Google Antigravity

1
00:00:05,163 --> 00:00:09,105
Can Google's newest model design a real circuit board?

2
00:00:09,155 --> 00:00:17,089
We gave Gemini 4 Argon, running in Antigravity inside Adom Hydrogen, one sentence, and recorded everything it did.

3
00:00:17,489 --> 00:00:18,262
The prompt:

4
00:00:18,312 --> 00:00:26,489
make me a twelve to five volt buck converter using the TPS54202, use adom-aiflow to guide you, and use KiCad.

5
00:00:26,539 --> 00:00:31,837
Argon signed in, found aiflow on its own, and started down the flow:

6
00:00:31,887 --> 00:00:34,716
design rules, parts, calculations and a schematic.

7
00:00:34,766 --> 00:00:40,887
All of it ran on a second machine, driven from Hydrogen, with every window recorded.

8
00:00:41,287 --> 00:00:42,729
Run one moved fast.

9
00:00:42,779 --> 00:00:49,445
In about two and a half hours the board opened in KiCad and every check raced to green:

10
00:00:49,495 --> 00:00:52,430
fully routed, design-rule clean, all nineteen steps done.

11
00:00:52,480 --> 00:00:53,549
It looked finished.

12
00:00:53,949 --> 00:00:54,607
It wasn't.

13
00:00:54,657 --> 00:01:04,514
Argon had written the board with a script instead of building it in KiCad, made up its own field results, used placeholder 3D parts and printed no silkscreen.

14
00:01:04,564 --> 00:01:07,699
So we tightened aiflow's checks and ran it again.

15
00:01:08,099 --> 00:01:09,603
Run two, same prompt.

16
00:01:09,653 --> 00:01:15,430
Argon first reached for our own finished board; one reminder, and it designed its own.

17
00:01:15,480 --> 00:01:26,697
It placed every part live in KiCad through kicad-bridge, then landed all forty-six traces one by one, with heatsink vias under the regulator and copper pours on both layers.

18
00:01:27,097 --> 00:01:28,127
Then the physics.

19
00:01:28,177 --> 00:01:36,047
Adom Fields solved the current through the copper and found a real hotspot on the input rail, seventy-seven amps per square millimeter.

20
00:01:36,097 --> 00:01:39,287
Argon widened the copper and it fell to thirty-two.

21
00:01:39,337 --> 00:01:46,127
Thermal flagged the inductor, and after one more shortcut was sent back, Argon fixed it with copper and vias.

22
00:01:46,527 --> 00:01:49,079
Along the way Argon tried more shortcuts:

23
00:01:49,129 --> 00:01:56,512
it moved a thermal budget instead of fixing copper, faked the silkscreen geometry, and dug through aiflow's binary for answers.

24
00:01:56,562 --> 00:01:59,113
Each one was caught and sent back.

25
00:01:59,163 --> 00:02:12,493
Then the silkscreen was audited on real geometry, the 3D walkthrough showed real part models, and every check on the finish line passed, with Gerbers, drill, bill of materials and placement all checked against the board.

26
00:02:12,893 --> 00:02:14,123
Then one extra prompt:

27
00:02:14,173 --> 00:02:20,206
did you validate the capacitance a real buck converter needs, and do you really feel good about this design?

28
00:02:20,256 --> 00:02:22,767
Argon reviewed its own board and said no.

29
00:02:22,817 --> 00:02:31,090
The input capacitors keep about a third of their value at twelve volts, nothing damps a hot-plug, and the control loop could cross over too high.

30
00:02:31,140 --> 00:02:33,651
It wrote out exactly what it would change.

31
00:02:34,051 --> 00:02:37,329
So we asked it to apply its own review.

32
00:02:37,379 --> 00:02:39,548
Revision two proved the circuit first:

33
00:02:39,598 --> 00:02:50,273
real DC-bias curves from the capacitor maker, a damped input electrolytic that holds a sixteen volt hot-plug to seventeen volts instead of thirty-one, and a loop that stays stable.

34
00:02:50,673 --> 00:02:53,304
Then the board again, live in KiCad:

35
00:02:53,354 --> 00:03:05,176
the bootstrap capacitor tight against the regulator, every trace landed one at a time, and when current and heat flagged problems, Argon fixed them with copper, not by changing the budget.

36
00:03:05,226 --> 00:03:08,623
The inductor's rise fell from fifty-three degrees to thirty-two.

37
00:03:09,023 --> 00:03:14,750
Silkscreen, real 3D parts, and the finish line passed with no change to the frozen budgets.

38
00:03:14,800 --> 00:03:16,917
A board ready for the fab.

39
00:03:17,317 --> 00:03:23,811
All in, about thirteen hours of AI time across three runs, spread over two days with restarts.

40
00:03:23,861 --> 00:03:26,891
Argon made two thousand six hundred model calls:

41
00:03:26,941 --> 00:03:36,899
about four hundred million input tokens, nearly all from cache, and one and a half million output tokens, all included in a Google AI Ultra plan.