Is Text-to-Video the next BIG thing?
026. Meditations on future of AI: video
Hey, friends—welcome back to online-offline, a newsletter about technology, culture, and the future. Hello to new subscribers! I’m glad you’re here. If you haven’t already, please support my work by subscribing below. Now, let’s dive in.
I recently read that writing should lead to more questions instead of providing answers. Instead of sharing a key insight today, I’m posing a few questions.
Who’s making generative videos & products?
I’m obsessed with the next era of generative AI: video. In the past few months, I’ve been playing around with a few AI video tools—Runway and Pika for example—and the fun I could be having is capped by the 4-second output and tricky prompt engineering. I just wanna make cool shit without learning video editing and whatever else, but that’s not possible today. Making a 30-second movie trailer? Possible but the clips will look like a bunch of stock imagery spliced together. A short film? Also possible, but only if you use other tools to generate the story, audio, music, etc. But this is only the beginning. There’s innovation on all levels of the gen. video stack from infra to apps, and things are moving quickly enough that the TikTok FYP could look very different in a few years. I’ll save you an entire essay because I already wrote one here: Generative Video: The Next Frontier in AI.
EDIT: I was supposed to publish this essay yesterday but I’m glad I waited an extra day because OpenAI just launched Sora, their text-to-video model. It looks incredible and the outputs can be minutes-long?? And it has varying camera movements and angles??? Can someone get me access to test this out? I’ll dedicate my first AI short film to you. TY!
Prompt for the video below: A beautiful homemade video showing the people of Lagos, Nigeria in the year 2056. Shot with a mobile phone camera.
Will AI make content creation cheap?
Everyone seems to think that generative AI will lower the cost of content creation and I’m inclined to disagree. I’m thinking about my friends who are building consumer apps and paying $$$$ ($1/user/day) to OpenAI—eventually, they debate the open-source route and fine-tuning a free model. However, reaching meaningful scale while maintaining quality output is a challenge. Value may accrue to incumbents (eg. Adobe, ByteDance) that can manage the cost of inference and survive long enough to see the cost of computing decrease. Either way, generated video is miles away from standing on its own, let alone in competition with the most viral TikToks. Runway and Pika are great, but you can’t generate more than 4 seconds of video. How can we talk about cheaply made content when the best AI video models haven’t mastered space and time?
The caveat here is AI use in traditional media and entertainment where it can generate efficiencies and cost reductions in comparison to CGI for example. Still, the 2023 actors and writers strikes are evidence that movie stars will still demand big checks, at least for a while.
Related, how will IP play out?
Getty Images is suing Stability AI, New York Times is suing OpenAI and Microsoft, UMG is suing Anthropic … Basically, each mega-AI company has been sued at least once.
Should AI models be able to train on copyrighted material? Who owns the model outputs? While humans design a model’s architecture, they don’t optimize model weights, so can weights be legally protected?
The question of IP goes beyond creative fields. What happens when you use AI to produce new compounds, new medications, or to design new hardware? How is that IP valued? As of February 13, AI-assisted inventions are not categorically unpatentable but they hinge on significant human contribution according to USPTO. The guidance and enforcement is relatively vague—how does one measure human versus AI contribution?
These open questions should be solved sooner than later. In fact, there are over $7 trillion reasons to prioritize it, especially in the US where industries that intensively use IP accounted for 41% of domestic economic activity/output and 44% of all employment in 2019. If an economist or lawyer has any opinions on this, please reach out. I have about a million questions for you.
And even more questions
You can generate pretty good fake IDs using AI—no, I won’t tell you how or where! Are NFC and cryptography the solutions here? Will we finally get the AI and crypto merger we’ve been dreaming of?
What new jobs will emerge? Everyone is a DJ these days, but that wouldn’t have happened without the invention of the phonograph, sound systems, mixers, and much more. Each innovation makes way for new avenues of production. AI, similarly, will create new jobs.
Will human creativity and production become more meaningful? Will we see the next Michelangelo and Sistine Chapel? Will IRL make a comeback? (I hope so)
Will someone get me access to Sora by OpenAI? (I really hope so)