Getting started with Photon Datum Chat

Photon Datum Chat is the private chat and agent workspace for members. It runs on infrastructure we control, behind single sign-on, and it is in early access.

Before you begin

Membership is not open yet. When it is, you will need a membership that includes the AI agent workspace, and you will need to be able to sign in to the portal. If you have not done that yet, start with Getting started with the Photon Datum Portal.

Open the workspace

1. Sign in to the portal.
2. Open the AI agent workspace from the portal. The portal manages your access, so there is no separate sign-in.
3. Start a conversation.

If the workspace does not open, go back to the portal and try again from there. Please do not save the workspace’s own address as a bookmark: members always arrive through the portal, which is what manages your access.

What it is good at, and what it is not

The workspace runs small language models. A small model runs on modest hardware, keeps your data on machines we trust, and costs little to run, and it is good at narrow jobs that can be checked: looking up a fact, summarising a short text, drafting, or sorting items into groups.

It is unreliable at counting and at judgement. In a small test on one laptop, with four model sizes and the same five tasks checked by software, the nine-billion-parameter model answered 40 of 50 correctly, and every miss was counting. A two-billion-parameter model looked up a single fact correctly every time but could not count. This is a small test, not a benchmark, so read it as a reason for care rather than a rating. The design follows it: counting and judgement are handed to tools and checks rather than to a model.

Ask it well

• Give one narrow job at a time. “Summarise this paragraph in three bullets” works better than a long request with several parts.
• Say exactly what you mean. Name the file, the person or the date instead of “it” or “that one”.
• Check any list or number. If the answer is a count, a total or a list that matters, check it against the source before you rely on it.
• Start again when it drifts. A new conversation is cheaper than arguing with one that has gone off course.

Your data

The workspace runs on our own private infrastructure, behind single sign-on, with data kept under our control. It does not send your work to a third-party AI service. The Privacy page is still in preparation and will say in full what we store and for how long.

What to expect

• The models run on infrastructure we control, not on a third-party service.
• The workspace is in early access, so features can change.
• Agents that act within limits, and background helpers that do longer research, are in development. The agent page below says where each one stands.
• The AI agent workspace page describes what is available soon, what is in development and what is on the roadmap.

If something is not working

• It will not open: return to the portal and try again from there.
• A reply is slow: a larger model takes longer to answer. A shorter, narrower request is faster.
• It is still stuck: contact us and say what you were doing and what you saw.

If you work in Visual Studio Code

Synod for Visual Studio Code brings a chat and agent panel into your editor. Its guide is Getting started with Synod for Visual Studio Code.