GTM Sauce

Akool · AI video and avatar maker

Train small AI models for one task each instead of one giant model

Keeps costs down; Coca-Cola picked it after 3 months of benchmarks

Workedfeature-developmentpositioning-niche

What they did

Instead of building the largest general video foundation models, Akool designs and trains smaller models from scratch for specific tasks: human/avatar generation, and a swap series (face, character, motion, head, hair). Different architecture and training data give better quality on corner and boundary cases. Training runs on rented cloud servers (mostly Lambda, plus others depending on workload). Large foundation models are only fine-tuned.

What happened

Controls cost and budget; Coca-Cola came inbound and chose Akool after benchmarking it as best on the market
Stage
2022-present
4:02
“We are not building the largest video foundation models. We are building like video models for specific tasks.”
Jeff Lu
“and these models are much smaller than the large foundation models that generate everything. So, that's how we uh control the cost and uh control the budget and make things work.”
Jeff Lu
11:31
“So, uh that's inbound, and they found us, and they uh they see the uh our technology is the best on the market, and uh so they benchmarked a lot, so uh they decided to work with us.”
Jeff Lu

Related topics

Get tactics like this every Monday

New tactics from the week's founder interviews, each linked to where it was said.

More from this episode