Three months hand-building a pocket AI terminal, from blown-up boards to a working RPG
OpenClaw in Your Hand: Building a Physical AI Terminal - Lech Kalinowski, Callstack · Lech Kalinowski
25 min total·Actually worth watching closely: ~6 min·2 must-watch clips
- 0:02 – 4:04Listen
It starts with a strange device in the basement
The speaker opens with a weekend basement cleanup and a dusty, 80s-looking device with a keyboard and odd displays that turns out to have a shell, and it is on — a way into the real question: what should a device look like if it is born to talk to an LLM. He then introduces himself and the months of work behind the build at a technology incubator.
The first thing you think about with an LLM is text — not audio, not a camera.
This stretch is the speaker telling a story and laying out his motivation, with the slide parked on the title page. You lose nothing listening with your eyes closed.▶ Jump to 0:02Speaker · Lech Kalinowski - 4:07 – 7:05Skim
Two screens, two jobs: the live surface and the stable one
Introduces the core interaction design — the one-color OLED holds whatever you are typing and whatever is changing, and once you hit enter the result is rendered onto the electronic paper, where it stays. Plus how the firmware draws pages straight from fixed static buffers and one-bit images held in memory: no markdown engine, no dynamic allocation.
Give each screen its own job and e-paper's fatal slow refresh stops mattering, while power draw stays extremely low.
One interaction diagram carries the split between the two screens; once you have read it, that is enough. The long middle stretch is spoken detail on how stripped-down the firmware is — fine to skim through.▶ Jump to 4:07Speaker · Lech Kalinowski - 7:06 – 9:46Skim
The parts list and the prototypes that got blown up
Breaks down the build: an ESP32 dual core microcontroller plus an OLED display, a keyboard, an encoder and a separate power supply unit. The focus is why power is the hardest piece — unstable current and voltage and a regulator that kills the OLED blew up a prototype and the display, and getting replacement parts took a couple of weeks.
Most of the traps in homemade hardware are in the power supply and the quality of your parts, not in the software you were worried about.
The slide is a parts list and a wiring diagram — one glance tells you which board he is talking about. The blown-board story is spoken, so there is no need to watch frame by frame.▶ Jump to 7:06Speaker · Lech Kalinowski - 9:51 – 13:20Listen
API compatibility and a real agent task
Open source models do not all follow the same API conventions, and wiring them up meant hitting a lot of walls; the fix is a layer that exposes one consistent interface. Then a demo of a command like "write a Java example and store it on my local machine" actually being carried out from the terminal.
Inference has to run on the backend and the device is only a thin terminal — no model today really runs on an MCU this small.
Hard-won advice on API traps; the screen is just the speaker and a few lines of text, so listening is more efficient than watching.▶ Jump to 9:51Speaker · Lech Kalinowski - 13:20 – 16:54Listen
The overlooked niche: a quiet device
The speaker argues that everyone is building AI hardware around microphones and cameras, while almost nobody builds for quiet places — a device that disturbs no one and is used only to read and write plain text. He has filed a provisional patent in that field.
The most valuable claim of the talk: the gap in the market is not compute, it is how quiet the setting is.
Pure argument with nothing to look at — good to leave running like a podcast.▶ Jump to 13:20Speaker · Lech Kalinowski - 16:54 – 19:58Watch
Hands-on demo: boot, tutorials, agent execution
A few core takeaways first, then the real thing: power on, connect to the local Wi-Fi network, the welcome screen, a help command listing everything the device can do, a walk through the built-in device and gaming tutorials, and finally the OpenClaw agent checking disk space on the backend host.
A complete device that actually works — not a render, not a concept.
Keep your eyes on that little screen throughout — the boot sequence, the command echo, the disk numbers the agent returns all appear line by line. Miss the picture and you have missed the talk.▶ Jump to 16:54Speaker · Lech Kalinowski - 20:12 – 23:12Watch
Four RPG worlds, all model-generated
The RPG mode: pick from Aldergan, Neon Abyss, the Hollows and Void Reach, and once you are in, the NPC personalities and memory, the maps, the skills and even the scene pictures are all generated on the spot — the pictures converted into one-bit bitmaps for the microcontroller to render.
Not one line of the game content is pre-made; the model produces all of it end to end.
Choosing a world, entering a scene, black-and-white dot-matrix art painting itself onto the small screen — the force of this stretch is entirely in what is happening on screen, and no description substitutes for it.▶ Jump to 20:12Speaker · Lech Kalinowski - 23:15 – 24:33Listen
Top-end compute behind a text game
The closing point: a game this plain, with no three dimensional graphics, runs on the most advanced Nvidia processor mounted in the DJX Spark. The speaker frames the whole device as a Game Boy — a device to play with LLMs.
What really decides the experience is the context, not the parameter count.
Closing remarks over the final slide — one minute and you have the whole thing.▶ Jump to 23:15Speaker · Lech Kalinowski