A hundred drawers, and Auntie Tan knows every one by hand — the same way she knows every customer.
That isn't a database of dropdowns. It's the way a good shopkeeper knows you — the aching knee, the baby on the way, the red-date tea you always take. Guanxi 關係.
Now imagine your shop remembering every customer like this.
Frontier builds these →On this page, three companies receive what you say or type: OpenAI, Deepgram and Cerebras. Our web host, Cloudflare, carries it to them.
When you tap the mic and choose to start talking, your voice is sent live to OpenAI so the assistant can hear and answer you.
If you type to the assistant instead, what you type is sent to OpenAI in the same way.
Each time you speak, a recording of what you said, starting a moment before you begin, is also sent to Deepgram, a speech-to-text company, so we can check the assistant heard you right. If Deepgram cannot answer, OpenAI makes that check instead.
If you talk over the assistant, the part of its reply that played is sent to Deepgram and OpenAI to work out where you stopped listening. That recording is the assistant's voice, not yours.
If the assistant fumbles an action, up to the last 40 lines of the conversation are sent in writing to Cerebras, so it can work out what the assistant meant to do.
Deepgram keeps the recording only while it transcribes it, does not use it to train its models, and does that work in Australia.
Deepgram's own description of what it keepsCerebras says it does not keep what it receives and has no right to train its models on it. It may do that work in the United States or other countries.
Cerebras's privacy policyWe use your voice to run this demo and for nothing else.
We do not use what you say or type to train models.
OpenAI does not use it to train its models either.
OpenAI's description of how it uses API dataSpeaking here does not add you to a mailing list and does not enrol you in our research. Those are separate decisions and we would ask you separately.
Your audio, and anything you type, leaves New Zealand.
We do not store your voice, or a transcript of what you say or type.
OpenAI keeps a copy of what you say and type for up to 30 days, only so it can check for misuse, and then deletes it, unless the law requires it to keep it longer.
OpenAI's description of how long it keeps API dataWe keep three things. None of them holds anything you said or typed.
A record of each session's timings and counts, for up to 90 days, only to improve this demo.
A count of the sessions your connection has started, filed under a scrambled form of your internet address, for about a day after your last visit.
How many tokens each typed turn used, for 30 days, so we can see what the demo costs.
This demo is run by Axiom Intelligence.
You can reach us about any of this at regan@axiomintelligence.co.nz.
The same address is listed on our company siteTo ask what we hold about you, write to that address and say roughly when you visited.
You do not have to speak. Just watch keeps the mic off and sends nothing: you can read this page without talking to the assistant.
Building. Not true yet.Each session ends after three minutes at most. It can be shorter if this demo's voice time for today is nearly used up.
Each demo has up to twelve minutes of voice a day, shared by everyone, and all demos together stop after 50 sessions a day. After that, voice is off until midnight, New Zealand time.
If you talk for the full three minutes, Regan gets a note that someone did. It does not say who, or anything you said.
Building. Not true yet.Please do not use the mic if you are under 16.