Communication is data.
Messages are a medium for transmitting data.
Northwind dashboard is up next. Spec is in NW-142.
I’ll pick it up Thursday.
The product called Praxos is a messaging application. Praxos Intelligence, the company, is a data company. Underneath the messaging application, there is a layer for moving data between people, between machines, and between the two.
An explosion in messaging
For most of history, messages were mostly written by people to be read by other people. That period is ending. Developers talk to models as if they were colleagues, and we believe that everyone else is about to follow. Behind every person at a keyboard or a phone screen there will be agents working. To work effectively, those agents will need to talk to one another, to other people's agents, and to other people.
We expect an explosion of communication in the next few years. Most of it will happen between machines acting for people. We have not yet designed for this future. Someone has to decide where these messages go, who receives and sees them, where they are stored, and how others can act on their contents.
Two people
1 conversation
Each with three agents
28 possible conversations, 1 of them between people
Two teams of three, each person with three agents
276 possible conversations, 15 of them between people
We want to lead this endeavor.
Three exchanges
We think about our work as the management of three kinds of exchange.
Computers with computers
Every tool keeps the same facts in its own shape. A customer is a row in one system, a thread in another and a PDF in a third. Each tool is a small kingdom with its own language, and they trade badly. The best thing a language model can do for the world is translate between these schemas, moving data from one shape to another without losing the sense. But a translator needs something firm to translate into. Cloud software has no deterministic layer behind it, and we have never understood why nobody builds one. So we are building one: data from many systems, mapped into established ontologies, so that changes can be predicted and relationships can be navigated. Provenance is part of it. Data that can say where it came from can be graded and valued, in quality as well as in quantity.
People with people, through computers
Between any two companies, the real connection has always been two people forwarding things. "There never was a direct link between my team and theirs. This was not a technology or platform problem. It was a people problem." We would like to see companies share what they know without texting each other or sending over PDFs, with agents carrying the routine traffic and a human stepping in only when a human is needed. When we bridge a Slack channel into Praxos, the person on the receiving end cannot tell the difference. We built it that way on purpose.
People with computers
An agent that cannot join the conversation is a tool. One that can join, and inherits the same history the humans have, is a colleague. Yet a colleague that speaks only to machines is no use to us. The output of an AI must follow a pattern a person can read, which is why we think AI will always need a human-facing interface, and why that interface is conversation.
Why a messaging app
Because working with data is equivalent to working with words. Code is words. An Excel spreadsheet is words and symbols. A PDF manual is words, symbols and drawings. Decisions are words, voiced or written, and made in chats, in calls, in review comments, in the margins of a pull request and everywhere in between.
As a corollary, memory is a byproduct of words. Words change the nature of our work, and as it changes, a chronology with relationships and changes grows over time. Thus, if the place where people talk is also the place where the record is kept, the record costs nothing to create and maintain.
This is why we say the product is a messaging app and the company, a data layer.
Where this goes
Work is becoming a matter of parallel threads, each held by a different agent or a different person. The more of it there is, the less any single head can hold, and the more people will find themselves acting as routers.
If we do well, the people who work in ten years' time will talk to one another and to their machines in one place. They will not remember what it was like to copy and paste; or, to wait for someone to get them a piece of data they need. If two or more people talk to each other, it will be because they have something new or exciting to talk about. Maybe, because they want to pass time. A lot of the coordination and “fake work” will be gone.
Today
One head in the middle, acting as a router.
If we do well
Threads meet in one place. People talk when there is something new to talk about.
