The AI's summary ability is the most basic and most frequently used scenario when AI emerges.
In daily life, I can quickly read so much content. Maintaining a high-intensity output every day also relies heavily on the AI's summary ability.
Nowadays, almost every AI assistant has a summary function, whether it's a web version or a browser plugin.
From my experience, there are very few products that do well in summarizing long content.
How can we better summarize long texts?
Summarizing and paraphrasing content and then understanding and recording it in our own context is an essential path for us to learn knowledge.
The purpose of summarization is to compress knowledge in our own way and make it accessible when needed.
Tiago Forte is the founder of Forte Labs. He has been helping people create their own "second brain" in a systematic way.
He proposed a method called "progressive summarization" to help people establish searchable and reusable notes.
This is a technique for designing easily discoverable notes by compressing information layer by layer.
Progressive summarization consists of a total of five stages:
Layer 1: The starting point of progressive summarization, like the cornerstone on which everything else is built, is basically the original text.
Layer 2: This is the first round of real summarization. Look for keywords, phrases, and sentences that seem to represent the core or essence of the discussed viewpoints. In this part, value judgment needs to be done to see if it's worth recording and preparing for layer 3. For me, it's also whether it's worth posting.
Layer 3: Use highlighting. Among all the bolded paragraphs, find a smaller number of highlighted paragraphs. The pursuit is the "cream of the crop." Highlighting is only done where it's truly distinctive or valuable. And layer 3 is only added when reviewing notes.
Layer 4: Although it's still summarizing, it has gone beyond emphasizing others' words and uses our own words to record.
Layer 5: For a very small number of resources, that is, those that we want to immediately become a part of our thinking and working methods, we integrate them. After analyzing from all angles of layers 1 to 4, we add our own thoughts and creativity to turn them into something else.
What requirements should a good AI summary tool meet?
Progressive summarization basically covers all stages of compressing and summarizing long content that we need. So we can evaluate and design AI summary products based on this principle.
So a good AI summary tool needs to meet the following requirements:
Have a long enough context, support uploading large files and multiple files, and support various document formats or content (this is related to the model and RAG ability).
Need to ensure comprehensibility, reasonable layout, not lose key information, and the summary needs to be logically complete (most summary tools have serious problems here. LLM is lazy and repeats the title).
Ensure discoverability. It can't be so long and messy that almost the entire text is spit out (this part occurs less frequently. The current problem of LLM is still outputting too little rather than too much).
In the third layer of summarization, it needs to identify the most valuable content and display it concisely for users to select.
The summarized content and the content that users communicate with AI according to document items need to be convenient for retrieval and review to facilitate completing the third and fourth layers of summarization.
Case analysis - How to achieve these requirements?
A few days ago, when testing these AI tools, I threw a long document to Tencent's Yuanbao and asked it to summarize. As a result, I found a button saying "Read this document in depth" following the output.
After trying it, it gave me a big surprise. For the tasks of long content summarization and in-depth reading, I now almost only use Yuanbao.
Let's look at the first part of the requirements above. Yuanbao's in-depth reading supports single files up to 100MB in size and can handle up to 256K tokens, which is close to 500,000 words of context. This may be the product with the most supported context I've seen except for Gemini.
In common document summarization and paper summarization scenarios, there is almost no need for more than this magnitude.
Here we use the paper about their own AI system released by Apple a few days ago for testing. This paper has as many as 47 pages.
After entering the in-depth reading page, the top layer is mainly divided into two parts, namely intensive reading and the original text. Intensive reading consists of three parts, namely core overview, paper evaluation, and key questions and answers.
The core overview part corresponds to layer 2 of progressive summarization. It shows you the complete content and logic of the entire article, allowing you to quickly judge the value of the content. Moreover, it doesn't have the lazy problem that common AI summary tools have. It even extracts the pictures in the PDF and does a mixed layout of text and pictures to help you understand.
Then the parts of paper evaluation and key question answering correspond to layer 3 of progressive summarization. By making LLM reflect actively, it selects the most valuable parts of the content for users to choose.
For example, the content of dataset collection and processing, performance improvement in the post-training stage, and the balance between low latency and capability that LLM actively finds here are all relatively crucial parts of Apple's AI system.
If I were to take notes, I might also note from these points, especially the part about achieving low latency.
In addition, Yuanbao also provides a remedial solution in case LLM fails to find the corresponding high-value content. You can choose to ask questions in the upper right corner and directly ask Yuanbao about the corresponding content.
Moreover, the content of this question is bound to the document. When you come back next time, you can directly view the content you asked last time without having to find the chat record and scroll for a long time to find it. And the chat record for in-depth reading is independent from other conversation records.
Compared to other AI assistants, it also does a good job in content retrieval and review.
In addition, you can also switch to the original text reading by clicking the top TAB. The shortcuts for selecting content when reading are also very appropriate. You can translate the content and directly call AI search to search for related knowledge that you don't know.
At present, the only problem is that the trigger entrance is a bit deep. Separately, it's actually quite good. Otherwise, every time in order to trigger it, you have to go through the main process first, which not only wastes tokens but also wastes time waiting for the first output.
In addition, it would be even better if the summarized content supports highlighting and then being collected in one place. It can also increase user retention.
In fact, we mostly deal with web pages on a daily basis. I hope a browser plugin or a way to summarize link content can be developed soon. I tried giving it a web link directly. It can summarize, but it cannot trigger in-depth reading.
If you find this helpful, you can send the content to your friends who you think need it🙏.