"Anyone who has ever used ChatGPT's artificial intelligence (AI) knows the problem: the list of queries made is getting longer and longer and is difficult to search. And why do you have to constantly tell the AI what to pay attention to? Certain settings and a note app can help.
After the initial fascination with ChatGPT last year, the hype has subsided for some: the machine makes mistakes in the details. It is important to tell her exactly what she should pay attention to. It helps to imagine the following: The digital journeyman is like a 14-year-old student intern in the company. You have to tell him exactly what he should do and how.
A prompt, the instruction to the machine, could then be:
"You are the leader of the 'Expansion to France' project. Create a short, clear and unambiguous summary as an 'Executive Summary' for the company management. Your summary should be written professionally, formally and objectively. Support your statements with relevant data and examples and focus on the three most important points from the attached project report. Write a maximum of six sentences per paragraph and limit your summary to a maximum of five paragraphs. Make sure your language and style are appropriate and coherent and that your information is in context. Relate solely based on information from the detailed project report below."
ChatGPT then complains in the paid version GPT-4 that the "attached" text is still missing. In the free version 3.5, however, the machine simply chatters away meaninglessly without knowing the project report - not a good start for the digital colleague.
The four version can do that better. If you then copy the project report into the input field, ChatGPT usually delivers as ordered. Provided that the length of the report does not exceed any limits. This is 4096 so-called tokens, which are fewer than 4096 words, in the case of long words syllables, sometimes also individual characters. If the document is longer, you can use special AI programs for larger documents, which we have already presented. These days, the provider OpenAI is also rolling out a way to upload PDF documents directly to GPT-4. Previously this was done via a special “Advanced Data Analysis” setting.
ChatGPT does not always work correctly. For example, the machine sometimes ignores the requirement of six sentences in our prompt. The guy understands a subtle question, “How many sentences were there?” and apologizes. To a clearer “Sit, six!” ChatGPT then responds as requested. Also the note “Make the sentences shorter and more precise!” the digital intern accepts without complaint. Quality comes from torment.
Chats with the machine accumulate over the months. These chat histories are difficult to search on ChatGPT. The above-mentioned prompt for an “executive summary” can then be saved separately in a notes app on your computer or smartphone, for example. From now on, I keep copying such prompts for summarizing texts back and forth. But also others:
"Please check the plausibility of the following three texts. Find the errors. And make suggestions for open questions that the author should investigate."
"Make me a marketing plan for the project. Name the most important target groups and give me an overview of the best types of advertising to address these target groups."
"Create a table with the columns: keyword, Facebook post, tweet, LinkedIn post, based on the checked texts. Use emojis on platforms where it is appropriate and common. Do not address the readers directly. Sound serious and neutral, not advertising. At the end of each post, make a reference in brackets to the best broadcast time (day of the week and time) in order to reach as many readers as possible on the platform. Show me the table. Then give me the table again to copy out, but this time as a .csv file. Use semicolons instead of commas to separate the cells."
The answers to these prompts are rarely perfect the first time. But they usually lay a good foundation for post-processing. That's how it is with 14-year-old digital journeymen: not yet dry behind the ears, but malleable - and always subject to critical questioning. For example, we cannot confirm the best broadcast times for social media publications.
If you want to go deeper into prompt engineering, Jina.ai from Berlin will find a powerful tool for improving prompt designs. The Promptperfect application generates more detailed stage directions for the machine - including for image generators - from initial ideas for a prompt. The service starts at around $10 per month.
At ChatGPT you can specify how your digital buddy should work in two places. Firstly, with the “Custom Instructions”. On the other hand, “Plugins” (extensions). It works like this.
Custom instructions tell the machine who you are. "I am a journalist who likes to write clearly and simply. I write German and am very meticulous with the facts." This could be entered under "Settings" (via the menu with the three dots) and be even more detailed.
And how do you think ChatGPT should respond? The service also provides for such a field. "Don't make up false things or speculate. Stay neutral and objective in everything. Use the temperature 0.1 or 0.2."
The latter, the “temperature”, describes in AI terms how predictable the machine should be on a scale of 0 to 1. The smaller the value, the more closely the machine adheres to the specifications. On the other hand, if you want to develop crazy ideas for a marketing campaign, for example, you would rather specify a temperature of 0.8 or 0.9. Motto: "Make me a marketing campaign with ten crazy ideas for the project. Use the temperature 0.9."
As a result, the machine then suggests, for example, lighting the Eiffel Tower with the company logo, flash mobs or retro advertising in the form of nostalgic French themes with a modern twist.
If you avoid the word “crazy” in this prompt and specify a temperature of 0.1, the digital journeyman recommends more conservative methods: market research, online marketing, press work.
At ChatGPT, plugins are extensions that can include a live search on the Internet, allow the generation of images or the ability to upload your own files such as PDFs. This is only possible with GPT-4, the paid version.
And in the end, how do you find that ingenious prompt with which you elicited the marketing campaign for the Austrian market from the machine a quarter of a year ago? The complicated method is to scroll back endlessly in ChatGPT. It's quicker with the following trick: In the settings using the three dots, select the menu item "Data controls" > "Export data". If “Chat History & Training” is activated, you will receive a link to download your past activities by email minutes later. Once downloaded, this data will be displayed as a locally saved web page. It can be searched for a keyword in the browser using CTRL-F or Command-F on the Mac.
If “Chat History & Training” is deactivated, the service deletes the data after 30 days. Something else becomes clear from the setting: When activated, ChatGPT can certainly use the prompts entered by humans to further train its language model.
When asked about this, the guy reacts like a bad intern: "I apologize for the confusion." So be careful about everything you discuss with the guy." [1]
1. Wie man ChatGPT effizient nutzt. Frankfurter Allgemeine Zeitung (online) Frankfurter Allgemeine Zeitung GmbH. Oct 31, 2023. Von Marcus Schwarze
Komentarų nėra:
Rašyti komentarą