This Insane AI Video Search Technology Selected by NVIDIA and Snowflake | Twelve Labs, Jae Lee | EO

Twelve Labs builds the search layer for the world's video

Twelve Labs is building the video understanding layer for the internet. Jae Lee and two friends from the Korean Cyber Command spent a year and a half hunting the next AI frontier and chose video, betting foundation models would be accepted before a market existed. Text and image models were already a capital race, but video understanding had open research problems, so a small team could win. A competition victory bought credibility and a seed round.

Takeaways

• Video is the internet's unexplored infrastructure data. Lee puts 80% of the world's data in video, with nothing adequate for developers or enterprises to make sense of it. Text and image foundation models were settled by capital, but video understanding still had open research problems.
• Twelve Labs maps language onto video. Its foundation models connect human language to what happens in footage, producing emergent capabilities such as searching and summarizing. A little over 20,000 developers actively use the search API, and customers include large media, sports and law enforcement organizations.
• A competition win brought seed investors in. Starting with about $2,000, the team spent roughly $200,000 on compute over ten days to win ICCV's video understanding competition. Index Ventures and Radical Ventures reached out, and Lee took a first call at 3:30 AM Seoul with no deck. Twelve Labs has raised about $30m in seed.
• Customers who pay but never use are warnings. Cofounder So Young, who runs business development, says one customer paid and never used the product. She argues teams should optimize for hearing the no early and work with true innovators in fields like law enforcement and e-learning instead of pitching ROI to everyone.
• The narrow-problem startup mantra was a mistake. Lee regrets chasing ideas like TikTok 2.0 and Gong 2.0 that did not excite the team, and warns founders against accommodating all their mentors' advice. His hiring test is whether he can learn from a candidate's core values.
• NVIDIA invested beyond money. A multiyear Oracle Cloud Infrastructure partnership gave Twelve Labs state-of-the-art NVIDIA chips. After five to ten minutes with Jensen at an Oracle partner event and a spot in NVIDIA's 2023 GTC, NVIDIA's venture team approached about a partnership beyond financial investment.

The first year, from a bagel shop

1. SJ was discharged on a Thursday and came back to the base that Saturday with laptops.
2. The team worked out of Las Bagel, a bagel shop in front of the base, for six months.
3. A friend let them use his Seoul office, which went bankrupt three weeks later and sent them carrying desktops and GPUs back out.