How to Thrive as a (Technical) Writer in a World Where LLMs Think They Can Do Better
Do LLMs think?
Can LLMs think?
Wait… what are LLMs?
Am I even thinking by asking these questions? Or am I discombobulated by attempting to give Artificial Intelligence — AI — models (and algorithms) way too much credit?
The Beginning of the New Word Order
The world woke up on the morning of 30 November 2022 to discover that this new phenomenon — ChatGPT — had been released upon unsuspecting netizens, and it had every intention of world domination, at least as far as language creators were concerned.
I’m sure by now that most people know what ChatGPT is. Still, for those that don’t: ChatGPT “is an Artificial Intelligence chatbot that uses natural language processing to create humanlike conversational dialogue.”
Behind the scenes, ChatGPT is a large language model (LLM), a “type of AI algorithm that uses deep learning techniques and massively large datasets to understand, summarize, generate, and predict new content.”
As an aside, I’m quoting from websites like TechTarget to ensure I get these definitions 100% correct, and, if I’m very honest, my brain is lazy. It’s the end of a workday filled with gazillions of words, so it’s easier to quote these definitions than to paraphrase them myself… or use an LLM to paraphrase them for me. Also, quoting judiciously from authority sites adds link juice to content — which is good, right — or not?
But I digress.
While the technology behind LLMs and Artificial Intelligence is mind-blowingly fascinating, it’s worth noting that, at a fundamental level, LLM indexes and algorithms are number-based. In layperson’s terms, text is converted into vector embeddings (numbers on a graph), and these numbers are compared to each other to determine which words belong together.
Yes, this is a gross oversimplification and can even be construed as a total misrepresentation, but bear with me here. All will become clear — hopefully sooner rather than later.
Wait… how do you use numbers to create language?
Ah — now we are reaching the raison d’etre of this text — well, the start of it at any rate.
But, once again, I digress — it seems that digressions are a normal part of my writing process. This is not always good and begs the question: Would the all-powerful LLM digress as much as I do?
Let’s see…back to the topic at hand.
Back in August 2023, I was exposed to Vector Databases, specifically MyScale, in a professional capacity. This is where I learned that LLMs can hallucinate and need interventions like RAG — Retrieval Augmented Generation — to improve the accuracy of their answers to given questions.
My favorite example of an LLM hallucination is the one quoted in the article titled: “Teach your LLM to Always Answer with Facts, not Fiction:”
Wait… what are LLM hallucinations, you might ask.
The authors of this article asked ChatGPT this question and got the following answer:
A Late-Life Migraine Accompaniment?
Charles Bonnet Syndrome?
Is Charles Bonnet Syndrome real? Is it a neurological disorder, or is it a guess that anyone named Charles has migraines when wearing a bonnet?
It turns out that Charles Bonnet Syndrome is real, while Late-Life Migraine Accompaniments are not.
Wow. It would seem that the LLM is not all-powerful and has no potential to take over the tech writer’s job — yet… or ever?
Well, no, never.
I don’t believe that Generative AI, specifically Natural Language Processing in the form of LLMs, will ever be able to do the same job as a human writer. How do mathematical equations and numbers imitate human emotion?
Nonetheless, they can be incredibly valuable in assisting tech writers in writing unique, never-seen-before-quality content.
However, the caveat here is: When used correctly.
As described above, when LLMs are left to their own devices, they face two challenges:
- Hallucinating
- Writing in a robotic, emotionless style
Writers, specifically tech writers, must pay careful attention to both challenges. Undoubtedly, LLMs can play a significant role in reducing research time. I delve into academic research as part of my research process. It stands to reason that this takes time, especially when researching topics that are less popular or are new. Fortunately, there are now helpful tools like Elicit.com that significantly cut down my research time by analyzing research papers based on a question I ask.
I also have a professional ChatGPT subscription for simple research, such as asking it to find statistics, kickstart ideas, and develop real-life use case scenarios or examples. But… again, I can’t emphasize this point enough —there is a caveat.
“ChatGPT can make mistakes. Consider checking important information.”
This statement is OpenAI’s disclaimer at the bottom of ChatGPT’s web user interface. Fortunately, I checked some statistics I asked for a short while ago:
Here is the chat thread — or at least part of it.
Me: Microservices are an integral component of cloud native computing and are becoming increasingly popular — as well as the de facto framework in software development — please use the web to find statistics for 2022 and later to support this statement — please state the website you used and give the URL.
Note: This is a “lazy” prompt. It can — and should — be structured much better.
ChatGPT: A survey highlighted by DZone in 2022 revealed that 63% of enterprises are adopting microservices architectures. This trend reflects microservices’ critical role in digital transformation, improving efficiency, customer experience, and cost savings. Despite its popularity, challenges like visibility into processes and team communication remain. For more details, visit DZone’s website at https://dzone.com/articles/new-research-shows-63-percent-of-enterprises-are-a.
Fortunately, I decided to look at these statistics and saw that they are 2018 statistics and not 2022 figures. The chat thread continues:
Me: These are 2018 statistics!
ChatGPT: I understand your request for more recent statistics from 2022 onwards. However, the information provided earlier was based on the latest available resources in the browsing tool at that moment. For the most current data on microservices adoption rates and trends beyond 2018, I recommend checking reputable industry reports and technology surveys from sources like Gartner, Forrester, or Statista, as these platforms frequently update their research findings with the newest data.
Me: Why did you tell me that 2018 statistics were 2022 statistics?
ChatGPT: I apologize for the confusion earlier. The reference to statistics being from 2022 was incorrect; the statistics provided were indeed from an earlier survey conducted in 2018, as mentioned in the DZone article. My intention was to share relevant information on the adoption of microservices, but I misspoke regarding the date. For the most current statistics and insights on microservices adoption, consulting recent industry reports and research articles is advisable. Thank you for your understanding.
Wait…what? A machine apologizing? You can be forgiven for thinking LLMs are human.
The Firecrab Take
The biggest takeaway here is that Artificial Intelligence, specifically Generative AI (in this context), is a valuable tool that can be added to the technical writer’s toolbox. However, it must be used judiciously and circumspectly. Using it indiscriminately and giving it too much power can end up with disastrous results.
The temptation is there to tell ChatGPT 3.5 (the free version) to write an article on any topic and not to check whether this article is factually correct. For instance, I have just asked ChatGPT 3.5 to write an article on LLMsx — I pressed enter in the message box before checking the message prompt. Oh well… let’s see what ChatGPT returns.
Note: The “x” in the word “LLMsx” is a typo — it should not be there.
This is the result:
Umm — is there such a thing as a Super Extended Large Language Model?
Or have ChatGPT and I developed a new model?
If so… who is the copyright holder of this new concept?
It turns out that there is an LLMSX. It’s the acronym for the Franklin Multi-Asset Moderate Growth Fund, which is listed on the NASDAQ stock market. It is not a Super Extended Large Language Model.
