---
title: "What we're building"
description: "How today's AI was made, why robots need first-person data, and why your attention to detail is the product. The shared briefing for every recorder and SPL."
---
Read this first · Shared briefing
This section is the same for everyone at opentez, whether you record footage yourself or
lead a team that does. You don't need any background in AI to understand it. By the end
you'll know what we collect, why it's valuable, and why your care with the details
directly determines whether we can pay for it.
## How today's AI was actually built
Modern AI models like ChatGPT and Claude were built in two stages.
The first stage is **pre-training**: the model reads an enormous amount of text from the
internet and learns to predict the next word, over and over, until it absorbs the
patterns of language and a great deal of world knowledge. Pre-training makes a model
fluent, but not yet genuinely helpful or reliable.
The second stage is **post-training**. This is where humans step in. People are paid to
show the model good answers, rank competing answers, and demonstrate exactly how a task
should be done. The model learns from these human demonstrations and preferences. This is
the stage that turned a raw text-predictor into something useful, and it is why an entire
industry of companies (Scale AI, Surge AI, Mercor and others) now earns roughly ten
billion dollars a year selling one thing to AI labs: high-quality human data.
**The one idea to hold onto.** A model is only ever as good as the human data it learns
from. Clean, careful, well-labeled data produces a capable model. Sloppy or improperly
sourced data produces a weak model or a legal problem. That principle is the whole
reason opentez exists, and the reason your attention to detail is the product.
## The next revolution: robots that learn by watching
Text was the easy part because it was already sitting on the internet. The next frontier is
**physical intelligence**: robots that can fold laundry, tile a floor, frame a wall, or
wire a panel. The models being built for this are called **VLAs
(Vision-Language-Action models)**.
A VLA takes in what a camera sees (vision) and an instruction in plain language ("tile
this section of floor"), and outputs the actual physical actions a robot should perform.
Recent systems you may hear about include Google's RT-2, OpenVLA, Physical Intelligence's
π-series, NVIDIA's GR00T, and Figure's Helix. They all work this way. Researchers describe this
moment as robotics having its own "LLM moment."
But there's a catch, and it's the entire business opportunity: a robot cannot learn to
tile a floor from Wikipedia. Language models had the whole internet to read. Robots have
almost nothing. Physical skills live in people's hands and habits, not in text. That data
has to be captured deliberately by real people doing real work, and right now it barely
exists.
## Egocentric data: the thing we collect
**"Egocentric"** simply means first-person point of view. Instead of filming a worker from
across the room, the camera sits on the worker's own head and sees what they see: their
hands, their tools, the material, and where their attention goes as a task unfolds. This
is exactly the viewpoint a robot needs, because it mirrors what the robot's own camera
will see when it tries to do the job.
This isn't theoretical. Meta's Ego4D and Project Aria research programs proved that
first-person video of everyday activity is gold for teaching machines, and university labs
have shown that a person simply wearing a head camera and doing tasks produces training
data comparable to far more expensive robot setups. opentez turns that proven idea into a
real, paid, at-scale operation focused on skilled trades, the hands-on work robots most
need to learn.
## What makes a clip valuable
| A valuable clip has… | Because… |
| --- | --- |
| **Hands visible in frame, doing real work** | Hands and tools are the single most important signal a robot learns from |
| **A clear first-person angle from a head-mounted phone** | It matches the viewpoint the robot will have |
| **Continuous, focused activity** | Robots learn the natural flow of a task, not idle time |
| **Real, diverse trades and environments** | Diversity is what lets a model generalize instead of memorizing |
| **Clean provenance: consent on file, no minors or exposed PII** | opentez only accepts data backed by a clear, documented chain of consent |
That last row is the quiet one that matters most. At opentez, having the most footage is
not enough. We must prove exactly where every clip came from and that everyone in it
consented. That documentation is our
moat and, if we get it wrong, our biggest risk. It's why the rules in the next sections
are strict, and why we can only pay for data that follows them.
## Where you fit
You are the source of the ground truth. Every robot trained on opentez data is, in a real
sense, learning from you and your crew. Two roles make that possible:
Capture the footage while wearing the headband, doing the work, and uploading clean clips.
Build and manage teams of recorders, guarantee the quality and legality of what comes
back, and deliver it to opentez in bulk.
The rest of the handbook covers your role specifically.