What LLMs mean for the future of writing

Article Summary provided by ChatGPT

Are you ready for a writing revolution? Large Language Models (LLMs) are changing the game, allowing for supercharged productivity and conversational documents that can be personalized for each reader. As a writer, you may be surprised to learn that 90% of what you write has already been written before, but fear not – LLMs can help you make that 10% shine.? They can clean up your style, proofread, and even play devil’s advocate. With the ability to converse and condition on vast amounts of written works, your literary identity can turn into an interactive, intelligent, conversational agent. The future of writing is conversational not static, allowing readers to interface with your body of work at any level of summary, ask questions, and even engage with other historical figures. Writing has already changed, and the possibilities are endless – are you ready to take advantage of it?

Thanks for the nice introduction ChatGPT! Let’s get going….

People are both excited and terrified by Large Language Models (LLMs). Some will tell you everything is going to be awesome and some say the end is nigh for human work and happiness. The truth is that no one really knows exactly what the effects will be and all we can really say is that major change is coming. For some of us that use ChatGPT and other tools like it, the change has already begun. That’s how we know it’s coming to you.

Let’s discuss one change that is almost certainly happening. ChatGPT will change how people write and will drastically increase writing productivity. That means if you don’t like to write, you are in luck. You will be able to do less of it. If you like writing, it means you will be able to write a lot more but differently.

Let’s discuss each in turn starting with people who don’t like to write. Really this includes everyone because there are some things that few people enjoy. One example is writing emails to give people information. Ugh. Boring.

Escape from office tedium hell

Example Email: Hey, David. We have started the official hiring process. We just need some info.? We need your address, phone number, SSN. Also we need the address of the last 5 places you lived, going back 15 years. And we need a list of all the people who lived with you in each of those places for a possible background check, ideally with phone numbers. Also, what was the date of your diplomas, both of them? Look forward to working with you. Bill.?

Ok, lovely. Now I have to spend the next two hours tracking down all this info and writing a reply email. I like writing but I don’t like this kind of writing.?

I’d estimate in about 2-3 years, replying will look like this.?

Me: GPT, please reply to this

GPT: Got, it. Did anyone besides your wife and first son live with you in Glendale?

Me: Nope

GPT: Got a phone number for Paul Heckte? Nothing in your contacts. Nothing on the web.

Me: No, just leave it blank.

GPT: Ok, I have everything else. How does this look?

(I look it over).

Me: Looks fine. Go ahead and send.?

GPT: Ok, please authenticate

(I put my finger on the fingerprint scanner)

GPT: Sent! Anything else?

Me: Nope

That took less than a minute. I saved almost 2 hours of toil. If you don’t like wasting time sending information that you sent a dozen times before, but can no longer find, you are going to like these things in the near future.

In fact we waste a ton of time writing to each other asking questions, writing and reading responses. Much of that can simply be automated.

Or you will be able to do things like this.

You: GPT, please read the Annual Technology Briefing. Let me know if there is anything in there that I’d be interested in.?

GPT: It’s mostly about the new data platforms that went live. They are also increasing investment in generative technologies. Also putting together a training course in gen-tech. They are switching over to the new marketing system very soon and warn about possible downtimes. They mention your group and your work in the Spotlight section and it’s quite flattering. Written by Wanda. Wanna hear?

You: No, tell me later. Is that the Fintel marketing platform??

GPT: You mean, Falintel. No. They went with the other one, Brenner Systems.??

You: Ok, Just send Wanda a thank you message and tell her I am looking forward to working with her new team.?

GPT: Got, it. This look ok?

You: Yup, send

(authenticate)

GPT: Anything else

You: Nope

I’m guessing most people are going to greatly enjoy not having to read everything, searching through emails and files and writing tedious emails. If it feels like you now have your own secretary, that’s because you do. Congrats on the promotion and welcome to the future.

But what if you are a writer? Now things become more complicated and the change is a lot more significant. To understand why, you have to understand a few things about LLMs.

Large Language Models (LLMs) in short

LLMs are essentially neural networks trained on vast quantities of writing. They can essentially be trained on everything people have ever written that’s in the public domain. In addition, they can be specialized for any purpose by including other information that is personal or proprietary. In the cases above, it’s clear the assistant GPT (Generative Pretrained Transformer) has been exposed to personal information, company information and has access to email, text or however we communicate in the future.

What they do is essentially construct a probability distribution of writing in general. They will assign a high probability to things that are often repeated like the Declaration of Independence and a very small probability to things like: “blab honker velum weiner country”. While it’s possible to write this last phrase (as I just did), I’m pretty sure I’m the first one in history to actually say it. Yea me! That’s what LLMs can tell you.?

That might not seem so useful but they can also calculate conditional probabilities. That is, they can calculate the probability of any text GIVEN that, some particular phrase comes before. That phrase can be a conversational phrase like, “Hey, GPT. Hope you’re having a good morning” or it can be a question. That’s called a prompt. It then figures out the most likely set of phrases to come next. And, perhaps amusingly, it actually does that one word at a time. It calculates the most probable word, gives it to you, adds that to the prompt and computes the next word. It continues giving you the words until some internal criteria is satisfied and then it stops. Now, it’s your turn to talk if you want. So it’s essentially just a conversion with a robot but a robot that has read nearly everything, knows what people tend to say and says those things.

Of course it hasn’t really read everything but generally has covered a lot of territory including reading things like Wikipedia which covers almost everything at least at summary level. The amount of data that can be crammed down its throat is up to the creators. The more data, the more it costs. Generally those costs will continue to drop as the technology develops further. According to the rumor mill, it costs around a million dollars every time they are re-trained from scratch.

If you are a writer?

?

Writing is going to change immensely. Now that you know what an LLM is, you should be able to see why. First the bad news.

You are an unoriginal hack! People like to think that their writing is sparkling with brilliance and originality; just like mine! It may be a brilliant idea that you are expressing but 90% of what you are actually going to write has been written by someone else before. You just never knew that until now. Sorry to bring you that bad news.

Now, I suppose most will find that shocking. I found it shocking. But it’s true and the fact that LLMs can write like professionals demonstrates that it is true. Now, I have to be a little more precise. No one has written the exact same words as you. The chance of anyone writing the exact same words probably heads to zero after 10 to 20 words. But AI isn’t a simple lookup table. It generalizes. It can recognize that things written in different ways are very similar.

But it isn’t all bad news. Maybe 90% of it is the same but there is that 10% which might be novel. Or it may even be much smaller. You see, writers don’t actually need to write original material. They just have to write things that feel original to the people who they imagine will read them. We can use the used-car euphemism; new to you. Sometimes we know it’s not original. In the extreme we are just plagiarizing and hope we don’t get caught. More often we get inspiration from others and write something we think is original enough. If we are nice we cite the ideas we have borrowed. Most of the time however, we just aren’t aware that it has been written because we can’t read everything. Computers however can.

So the thing is this. Let’s say 95% of your piece of writing is unoriginal and 5% is original. Imagine we have a way of separating the two. What that means is that writing is only 5% efficient. The rest has been written before and LLMs therefore know how to write it.

So should we writers be depressed about this? Only for a day or two. Then you get over it and realize what it also implies. If we are only 5% efficient and we can pair up with the electronic beast, maybe we can increase our productivity by a factor of 20. Now you’re thinking!

So, yes this thing is not going to replace writers. It is going to supercharge writers. If you wrote one article per week, now you can write 20. If you wrote a book every 5 years, now you can write 4 books per year. There are going to be 20X as many books soon. Maybe now Barnes & Nobles will finally be profitable!

Perhaps you detected a little bit of sarcasm. If people write 20X as many books, you still have to have people willing to read 20X as many books. Hmm. Seems like we have run into another problem.

What’s really going to happen to writing

If we’re just thinking about writing faster we are missing the big picture. If you’re a writer like me you know all about the battle against bloat. We try to write as succinctly as possible because people don’t like to read long articles. If writing succinctly was professional football, I’d be the New York Jets. It’s hard! What makes it hard is this. You want to ensure you get all your important ideas into the article. You also don’t want people to misunderstand you. If you’re making an argument, particularly a controversial one, you had better have your ducks in a row and have killer arguments and shoot down every possible counterargument before they even think of them. You also don’t want to let down the people who love your writing but don’t have all the technical background. That’s why I took the time to briefly explain LLMs.?

So the problem is really this. You don’t actually know who is going to read your article or who is going to like it and repost it. A whole host of people could come along and read it. How are you supposed to write an article which is the right depth, right tone and right length for everyone when people are so different? You can’t, right? You have to just make a choice, roll the dice and hope you are not about to embarrass yourself or get ignored.

This is where LLMs come in to save the day. It’s not just that they can help you write more succinctly. The big idea is that your article can truly come alive and become a conversational agent themselves rather than a static document.

Intelligent documents

It works like this. You write your 30 page article that only your biggest fan is going to read. Then you post it wherever you post material. When people see it they can read it. But now they can read it at any level of summary they want. And the exact way it can be summarized can be personalized for each of them.

So let’s say I am going to read someone else’s article on generative AI and how it will change education. My AI helper knows that I know enough about generative AI that I don’t need to read the introductory paragraphs. It can hide those. I also tell it I just want a very short summary. My AI helper rewrites it into three poignant sentences that explain the thesis. If it sounds uninteresting I pass. If it does sound interesting, I can tell it to expand to a one page summary. I read that one page. If I like what I am reading, I can expand it to the full article and read from the beginning.?

Or I can just start asking questions. I can essentially have a conversation with the document which is similar in a way to conversation with the author themselves. It can essentially predict what the author would say if they were available to chat with you in real-time. If the author has written many published articles on the subject, the agent can also read all of those and condition the LLM to respond with the context of their entire body of published works.

Your virtual literary identity

So effectively what is happening is this. Everything you write publicly can become a conditioning state for the LLM which means that your literary identity actually turns into a conversational agent. Your written works can now speak for themselves, literally. You don’t have to be succinct anymore. You can write in as much detail as you want. In addition, you don’t need to explain introductory things that don’t have anything to do with your thesis but are still a prerequisite for understanding it. If the reader asks, the agent will tell them what they need to know.

The new writing process

You also don’t need to focus as much on style. The AI can clean up your style, fix your spelling and grammar and rearrange stuff to improve the flow. It can act as an editor. It can also proofread. You can make it play devil’s advocate and look for flaws in your reasoning or identify incorrect facts. That can prevent you from publishing embarrassing things where you totally got something wrong and now have to live with it for the rest of your life.

Writing doesn’t need to have an intended structure anymore. Are you writing a book? An essay? A blog post? A bunch of aphorisms? It doesn’t matter any more. What matters is that you come up with new ideas and write them down clearly. You can put them together in many ways later. Likewise the reader can rearrange them any way they want as well. Or it can just interface conversationally.

The future is actually here

If you are rolling your eyes thinking that I am just speculating wildly, you are in for a surprise. ChatGPT can already do many of these things. Let’s see. Here is a snippet where I told ChatGPT to condition on the conservative economist Milton Friedman and his works and I ask him about regulation.

Hmm, sounds interesting. But surely this guy doesn’t think nuclear power plants don’t need regulation. Right? Let’s just ask.?

Ok, maybe I’m being convinced. But I want to hear a counter argument to see how it holds up.

Ok, I’ll stop there. Now this may or may not be impressive to you but also consider that ChatGPT is new. It probably hasn’t read Friedman’s or Krugman’s books. But in the near future, it or LLMs like it will. Note that this general purpose LLM is not in any way optimized to do what I just did. It just kinda worked. It’s only going to get better.

The future is conversational

Do you have a favorite writer? Is someone your intellectual hero? Perhaps you are a linguist and you’d love to get some one-on-one time with Noam Chompsky. Or maybe you want to talk to Einstein. Or maybe you want to have a crazy conversation with Kanye West. Now you will be able to do just that. You still might want to read the fixed-format writing they have done but maybe you just want to cut to the chase and ask something more specific or ask them to describe something you don’t understand.

Now, don’t take this too literally. I’m not saying we can now give everyone their own personal Einstein in a can. These agents conditioned on public writing are not real replacements for the people. If you can book Chompsky for a coffee meeting, then you probably should do that. But it might be a good enough proxy to guide you through their work and more efficiently find what you're looking for. I think it is highly likely that reading in the future will primarily be conversational.

But what’s the point?

Depending on your fondness for change, you may be loving this or hating it. To be honest, we all hate it at first. Rapid change is always scary. But keep in mind what the actual point is. The point is better, more efficient communication and faster learning not to mention less tedium. Unless that frozen head thing works out, this is mostly going to benefit our children. Just imagine how much faster they are going to be able to learn. They will be able to do college in two years instead of four. Or, heck, who needs college when you have your own personal virtual professors? (I have 6 kids, let me dream). We are going to have a vastly more educated society; or is it a society that can self educate on the spot as needed. It’s hard to know exactly but I think the benefits will be enormous.

Conclusion

I hope by now I have convinced you that writing is going to change dramatically and this is already practically upon us. I am using it to write this article. Google’s AI is correcting or suggesting spelling and grammar fixes in real time as I type. When my fingers stop typing, I’m going to turn it into a prompt (or a couple) and feed it into ChatGPT. I can ask for improvements to the style. I can ask it to find criticism. Maybe it can even help me with succinctness. In other words, my life as a writer has already changed and, very soon, so will yours.