mnmlmnl

TL;DR: I used ChatGPT extensively to edit The missing discussion of LLMs potential. It was cool. I will do it again.

I have been using ChatGPT to help me edit some of my writing, but never for a full blog post. I am not a native English speaker and decided to write (and improve my writing) in 2021. A big part of writing is editing. Editing is notoriously difficult, but it is hard for an author to distance themselves enough from their writing (and thinking) to make the choices that would benefit a reader the most. This is true in music just as it is in writing. While an easy way to get some distance is to let pieces rest, this is not realistic in a professional context, nor do I want to leave drafts of blog posts lying around too long.

An often untold benefit of editing is editing other people's writing and having other people edit your own writing. It is a drastically efficient way of getting a sense of what makes for good writing and what doesn't, and it provides mutual benefit. It is also hard to see one's own words being modified, discarded, and criticized. Learning how to criticize and how to accept criticism is a big part of art education. It should be more prevalent in the education of programmers since code reviews and refactoring is a crucial aspect of programming.

One thing ChatGPT is particularly impressive at is stylistic transfer (rewrite X in the style of the declaration of independence was a viral hit). I figured this would allow me to see my own writing from different perspectives, in a very quick feedback loop.

Editing an article

I usually start my blog posts by freewriting on my freewrite traveler or by collecting some of my threads on [mastodon]. I'll do a first editing pass, playing with the order of arguments, finding additional points to raise in my Zettelkasten, and rewriting sentences to create flow. Lately, I experimented with pushing these second drafts out quickly, which was a good exercise to overcome my perfectionism, although I think I veered too much in the other direction.

I have been using professional editors in the past, and the value they add is very different. They can understand what I am trying to achieve—the point I might be trying to make—and suggest writing a completely different article instead or discarding everything but a paragraph and starting again. These are judgments I wouldn't trust ChatGPT with (although I might decide to use it for ideation; in that case, I would still be operating within the confines of my own limitations).

In this case, I have been working on an extremely long article of using LLMs and stable diffusion for retrieval practice and decided to extract irrelevant rambling into their own blog posts. I asked ChatGPT to suggest different ways of rephrasing individual paragraphs, ultimately settling on the “style of a sober Atlantic article.”

Here's the raw dump of the session, for your perusal. I think it has overall been very beneficial to the wording of the article, although only time will tell. I can definitely see myself becoming bored of its style after using it for a couple of articles. I could see myself deciding that I don't need it anymore except for sentences that give me real trouble. I could see myself building a tool that streamlines the process so that I can use it for everything I write. I could myself pushing it much further than I did here.

For a full comparison of after / before, see this entry in my vault.

Using ChatGPT as an editor, a raw log

Opening paragraph

Original paragraph:

My favorite software tool of 2022 is Github Copilot, which has transformed how I write software. While I studied machine learning in general and neural networks in particular in 2017, I got frustrated by the hype-to-usefulness ratio and decided to move on to other things. Ironically, this was right as the Transformers paper came out. GPT2, GPT3, and then ChatGPT definitely showed me how fast the field could move, and ChatGPT made me realize that a whole new page had been turned.

Prompt for GPT:

Rewrite this paragraph slightly to make it a bit more lyrical, in the style of a new yorker article:

Answer:

As I write this in the waning days of 2021, my thoughts are already with the software tool that has captured my heart and changed my life: Github Copilot. It was back in 2017 when I was deeply immersed in the study of machine learning, and particularly neural networks, that I began to feel a sense of frustration with the disconnect between the hype and the practical applications. I made the decision to move on to other pursuits, little realizing that this was the very moment when the Transformers paper was about to shake the field to its core. The rapid succession of GPT2, GPT3, and ChatGPT that followed only served to underscore the breathtaking pace of progress, and it was ChatGPT that truly opened my eyes to the fact that a new chapter had begun.

Prompt:

Rewrite this paragraphh in the style of a critical the atlantic article:

Answer:

In 2022, it is Github Copilot that stands out as my preferred software tool, fundamentally altering the way I approach writing code. As I delved into the study of machine learning and, specifically, neural networks in 2017, I was increasingly disappointed by the disproportionate emphasis on hype over tangible usefulness. I ultimately decided to turn my attention elsewhere, only to be caught off guard by the release of the Transformers paper. The subsequent emergence of GPT2, GPT3, and ChatGPT served as a harsh reminder of the lightning speed at which the field was advancing, with ChatGPT serving as a particularly poignant signal that a new era had dawned.

Prompt:

Rewrite this paragraph in a more sober, yet personal style:

In 2022, it is Github Copilot that stands out as my preferred software tool, fundamentally altering the way I approach writing code. As I delved into the study of machine learning and, specifically, neural networks in 2017, I was increasingly disappointed by the disproportionate emphasis on hype over tangible usefulness. I ultimately decided to turn my attention elsewhere, only to be caught off guard by the release of the Transformers paper. The subsequent emergence of GPT2, GPT3, and ChatGPT served as a harsh reminder of the lightning speed at which the field was advancing, with ChatGPT serving as a particularly poignant signal that a new era had dawned.

Answer:

As I reflect on my experiences in the world of software development, there is one tool that has had a particularly profound impact on my work: Github Copilot. When I began studying machine learning and neural networks in 2017, I found myself growing disillusioned by the emphasis on buzz over substance. I decided to shift my focus elsewhere, only to be startled by the arrival of the Transformers paper and the subsequent rapid emergence of GPT2, GPT3, and ChatGPT. These developments drove home the point that the field was moving at an extraordinary pace, with ChatGPT serving as a particularly noteworthy marker of a new chapter. Through its influence on my approach to writing code, Github Copilot has been a game-changer for me.

Ultimately, I settled on:

I studied machine learning in general and neural networks in particular in 2017 but got increasingly frustrated by the disconnect between the hype and actual results when put into practice. I decided to move on to other things. Ironically, this was right as the Transformers paper came out. The subsequent emergence of GPT2 and then GPT3 (alongside Codex) was a reminder of how fast the field could move, and ChatGPT made me realize that a new chapter had begun. My favorite software tool of 2022 is Github Copilot, which has fundamentally transformed how I write software.

Second paragraph

Original paragraph:

I don't know if the technology itself is just incrementally improving or if an actual threshold has been crossed. ChatGPT dug itself out of the uncanny valley that was holding me back from using natural language models (while I was immediately comfortable with more pragmatic applications like Copilot, in part due to the care around its prompt engineering ). Shortly after its release, I decided to give ChatGPT a spin as a daily tool for both work and personal research—just as I would a new IDE or programming language; I am still on that journey, and while I feel I haven't really left home base yet, I have already discovered many amazing things.

Prompt:

Rewrite the following paragraph in the style of a sober atlantic article

Answer:

It is difficult to say whether the recent advancements in technology are simply incremental improvements or if we have reached a true turning point. For me, ChatGPT represented a crucial step out of the uncanny valley that had previously hindered my use of natural language models. Despite feeling immediately at ease with more practical tools like Copilot, thanks in part to the attention given to prompt engineering, I was drawn to ChatGPT and decided to incorporate it into my daily work and personal research in the same way I might with a new IDE or programming language. As I continue on this journey, I feel as though I have only just begun to scratch the surface of what is possible, yet I have already made a number of exciting discoveries.

Great sentence:

“It is difficult to say whether the recent advancements in technology are simply incremental improvements or if we have reached a true turning point.”

In general the whole paragraph is better than what I wrote, and I settled on:

It is difficult to say whether the recent technological advancements are simply incremental improvements or if we have reached a true turning point. ChatGPT dug itself out of the uncanny valley that was holding me back from using natural language models. Despite feeling immediately at ease with a more practical tool like Copilot, thanks in part to the attention given to prompt engineering, I was mesmerized (as so many of us were) when ChatGPT came out; I decided to incorporate it into my daily work and personal research in the same way I might with a new IDE or programming language. As I continue on this journey, I feel as though I have only just begun to scratch the surface of what is possible, yet I have already made a number of exciting discoveries.

Creating a subheading

I then decided to ask it for alternative headings, which was less interesting (although, scroll further down for a further attempt).

Prompt:

This is a subtitle in the article I am writing, but I would like it to be shorter and a bit more impactful, while retaining its meaning: Amidst the criticism, I am missing a discussion of their potential

Answer:

“Ignoring the potential amidst the criticism”

Prompt:

no, try again

Answer:

“Overlooking potential amid the criticism”

Prompt:

no, try again, emphasizing that I a missing a discussion of the technology's potential

Answer:

“The missing discussion of potential amid the criticism”

Next paragraph

Original paragraph:

I think LLMs and image-generation networks are powerful tools for learning and helping humans process information. I am always a bit dismayed by the debates around these technologies in my social bubbles, not because the points raised (biased training, generating wrong but plausible-sounding bullshit, non-consensual training on copyrighted data, training and research being mostly in the hands of corporations, cost, and energy concerns, impact on knowledge workers being replaced by machines) aren't important, but because the debate never really focuses on actual benefits to its users, besides the wild overblown promises of the VC/tech/business people.

Prompt:

Rephrase the following in the style of a sober atlantic article

Answer:

As someone who believes that LLMs and image-generation networks have the power to facilitate learning and aid in the processing of information for humans, I am often disheartened by the debates surrounding these technologies in my social circles. It is not that the issues raised – such as biased training, the generation of misleading but credible-sounding misinformation, non-consensual training on copyrighted data, the concentration of training and research in the hands of corporations, cost and energy considerations, and the potential for displacement of knowledge workers by machines – are not important. Rather, it is that these discussions rarely focus on the tangible benefits these technologies can offer to their users, beyond the extravagant and exaggerated claims made by venture capitalists, technology firms, and businesspeople.

I ended up with:

I think LLMs and image-generation networks are powerful tools for learning and helping humans process information. I am often disheartened by the debates surrounding these technologies in my social circles. Not because the issues raised—such as biased training, non-consensual training on copyrighted data, the concentration of training and research in the hands of corporations, cost and energy considerations, and the potential for displacement of knowledge workers by machines—aren't important. Rather, it is that these debates rarely focus on the tangible benefits these technologies offer beyond the exaggerated claims made by venture capitalists, technology firms, and businesspeople.

Asking it for its “reasoning”

I decided to ask it a couple of questions about what it was “thinking.” I have been working with ChatGPT enough to know that it has absolutely no real understanding of what it is doing but instead is just vibing. Yet, these vibes have structure, and vibes are what I am interested in, so it is instructive to ask these kinds of questions at times. It is important to remember that they might sound like some “human reasoning” was going on, but that is just an illusion. The reasoning these models have is ultimately very alien to us.

Prompt:

Why did you do the changes you did?

Answer:

I made the changes to the original paragraph in order to reflect the more sober, measured tone of an Atlantic article. This involved rephrasing some of the more casual language, such as “my social bubbles,” and avoiding words and phrases that might come across as overly emotional or subjective. Additionally, I aimed to add clarity and specificity to certain points, such as by specifying the types of biases and the sources of overblown promises.

Note how the second part is just nonsense if you need a reminder that this stuff is not “intelligent” in any reasonable sense of the word.

Prompt:

which phrases came across as overly emotional or subjective in the original paragraph?

Answer:

Phrases that might be considered overly emotional or subjective in the original paragraph include “always a bit dismayed” and “wild overblown promises.” These phrases convey a sense of strong personal feeling, which might not be appropriate in a more sober, objective style of writing such as that of an Atlantic article.

Prompt:

Are there other changes you made that you think were particularly important?

Answer:

One change that I think was particularly important was the addition of specificity to certain points. For example, the original paragraph mentioned “biased training,” but did not specify the type of bias being referred to. By specifying that the biases in question relate to “training,” I was able to add clarity to this point and make it more concrete for the reader. Similarly, the original paragraph mentioned “wild overblown promises,” but did not specify the source of these promises. By specifying that the promises in question are made by “venture capitalists, technology firms, and businesspeople,” I was able to provide more context and help the reader better understand the issue being addressed.

sic.

Trying to get a subtitle again

Newer LLMs are usually quite good at summarizing text, and it is one of my main uses for writing documentation or writing emails. I know I tend to write too much, and I'll often distill things through ChatGPT, and shorten my writing (rarely actually using ChatGPT's output, because it is very often not exactly what I am trying to express).

Prompt:

What do you think the main point of the paragraph was?

Answer:

The main point of the paragraph seems to be that the debates surrounding LLMs and image-generation networks often do not focus on the actual benefits these technologies can offer to users, but rather on other issues such as biased training and the concentration of training and research in the hands of corporations. The writer expresses frustration with this lack of focus on the potential benefits, and suggests that the extravagant promises made by certain groups may be contributing to this lack of emphasis.

Prompt:

What would be a good subtitle for this paragraph?

Answer:

A good subtitle for this paragraph might be: “Failing to Consider the Potential Benefits of LLMs and Image-Generation Networks.” This title captures the main point of the paragraph, which is that the debates surrounding these technologies often overlook their potential benefits, and conveys the sense of frustration expressed by the writer.

Editing a short paragraph

This shows that the conversational interface to ChatGPT is quite interesting. In many ways, the whole process feels like talking to a well read but utterly stoned friend who is also a pathological liar. I actually have a few real-life friends like that, and I always enjoy spending time with them, because their company is very intellectually stimulating. Having to take everything they say with a huge grain of salt is part of that stimulation.

Prompt:

Can you summarize the following in a single sentence, in the same sober style of an atlantic piece? It seems that most people loudly arguing against LLMs have never tried to work with them. I get it that people are worried about low-quality posts on LinkedIn and flooding of disinformation or plausible-sounding spam, but that phenomenon has already reached a scale that overwhelms the individual.

Answer:

Most individuals who vocally oppose LLMs have not personally worked with them, despite concerns about low-quality posts on LinkedIn and the proliferation of disinformation and plausible-sounding spam that have already reached a level that exceeds the ability of individuals to combat.

Prompt:

Actually, summarize it in two distinct sentences, to separate the two facts

Answer:

Most individuals who vocally oppose LLMs have not personally worked with them. There are concerns about low-quality posts on LinkedIn and the proliferation of disinformation and plausible-sounding spam, but these issues have already reached a scale that exceeds the ability of individuals to combat.

I ended up with:

It seems that most people who vocally oppose LLMs have never personally tried to work with them. Most concerns I see are people worried about low-quality posts on LinkedIn and the proliferation of disinformation or plausible-sounding spam, but these issues have already reached a scale that exceeds the ability of individuals to combat.

More paragraph editing

I lost stamina towards the end, and decided to wrap things up:

Prompt for the final paragraph of the article:

Rewrite the following in a the style of ezra klein on vox.com: If we want to do proper technology criticism, we have to put in the effort to evaluate it fairly and intensively. There are tremendous amounts of interesting aspects to these models when used for personal productivity and great potential for them to be used for self-actualization. The more I play with them, the more I see their limitations, but the more I have scenarios where they turn out to be extremely useful, if not game changers.

Answer:

If we want to engage in truly effective technology criticism, it is essential that we put in the time and effort to assess these tools objectively and thoroughly. There are many fascinating aspects to LLMs and image-generation networks when it comes to personal productivity, and there is great potential for them to be used as aids in self-actualization. The more I experiment with these tools, the more I am able to discern their limitations, but I also keep encountering situations where they prove to be remarkably useful, if not downright transformative.

Prompt:

Do it again, but just for this sentence: If we want to engage in truly effective technology criticism, I think it is essential that we put in the effort to evaluate these tools thoroughly, if only to give solid counterarguments.

Answer:

In order to engage in truly effective technology criticism, it is crucial that we invest the necessary time and effort to evaluate these tools thoroughly, even if only to provide well-supported counterarguments.

Prompt:

Rewrite the following sentence as a punchy closing sentence to the few previous paragraphs: There are many things I find problematic about the current rollout of these models, but that is not going to keep me from studying them. For example, I have my own personal concerns about the misuse of LLMs for disinformation.

Answer:

Despite my reservations about the current deployment of these models, I remain committed to examining them closely – even my own personal concerns about the potential misuse of LLMs for disinformation will not deter me.

Prompt:

How would David Resnick write the following sentence: I find many things problematic about the current rollout of these models, but if anything, that is even more of an incentive to study them.

Answer:

It is precisely because there are so many aspects of the current rollout of these models that I find problematic that I am all the more motivated to study them.

Prompt:

How would Ezra Klein write it?

Answer:

There are many things about the current rollout of these models that give me pause, but rather than dissuade me, those very concerns only strengthen my resolve to delve deeper into their study.

What I ultimately settled on:

If we want to engage in truly effective technology criticism, I think it is essential that we put in the effort to evaluate these tools thoroughly, if only to provide well-supported counterarguments. There are many fascinating aspects to LLMs and image-generation networks when it comes to personal productivity, and there is great potential for them to be used as aids in self-actualization. The more I experiment with these tools, the more I am able to discern their limitations. Still, I also keep encountering situations where they prove to be remarkably useful, if not downright transformative.

I find many things problematic about the current rollout of these models, but if anything, that is even more of an incentive to study them.

Conclusion

I hope this raw overview of what I did for the blog post I just published was instructive. This is all an experiment, and I would love feedback. Do you think the editing was valuable? Do you do something similar? How does this compare to human editing? What are the downsides you see?