What I learnt giving my second online AI Fluency session
Here’s the recording of the session:
I did the second session of my AI Fluency online trainings yesterday. It went well, apart from a couple of technical issues. It’s amazing how much tech is needed for these sessions. I could probably use a lot less but I want them to be as good as possible.
Technical problems
My original intention was to use OBS Studio and the macOS Background removal plug-in (or this alternative, which works in a technically different way and is also good) to overlay myself on the slides and to record the OBS stream and use the OBS virtual camera with Google Meet. (I didn’t have the problem that I had the previous week, of Google Meet not being able to see the virtual camera). However on testing this earlier in the day I ran into two problems:
- A bit of my grey hair is a similar grey to the background, leading to an unfortunate hole in my head (I could fix this by lighting the background white wall better).
- During the test the camera input froze and I had to fix that by restarting things. I don’t want that to happen during a live broadcast.
So just before the call I decided to do what I’d done on the previous call — record my desktop and the video camera stream using Quicktime. However, a minute into the live recording Google Meet dropped the screen sharing and I got an OSX message saying my hard drive was almost full. So I just stopped the two recordings and shared my screen again. So the main lesson of the session was: make sure to have plenty of hard disk space to record during the session. To be on the safe side, say 100Gb free.
In fact, I’ve discovered that when using Quicktime during the recording it saves in a Lossless format, so 4K uncompressed video generates approximately 3GB per minute! That’s 60Gb for a 20 minute video. So a better thing to do is to save the Quicktime 4k camera output to an external drive with plenty of free space.
Also next time I will make the Sony Alpha 7iii export 1920x1080 rather than 4k. The difference in quality in the end result will probably be very small, whilst it will make the process much more efficient.
One other small learning point: Put your phone in silent model when presenting.
Using a teleprompter
A few years ago I purchased a Neewer Teleprompter, but never really used it. I decided to try using it this session. Apart from the physical setup, I did the following:
- I created a white text on black background landscape document for the text to appear on the teleprompter. I created one page per slide.
- I set up my Stream Deck with keys to change to previous and next slides, and to go to the previous and next pages in the PDF document using Preview. This worked really well.
- I used OSX BetterDisplay to flip the iPad display of the iPad I was using to display the teleprompter text. This was a bit of a fiddle to get right, but it works.
- Preview handles keystrokes different depending on whether it has been made full-screen using the green window button, or selecting “Full screen” from the menu. For the Teleprompter slides, put Preview into full screen by selecting “Full screen” from the menu.
Having done all of that the teleprompter worked really well. I just had main bullet points displayed, and a few facts that I didn’t want to forget. However, I found that I didn’t really read the text displayed on the teleprompter at all. I’m not sure if that’s because I’m not used to it, or because I just don’t need it. I tend to talk about things I’m interested in, so don’t need a prompt. Sometimes I do forget to include things in presentations, so the teleprompter may help me avoid that. More experimentation needed.
Setting up with BetterDisplay
Once the iPad is added as an extended display:
- Open the main settings window.
- Click “Create new virtual screen”
- Choose “Match aspect ratio of and associate to a display
- Associate the virtual screen with this display
- Go to the topdown settings
- Change the resolution for the iPad to the lowest
- Stream virtual display to the iPad (itself)
- Select the horizontal flip
- Put the PDF on the iPad using Preview
- Select “Full screen” from the menu.
Final Cut Pro disk space usage
FCP was the cause of my low disk space. Having edited the first video last week, it has filled my hard disk with its caches and pre-renders. I spent some time fiddling about working out how to move all it’s various files to an external drive before deciding on a simple method for next time: Do one FCP project (i.e. one session) per Library and just draft the whole .fcpbundle file to an external drive for storage once it’s done. This will free up a lot of space on the local hard disk and keeps the process very simple.
Session feedback
Giving the talk went fairly well. I still say “umm” and “err” too much, but I can work on that. At the end of the session I asked for feedback via a Tally form. I got good feedback from the session, with seven responses. Here’s a summary (by Claude Sonnet):
Feedback Summary:
Strong positives across all responses:
- Clear, accessible explanations for beginners
- Appropriate pacing (“digestible pace”)
- Visual demonstrations effectively illustrated concepts
- Your illustrations are well-received
⠀Content effectiveness:
- Base model concept landed well
- Tokenization demos were particularly valuable
- “Lossy compression” analogy resonated
- OpenRouter tool sharing appreciated
⠀Areas needing clarification:
- Token splitting logic still confusing (“why tokens are sometimes split words”)
- Next token prediction concept needs more emphasis
- Connection between base models and structured AI systems unclear
⠀Technical issues noted:
- Recording failure disappointed attendees
- Need cleaner video starts (remove waiting time/false starts)
- Consider 1.5x speed for recorded versions
Engagement level: Seven responses from a session suggests solid engagement. The detailed feedback indicates people are thinking critically about the content rather than passively consuming.