The Death of Dead Data: How 12 Labs Taught AI to Watch Video

Video accounts for over 80% of all data on the internet. Yet, for the current generation of AI, video is effectively “Dead Data.”

If you want to search a video today, most AI models have to rely on transcripts or static thumbnails. They can’t actually understand what is happening over time. They are blind to motion, context, and temporal logic.

Enter 12 Labs.

Founded by a team of young South Korean founders led by Jae-Kyu Lee (JK), 12 Labs is building the multimodal foundation models that allow computers to see, listen, and understand video as a continuous stream of information.

The Problem: The Transcript Trap

Most “Video AI” is actually just “Text AI” in disguise.

To search a thousand hours of security footage or sports broadcasts, legacy systems convert the audio to text and search for keywords. If a goal is scored in a soccer match but the announcer doesn’t say “GOAL,” the AI misses it. If a specific person walks across a warehouse floor without speaking, the AI is oblivious.

This is the Transcript Trap. It makes petabytes of enterprise video data unsearchable and useless for real-time decision-making, contributing to the same corporate amnesia that plagues the modern enterprise.

The Playbook: Multimodal temporal understanding

JK and his team realized that video isn’t just a series of pictures; it is a complex, temporal language.

Ignoring the Chatbot Hype

While the rest of Silicon Valley was rushing to build wrappers for ChatGPT, the 12 Labs founders stayed focused on the unsexy world of infrastructure.

They realized that the world didn’t need another chatbot; it needed an engine that could index the visual world. Like the founders of Windsurf, they had the intellectual honesty to ignore the hype and solve a harder problem.

The Mathematical Frontier

12 Labs built a proprietary “Video-to-Vector” engine. Instead of transcribing words, they convert visual actions, objects, and sounds into mathematical coordinates (vectors).

1. Natural Language Search: You can ask the AI, “Find the moment the red car swerved to avoid the cat,” and the AI finds the exact frame, even if there is no audio or metadata attached to that clip.
2. Action Summarization: It can “watch” a 10-hour stream and provide a bulleted list of every key action that occurred, effectively killing the workflow friction inherent in manual video review.

Founder Lessons: The Infrastructure Insurgency

1. Attack the “Dark Data”

The highest-leverage founders don’t compete for the data that is already clean and indexed. They go after the Dark Data, the massive piles of information that are currently impossible to process. Just as Terranox AI digitizes the earth’s crust, 12 Labs realized that video was the biggest untapped resource in the enterprise.

2. Global First, Local Second

Though based in Seoul, 12 Labs operated with a “Global First” mindset from Day 1. They moved their headquarters to San Francisco and attacked the US market immediately, a move mirrored by the Paystack Mafia in their bid to build for a global audience. They understood that infrastructure-level AI requires the gravity of the Silicon Valley ecosystem to scale.

3. Solving for “Utility” over “Entertainment”

12 Labs didn’t build a video generator. They built a video understander. They prioritized the “Work” over the “Art,” following the 5-step algorithm of focusing on utility and deleting the fluff. As a result, they have built a massive moat in the enterprise sector while others are fighting over consumer entertainment.

Final Word

The next era of AI won’t just talk to us; it will watch over us.

By teaching computers to understand video temporally, 12 Labs is ending the era of Dead Data. They are part of the broader Hard Tech insurgency, proving that the most valuable founders aren’t the ones following the trends, they are the ones building the eyes that allow the AI to see the world as it actually exists.