Find out how we added automatic topic-detection and note outlining to jamie.
A couple months back, our team at jamie started to work on one of the most exciting product ideas we had in a while. An AI assistant that automatically generates meeting notes for you. We have announced it to the world in late October and have gotten hundreds of signups within a few days. That's when our journey of building the world's best meeting summary software has started. In this blog post, I want to share an update on how far we've come and what exciting new things we have been building to make summaries better for our users.
When we started building the first version of jamie, we had a narrow scope. Our goal has been to get an early version into the hands of users as quickly as possible to test whether people cared about it. That's why we focused on the following feature set:
With this in mind, our team started working pretty much day and night to get a first version of this automatic meeting summary assistant done as quickly as possible. After a few weeks, we were ready to test it the first version.
We gave this first version to a selected group of early users to see how well it does in the real world. The reactions were great. People were impressed by how well an AI can capture contents and rephrase them based on the context. jamie generated a long list of meeting notes combined with a quick overview of the highlights discussed in a meeting. Something we thoughts was already quite impressive. Nevertheless, we quickly learned that an essential part was missing: topic outlines of notes. For jamie to be useful beyond the first sight, this had to be added.
If you have written meeting notes before in your life, you know how important it is to have a good outline in your notes. For example, if 3 completely different topics were discussed, it would be best to have 3 headlines below which the notes are written in a sorted way. No one wants a long list of meeting notes without structure. That's why we started to think how we can engineer a solution around this challenge.
After learning that structured meeting notes are a must for our product to be useful in this field, we spent weeks exploring different ways to implement this. The most important consideration has been to make it seamless for users. Our aim was to build a magical experience that delights our users. Luckily, thanks to some clever engineering on top of GPT-3, we were able to find a way to make this experience come true.
In essence what we are doing in the background when generating summaries for meetings is this:
What sounds simple on first sight is powered by one of the most impressive large language models (LLMs) the world has seen to date: GPT-3. By fine-tuning and engineering clever algorithms around natural language processing with GPT-3, we have been able to implement this functionality for every user going forward.
As mentioned, the idea of having an automatic outline for meeting notes has now become reality. Hence, we rolled out this new feature for all users as of this week.
So what do you need to do to take advantage of these changes? Well, nothing. From now on, all generated summaries support this topic-based outline without you needing to change anything. However, keep in mind that outlines are only generated if the meeting is sufficiently long (around 10 minutes). Luckily, in a real-world setting, this should be the norm.
Additionally, we have improved the executive summaries that are being generated based on the meeting notes. They now capture the essence of what has been discussed more accurately. The default output generated by jamie are meeting notes and an executive summary. So no matter what you are looking for in your workflow, jamie has got you covered.
Further, you can quickly change or rephrase topics or meetings notes right in jamie. This way, you can cut down the time you spent on writing meeting notes from 10-20 minutes to less than 120 seconds.
A question we get a lot is what we will do going forward to improve the quality of summaries even more. There are two activities we are focusing on.
Firstly, we are continuously optimising our algorithm for the summary generation based on user feedback. Every user can rate the quality of a summary which we are using as input to optimise the way we are using GPT-3 and the output we are generating. This means that jamie will get better over time continuously with increased usage.
Secondly, we are working on an implementation to identify speakers in a meeting based on their voice. This would allow for more accurate notes and clearly depicting who said what in a meeting. This is another core improvement we are expecting to ship within the next months.
If you've gotten interested in trying out jamie and seeing how far this technology has come, feel free to sign up and try it for yourself. We offer 5 free meetings for every users so you can convince yourself of the quality.
Fun fact: you can also use jamie to summaries lecture, YouTube videos, podcast, or whatever you want. So get creative with the use cases ;)