WEBVTT

NOTE Narration transcript. Text is the script synthesized by minimax (speech-2.8-hd, voice English_expressive_narrator); each cue was synthesized as its own audio file and is timed by that file's measured duration. Verified at 103.9% of expected length, 139.6 wpm (median cue 142.2 wpm). Script source: public/videos/the-02-percent-synthesis-problem.vtt at commit 4641d96 (pre-overwrite narration prose)

1
00:00:00.000 --> 00:00:05.380
Okay, so AI predicted three hundred eighty thousand stable new materials.

2
00:00:05.680 --> 00:00:07.290
Huge, right?

3
00:00:07.590 --> 00:00:13.510
But by late 2023, only seven hundred thirty-six had actually been synthesized.

4
00:00:13.810 --> 00:00:16.030
That’s zero point two percent.

5
00:00:16.330 --> 00:00:19.400
And no, this isn’t just labs moving slowly.

6
00:00:19.700 --> 00:00:25.800
A crystal can look perfect on a computer and still have no practical chemical path into existence.

7
00:00:26.100 --> 00:00:28.060
Prediction is scaling.

8
00:00:28.360 --> 00:00:30.660
Synthesizability isn’t.

9
00:00:30.960 --> 00:00:33.430
Most screens skip four brutal filters.

10
00:00:33.730 --> 00:00:38.630
One: can real precursors reach the target instead of some competing phase?

11
00:00:38.930 --> 00:00:44.050
Two: are the forces accurate away from a tidy, equilibrium crystal?

12
00:00:44.350 --> 00:00:51.690
Three: what happens when the material has vacancies, disorder, or stacking faults—like real materials do?

13
00:00:51.990 --> 00:00:54.850
And four: what does failure cost?

14
00:00:55.150 --> 00:00:58.300
Usually weeks of lab time and thousands of dollars.

15
00:00:58.600 --> 00:01:00.950
This is not a tiny accuracy problem.

16
00:01:01.250 --> 00:01:09.230
Miss an ion-migration barrier by one hundred milli-electronvolts, and the predicted hopping rate changes about fifty-fold at room temperature.

17
00:01:09.530 --> 00:01:12.420
You can literally call a conductor an insulator.

18
00:01:12.720 --> 00:01:15.150
The answer isn’t a bigger generator.

19
00:01:15.450 --> 00:01:25.770
It’s a correction and verification layer: an environment error field that measures where universal potentials go wrong, then corrects them at runtime with proper forces.

20
00:01:26.070 --> 00:01:27.940
Better candidates rise.

21
00:01:28.240 --> 00:01:31.220
Bad ones get flagged before the furnace turns on.

22
00:01:31.520 --> 00:01:38.350
That matters because batteries are tied directly to roughly twenty percent of the emissions cuts needed by 2030.

23
00:01:38.650 --> 00:01:47.140
Cobalt-free cathodes could also break a supply chokepoint: about seventy percent of cobalt comes from the Democratic Republic of Congo.

24
00:01:47.440 --> 00:01:51.400
And every decade of delay locks in emissions we cannot recover later.

25
00:01:51.700 --> 00:01:59.200
The goal is simple: predictions that flow to synthesis, characterization, and scale—not another generation cycle.

26
00:01:59.500 --> 00:02:04.413
Read the full article to see how the zero-point-two-percent gap closes.