DSNB · Explain
Understand DeepSeek at the depth you want
DeepSeek is a story you can read at four different depths. Each topic below has the same four tabs — a one-paragraph version for a 10-year-old, a longer version for a curious teenager, a working explanation for a software engineer, and the underlying research framing for someone with a machine-learning background. Same story, four resolutions. Pick the depth you want, switch up or down anytime.
From DeepSeek's perspective
We chose to write this site this way because DeepSeek itself was built on the premise that good ideas should be explainable. Open weights, open papers, and a refusal to hide behind marketing language are all the same principle: a story worth telling should be tellable at every level.
Pick a depth
Kid
Plain language. One short paragraph. Analogies in place of jargon. Aimed at a 10-year-old reading on their own.
Teen
A curious teenager with internet access. Some jargon used, but defined inline. About one page of reading.
Engineer
A software engineer who has shipped real systems but doesn't do ML for a living. Concrete numbers, no hand-waving on tradeoffs.
Researcher
Someone who has read transformer papers. Architectural details, paper citations, what's novel vs. derivative.
Start here
- Depth: KidRead full →
What is DeepSeek?
A Chinese AI company that gives away its best models for free, born from a quant hedge fund.
- Depth: TeenRead full →
Why open-source instead of selling APIs?
Open weights let anyone deploy DeepSeek-grade AI on their own machines — the move that broke the price ceiling on frontier models.
- Depth: EngineerRead full →
How did $5.6M train a GPT-4-class model?
DeepSeek-V3 was trained for ~1/10th the cost of competitors through ruthless engineering across data, architecture, and numerical precision.
- Depth: ResearcherRead full →
What is Multi-Head Latent Attention (MLA)?
DeepSeek's attention reformulation that compresses the KV cache by ~93%, making long context windows affordable at frontier scale.