AI Meeting Notes: The Hardest Part of a Chinese Lab Meeting Isn't Understanding—It's Keeping Track of Who Decided What
Our lab meeting is once a week, two and a half hours, with seven or eight people, all in Chinese. It sounds like there shouldn't be any problem.
But after every meeting I always have this feeling: I understood everything the whole time, yet once I walk out of the room I can't clearly say "what exactly was decided just now, and who agreed to do what."
The difficulty of a Chinese-language lab meeting isn't the language; it's attribution.
Three Things That Get Lost in Meetings
Who raised it. Two people may speak one after another on the same topic, and their views may differ. After the meeting, I only remember "this direction was discussed," not who raised it first.
Whether it was decided. When the advisor says "we can try it first," it sounds like agreement, but it actually means you should judge for yourself; when they say "let's look at this again," it literally means putting it aside, but it may actually mean opposition. This kind of judgment isn't in the literal words.
Who took on the task. The five minutes before the meeting ends are the most chaotic; several people talk at once, tasks get divided up through back-and-forth pushing, and no one writes them down.
These three things can't be solved by "understanding."
I Close Things Out with a Three-Column Table
The transcript and segmented summary come out first—I use MuchSci's AI Meeting Notes to get the result of real-time listening and translation plus automatic segmented summaries—and then I go through it myself using three columns:
· Topic column: how many matters were discussed in this meeting, and who raised each one. If the topic changes midway through the meeting, I split it manually: the segmented summary is split according to speech pauses, and when the advisor suddenly shifts from discussing A to B, these two parts often end up in the same segment; for that segment, I listen again alongside the translation to find the cut point.
· Conclusion column: which category each matter ultimately falls into—"decided," "continue discussing," or "hold for now." I have to judge this column myself; the segmented summary only says "X was discussed," not whether "X was decided."
· To-do column: who, by when, and what to deliver. Write it down with specific people, not "follow up later."
I don't let a tool handle attribution for me. The reason is straightforward: I've checked several meeting records, and the segmented summaries don't contain reliable speaker attribution. So I fill in who said what and who took on what myself, using the names; for anything I can't fill in, I ask directly in the group chat within ten minutes after the meeting instead of guessing.
Where to Store the Three-Column Table
The lab meeting happens once a week, and after a dozen or so meetings pile up, looking through them becomes a mess. I organize the "To-do" column by person and week into MuchSci's Personal Database, mainly to check two things: who didn't submit this week, and who has been assigned tasks for two weeks in a row. When I need to show the table to others, I use MuchSci's Office Assistant Agent to format it into an editable version, which is easier to use than a screenshot.
What I Don't Record
Lab meetings often involve unpublished data, details down to sample numbers, and other people's project progress. My approach is: content involving other people's data does not go into shared spaces; my own records are stored only in MuchSci's Personal Knowledge Base under my own account, and I only put in the parts relevant to me. Before the meeting, I say that I will be taking notes today, and if someone says not to record a certain segment, I skip that segment.
This is also the line I draw for this kind of graduate student AI assistant: it can help me keep the words, but it cannot make decisions for me about how I deal with people.
Cost
The lab meeting is a fixed weekly expense, so I count it as a fixed item in my monthly spending; the occasional extra-long meeting counts as a variable item. Chinese meetings don't need translation, so credit consumption is somewhat lower than for English meetings, but after two and a half hours it's still more than I expected. So I skip the topic segments unrelated to me and don't listen closely to them, and only listen closely to the part I'm responsible for.
Now after a meeting I only keep that three-column table. Not understanding the meeting is one thing; after the meeting being unable to say clearly who decided what is another—the latter is what actually causes delays.