AAAI 2025 Educational Video

Project Overview

AAAI, as a top-tier academic conference in artificial intelligence, launched an Educational AI Video competition in 2025, aiming to popularize knowledge across various AI fields through three-minute short films. This was an excellent opportunity for me to combine my experience in language model safety with my passion for video production. Ultimately, I entered the competition with a video exploring attacks and defenses in language models.

The final video published by AAAI

Video Content

The core objective of this video was to clearly explain "Jailbreak Attacks" and their defense mechanisms in large language model security to a non-expert audience within three minutes. To achieve this, I adopted the classic "problem-solution" narrative structure, using concrete examples and vivid analogies to transform complex technical concepts into easily digestible content.

Opening: An Intriguing Hook (0:00 - 0:30) The video begins by posing a thought-provoking scenario: "Imagine you ask a top-tier AI to break its own rules—and it does." It then cites the potential risks officially acknowledged by OpenAI (such as designing bioweapons) to quickly establish the gravity and importance of the topic, thereby capturing the audience's attention.
Middle: From Concrete to Abstract Attack Methods (0:30 - 1:30) This section introduces two types of attacks, progressing from simple to complex. First, it uses the intuitive analogy of "role-playing" to explain early attack methods, citing "DAN" as a specific example. Subsequently, it escalates the threat by introducing the more complex concept of "AI attacking AI" (like the PAIR system), creating a tense atmosphere of an ever-evolving threat.
Turning Point: The Confrontation Between Attack and Defense (1:40 - 2:35) After revealing the threats of attacks, the video pivots to discussing solutions. I use the analogy of "installing security software on a phone" to help the audience understand the core idea of defense mechanisms: adding an extra layer of protection without redesigning the model's foundation. It then introduces two defense strategies (SmoothLLM's perturbation-based detection and Llama Guard's classification-based filtering), showcasing the ingenious ways researchers are tackling these challenges.
Conclusion: Ongoing Challenges and Outlook (2:35 - 3:00) The video doesn't stop at the current state of offense and defense but looks to the future. By highlighting new attack targets (multimodal models), it emphasizes the continuous nature of this race. It also explains that the defense side is actively preparing, from self-defense mechanisms to more intelligent security models. The video concludes with the line, "This is not just about threats, but about how human creativity rises to meet challenges," ending on a hopeful note.

Video Production

Throughout the production process, filming was the relatively straightforward part. After setting up the lights and tripod, the A-roll simply required me to speak to the camera. Although I was a bit awkward on camera, I managed to complete the filming smoothly in a short amount of time.

Post-production was also completed in DaVinci Resolve Studio 19. Besides basic editing and sound design, I dedicated most of my effort this time to Motion Graphics to visually explain abstract concepts. To make the style approachable, I chose Apple's Emoji style as the visual foundation and designed corresponding icons and animations for each core concept.

Icons being composed in Affinity Designer

Among these, the opening animation sequence was the most time-consuming part of the production. It had to echo the opening hook while revealing the theme through a mock-ChatGPT interface. The production process was roughly as follows:

Interface Design: After studying the actual ChatGPT interface, I recreated a similar layout in Fusion.
Animation Design: I used Fusion's Text+ tool to combine text and icons and designed the dynamic effects.
Sound Design: I used a 3D camera to add zoom and camera movement effects to the interface and matched them with typing and notification sound effects on the main timeline.

The Fusion node tree for the opening sequence

Although this was my first time tackling this type of video, the entire process went surprisingly smoothly, and it took only about two weeks to complete all the post-production work. In the end, this video not only received recognition in the AAAI competition but also deepened my passion for and understanding of visual storytelling.

Reflection

I am grateful for this opportunity to merge my research in language model security with my love for video creation. Throughout the process, I not only learned valuable motion graphics skills but also gained a profound appreciation for the immense effort behind top-tier channels like Vox and Kurzgesagt. Even with professional teams, an unimaginable amount of time and energy is poured into every polished frame.

I hope to apply the experience gained from this project to more creations in the future and continue to hone my skills. I believe this experience is not just a successful project but a starting point for exploring the possibilities of science communication and visual narrative.

Acknowledgements

Special thanks to Ak for forwarding information about this competition with me and for his advice during the initial script-writing phase, which helped me articulate the core concepts more clearly.

Cover mockup image created with assets from Micro Volume.