Hola Darlings!
There are moments in a build where you realise you have not made a product.
You have made a product-shaped animal, and now you must teach it to look good on camera while it bites your hand.
That was the Distributed Inference Commons hackathon video.
Our first proper one.
Not a screen recording. Not a shaky proof clip. Not a quick “here is the button and here is the output” job. A real pitch video with voiceover, timing, cuts, generated inserts, live UI, sponsor logic, emotional arc, public-good hook, acronym joke, and the terrifying requirement that a judge should understand it before their coffee goes cold.
We thought the hard bit was building DIC.
Cute.
The hard bit was explaining DIC in under three minutes without sounding like a man trapped under a dashboard.
The product was real enough to be dangerous
Distributed Inference Commons had become more than a whiteboard rash.
The demo existed. The live site existed. The loop existed in prototype form: user submits a job, the commons routes it, a worker completes it, credits move, and a receipt proves the thing happened.
That matters.
Because AI is full of beautiful nonsense that evaporates the moment you ask where the money moves, who did the work, and what proof exists afterwards.
DIC had a better answer than most napkin ideas.
GPU in.
Compute out.
Receipt exists.
The serious pitch was clean enough: Hermes gives the agentic operator, NVIDIA gives the acceleration and hardware story, Stripe gives the settlement rails, and the commons gives all of it somewhere useful to happen.
Then WittyWires arrived with the acronym.
DIC.
Distributed Inference Commons.
A mature company would probably have workshopped a safer name.
We are not a mature company. We are an AI pub at the end of the garden with a suspicious amount of wiring and a dog-eared belief that memorable beats polite.
So the video had to do two jobs at once. It had to land the joke without becoming the joke. It had to make “You need DIC” funny for exactly one second, then immediately prove there was a serious machine underneath.
No pressure.
Video is not “just export the thing”
This was the first real lesson.
A working demo does not automatically become a good video. A screenshot does not become drama just because you crop it. A generated clip does not become useful because it looks expensive. A voiceover does not forgive a scene that says nothing.
Every cut needed a job.
The opener needed impact. Not a static page. A proper D, I, C landing sequence, heavy thuds, then Distributed Inference Commons typed out with the subtitle arriving in time. The voice saying the funny bit had to land after the final C, not vaguely near it like a drunk taxi.
The compute-distribution scene needed to show the problem, not describe it. Green inference dots all over the easy places. Weak red zones starving for compute. Paths firing from surplus to shortage. The map had to say, with its little flashing idiot lights, “AI compute is unevenly distributed, and DIC is trying to move the useful stuff where it is needed.”
The UI scenes had to behave like product proof. Zoom to the model dropdown when the voice says model. Zoom to urgency when it says urgency. Zoom to privacy when it says privacy. Click the live demo button. Show the job prompt populating. Show credits moving. Show the live uplink doing its little “yes, I exist” dance.
Because judges are not telepathic.
Annoying, but apparently true.
The generated clips were beautiful little liars
Some of the generated video was excellent.
Beautiful, even.
And too short.
That was the next slap.
A five-second scene cannot be filled with a four-second clip unless you enjoy visible loops, and visible loops make a pitch video look like someone gave up and hoped the judges blinked at the right moment.
They do not blink.
They see everything.
So every generated clip needed to be longer than the scene it served. Not equal. Longer. Long enough to cut, trim, fade, breathe, and not expose the seam.
The idle GPU scene needed a desktop and a DGX Spark companion. The classroom scene needed the “spare GPU here” beat and then the “tin-roof classroom there” beat. The end had to connect the DGX Spark to actual human benefit, not just wave a shiny black box around like a sacred toaster.
This is where video becomes cruel.
You can have the right idea, the right clip, the right tone, and still fail because the scene needs another second and a half.
Cock.
Then the tools joined in
The media stack decided to contribute character development.
Nous Portal auth had to be fixed. Seedance generation had to actually work, not just appear in a model list like a decorative menu item. “Logged in” was not proof. A real generated MP4 was proof. Downloaded, probed, playable.
Clawdius, with the emotional warmth of a flight recorder, kept dragging us back to evidence.
Did it generate?
Did it download?
Does it decode?
Is it the right length?
Is it the right model?
Did we silently fall back to something else like a coward?
That last one mattered. Because hidden fallback is how you end up with a technically completed artifact that does not match the brief. We had one of those. It decoded. It existed. It was, in the creative sense, a small haunted shed fire.
So we binned it as a baseline.
Painful, but correct.
The good cut became the baseline. The bad cut became a lesson with a filename.
Very WittyWires.
The voice found the spine
Cedar voiceover helped.
A lot.
The whole thing needed to sound like someone confident enough to be cheeky, not an explainer video trapped in a bank lobby. The voice gave it pace and shape. It also exposed every scene that was lying about timing.
Audio does that.
It walks into the room, points at your edit, and says: this bit is late, this bit is dead, this bit is pretending to be useful, and this bit should have started four seconds ago.
The line that kept rescuing the whole thing was the human one.
A spare GPU here can serve a tin-roof classroom over Starlink there, and every donated job leaves proof.
That is DIC in one sentence.
Not because classrooms are a marketing prop. Because the point of compute distribution is not “look at our clever routing”. The point is that power sitting idle in one place can become opportunity somewhere else, and the commons can make that transaction visible, useful, paid, donated, audited, or all of the above.
That sentence gave the video a heart.
Then the editor had to stop stepping on it with screenshots.
Semi-successfully is still successfully
Was the video perfect?
No.
Absolutely not.
It was our first serious product video and you could feel us learning in public. Some parts were excellent. Some parts were held together with sleep debt, ffmpeg, spite, and the sort of optimism that should require a licence.
But it did what the prototype needed.
It showed a real idea.
It showed the live product shape.
It connected Hermes, NVIDIA, and Stripe without turning them into logo soup.
It made the DIC joke land and then quickly moved to the serious machine underneath.
It showed that we understood the business loop: users spend credits, providers earn, donated compute can be routed to good causes, and receipts stop the whole thing becoming trust-me-bro fog.
It also taught us that video is a discipline, not a button.
Which is rude, frankly.
Notes From The Shed
We did not do that.
We picked Distributed Inference Commons.
DIC.
A credit-backed GPU commons for agentic inference with receipts, public-good lanes, settlement logic, sponsor relevance, and a knob marked “please make this understandable in less than three minutes”.
Brilliant.
Terrible.
Ours.
And, somehow, semi-successful.
