Writing in a world of LLM

Writing is scary and difficult for a multitude of reasons. Lately I’ve been thinking a lot about the idea that with the LLMs (Large Language Models) having a sizeable impact on the world, the act of writing itself might become irrelevant.

My gut tells me this is not true, but I want to expand on that in this piece.

My writing has flaws and bad habits, but it also carries my style and a vibe that should be kinda unique. And I like that! I try to improve it by writing more at work, on this blog and through some of my hobbies. My style is refining and evolving constantly, but sometimes (especially when I’m having trouble putting down what I want) I’m asking myself: “what’s the point?!”

I could go on ChatGPT right now, write four terrible sentences containing the gist of my feelings and it would write a better piece than I ever could. Hell, also give it other articles from this blog, my tweets and some of my writing from my job, and it would also match my style.

So, why do I still try to write myself? To be completely honest, I’m not sure. And sometimes I fear that I may be losing my time. But I’ll try to answer my own interrogations in this post.

The easy win

One thing that I found true about writing is that it helps me organize my thoughts. We often hear that you don’t understand something fully until you can teach it, and I think something similar could be said about writing. If you can’t express it clearly, it’s probably a bit too messy for you. Ben Kuhn wrote a bit more about it in Writing to clarify thoughts.

I started this post with vague ideas, and they came to life in a deeper level of understanding as I wrote and rewrote them. I’m sorting them from a messy blob of data into readable sentences (I hope). This is not a development that would have occurred if I didn’t write this piece myself, even if the generated post could have included some of the understandings I’ll make, I wouldn’t have learned them as a result of the process.

Ok, that was the simple one. Let’s dig deeper.

Predicting the future

I’m going to go astray for a bit, but to understand the place of written content in the future, I want to take a look at some possibilities.

I have no idea how much more advanced LLMs will become (This is what makes this so daunting to me). Whatever issues they currently have will get fixed, patched, erased by another version, whereas my writing style evolves at a much slower pace (To be fair, if I could output content at a speed of 100s of tokens per second, maybe I would improve much faster).

So what predictions can I try to make? How will written content change in a near future?

I - Companies/Product/SaaS blogs - SEO abused for aquisition

Some blogs are already fully automated (by seobot, drafthorse). With AI writing titles, outlines and content produced on a daily/weekly basis with no human reviews needed, you can get thousands of articles written in a day. This floods the Web, but this content mostly serves one goal: facilitate acquisition by appearing on more Search results. As long as Google Search works like it does now, it will keep happening.

Prediction #1: Search will drastically evolve to avoid this trend, with a revolution of SEO as we currently know it. All this content is often what I would consider low quality. Not specially because it’s badly written (as mentioned earlier, LLMs will get better) but mostly because there is a limit to how much relevant stuff you can say about a topic. Most of theses articles will end up looking very similar to each other and will not add any value at all. (Google might react way too slowly tho, maybe a new contender will emerge?).

Prediction #2: This kind of content (Why you should use our product, How to use our SaaS, Tips for X, Differences between Competitor Y and us…) will be replaced by a “chatbot like interface”. By chatbot like, I mean a kind of dynamic page where the user query (maybe directly from Search) would be executed and generate something like a blog post, live, for your need. (Kinda sick.)

II - Social Media & Individual Interactions - Fake content farms for advertising and attention market

Social Medias fought against bots and generated content for a long time. Even if content from LLMs improve by a lot and are “similar” to the style of the poster, I don’t think I’m taking a lot of risk by saying that they won’t be welcomed on most platforms.

(Mini Prediction #3: There’ll be some platform that accepts and even encourage the use of LLM trained on your data to generate content, they will then let the LLMs produce content and interact, find the individuals or communities with the most chances of matching what you research and then introduce your real self to that. Matchmaking 2.0).

I don’t know how they can stop/limit this, though. On Twitter, someone is selling a tool to automatically generate answers to other people. Someone else is using AI to post automatically 10 tweets a day. How long before half the interactions of the platforms are AIs talking to each other while some real users get lost trying to make meaningful connections? The perception, impact of theses changes may vary greatly for people. The ones consuming more, without really interacting with others may not dislike/realize the rise of automated content. And after all, if it fills their needs, why bother? But the others may desert the unregulated Social Medias in search of a place where they can exchange with other humans.

AI Companions(romantic/friendship) are also becoming more popular, with AI Girlfriend or Meta characters. Theses are human-like connections that are always available and behave in a pre-determined way. This gave them an edge over your real friends or partners.

III - Work Situation - Office politics gamed for influence

In work situations, a huge amount of interactions takes place in writing. Be it project reportings, RFCs, document or code reviews, written standups… I assume some people will always read a shorter version produced by a LLM instead of the original one. I also think a lot of theses will now be generated to appear more profesional, knowledgeable or successful.

Prediction #4: Big companies, or heavily political ones, will find themselves with a lot of individuals relying on generated content to navigate the spheres of influence. All companies (hopefully) will try to fight this. They will use the wrong methods: a ban on generated content (not verifiable), return to the office (won’t be enough), more micromanagement (poor workers) instead of relying on incentives and culture. Political companies will become even more political.

IV - Other - Drastic changes in the web and how we interact with people

We’re already encountering stupid situations. I can read a recap produced by my browser using GPT of an article written by GPT. This is funny, and absurd. And I believe it will become a lot more absurd!

As mentioned earlier, changes of Search may improve some of this. For emails, however, I think this was just the next step in the already heavy automated/spammy world. I guess we will all have to differentiate more between trusted sources and unknown ones?

Mini Prediction #5: Something will emerge to reverse the system of emails currently existing. Instead of being contacted by anyone knowing your email, you’ll explicitly allow some emails to contact you (with limitations like “only 3 times per week” or “only Monday”). It could also just be a democratization of the public/secret key system where you would entrust some people, systems or apps with a key to contact you.

I’ll also add some examples of possible changes suggested by Drew DeVault:

• The complete elimination of humans in customer-support roles
• Widespread propaganda and astroturfing, both in politics and advertising
• More convincing spam and phishing content, more scalable scams

The place of written human content in this future

With all of this in mind, I do believe that some written content currently done by human will be replaced. But in plenty of the parts that are automatable, many won’t be because they lose all interest the moment they are:

• The company blogs represent the brand: they are used for hiring, to showcase technical excellency, to transmit the company culture => no value if automated
• Social media interactions with other users: are used to meet people, discover their personality, have fun => no value if automated
• Engineering RFC: are used to get team approval && management buy-in => no value if automated

And there are plenty of others topics I didn’t mention here, personal blogs, comic sketches, fiction books, role playing game script. Sure, you’ll get some automated ones, but I think the better ones, will still be written mainly by a human (maybe with some help of AI).

The bonds created between a writer and a reader are a lot of what makes writing and reading blogs and books enjoyable. You won’t form a similar bond with content written by machine (or you might, until you learn it’s not written by a human).

The ability of a dungeon master to tell a story with the help of the players, that also works mechanically in the system played with unplanned exchanges, decisions and the human factor should stay more fun?

And the experience deepens once you start exchanging with the author of the written content through other means, social media, email, video call, at conferences or around a cup of coffee.

In a world where writing becomes automated, good writing will be even more valuable.

_{Thanks to Anne-Sophie Tanguy, Ludovic Fleury & Guillaume Navarre for their insights and reflections}