Research Video Collaboration
Our team used a few different tools to collaborate and produce our final research video project on how LLMs work. For planning, communication, and collaboration, we used Slack with a Canvas page to keep track of project details and learning resources.
To get an idea of how to approach the project, we used ChatGPT to generate an initial script for a video on the topic. From that script, we split up the video into sections and assigned sections to individual team members. From there, we edited and iterated on the script, focusing on our individual sections and contributing our own understanding of the topic.
To produce the actual videos, we used two tools: Google Vids and Adobe Express. The assignment involved making two separate videos: one for a professional/academic audience and one for a more general audience, so that it could be understood by a kid. We found that Google Vids was better for making the longer professional presentation, and Adobe Express was better for making short-form content.
The process was relatively smooth, though there were some bumps in the road. For example, we initially planned on using Google Slides and recording our presentation with Zoom, but then we found Google Vids, which turned out to be a better tool for the job. Google Vids has a feature to import Google Slides though, so it worked out pretty well.
Weekly Summary
This week involved reviewing and reflecting on a TED talk of our choice. I found a talk by Yejin Choi called Why AI Is Incredibly Smart and Shockingly Stupid. In the talk, Choi explains how AI can sometimes fail at tasks that humans can easily accomplish. For example, when asking a question with a straightforward answer, GPT-4 gives a convoluted and incorrect response:
The solution to these kinds of problems, Choi argues, is to teach AI to learn through a process called Symbolic Knowledge Distillation. In this process, a larger model “teaches” a smaller model, which results in a discrete, human-readable knowledge graph with corresponding neural weights. The resulting model is more efficient and understands commonsense reasoning better than its teacher model.
The promise and peril of AI podcast from Harvard raised some interesting questions about ethics and AI. The discussion brought attention to the uncertainty surrounding AI and its effects on society, despite the many benefits it presents, as well as the potential need for regulation. Embedding ethics into computer science education at Harvard was mentioned as a way to encourage ethical thinking in technologists. Another interesting point was the potential for AI to both level the playing field by empowering lower income workers to improve their skill-set, while at the same time amplifying existing inequalities by reducing the need for entry-level workers.
This week also involved some lessons on presentation skills. Slides decks like PowerPoint can often be useful for presentations, but delivering an effective presentation requires more than just reading bullet points. Giving a good presentation communicates an idea effectively, which involves knowing your audience and making sure to explain things in a way that they will understand.
Toastmasters offers some helpful advice on giving a technical briefing: know your audience, state the purpose of the presentation, organize the material in a logical fashion, and summarize the main points. I’ll make sure to keep this in mind as I develop my presentation skills.