Runtime ≈ 2:45 · ~390 spoken words · unhurried pace
Plain text is what you say, word for word. Italics in brackets are stage directions — do not read them out.
[Terminal is typing. Say nothing for the first three seconds. Let it build.]
That took four seconds. A payment system — API, database, tests. It's real code. It probably works.
[Screen turns: BUT… Who verified it?]
Nobody did.
[Boot sequence powering up the systems.]
AI writes software faster than any human can review it. The bottleneck isn't writing code any more. It's trust.
[Title card appears. Click ENTER COMMAND CENTER.]
This is Code Arena. Not a review dashboard — a command center for autonomous software verification.
[Gesture at the queue.]
Three pull requests waiting. The system has already decided which one goes first, and which agents to send.
[Gesture at the infrastructure panel, right side.]
And every layer on the right is a real integration. Green means it's calling the live API right now.
[Click ⚔ ENTER ARENA on PR #4921.]
Greptile maps the entire repository first. Eighteen hundred files, ninety-eight percent understanding. These agents read the codebase — they don't guess from the diff.
Modal spins up isolated sandboxes. Every attack runs in a real container, never against production.
[Fleet launches.]
Four agents.
[First hit lands.]
Bug Hunter — double refund vulnerability. Health drops fifteen.
Now watch this.
Security finds a missing authorization check. Then it finds the same pattern in four more services — and three of them are files this pull request never touched.
[PAUSE. Two full seconds. Let it land.]
That's the difference. A human reviews the diff. This reviews the codebase.
Performance simulates ten thousand requests and finds the database bottleneck — p99 goes from three hundred milliseconds to over two thousand.
Then OpenAI reasons across all twelve findings, and the coding agent rebuilds the code. Watch the city repair itself.
But fixing the code isn't enough to ship it. The build goes into a simulated production environment — ten thousand users, database stress, latency injection.
Traffic spike, passes. Database recovery, passes.
[PAUSE — wait for the red row to appear.]
Payment retry — fails.
Every patch was correct. Every test passed. It still falls over, because the problem isn't in the diff. It's in the architecture the diff was written into.
[The gate appears. The arena stops. Let the silence sit for a beat.]
So the arena stops.
This is the part that matters. AI did all of the execution — twelve fixes, on its own authority. It escalates exactly one thing to a human, with a recommendation and ninety-four percent confidence.
[Click ✓ APPROVE.]
Approved. The agent applies the migration, re-runs the failed stage, and it passes.
[Verification → score → the report opens on its own. Scroll slowly.]
Verified. Ready to ship.
And the developer gets this — not a score. Every problem, exactly where it is, and the fix. The simulation results. What the AI recommended and what a human chose. And a full replay, so every change is explainable.
[Click ▶ FINAL VISION. Camera pulls away over the city.]
Software engineering is moving from humans reviewing code…
[PAUSE.]
…to AI systems proving it.
AI doesn't just write code. AI proves code deserves to exist.
[Stop talking. Hold on the final frame for three seconds before you cut.]
- Three pauses carry this video: after the variant reveal, after "Payment retry — fails", and before the last line. Don't rush them — they do more work than any sentence in the script.
- Say numbers as words: "eighteen hundred", "ninety-four percent", "p99" as "p-ninety-nine".
- Don't explain the 3D. It explains itself, and narrating visuals reads as filler.
- If the wifi is bad, drop the "calling the live API right now" line rather than pointing at a panel that says SIM.
- Record one clean take of the closing three lines separately. It is the line people remember, and it is worth having a spare.
Keep four beats only:
| Time | Beat |
|---|---|
| 0:00 — 0:10 | The hook |
| 0:10 — 0:25 | The variant chain |
| 0:25 — 0:48 | Simulation failure + the human decision |
| 0:48 — 1:00 | The closing line |
Record the arena at 2× and cut the scan, repair and report sections entirely. Those four beats are the whole argument.
-
npm run dev, fullscreen, browser zoom at 100% - Run the demo once end to end so shaders are compiled and warm
- Close other GPU-heavy tabs
- Decide: driving it yourself, or ▶ AUTO DEMO · 2:30
- Have
/reportopen in a second tab if you want to cut to it - See
DEMO.mdfor click-by-click recovery steps