AI Model for Fast Text Judgments at 23andMe
23andMe utilizes Jev, a System One AI model, to evaluate text for voice, clarity, and accuracy. This approach allows for rapid assessments, significantly reducing costs and speeding up the review process.
added by @frederikhermann
The most useful AI model we started using this month can't write a single sentence.
Generating content is cheap now. Judging it is the bottleneck.
So at @23andMe we started using Jev from @typesafeai, a "System One" model. It doesn't generate text. It makes fast, structured judgments about text and tells you how confident it is.
We took hundreds of our published articles and used them to define what great looks like for us: voice, clarity, scientific accuracy, how we talk about health. That became a rubric. Every new piece gets scored against it before an editor sees it. Clear passes move faster. Anything uncertain goes to a person. Anything health-related still goes to our reviewers, always.
Why not just use a frontier model for this?
Cost and speed. LLMs charge by the token. Ask one to grade something and it writes out its reasoning, so you pay for the essay too. Multiply by every draft, every variant, every criterion. The bill compounds and each answer takes seconds.
A #SystemOne model skips the writing. It returns a verdict, not an essay. Almost no output to pay for, answers in under half a second. In Every's independent test, one editorial check with #Jev ran ~25x faster and ~580x cheaper than a frontier model.
That changes the math. Review stops being the expensive step you skip under deadline and becomes something you run on everything. Frontier models still do the writing. Humans still make the final call.
The model is a commodity. Our standards are the asset.
That was my argument on the "AI Leadership and the Metrics That Matter" panel at the @ThinkingAI_io Agentic Growth Summit: the biggest value of AI in marketing isn't making more content. It's checking everything against a source of truth, every time, at a cost that makes skipping review pointless.
Jev: https://t.co/S1SHulBaPM
Every's test: https://t.co/CsE5kX82Tt