How I use Large Language Models

ChatGPT, Bing, and Google Fi have greatly improved my workflow and other aspects of my life.

Let’s take a break from All the Kingdoms of the World this week. Today, I relay my experience using large language models. I have some tips and tricks to help you. At present, LLM discussions suffer from the loud voices of grifters and sensationalists. You should believe about 10% of what they say. Or you can follow Ethan Mollick, and you’re halfway to good information from him alone.*

Two quick points. Following Mollick’s advice, I use Bing on Creative to answer all my web queries. It is so good I never Google anymore. I also find the Bing app pretty intuitive on my smartphone. I use Google Fi on long solo drives. You can talk to it in real time and learn a lot.

I pay to use ChatGPT 4. I use it to improve my research. Here’s what I’ve found.

(1). GPT-4 can create excellent documents from standard writing formats. Examples include recommendation letters and grant proposals. OpenAI has trained it with millions of examples. Letters of recommendation need to be personal. But GPT-4 can help produce the first drafts of a grant proposal. Or at least a skeletal outline. That has saved me a good bit of time.

(2) GPT-4 writes good paper abstracts from an introductory section of a paper. In general, the more content you give GPT, the better results it produces. The more complex the content, the more it can enhance the product. I can put a long introduction into ChatGPT and ask it to make a shorter abstract, and it does a decent job. Note: ChatGPT can only count words if you finagle it a bit. But it can replicate a good abstract based on millions of examples. Writing abstracts is now less tedious.

(3) GPT-4 is bad at generating objections to arguments. Yet, it is not terrible. Ask GPT-4 to generate objections by providing a short argument or paper section. It may give you five or so, four of which are no good. But sometimes, one of them is good enough to see humans wanting to advance it. GPT-4 surprised me here. I expected it to be 100% bad, but it is more like 90% wrong. In the academy, asking others to read your work takes effort and commitment. But there's a faster way to improve your papers. You can also ask GPT-4 to repeat your argument, which is helpful.

(4) Academics hate meeting word limits for submitting articles to journals. But GPT-4 is fabulous at cutting words from a paragraph or two. Follow my advice in the next point with a few paragraphs. You’ll soon approach the word limit. (If you’re trying to cut 100-300 words out of a 10,300-word piece).

(5) I write rough drafts before editing and adding citations. ChatGPT excels at editing rough drafts. It knows how to reduce the reading grade level of text. Some reading grade-level algorithms are old. For a century, grade schools have needed ratings to assign grade-appropriate books. So, if you have a paragraph, use the Hemingway Web app to generate the reading grade level. Then drop the paragraph in GPT-4 and tell it to reduce the grade level to grade 8 or 9. It helps me see tons of extraneous text, sometimes as much as 40% of my paragraphs. Also, do go paragraph by paragraph. The more text you give GPT-4 at one time, the worse its edits. It’s also good at the sentence level, but I focus on paragraphs.

Here, I’ve had great success. I can often reduce the number of words in a rough draft by 30-40%. I took one rough draft from ten thousand to six thousand words in a few hours.

Now, the text it produces is weird and child-like. You can’t rest with that text. It breaks the text into small portions, too small for ordinary readings. Combining short sentences into long ones is more straightforward than shortening long sentences. So, the text atomization saves time.

ChatGPT could be better at varying sentence lengths. If you don't connect it to Wolfram Alpha's web app, it won't count the words it makes. Even with WA, the counting is very slow. So you’re going to have to vary sentence lengths yourself for now. I use Pro-Writing Aid to measure sentence lengths and edit within that app.

(6) GPT-4, even Bing’s version, is still not great at finding sources, though it is much better than a few months ago. So, it is OK to use Bing for citations (owing to its live internet connection). I still don’t use ChatGPT to produce sources. However, it can often find older sources in the public domain. OpenAI likely trained it on those texts.

(7) Generally, the more you know about a subject, the more ChatGPT can help you write about it. You can find mistakes fast, and the tech gets smarter as you ask more intelligent questions. Specific questions are better than general questions.

(8). Besides writing, I use Bing on Creative to answer other questions. For example, I can learn how to solve annoying problems in my house. I can find remedies for minor health issues, improve my diet, and plan workouts. I can even discover fun games to play with my kids when I'm out of ideas. Asking GPT-4 for help is cheap and could have a big impact.

* I also recommend ignoring Mushtaq Bilal. He has sent me down many rabbit holes. Moreover, he advises people on how to improve their academic writing, but that skill doesn’t show up in his CV. He has 35 citations. By contrast, Ethan Mollick has 10269. Bilal also insists on payment for all his speaking events. And he reacts with indignation when he’s invited to speak for free. I know because he tweets about it. I find this intolerable.

[Please subscribe, but do NOT upgrade to paid! I still can’t change the button.]