AI Lecture Notes: After a Public Lecture, Why Are the Notes Still Unusable?
Last month I attended a two-hour public lecture on a topic adjacent to my research. I sat in the third row, filled four pages, and felt I had gained a lot as I left.
A week later, when I needed to write the background section of my thesis proposal, I opened those four pages. Only three sentences were usable: one was the speaker's conclusion, one was the number of a figure, and one was the name of a database he mentioned. The basis, conditions, and scope of application—none of it was there.
The notes were unusable not because I wrote too little, but because what I wrote could not be connected back to an argument.
Three Ways My Notes Failed
At the event, I could only note conclusions, not the basis. A conclusion the speaker derived in two minutes became seven words on my page. Seven days later, looking at those seven words, I had no idea how they had been reached.
A memorable quote out of context is just decoration. The sentence I copied was "holds under a certain condition" in the lecture, but once on paper it became unconditionally true.
Chinese and English terms did not match up. The speaker spoke in English, and I took notes in Chinese. When I later tried to recall the original English term, I could not, and I had no idea how to search for it.
I Do Only Two Things During the Lecture
First, mark only anchors. When I hear something related to my research, I write one word plus an approximate time on the paper, such as "Figure 3 / multi-omics / around 40 minutes," and stop trying to capture complete sentences. Ten anchors on one page are more useful than four pages of sentences.
Second, process it immediately after the event. The first thing I do back at my desk is use MuchSci's AI Lecture Notes to transcribe the entire lecture—while my memory is still fresh, I know roughly where the anchors I marked are, so they are quick to find. For English passages I am unsure about, I check them against MuchSci's Paper Translation.
For a two-hour recording, transcription plus listen-along translation and automatic segmented summaries usually take less time than flipping through those four pages.
Afterward, I Follow Four Steps
Step one: read the segmented summary first and mark the two or three sections relevant to my research. I do not look at the details of the other sections; I only check where the segments are divided and skip them once I confirm they are not what I need.
Step two: listen closely to the marked sections. I listen to the original audio alongside the translation, focusing on "under what conditions this conclusion holds." This step cannot be skipped; conditions are often mentioned only briefly in speech and are easily dropped from summaries.
Step three: verify the literature mentioned in the lecture. For the papers whose author names and years the speaker announced, I use MuchSci's Paper Search Assistant to find the exact sources and copy the DOIs accurately. It searches Crossref's 180 million records, which is much faster than my old habit of guessing through search engines—but I still manually adjust the final citation format according to my school's template.
Step four: save it and write three lines of my own judgment. I save all the materials from the lecture in MuchSci's Personal Knowledge Base, and separately write three lines: how this public lecture relates to my research, which point I can use, and which point I disagree with. Without these three lines, a month later it is as if I never attended.
For the handouts or reference materials the speaker shares, I put them together with this record in MuchSci's File Asset Library, saving myself from searching everywhere six months later for "where that thing from back then is."
Boundaries and Costs
· I do not share the speaker's unpublished data externally, nor do I put it in any shared space.
· I do not quote sentences from the translation. If I need to quote, I go back and listen to the original audio and write it according to the original wording.
· I verify numbers and figure conclusions from the lecture against the original text.
· This kind of AI assistant for graduate students should not be treated as a tool that "attends the lecture for me." It saves things for me, but I still have to make the judgments myself.
· In terms of cost, transcribing and closely listening to a two-hour lecture consumed more credits than I expected. Later I split my approach into two types: for lectures far from my research, I only do transcription and segmented summaries, without close listening; only for lectures close to my research do I do the full process.
Now I have only one criterion for judging lecture notes: can I state "under what conditions this conclusion holds"? If I cannot, it means the notes are still just paper, not material.