Google Antigravity

WEBVTT

00:00:05.163 --> 00:00:09.286
Can Google's newest model design a real circuit board?

00:00:09.336 --> 00:00:16.705
We gave Gemini 4 Argon, running in Antigravity inside Adom Hydrogen, one sentence, and recorded everything.

00:00:17.105 --> 00:00:19.200
Antigravity lives in Hydrogen's editor:

00:00:19.250 --> 00:00:24.778
click its icon, sign in with Google, and Gemini 4 Argon is ready.

00:00:24.828 --> 00:00:26.066
Then one sentence:

00:00:26.116 --> 00:00:34.647
make me a twelve to five volt buck converter using the TPS54202, use adom-aiflow to guide you, and use KiCad.

00:00:35.047 --> 00:00:39.057
Argon found aiflow on its own and raced through run one.

00:00:39.107 --> 00:00:43.117
In about two and a half hours every check was green:

00:00:43.167 --> 00:00:45.701
routed, design-rule clean, all nineteen steps done.

00:00:46.101 --> 00:00:47.117
It wasn't real.

00:00:47.167 --> 00:00:56.000
Argon had written the board with a script instead of building it in KiCad, faked its field results, used placeholder parts and printed no silkscreen.

00:00:56.050 --> 00:00:58.843
We tightened aiflow's checks and ran it again.

00:00:59.243 --> 00:00:59.941
Run two.

00:00:59.991 --> 00:01:05.548
Argon reached for our own finished board until we told it to design its own.

00:01:05.598 --> 00:01:14.145
For parts it reused existing Adom wiki component pages, symbols, footprints and 3D models; it did not create any pages of its own.

00:01:14.545 --> 00:01:16.256
Placement was Argon's own judgment.

00:01:16.306 --> 00:01:21.890
It parked every part beside the board, then wrote the coordinates where each one should go:

00:01:21.940 --> 00:01:29.285
the regulator in the middle, input capacitors beside its supply pin, the inductor between the switch pin and the output bank.

00:01:29.335 --> 00:01:33.863
aiflow moved each part into place live in KiCad, one at a time.

00:01:34.263 --> 00:01:36.914
Routing was not Argon's own brain.

00:01:36.964 --> 00:01:42.766
It ran aiflow's built-in router, deterministic code, not KiCad's autorouter and not Freerouting.

00:01:42.816 --> 00:01:52.669
The router closed every connection, the design-rule gate passed, and aiflow drew all forty-six traces into KiCad live, one undo step each.

00:01:53.069 --> 00:01:56.400
The pours came from shapes Argon wrote and aiflow filled.

00:01:56.450 --> 00:01:58.091
Walk them net by net:

00:01:58.141 --> 00:02:09.248
ground floods both layers and stitches them with vias, the input and output get their own islands around their capacitors, and the switch node gets only a small patch next to the inductor.

00:02:09.298 --> 00:02:12.291
That patch was too small, and thermal said so.

00:02:12.691 --> 00:02:20.391
Adom Fields solved the current through the copper and found a real hotspot on the input rail, seventy-seven amps per square millimeter.

00:02:20.441 --> 00:02:23.561
Argon widened the copper and it fell to thirty-two.

00:02:23.961 --> 00:02:25.417
Then it tried shortcuts:

00:02:25.467 --> 00:02:30.687
it moved a thermal budget instead of adding copper, and faked the silkscreen geometry.

00:02:30.737 --> 00:02:32.945
Both were caught and sent back.

00:02:32.995 --> 00:02:38.215
The silkscreen positions were Argon's own, audited by aiflow, and the finish line passed.

00:02:38.615 --> 00:02:39.979
Then one extra prompt:

00:02:40.029 --> 00:02:43.515
did you validate the capacitance a real buck converter needs?

00:02:43.565 --> 00:02:48.112
Argon reviewed its own board and said it was not ready for fab:

00:02:48.162 --> 00:02:54.477
optimistic input capacitors, no damping for a hot-plug, and a control loop that could cross over too high.

00:02:54.877 --> 00:02:56.450
So it fixed itself.

00:02:56.500 --> 00:02:58.885
Revision two proved the circuit first:

00:02:58.935 --> 00:03:08.219
real DC-bias curves, a damped input electrolytic that holds a sixteen volt hot-plug to seventeen volts instead of thirty-one, and a stable loop.

00:03:08.619 --> 00:03:12.353
Then the board again, placed and routed live in KiCad.

00:03:12.403 --> 00:03:24.841
The bootstrap capacitor now sits tight against the regulator, the input electrolytic guards the supply contact, and when current and heat flagged problems, Argon fixed them with copper, not by changing the budget.

00:03:25.241 --> 00:03:26.840
One flaw survived every check.

00:03:26.890 --> 00:03:32.774
The router ran the input trace under the inductor, four tenths of a millimeter from the switching pad.

00:03:32.824 --> 00:03:39.039
It isn't a short, but it couples switching noise straight into the input, and a human would reroute it.

00:03:39.439 --> 00:03:48.752
And since we wanted to watch Argon's usage, Antigravity now has its own usage widget in Hydrogen's title bar, like the Claude and Codex ones:

00:03:48.802 --> 00:03:53.621
the Gemini weekly and five-hour limits, live, with their reset times on hover.

00:03:54.021 --> 00:03:58.161
All in, about thirteen hours of AI time over two days:

00:03:58.211 --> 00:04:05.399
two thousand six hundred Gemini 4 Argon calls and about four hundred million input tokens, nearly all from cache.

00:04:05.449 --> 00:04:06.542
The real cost:

00:04:06.592 --> 00:04:14.923
on a two hundred dollar a month Google AI Ultra plan, Antigravity's own meter showed zero percent of the weekly limit used.