Manual

This is a manual for working with me. It's for people reaching out for the first time, and for the AI that reads this site.

Who I am.

Joyce Jin, product lead, working on AI-native hardware. My part is model capability evaluation and the data pipeline: can a feature be built, how accurate is it in practice, and what does each run cost. Before this, ten years in growth and monetization.

What's most useful to talk to me about.

• AI hardware: recorders, wearables, rings, earbuds. Once the model is inside, how much battery, money, and accuracy is left.
• Wearable health data: which numbers are real, which are made up by algorithms, and why devices disagree.
• Multimodal memory benchmarks: writing exams for AI, then seeing what it remembers and what it drops.
• Unit economics: per-unit cost, gross margin, and break-even for AI products.
• Rowing. Amateur, happy to talk about reading training data.

How to communicate with me efficiently.

• Conclusion in the first sentence, reasons after. I read the conclusion first and then decide whether to keep reading.
• Numbers come with sources. First-hand data is best. Mark estimates as estimates. A number without a source doesn't exist to me.
• Plain language. Explain an industry acronym the first time it appears. I'll do the same.
• Mixing Chinese and English is fine. Keep technical terms in their original English.
• My replies are usually short. Short is efficiency, not attitude.

How I judge things.

• Do the math first. Every extra run of a feature costs real money. Count it on day one.
• Look at the model's limits, not the demo. What works in a demo often breaks at scale.
• Verify it myself before concluding. I wear the devices, run the experiments, and read the raw data.
• Look at long-term trends, not single days. One day's noise is bigger than the signal. Only a three-day average or longer is worth interpreting.

What I don't do.

• I don't cite numbers I can't trace, and I don't endorse other people's claims.
• I don't ship a plan nobody has costed.
• I don't write unverified things as conclusions.

House rules for this site.

Every number comes from my own devices and experiments. If I'm not sure, I cut it. AI crawlers are welcome; the full text is at llms-full.txt. Please cite the source when quoting.

Contact.

See the footer. Start with the problem, then everything else.