A 35-Year T Cell Researcher Gave GPT-6 Astra One Sentence.

Derya Unutmaz has spent most of his career studying T cells. He gave GPT-6 Astra a single prompt: "Create a 5 minute educational video about T cells."

Watch it before reading on, because the reaction is the finding. A researcher of more than 35 years in the field said he could not have explained it that well, and is now planning a full lecture series the same way.

This is the most impressive demo of launch week and almost nobody outside biology saw it.

The same week, a closed piece of hardware

Claire Vo owns a Divoom Bluetooth pixel speaker. Closed device, no API, undocumented protocol. She calls it her "personal Everest" and has been trying to crack it across model generations. Earlier models, in her words, "barely let me flash a 'hello world' message."

Astra built her a web app that connects to the device and lets her draw on it live with a mouse. Then a CLI so the speaker could talk back. Then, given a prompt about her latest podcast episode, it found the episode information and streamed a scrolling message and an animated graphic to the speaker's screen.

These are the same problem

A protein-signalling pathway and a Bluetooth protocol have nothing in common except the thing that matters here: the documentation did not exist.

Divoom never published the protocol. And 35 years of immunology intuition is not written down anywhere in the shape a teaching video needs, which is exactly why making one is hard and why Unutmaz was impressed. In both cases Astra had to construct the missing map before it could produce anything.

The 3D work in Part 1 is the third instance of the same move. A 2D sketch does not document a steam train's parts, and Astra returned 3,295 of them.

Anatomy, twice

Put the immunology run next to the 50-bone skeleton from Part 1. Same competence, different domain. One arranges bones so a wolf runs correctly. The other arranges concepts so T cell memory formation makes sense to a student. Both need a working internal model of a system, not a lookup.

Spatial reasoning showed up as its own claim too.

Why this matters more than the games

Every kart racer and FPS map from launch week is a build in a domain where success is obvious and the model has seen ten thousand examples.

These are different. In each one the person holding the stopwatch was a specialist evaluating work inside their own speciality, and the specialist was impressed. That is a harder test to pass and a harder result to wave away.

It is also the one your company should care about. You do not have a shortage of kart racers. You have a shortage of the thing Unutmaz has: knowledge that lives in one person's head and has never been written down in a usable form.

The operator read

Find your undocumented system. Every company has several. The pricing logic only one person can explain. The integration nobody has touched since the engineer left. The onboarding knowledge that lives in a senior rep's habits.

The old answer was to schedule interviews and write it down, which never happened. The new option is to point a model at the artifacts, let it construct the map, and have your specialist correct it. Correcting a draft is a job people will actually do. Writing from scratch is not.

One caution, and it is the whole of Part 4: each of these was one attempt, by an expert, on a task they chose.

Part 3 of Astra, Week One.