, active way we do when generating prose de novo. Second, as I suggested above, I doubt that human beings, even experts, have a good sense of all the factors that shape word choice -- everything they’re being sensitive to. You would have phrased it slightly differently, and even if you don’t know that, or why, a different signal is sent and received."
This is the best articulation I've seen of why simply reviewing and copy-editing does not provide remotely the same value as writing from scratch. I spent a considerable amount of time over the past two weeks reviewing and improving a work document that was the output of an LLM. Given the number of people involved and the final level of effort, I'm firmly convinced that writing it manually would have been faster and resulted in a higher quality product. Getting the wording right matters.
I really did like the author's framing there, but I think there is a different, simpler way to put it:
Why would you think that asking someone else (that is, another human) to write something (and then reviewing it) is the same thing as writing it yourself?
You may trust the other writer's opinions and knowledge, but it will not have the same tone, structure, word choice, understanding, or narrative flow as it would if you were to write it yourself.
And when it's an LLM, you should not trust it's "opinions" and "knowledge", because it does not have either of those things. The appearance of those things is just that, an appearance.
> you should not trust it's "opinions" and "knowledge"
How do you recognize someone as having true opinions and knowledge from someone having just appearance of them?
Except that there are such things as speech writers, whose job is literally to write prose in the voice of the person who’s going to read it.
And there are countless other jobs out there where people are the “voice” of another party. From those who manage social media presence, to PR firms, to copyrighters, all who put statements out on others behalf.
There was already an industry of professionals whose job it was to have others write things on our behalf, and that existed long before LLMs were a thing. What LLMs did was make that that service available much more cheaply.
Speech writers do not write in the voice of someone else, they create the public image of say a president who is inarticulate and make them seem polished and articulate, if anything they lend their voice and ideas to someone else. That’s why they are hired.
And no LLMs are not producing ‘the same thing’ as a speechwriter. If you are unable to tell the difference I’d suggest doing some a lot more reading of human books. LLMs produce stultifying pablum without coherence or style.
> Speech writers do not write in the voice of someone else, they create the public image of say a president who is inarticulate and make them seem polished and articulate, if anything they lend their voice and ideas to someone else
Nope. The exact opposite in most cases. Eg Jon Favreau (Obamas speechwriter) has said:
“As a speechwriter, your ego has to take a backseat. The goal is to make the speaker sound like the best version of themselves, rather than to showcase your own cleverness."
> And no LLMs are not producing ‘the same thing’ as a speechwriter.
I didn’t say it was the same. I said it was an existing industry that AI has stepped into and AI is taking the low end of market.
It’s just like how AI has captured the low-end of many tech roles too.
> If you are unable to tell the difference I’d suggest doing some a lot more reading of human books.
“Some a lot more”? Very ironically timed editing error there. You can bet an LLM wouldn’t have made that mistake ;)
Re speechwriters, he’d never get another job if he didn’t say that.
AI has replaced neither full tech roles nor speechwriters.
I think that’s a good example of the kind of trivial error humans make all the time when not rereading/editing. When humans make mistakes, it’s usually that kind of mistake, which IMO doesn’t really matter in an internet comment.
Unfortunately LLMs make gross errors of style and content and often just don’t make any sense in long form text. That’s a very different category of error.
Speech writers and shadow writers in general are writing on behalf of other people, but they're not writing the same thing. They can't! Usually they're writing something better, with the proper voice and authentic to the views of the person, but still not the same as what they would have written.
Depends on the individual. It’s usually a highly collaborative process. But in general, you’re correct that you’d expect a good writer to write better than the person who hired them.
There's knowledge in books, and it's ingested all the books, so it does hold knowledge. But I guess that's not how you mean it.
The key difference is that while an encyclopedia holds facts themselves, LLMs trained on that source material encode something more like a highly probable facsimile of those facts - the original fact was lost, LLMs are lossy, but can often be generated again with a decent level of accuracy by churning through stats about words, concepts, and relationships between them.
The whole catch is that they can often be regenerated. But LLMs (on their own, in their parametric memory - which is the result of training) don't have any conception of whether what they've generated is a real reproduction of some training material or whether they've invented something false that seemed probable based on their encoded stats. When the probability produces something contrary to what was in the training material, you get hallucinations.
They're very, very good predictive text models and can be very, very powerful when hooked up to other tools or outside databases. But its fundamentally lossy technology and all the books having been fed in doesn't guarantee all of the knowledge from those books can be spat back out.
Your criticism of LLMs also applies to humans and so implies humans don’t possess knowledge.
Humans can learn texts by heart, even lots of text (I have some expertise on that, having studied opera singing). They can reproduce these texts accurately, deterministically and repeatably. An LLM is a statistical machine. It does not know any text by heart and it is by pure chance that it sometimes reproduces existing texts verbatim.
Ant tips for doing so? I've never been able to do that, even as a kid. I remember the meaning but not the exact words.
Break it into small chunks and practice. Practice more. Put the chunks together and practice even more. Like any skill, it isn't something people are magically good at beyond minor proficiency.
Every performance- singing a song, playing an instrument, performing a stand-up routine, giving a speech, performing a theatrical role are all things that are best done from memory but require practice.
There are pneumonic tricks you can use- I've seen some people do it for tricks like memorizing the order of a deck of cards- but it's less useful for long term recital because it helps with order but not comprehension or fast indexing.
What zdragnar says :)
Humans sometimes don't remember things 100%. LLMs can produce certain text and be run deterministically.
I don't really understand this side of the debate other than as a gotcha tbh.
If LLM use atrophies your brain and skills that's bad. If it has a repulsive writing style that's bad.
I'm not sure what the debate about whether an AI is a statistical parrot unlike humans accomplishes. Is relying 100% on a bad human speechwriter somehow better?
It comes up frequently when discussing whether LLMs actually have intelligence or qualify as AI, especially in discussions on the path to achieving AGI.
The primary complaint seems to be that LLMs are held to a higher standard than humans, though I don't particularly buy that line of reasoning.
It's funny. I always thought that writing was meant to inform, persuade, or entertain about the subject at hand.
But in professional settings, a lot more of the informativeness is about the author, and a lot more of the persuasiveness is I'm worth your time and money. So, if the author is an LLM, and obviously so, what exactly are you informing your audience of (about yourself), and what are you persuading them to do (with your article).
I think we now know.
It doesn't hold knowledge it has the ability to seive through a lossy latent representation of knowledge.
People using LLMs not in a gruntwork manner tend to confuse style and tone with facts and knowledge.
LLMs are great to make drafts if you give them the source materials. They're great at validation if you give them the tools. They are great at layouting if you give them linters.
Encode architecture and decisions in your workflow, then proofread what your agents have been working on. Not the other way around.
Better validation and testing means more work will transform from exhaustive decision work to automateable gruntwork.
A workflow I've recently discovered for myself that provides a kind of middle ground: I'll ask the LLM to write a first draft in a language that isn't the one the piece should ultimately be in. Then I'll use the LLM's first draft as a blueprint, to write a first draft myself in the intended target language. This fights my brain's temptation to just shut off, and forces every word choice to actually be mine. Then I'll hand that over to the LLM for editorial suggestions and iterate from there.
Excellent idea, I'll try this. And for those who don't speak a second language other than English, you could use leet.
Wow, love this! Could even match with a language I am actively learning so it helps that as well!
I use an antagonistic agent to review and refute its findings, works great.
> First, once the text is on the page, it’s easy to passively let the approximate word suffice, rather than thinking about word choice in the same effortful, active way we do when generating prose de novo.
Meanwhile, me, as an English non-native speaker, ended up discussing two sentences I want send to HR for ten minutes while applying to a job.
I do believe there's generally a bias to accept something that's already written. The much bigger reason though is why you let somebody else write it to begin with.
It might just be that not thinking carefully about every sentence/wording was the exact thing that made you use AI to begin with.
> I spent a considerable amount of time over the past two weeks reviewing and improving a work document that was the output of an LLM.
I had the same experience with texts where I have a very detailed expectation of the desired result. This is just a general limitation. For code, there's the saying "the precise description of the solution is already the code". Describing X is a simplification of X, oftentimes it's fine guess the gaps. But when it's not, it didn't help to describe X, you have to manifest X itself.
Writing is thinking. In many situations where we are called on to write, what is actually needed is thought.
Careful consideration is needed. Word selection, grammar, all while keeping the reader in mind. You're sending a message, afterall, so the message should be readable.
And then if you want to tell a message with a different tone, again, you must consider differently.
The considering can be strengthened with exercise.
I can't imagine worrying about signaling as I write: what a huge distraction.
The problem is: this type of thinking and writing takes a lot more time and attention.
How much consideration should you put into a message? It's a personal answer, but also one that can be constrained by time.
That’s all true, how you convey your ideas matters, and takes effort.
I was talking about the step before that: what are your ideas?
I get you. Just adding some more to chew: Expressing ideas.
How do you circle an idea in your own mind: your train of thought, in order to express an idea?
Vocabulary to think, and then if we need to share, we use careful to pick words with the reader in mind.
Your ideas are based on lots of inputs. Things you read, discussions with others, etc.
But just in your own head they lack clarity. I think this can be deceptive to people, it’s a blind spot. How often have you sat down to talk over some disagreement with someone only to find their own thoughts are a jumble of disjointed ideas that don’t make sense?
yikes, 'we use careful consideration to pick words' is what I meant. Eating my own words.
What if I am using Claude as a rubber duck where Claude is the crowdsourced version of all the human thoughts on that topic. It is rerouting all the thoughts ever recorded on that topic through the chatbox to me. Something synergistic could emerge.
I view it as two different things. You need sources of input to stimulate your own thoughts. Historically, talking to others, reading other works, etc. I can see a chatbot as part of that tradition.
But in your own mind, your thoughts are still incomplete. The act of writing them down (or, I imagine, in an oral tradition somehow committing to a specific narration) is very important. IMO, when you delegate that to an LLM, you are not really thinking. It's the same as if you explained your ideas to someone (eg in an interview), and then they ghost wrote the work for you.
It's probably annoying the way I oversimplify things, but you reminded me of something I read somewhere.
Some Jazz musician was asked to define Jazz.
He couldn't really, his answer was something like: 'I don't know. But I'll know when I hear it'.
To add something to the discussion directly: thinking requires vocabulary. Vocabulary is the currency of thought.
You can usually express an idea with a few thoughts, or many. The audience, and amount of details chosen should always be kept in mind. Writing helps you to remember vocabulary and word choice when expressing ideas.
While I mostly agree with you, what is the real difference between an editor saying, 'when generating prose de novo' sounds pretentious to an American audience so let's use 'when writing from scratch' instead, versus an LLM giving the same advice?
Not that it's necessarily better but it's the kind of thing an editor might pick up and so why would you reject the advice if it's a machine giving it rather than a human?
TFA is about having AI do the writing, not about having AI suggest editorial improvements. The parent comment argues against having the AI write the text and having the human merely review it, not against having the AI review human-written text (and the human deciding which of the AI’s suggestions they might apply).
See also this existing comment: https://news.ycombinator.com/item?id=49768564, which I completely agree with and which is complementary to the above root comment.
True, but I was responding to the comment rather than the article.
I expanded my comment since.
Sure , but that seems like you've agreeing that an AI suggestion is as good as an editor suggestion.
What do you see as the difference?
I don’t see any conflict with what the root comment wrote. There are probably some differences between human and AI editorial suggestions that could be discussed, but that wasn’t the topic here.
It was the topic I was discussing with the person I replied to.
If you want to discuss something else just let me know.
Same flow as translating from your native tongue to a foreign language.
Going from foreign to native is easier than native to foreign. The latter requires completely different brain paths and a lot more understanding of the language to actually get to something correct.