How to use GPT-4 to analyze and summarize any textual content

Although I have been primarily writing about generative artificial visual intelligence, like Stable Diffusion and Midjourney, I have also been thrilled about the new GPT-4 released recently. The GPT-4 is an Multimodal Large Language Model (MLLM) utilizing Natural Language Processing (NLP), so it is very skilled in summarizing texts and also images (this is where the multimodality refers to).

Basically, this means that you can insert any text into GPT-4 (and soon also images) and ask it to analyze the content, and by telling what you want to know from the text, it does the job in a few seconds.

In the next article, I will demonstrate this feature with a YouTube video, as this feature might not necessarily occur in everyone's mind at first. ??

To analyze YouTube video content with GPT-4

To summarize a YouTube video with ChatGPT-4, you need the transcript (i.e., what is said in the video). And as many might know, at least the videos in English have transcripts nowadays.

In short, copy+paste the transcript after your prompt to GPT-4, and voilá!

To provide a bit more detailed instructions and some extra tips and tricks, I wrote the following 4-step guide.

By following these four steps, I bet anyone somewhat skilled in English can get the job done. ??

1. Under the selected YouTube video screen, there are three dots. Click the dots, and you will get the how transcript button visible, press it. You can include the timestamps, but I noticed that the timestamps will drastically consume the tokens, so the length of the video GPT-4 can analyze will be much more limited. To eliminate timestamps, click the three dots next to “Transcript” and toggle timestamps (on/off).
2. Copy the transcript and jump into the ChatGPT (by the time writing this, GPT-4 is available only for paying customers), create a new chat and remember to choose GPT-4 from the top menu.
3. Now tell the GPT-4 that you are about to give it a video transcript and ask it to analyze the content. Do not rush, and pay attention to the prompt you are providing. Use good English, be consistent (i.e., do not use different terms from the same thing), and be logical in the process; first, analyze, then tell the format you want the info, then what info you want in what order, and then press Enter to submit the task for AI.
4. After the analysis, you can ask more questions about the content, or if the answer misses the point, you can edit the question presented earlier and it will try again. This might be a good thing to do, as GPT-4 remembers each chat (but cannot know the content of other chats); keeping the wrong data in the chat can mess up the results.

A cool trick for analyses is to know that ChatGPT can draw tables. To demonstrate this, I prepared the following example from the video “GPT 4: Full Breakdown (14 Crazy Details You May Have Missed) - Last One is Extra Wild”. I did what I previously explained in the 4 steps and gave ChatGPT the following:

Prompt (without the transcript, which is too long to be included here)

Analyze the “transcript” between [ and ], and then do a table about the 14 Crazy Details of GPT-4 provided in the transcript. Sort these 14 in order of importance if possible. If not possible, sort details in the order of appearance. In the table's first column, insert the feature name; in the second column (if possible), what the feature does; and in the third cell, how to use the feature (if applicable). After the table, write a summary of the transcript in general.

The answer, generated by ChatGPT-4

I copied the table ChatGPT-4 provided and took it to Google Sheets. The info also converts smoothly to Excel if that suits you better. This allows, for example, to adjust of the table's appearance and next-level data analysis. Depending on the data, you can, for example, generate pie charts, run more detailed statistical analyses, and so forth. ??

"The transcript discusses the release of GPT-4 and its various features and capailities. It mentions that GPT-4 powers Bing, has doubled context length, and has withheld model training details. The model shows improved performance in tasks like the bar exam and hindsight neglect tasks, as well as better factual accuracy, image-to-text capabilities, and multilingual support. However, the transcript also discusses safety concerns, such as generating undesirable content, realistic targeted disinformation, and emergent behavior like power-seeking. The model has been tested for potential self-improvement capabilities, which may have implications for future AI development."

And although the table and the summary are not perfect, they still provide a very good and fast overview of the video --> I need to repeat myself with this ?? emoji!

As, of course; this means you can do the same with any textual information; news, books, articles, research papers, etc., that fits within the limits of ChatGPT-4 current capabilities. These are presented shortly in the following.

Some limits and notes of ChatGPT-4 when analyzing texts

However, there are some* limits with ChatGPT-4:

• For regular users, the length of the text inserted in chat cannot be more than 8192 tokens (which I estimate to equal roughly 12 pages of text); for selected users, the limit is about 50 pages of text (I.e., 32,768 tokens). Source: https://openai.com/research/gpt-4
• In the ChatGPT-version, uploading documents or inserting images is impossible (at least on 19.3.2023). I assume we will see these features come out while the developers can get stuff done through API. This will allow GPT-4 to read and analyze research papers with graphs and images, for example.
• GPT-4 is not flawless and is also prone to hallucination. It is better than previous GPTs to admit that it doesn’t know, but there definitely is a need for fact-checking, as AI cannot be held responsible for the insights it provides.
• The shorter inserted text is, or the simpler the prompt, the more accurate the responses are. (Even a minor change in prompt can drastically affect the results received – i.e., as said, be careful and consistent with the language.)
• Also, the quality of the transcript and the nature of the video (like how much it relies on visual elements not appearing on the transcript) affect the analysis.
• Based on my tests, it also seems that GPT-4 can get confused, especially if multiple languages are used in the same chat. So even though that is possible, I recommend sticking to one language (from which English is the best option). And if the results do not seem accurate anymore after tinkering with prompts, it might be better to start from scratch.
• Keep in mind that the summary is a summary. For example, by watching the entire video GPT 4: Full Breakdown (14 Crazy Details You May Have Missed) - Last One is Extra Wild used as an example, you will learn tons of more valuable information from GPT-4. But when a general overview is good enough, seek no further.

* Please note that the list of limits or notes is not all conclusive, and all here is based on empirical tests and various sources, which are not all fact-checked.

Summary

(Chat)GPT-4 is transforming both content analysis and production rapidly. There is the time before NLP LLM and the time after. We have now entered the latter.?This might have enormous consequences, which I have addressed in this recent article, "Get Ready for Super Smart Robots in the Future!".

And If someone wonders why I share this info for free, it took around four hours to write this text; the reason is simple. This is just the tip of the iceberg. By sharing AI info, I hope to help people understand the huge disruption we are experiencing now. From which I am happy to come and tell more with a reasonable fee. Don't hesitate to contact me via LinkedIn DM to learn more. ??

#ai #artificialintelligence #gpt4 #chatgpt #openai #video