5 minute read

LLMSpeak

Introduction

My journey of writing started when I was six, my mother gave me a diary. The diary was a purple notebook with McMug and his friends on it. With my jagged handwriting, I have filled up all the pages with all my adventures, challenges and feelings. Since then I have written to make sense of my life and pass through the gruesome examinations. Through pen and paper and later a text processor, I could feel, think and express. In writing, I found a place to stay focused, a place to reorder my thoughts and a place of achievement. After all, modern life is too easy. Writing is mentally taxing. It is one of the hard things modern gentlemen can pick on to challenge, apart from Hyrox and marathon.

This hobby also gives me an advantage in my career. I could get wins by presenting my ideas and plans well. Some of the career opportunities I was given, I have partly attributed to my ability to write well. In some of my previous employment, sometimes my writing brought clarity to me and my colleagues so we could move forward strategically. For that I was grateful.

My strike of luck ended in December 2022.

Since ChatGPT is out in December 2022, my world was struck by a meteor. The LLM seems to be “writing” a whole article with a few prompts. Those prompts can even be super vague and incoherent! With a bit of “prompt-engineering” and pulling the slot machine of the next token generation a few times more, you can write a decent article.

I struggled to find meaning to write.

Limitation of writing with LLM

As time goes by, my experience of writing with LLM has been very different from what I heard from my friends and the social media. Both of the two circles praise how good and how effective it is to churn out a linkedin article to get impressions and likes, leads and customers.

Until recently I realised I am the outliner LLM cannot serve. I write to clarify and strategize, and then to communicate the result. For the first 2 purposes, LLM is not the best tool. LLM is a sales and marketing machine which is “mid” at best and too Chinese or American. They are backed by some preliminary research.

LLMs use humongous folders of text to train. The folder is technically called corpus. The corpus came from the internet from the get go, which is mostly sales and marketing materials. The LLM model will then be optimized further in the fine tuning stage and will be aligned with the most common use cases - commercially and academically. We use it to write copies to sell & promote. We use it to convince people to buy and care. A lot of time I am not writing to do so. I am writing to express myself and reorder my messy ideas. LLM turns my mind to sell sell sell, when my product, which is my point of view, is not even well formed.

Another trait of the LLMs as a result of getting their corpus mostly from the internet is the output tends to be very American English or Chinese. As a Cantonese native, I see a lot of homogenizing of our Cantonese. A lot of slang words are losing their usage. The output from the LLM is subtly modifying our Cantonese to an American or Chinese way.

Other than the corpus, LLMs are, after all, a statistical machine that outputs the most probable next words. No matter how sophisticated it is performing, the underlying mechanism remains the same. As time goes by, all your articles and scripts will be similar to all 2,735,894 articles on social media that hour. In the Gen Z slang, your creation will be mid.

LLM has a tendency to be an overprotective English teacher. There can be only a few allowed ways to assemble words and stitch them together. No mistakes are allowed. It allows no experimentation with words and how they can be uniquely assembled.

I coined the term: LLMSpeak. (Turns out Sam Altman also state that)

How to battle LLMSpeak

Although LLM seems to be here to stay, how can you as a writer battle against LLMSpeak? You should own your creation, even if it is from the LLM or not. With this responsibility of your work, you can be both responsible and creative.

Although you can use a lot of tools to humanize your article, and no one actually will look into it and spot it, as time goes by, I don’t think there can be consistency and quirkiness unique to you. And in a world full of mediocre and safe contents, keen eyes and readers with heart can feel the authenticity of your article.

Also you should actually write more! You can even leverage the limitless possibilities of LLM to generate different versions of your creation, you can taste those and give your personal judgement on which version suits you and your audience best.

For example, for this article, I have come up with a barely readable outline and passed them through 3 versions: A handwritten version, a generated version from Gemini, a generated version of Claude. The generated version I also use the humanizer skill to humanize them. Then I can see my style or limitations, depending on your lens.

A final reminder: remember why you write

You are writing for a human, even if they parse your piece to an LLM to summarize it. Treat your readers as a human. They deserve your full attention on making them understand you in an organized way.

You may not know when the reader is yourself. You don’t want to read an alien LLMSpeak with your name on it.

References

Nineteen eighty-four : George Orwell : Free Download, Borrow, and Streaming : Internet Archive

[2508.01491] The Homogenizing Effect of Large Language Models on Human Expression and Thought

Benchmarking Linguistic Diversity of Large Language Models

Different Versions of LLMSpeak

GitHub - blader/humanizer: Claude Code skill that removes signs of AI-generated writing from text

Tweet on spotting LLMSpeak Tweet on LLMSpeak from Sam Altman