WEBVTT

1
00:00:00.191 --> 00:00:07.517
Hello Internet, this is your captain speaking Lester Nare, joined as always by my co-host and our resident PhD Krishna Chowdhury.

2
00:00:07.898 --> 00:00:09.679
This is From First Principles.

3
00:00:10.080 --> 00:00:11.961
We have a great episode for you this week.

4
00:00:11.981 --> 00:00:23.111
We're going to cover four stories in AI, space and immunology, starting with a new story out of China's DeepSeek, which says its AI model costs just $294.99.

5
00:00:24.832 --> 00:00:26.173
$1,000 to train.

6
00:00:26.493 --> 00:00:29.215
This is really important, and we're gonna go a deep dive into why.

7
00:00:29.716 --> 00:00:33.739
Followed up by our second story, sounds like we may have discovered the Ocarina of Time.

8
00:00:33.799 --> 00:00:35.660
We may need to get Zelda involved.

9
00:00:36.040 --> 00:00:40.483
As scientists just made the first time crystal that you can see.

10
00:00:40.964 --> 00:00:42.345
I'm really interested in talking about that one.

11
00:00:43.165 --> 00:00:49.070
The third story is about hidden viruses in our DNA that could be medicine's next big breakthrough.

12
00:00:49.490 --> 00:00:58.618
And we're going to end with space with the brightest fast radio burst ever detected and how it could help us solve an enduring cosmic mystery.

13
00:00:58.758 --> 00:01:01.500
This is From First Principles.

14
00:01:17.345 --> 00:01:17.745
My friend.

15
00:01:18.807 --> 00:01:19.387
How's it going?

16
00:01:19.668 --> 00:01:20.288
Episode 9.

17
00:01:20.829 --> 00:01:21.449
Episode 9.

18
00:01:21.509 --> 00:01:22.811
10k followers on Instagram.

19
00:01:22.831 --> 00:01:23.752
Woo!

20
00:01:24.052 --> 00:01:26.355
We still haven't been shut down by the FCC yet.

21
00:01:26.375 --> 00:01:26.975
Yeah, yeah, yeah.

22
00:01:27.216 --> 00:01:27.936
It's coming, though.

23
00:01:27.956 --> 00:01:28.317
It's coming.

24
00:01:28.437 --> 00:01:29.538
We're talking too much science.

25
00:01:29.598 --> 00:01:30.018
It's coming.

26
00:01:30.038 --> 00:01:30.499
It's coming.

27
00:01:30.519 --> 00:01:32.481
They're going to defund us and then ban us.

28
00:01:32.601 --> 00:01:35.644
But before they do, we're going to talk about some great stories.

29
00:01:36.825 --> 00:01:39.206
The first one, we always end up touching on AI.

30
00:01:39.706 --> 00:01:42.707
There's a lot of research going on in AI in a variety of ways.

31
00:01:44.508 --> 00:01:51.990
And our first story was a story in Reuters that also had two nature papers that were related to it.

32
00:01:52.150 --> 00:01:52.310
Yeah.

33
00:01:52.850 --> 00:02:01.292
The headline on this is China's DeepSeek says its AI model has hit a cost of just $294,000 to train.

34
00:02:02.092 --> 00:02:09.313
The sort of byline is in a paper that is likely to reignite the debate over Beijing's place in the race to develop AI.

35
00:02:09.373 --> 00:02:09.593
Yeah.

36
00:02:10.053 --> 00:02:16.214
This sort of cost, which is significantly lower than the figures that we see in the US.

37
00:02:16.354 --> 00:02:20.595
And I know we've talked about DeepSeek and all of this quite a bit.

38
00:02:20.655 --> 00:02:20.935
Yeah.

39
00:02:21.935 --> 00:02:24.136
Uh, it's a, it's a crazy model.

40
00:02:24.557 --> 00:02:30.380
And the fact that the team like put out a peer review paper, you know, no one else has done that in the field of AI.

41
00:02:30.500 --> 00:02:30.620
Uh,

42
00:02:31.540 --> 00:02:34.562
like gone to nature and put out, okay, this is how we trained.

43
00:02:34.902 --> 00:02:35.922
This is how we did it.

44
00:02:37.403 --> 00:02:45.007
It is, you know, it is, they're actually, it's, it's kind of crazy that China is setting the precedent for open science in the field of AI.

45
00:02:45.527 --> 00:02:45.767
Right.

46
00:02:46.388 --> 00:02:47.608
It is very interesting.

47
00:02:47.828 --> 00:02:54.852
And not only one, but two papers in, in nature, main brand, not, not a nature sub, you know, sub category.

48
00:02:54.892 --> 00:02:57.813
I triple E. This isn't even like, you know, Neurips.

49
00:02:57.933 --> 00:02:58.594
This is nature.

50
00:02:58.614 --> 00:02:58.694
And,

51
00:02:59.454 --> 00:03:04.136
And what I thought would be helpful, we're going to do sort of two aspects on this story.

52
00:03:04.156 --> 00:03:05.097
Yeah.

53
00:03:05.517 --> 00:03:11.660
Folks might have seen the Deep Seek news earlier in the year when there was sort of a big outcry about it.

54
00:03:11.740 --> 00:03:13.521
Yeah, they released in January, I think.

55
00:03:13.581 --> 00:03:13.821
Yes.

56
00:03:13.921 --> 00:03:15.342
And then all of our stocks failed.

57
00:03:15.522 --> 00:03:16.963
Yes, it was a big deal.

58
00:03:17.003 --> 00:03:17.843
Yeah.

59
00:03:17.983 --> 00:03:22.766
But January wasn't the start of this great power war over AI.

60
00:03:22.886 --> 00:03:23.206
Right.

61
00:03:23.406 --> 00:03:28.949
And I thought it would be good to just go for a couple of beats on the sort of geopolitical context of

62
00:03:29.429 --> 00:03:31.732
that this research paper kind of came out in.

63
00:03:31.872 --> 00:03:32.252
Yeah, yeah.

64
00:03:32.473 --> 00:03:37.137
And to sort of set the stage, right, there's an AI great power competition between the U.S. and China.

65
00:03:38.619 --> 00:03:44.625
AI, as we know, it has existed for quite some time, you know, before all this generative AI stuff.

66
00:03:44.685 --> 00:03:44.986
Yeah.

67
00:03:46.207 --> 00:03:57.782
But in the sort of recent context, we're going to start in 2017, where China launched its next generation AI plan, which said it was going to lead the world in AI by 2030.

68
00:03:58.203 --> 00:03:59.885
They're famous for their five-year plans.

69
00:03:59.985 --> 00:04:00.186
Right.

70
00:04:00.266 --> 00:04:00.446
Right.

71
00:04:00.586 --> 00:04:03.527
And the idea is U.S. has been the global hegemon forever.

72
00:04:03.947 --> 00:04:07.347
China's looking at a way in which it can sort of destabilize that control.

73
00:04:08.048 --> 00:04:16.589
And Xi Jinping, leader of the Chinese Communist Party, said that advanced tech is the sharp weapon of modern states.

74
00:04:16.829 --> 00:04:17.529
Yeah, that's true.

75
00:04:17.589 --> 00:04:29.952
And as we've discussed on this podcast previously, there is a very strong relationship between the military intelligence and national security apparatus within nation states and frontier technologies.

76
00:04:29.972 --> 00:04:30.132
Yeah.

77
00:04:30.672 --> 00:04:31.773
AI just being one of those.

78
00:04:31.793 --> 00:04:33.373
Yeah.

79
00:04:33.673 --> 00:04:46.218
And then in 2021, what was interesting is the US National Security Commission on AI, which was chaired by the former Google CEO Eric Schmidt, warned that American dominance in AI was not guaranteed.

80
00:04:46.578 --> 00:04:48.939
So we already have this setup of that friction.

81
00:04:49.219 --> 00:04:49.459
Wow.

82
00:04:49.719 --> 00:04:59.123
Now, if we go into the Biden era policies, because the US government, as we know, as it relates to technology innovation, is usually reactive as opposed to proactive.

83
00:04:59.223 --> 00:04:59.463
Yes.

84
00:04:59.704 --> 00:04:59.804
Yes.

85
00:05:23.265 --> 00:05:28.913
The Biden era was the first administration to take policy in earnest in terms of AI specifically.

86
00:05:29.053 --> 00:05:29.333
Okay.

87
00:05:29.493 --> 00:05:36.283
And this started in August 2022 when Biden signed the Chips and Science Act, which many people have probably heard the Chips Act.

88
00:05:36.303 --> 00:05:37.164
Yeah, the Chips Act.

89
00:05:37.364 --> 00:05:42.589
$250 billion bipartisan bill aimed at boosting U.S. semiconductor manufacturing and innovation.

90
00:05:42.909 --> 00:05:46.873
This dovetails with the re-onshoring of manufacturing conversation post-COVID.

91
00:05:47.493 --> 00:05:56.141
There was $52.7 billion in supplemental appropriations that was specific to semiconductor-related programs for 2023 through 2027.

92
00:05:56.181 --> 00:06:00.104
The idea being that it was also going to create guardrails.

93
00:06:00.905 --> 00:06:08.427
preventing recipients like Intel and TSMC from expanding their advanced fabs in China.

94
00:06:08.847 --> 00:06:14.469
So we can see the theory of the case here was, can we create a choke point at the hardware level?

95
00:06:14.789 --> 00:06:23.452
Because these chips were viewed at the time as being the key aspect of the stack that was gonna be the driver of progress and innovation.

96
00:06:23.872 --> 00:06:29.876
And TSMC is in Taiwan, which is like now iffy if China's going to do something, right?

97
00:06:29.996 --> 00:06:30.356
Correct.

98
00:06:30.416 --> 00:06:35.719
China's been doing a huge military buildup and training around this idea of we're going to take Taiwan.

99
00:06:35.799 --> 00:06:36.200
Yeah, yeah.

100
00:06:36.240 --> 00:06:37.901
That's basically their thesis.

101
00:06:37.921 --> 00:06:43.304
We're going to take Taiwan and you don't want to be here when that happens to us, the U.S. Yeah, yeah.

102
00:06:43.524 --> 00:06:45.946
And so there are a lot of layers to this.

103
00:06:46.006 --> 00:06:47.507
The Taiwan point's a really important one.

104
00:06:48.273 --> 00:06:53.055
Now, 2022, we're still in 22 August, CHIPS Act drops.

105
00:06:53.215 --> 00:07:00.697
And then October 7th, 2022, the October 7th controls were established, which are export controls.

106
00:07:00.837 --> 00:07:08.020
Now, they named it this prior to the unfortunate October 7th of 2023, which is the Israel-Gaza conflict.

107
00:07:08.340 --> 00:07:15.426
But at the time, this was a restriction on exports of specific NVIDIA and AMD chips.

108
00:07:15.546 --> 00:07:25.715
The H100s, A100s, and the AMD Mi250s, which basically meant this was the whole, we have a small yard, high fence.

109
00:07:25.935 --> 00:07:26.496
Yeah.

110
00:07:26.576 --> 00:07:29.258
Making it really, really difficult to get these sort of frontier chips.

111
00:07:29.458 --> 00:07:32.162
Yeah, and the thesis was that, like, these chips are the future.

112
00:07:32.282 --> 00:07:32.482
Yes.

113
00:07:32.602 --> 00:07:34.225
And without these chips, you can't do shit.

114
00:07:34.365 --> 00:07:34.765
Exactly.

115
00:07:34.825 --> 00:07:36.688
That was the theory of the case at the time.

116
00:07:36.708 --> 00:07:36.848
Yeah.

117
00:07:37.629 --> 00:07:45.079
As a small side note, it's interesting that a month, almost just over a month later, on November 22, 2022, ChatCBT launched.

118
00:07:48.129 --> 00:07:48.569
Ah, okay.

119
00:07:48.629 --> 00:08:02.316
So all this conversation was happening prior even to the meteoric rise of the first real production level LLM generative AI accessible to millions of people.

120
00:08:02.356 --> 00:08:03.797
And the growth curve, as we all know.

121
00:08:03.817 --> 00:08:09.960
I mean, ChatGPT was the moment where now everyone was like, oh, AI is here.

122
00:08:10.040 --> 00:08:10.240
Yes.

123
00:08:10.740 --> 00:08:13.421
It's not just this thing people are saying and da da da da.

124
00:08:13.482 --> 00:08:14.342
It's not theoretical.

125
00:08:15.102 --> 00:08:16.863
I can literally chat with it.

126
00:08:17.003 --> 00:08:17.223
Yes.

127
00:08:17.343 --> 00:08:19.204
And it's like giving me amazing answers.

128
00:08:19.284 --> 00:08:19.504
Yes.

129
00:08:19.564 --> 00:08:19.785
Yeah.

130
00:08:20.025 --> 00:08:20.245
Yes.

131
00:08:20.425 --> 00:08:20.625
Yeah.

132
00:08:20.865 --> 00:08:22.986
So now we're moving into 2023.

133
00:08:23.166 --> 00:08:24.247
Right.

134
00:08:24.387 --> 00:08:37.193
So throughout 2023, now the Biden administration is not only trying to create the choke points domestically, but now we're going out to our allies all across the world to sort of expand that tightening.

135
00:08:37.573 --> 00:08:40.995
So this included in the Netherlands with ASML.

136
00:08:41.355 --> 00:08:53.378
And in Japan with Nikon and Tokyo Electron blocking them from providing China access to their lithography tools, which are used to actually build and generate these more advanced chips.

137
00:08:53.678 --> 00:08:54.158
Yeah.

138
00:08:54.218 --> 00:09:07.481
This was followed up by in August of 23, Executive Order 14105, which now also restricted investment from U.S. investors into any China AI or quantum firm.

139
00:09:07.761 --> 00:09:08.021
Okay.

140
00:09:11.682 --> 00:09:34.217
like choking all pathways right to hardware innovation yeah for hardware and for hardware innovation okay right so while this was all happening in 23 uh interestingly enough uh there was the very very sort of public falling out of the founder or ceo at open ai sam altman so many

141
00:09:34.677 --> 00:09:49.533
folks may remember there was that huge board battle i forgot about yeah dude he was ousted he was ousted yes and this was around open ai was a non-profit they were doing a lot of for-profit moves they wanted to kind of rejigger their structure

142
00:09:50.534 --> 00:10:04.324
And it kind of signified the realization across the Silicon Valley consensus that this thing was going to be bigger, more powerful, more impactful, more money generating than even they had predicted.

143
00:10:04.344 --> 00:10:07.006
It was happening quicker than they expected.

144
00:10:07.086 --> 00:10:08.307
So there was that power struggle.

145
00:10:08.627 --> 00:10:08.887
Right.

146
00:10:09.127 --> 00:10:11.189
Ultimately, the king returned to the castle.

147
00:10:11.649 --> 00:10:12.009
Yeah.

148
00:10:12.309 --> 00:10:15.591
And Sam Reid got his position as CEO.

149
00:10:16.111 --> 00:10:18.412
But all this we've talked about so far has been the U.S. side.

150
00:10:18.632 --> 00:10:18.832
Yeah.

151
00:10:19.152 --> 00:10:23.934
So in 23, China's response, what has China been doing as the U.S. That's right.

152
00:10:23.954 --> 00:10:27.455
So, yeah, we've been battling internally, so to speak.

153
00:10:27.675 --> 00:10:31.697
And to sort of create this regulatory regime around access to chips.

154
00:10:31.797 --> 00:10:31.917
Right.

155
00:10:31.937 --> 00:10:32.537
So what's China's response?

156
00:10:33.037 --> 00:10:52.682
So China had firms like Huawei introduce their own, you know, what they call frontier chips with the Ascend AI chip, as well as shocking the global community when SMIC released their 7 nanometer process.

157
00:10:53.482 --> 00:11:02.766
for smartphone chips, which again, the idea was TSMC has really had a monopoly on the seven, five, and I believe three now, nanometer chip sizes.

158
00:11:02.786 --> 00:11:05.387
Their latest is three, but like sub 10 was them.

159
00:11:05.907 --> 00:11:08.168
Right, and no one else could really do it.

160
00:11:08.208 --> 00:11:09.809
And that was like a really big bottleneck.

161
00:11:10.929 --> 00:11:17.795
both for, and there's both operational reasons and proprietary intellectual property reasons why that's difficult.

162
00:11:18.035 --> 00:11:27.363
They also then had their other big conglomerate technology companies, Baidu, Alibaba, begin to put out their own chat GPT type LLMs.

163
00:11:27.423 --> 00:11:29.404
So we had Quinn, we had Ernie Bot,

164
00:11:30.285 --> 00:11:32.347
But nobody really... No one really thought anything.

165
00:11:32.367 --> 00:11:33.067
Nobody really cared.

166
00:11:33.127 --> 00:11:33.828
No one cared.

167
00:11:34.068 --> 00:11:43.135
They were all trained on the cut-down NVIDIA H800 GPUs, which were still allowed to be sold to China, but they were massively sort of handicapped.

168
00:11:43.215 --> 00:11:43.456
Yeah.

169
00:11:45.537 --> 00:11:57.528
Interestingly, now we're sort of getting into 2024, which is more of the same, but we talked about the intersection of the national security state, the intelligence community, military intelligence, and frontier technologies.

170
00:11:57.688 --> 00:11:57.908
Yeah.

171
00:11:58.329 --> 00:12:02.753
And Marc Andreessen made this interesting note in a podcast where he...

172
00:12:03.393 --> 00:12:10.760
Basically was trying to call out the folks who are called decels or the folks who are really heavy on AI safety.

173
00:12:10.820 --> 00:12:11.601
Deceleration.

174
00:12:11.881 --> 00:12:14.964
Yeah, the decelerationists who are saying let's slow down, let's slow our roll.

175
00:12:14.984 --> 00:12:15.624
Got it.

176
00:12:15.724 --> 00:12:16.765
Elon is a famous one.

177
00:12:16.946 --> 00:12:19.408
Elon is arguably the most famous decelerationist.

178
00:12:20.068 --> 00:12:40.165
And Marc Andreessen's point was, if you believe that this threat is as existential as nuclear weapons, you're still allowing these open AI employees to go to an unprotected, unsecured building in the same way that – not in the same way that, like, nuclear facilities are protected from an operational security perspective.

179
00:12:40.205 --> 00:12:45.089
So he's like, if you believe this to be true, the entire operating infrastructure of these companies needs to change.

180
00:12:45.229 --> 00:12:45.469
Yeah.

181
00:12:46.330 --> 00:12:48.153
That's fair, actually.

182
00:12:48.253 --> 00:12:49.476
I never thought about that, but yeah.

183
00:12:50.077 --> 00:12:51.239
It's a fair argument, right?

184
00:12:51.259 --> 00:12:51.419
Yeah.

185
00:12:51.779 --> 00:12:53.082
And in 2024, on June 13th,

186
00:12:56.700 --> 00:13:14.968
After all this board kerfuffle in 23, the former head of the National Security Agency and commander of US Cyber Command, Paul Nakasone, was added to the OpenAI board, which was the first salvo in what I like to call the nationalization of the frontier models in the United States.

187
00:13:14.988 --> 00:13:15.248
Got it.

188
00:13:15.468 --> 00:13:18.630
Right now, the national security state has a seat at the table.

189
00:13:19.750 --> 00:13:25.293
Shortly after, OpenAI announced several projects with the Pentagon, the DOD around using it.

190
00:13:25.473 --> 00:13:30.095
So this is all background leading up to the Sputnik moment for China.

191
00:13:30.295 --> 00:13:31.036
That was a big one.

192
00:13:31.216 --> 00:13:44.002
Which is in January of this year, 2025, when the Chinese startup DeepSeek released their R1 model, which was claiming at the time a Chachapiti 4-level reasoning.

193
00:13:44.222 --> 00:13:46.123
So this is like our best, most frontier.

194
00:13:46.143 --> 00:13:50.145
This is pre-ChatGPT5, which we now have today in September.

195
00:13:51.326 --> 00:13:55.108
Fully trained on NVIDIA H800s.

196
00:13:55.228 --> 00:13:55.468
Yes.

197
00:13:55.708 --> 00:13:56.929
The not good chips.

198
00:13:56.949 --> 00:13:58.529
The not good chips, yeah.

199
00:13:58.649 --> 00:14:06.373
Using a mixture of experts techniques for the training alongside with reinforcement learning at what we now understand to be a sub $300,000 training cost.

200
00:14:06.694 --> 00:14:08.815
Now it's important to note the amount of money...

201
00:14:09.695 --> 00:14:22.525
that Anthropic OpenAI Meta has raised to train, Grok, to train their foundation models is in the order of tens of millions, hundreds of millions of dollars.

202
00:14:22.685 --> 00:14:22.885
Yeah.

203
00:14:23.025 --> 00:14:25.127
So this is an orders of magnitude difference.

204
00:14:25.187 --> 00:14:26.047
Orders of magnitude.

205
00:14:27.308 --> 00:14:37.738
And so what we ended up creating by doing this choke point on the hardware side was necessity is the father invention or whatever, right?

206
00:14:38.358 --> 00:14:44.644
China then said, okay, let's innovate on the training model and infrastructure to try to get the same efficacy.

207
00:14:44.744 --> 00:14:47.066
On the algorithms and the software.

208
00:14:47.447 --> 00:14:48.207
Not on the hardware.

209
00:14:48.668 --> 00:14:50.630
So this created a huge...

210
00:14:53.341 --> 00:15:00.187
surprise from a lot of US observers because the consensus was if we choke off the hardware they got nothing.

211
00:15:00.547 --> 00:15:02.268
And this totally blew that out of the water.

212
00:15:02.308 --> 00:15:07.873
So at first there was fascination which was immediately followed by the national security narrative taking over.

213
00:15:08.333 --> 00:15:28.260
An interesting point about what China did with DeepSeek was it was also, if we exclude LAMA, Meta's LAMA models, which are open source, open weights, but arguably weren't as competitive as R1, China was the first frontier scale open weight model.

214
00:15:28.300 --> 00:15:28.701
Yeah, yeah.

215
00:15:28.721 --> 00:15:29.841
They just published their weights.

216
00:15:30.161 --> 00:15:30.361
Right.

217
00:15:30.581 --> 00:15:35.083
Which the point here is then you can take it and make it into whatever you want to.

218
00:15:35.583 --> 00:15:39.666
So the app for DeepSeek became the number one most downloaded app globally.

219
00:15:39.966 --> 00:15:40.627
Oh, I didn't know that.

220
00:15:40.747 --> 00:15:46.370
It was then banned in India, Italy, Australia, Taiwan, because the national security narrative came in.

221
00:15:46.390 --> 00:15:48.452
And it was like, this is a propaganda tool.

222
00:15:48.812 --> 00:15:50.653
It's a data siphoning tool.

223
00:15:51.674 --> 00:15:52.715
Don't trust DeepSeek.

224
00:15:53.115 --> 00:15:58.399
But it strategically left out the point about the fact that you could download the model.

225
00:15:59.059 --> 00:15:59.299
Yeah.

226
00:15:59.399 --> 00:16:02.623
Run it locally and change the weights however you wanted to.

227
00:16:02.723 --> 00:16:05.726
So with DeepSeek, if you tried to look up Tiananmen Square.

228
00:16:05.866 --> 00:16:06.106
Yeah.

229
00:16:06.386 --> 00:16:07.327
You weren't going to get an answer.

230
00:16:07.588 --> 00:16:08.008
Really?

231
00:16:08.228 --> 00:16:08.429
Right.

232
00:16:08.809 --> 00:16:10.411
So they were, you know.

233
00:16:10.511 --> 00:16:11.031
Classic.

234
00:16:11.051 --> 00:16:11.872
Doing a little bit on the.

235
00:16:11.892 --> 00:16:12.393
Yeah, yeah, yeah.

236
00:16:12.433 --> 00:16:12.853
Of course.

237
00:16:13.033 --> 00:16:13.894
On the information control.

238
00:16:13.914 --> 00:16:14.615
But it's open weights.

239
00:16:14.875 --> 00:16:15.115
Yeah.

240
00:16:15.135 --> 00:16:16.737
So if you want the Tiananmen Square answer.

241
00:16:17.646 --> 00:16:18.727
You can do it on your own.

242
00:16:18.887 --> 00:16:19.607
Oh, interesting.

243
00:16:19.688 --> 00:16:20.008
I see.

244
00:16:20.168 --> 00:16:20.428
Yeah.

245
00:16:20.708 --> 00:16:24.611
The narrative really tried to avoid talking about that point.

246
00:16:24.671 --> 00:16:28.454
At the same time, people started calling open AI closed AI.

247
00:16:28.674 --> 00:16:28.954
Yeah.

248
00:16:28.974 --> 00:16:32.057
Because they had never had an open source model.

249
00:16:32.077 --> 00:16:34.638
They never had a research paper, a peer review paper.

250
00:16:34.678 --> 00:16:34.899
Right.

251
00:16:34.919 --> 00:16:35.179
Yeah.

252
00:16:35.339 --> 00:16:36.460
Around their process.

253
00:16:36.980 --> 00:16:46.852
So now this brings us to where we are today with these two new papers in Nature, which go a little bit deeper into how DeepSeek went through their process.

254
00:16:46.892 --> 00:16:48.314
And we'll sort of talk about this.

255
00:16:48.715 --> 00:16:54.562
But I think that background is important context to understand, like, why does DeepSeek matter?

256
00:16:54.862 --> 00:16:55.102
Yeah.

257
00:16:55.382 --> 00:17:03.286
Why is it uniquely different than chat GPT, than Claude, than, you know, Gemini, Gemini, all these other models.

258
00:17:04.207 --> 00:17:08.449
And why is it important that they actually have a peer review paper?

259
00:17:08.989 --> 00:17:14.272
We always tell a lot of the U S conversation around science research is, Oh, China's papers.

260
00:17:14.652 --> 00:17:15.753
They fake the data.

261
00:17:15.853 --> 00:17:16.133
Yeah.

262
00:17:16.153 --> 00:17:17.273
They're not really good.

263
00:17:17.553 --> 00:17:18.214
They're not where we are.

264
00:17:18.234 --> 00:17:22.536
And I think this, this has been a, a unfortunate, uh,

265
00:17:24.390 --> 00:17:33.382
viewpoint in the U.S. about China generally, not only with cars saying they just steal our stuff, also with this, that is a little bit outdated.

266
00:17:33.742 --> 00:17:34.483
It is, definitely.

267
00:17:34.503 --> 00:17:38.147
Because we're seeing with cars, the best EV cars in the world are made by China.

268
00:17:38.448 --> 00:17:39.469
It's not arguable.

269
00:17:39.629 --> 00:17:42.092
They make them cheaper, faster, better, stronger.

270
00:17:42.192 --> 00:17:42.413
Yeah.

271
00:17:43.173 --> 00:17:47.717
which we've now put tariffs on in order to say, oh, don't look here, there's no problem, Ford is still the best.

272
00:17:48.297 --> 00:17:59.084
And AI is taking a similar, in terms of our posture, a similar political stance that China doesn't really know what they're doing, this is not really a big deal, blah, blah, blah.

273
00:17:59.245 --> 00:18:04.788
I think we need to be careful about ascribing a narrative that may have been true 15, 20 years ago

274
00:18:05.589 --> 00:18:07.570
Yeah, that's certainly not true now.

275
00:18:07.630 --> 00:18:08.530
That's certainly not true anymore.

276
00:18:08.550 --> 00:18:10.011
We've covered Chinese research.

277
00:18:10.031 --> 00:18:14.212
We've covered the hexagonal diamonds that were first created by China.

278
00:18:14.313 --> 00:18:15.473
Everyone's been going after them.

279
00:18:15.553 --> 00:18:15.753
Yes.

280
00:18:15.793 --> 00:18:20.055
But two labs in China were going after them, and actually one of them succeeded, right?

281
00:18:20.095 --> 00:18:20.195
Yes.

282
00:18:21.095 --> 00:18:22.696
And the proof is in the pudding.

283
00:18:22.896 --> 00:18:23.016
Yes.

284
00:18:23.036 --> 00:18:23.516
They made it.

285
00:18:23.976 --> 00:18:24.977
They have all the tests.

286
00:18:25.017 --> 00:18:25.217
Right.

287
00:18:25.657 --> 00:18:26.858
And there's peer-reviewed papers.

288
00:18:26.918 --> 00:18:28.239
And it's out in nature.

289
00:18:28.299 --> 00:18:28.959
It's out in nature.

290
00:18:29.159 --> 00:18:29.659
Right.

291
00:18:29.800 --> 00:18:31.821
And we always talk about, well, this and that.

292
00:18:31.901 --> 00:18:35.763
So now that we've given that context, let's talk about these two new papers.

293
00:18:36.163 --> 00:18:36.403
Yeah.

294
00:18:36.523 --> 00:18:48.370
These papers are pretty incredible in the fact that they're the first peer-reviewed paper from a sort of premier AI large language model.

295
00:18:48.551 --> 00:18:48.731
Right.

296
00:18:49.031 --> 00:18:51.592
You have, like, obviously, ChatGPT, Gemini.

297
00:18:51.772 --> 00:18:51.912
Yeah.

298
00:18:52.933 --> 00:19:04.675
Grok, all of these large language models, but they haven't actually published anything that shows sort of, OK, that that, you know, I'm sure that the team at DeepSeek has kept some secrets to themselves.

299
00:19:05.036 --> 00:19:12.837
But the fact that they've actually gone out and published this peer review paper is, I think, quite a watershed moment for the AI community.

300
00:19:12.877 --> 00:19:13.077
Right.

301
00:19:13.097 --> 00:19:15.998
Because they're trailblazing this idea that maybe.

302
00:19:16.938 --> 00:19:23.625
AI is a part of open science and maybe it should be out in the open just like normal science, right?

303
00:19:23.665 --> 00:19:27.388
Where it's like you as a scientist, you find something cool or you make something cool.

304
00:19:27.708 --> 00:19:29.750
You put it out in a peer reviewed paper, right?

305
00:19:29.830 --> 00:19:37.978
And you get let experts actually talk back to you and then like see how things actually work, right?

306
00:19:38.118 --> 00:19:39.620
You want to get into the nitty gritty of it.

307
00:19:40.520 --> 00:19:54.003
In the technical details, this paper is quite interesting because it tells us some of the tricks that they used in actually making it possible to train something like DeepSeq on perhaps only H-800s.

308
00:19:54.443 --> 00:19:56.484
Again, maybe not, right?

309
00:19:56.624 --> 00:20:00.305
There's always a possibility that they're not totally forthcoming.

310
00:20:00.485 --> 00:20:00.685
Right.

311
00:20:01.005 --> 00:20:01.205
But...

312
00:20:02.525 --> 00:20:05.487
They've sort of laid bare some of their strategies.

313
00:20:05.827 --> 00:20:07.127
And I think that's quite interesting.

314
00:20:07.788 --> 00:20:12.250
I want to actually just make two small context notes that I think are interesting here.

315
00:20:12.710 --> 00:20:27.317
One is that the founder of DeepSeek, his story is that the reason they ended up creating R1 was actually he was doing the trading, like financial trade, like finance.

316
00:20:27.777 --> 00:20:28.037
Right.

317
00:20:28.198 --> 00:20:29.679
It was related to a hedge fund.

318
00:20:29.719 --> 00:20:30.660
It was a hedge fund plan.

319
00:20:30.740 --> 00:20:35.706
And they were like, we want a more powerful model to utilize for.

320
00:20:35.786 --> 00:20:37.147
So it wasn't even their core.

321
00:20:37.828 --> 00:20:40.971
It was an outgrowth of another need.

322
00:20:41.532 --> 00:20:45.576
And then I want to talk about this open sourcing concept of why China did.

323
00:20:45.596 --> 00:20:48.219
Why are they doing this open source, open weights?

324
00:20:48.479 --> 00:20:48.699
Yeah.

325
00:20:48.739 --> 00:20:49.060
Move.

326
00:20:49.360 --> 00:20:59.091
There was an interesting there's a podcast, the BG2 podcast with two U.S. investors, Brad Gershner and Bill Gurley.

327
00:20:59.391 --> 00:21:00.813
And they've been talking about this a lot.

328
00:21:01.233 --> 00:21:09.323
If you look at China's strategy generally as it relates to if you're in a world where the U.S. already has dominance in a lot of these technologies that are closed source.

329
00:21:09.883 --> 00:21:21.947
From a how do we remain competitive angle, this is actually what Facebook did by trying to have Lama be open, is the best thing you can do to battle closed source is to create an open source ecosystem.

330
00:21:22.148 --> 00:21:22.388
Right.

331
00:21:22.468 --> 00:21:31.671
That now becomes competitive with that closed source because you're going to then now get input from a wider array of experts that invest in your version of the future.

332
00:21:32.211 --> 00:21:34.432
And so for China, there was an incentive there.

333
00:21:34.632 --> 00:21:54.173
geopolitically to go open source being a second mover as the means by which to combat the closed source dominance of the us so oh this deep seek model is not an isolated incident of china using open source as a way to backstab u.s's closed source dominance

334
00:21:54.373 --> 00:21:57.816
Yeah, there's a strategy behind this.

335
00:21:58.116 --> 00:22:11.487
That is a larger part of how China is trying to spread its influence across these areas in which the current West's control and buy-in is waning.

336
00:22:11.707 --> 00:22:11.907
Yeah.

337
00:22:12.227 --> 00:22:15.249
I mean, I think they figure that like the world is a really big place.

338
00:22:15.850 --> 00:22:21.313
OK. And these developing economies are exactly that developing.

339
00:22:21.373 --> 00:22:21.593
Right.

340
00:22:21.933 --> 00:22:22.714
They're getting bigger.

341
00:22:23.174 --> 00:22:24.955
People are getting more affluent.

342
00:22:25.355 --> 00:22:28.717
Technologies there are going to start getting better and better.

343
00:22:28.857 --> 00:22:32.740
And they're going to require AI to keep up with the rest of the world.

344
00:22:33.020 --> 00:22:33.220
Right.

345
00:22:33.380 --> 00:22:35.421
And maybe they don't want to pay money.

346
00:22:36.442 --> 00:22:41.326
whatever royalties, right, to these closed systems.

347
00:22:41.426 --> 00:22:43.067
I think it's really cool.

348
00:22:43.707 --> 00:22:53.855
And I think the idea that DeepSea came out, right, with their model in January, I think it was also a watershed moment for the rest of the AI community in terms of...

349
00:22:54.575 --> 00:22:59.238
finding out the efficacy of something like a pure reinforcement learning approach.

350
00:22:59.458 --> 00:22:59.739
Okay.

351
00:23:00.139 --> 00:23:15.690
Rather than some of the techniques that ChatGPT and these other firms had used, which was a supervised fine tuning or this sort of human feedback reinforcement learning, which always requires a human in the loop.

352
00:23:15.870 --> 00:23:16.070
Yes.

353
00:23:16.190 --> 00:23:16.390
Right.

354
00:23:16.791 --> 00:23:17.811
What these guys did was...

355
00:23:19.092 --> 00:23:21.192
kind of a pure reinforcement learning approach.

356
00:23:21.212 --> 00:23:24.133
And that's something that they've laid out in this paper in nature.

357
00:23:24.273 --> 00:23:24.513
Okay.

358
00:23:24.893 --> 00:23:34.095
The idea is that sort of traditional LLM reasoning models, what they do is they require human input in order to steer them towards the correct answer.

359
00:23:34.295 --> 00:23:34.495
Right.

360
00:23:34.795 --> 00:23:37.996
In order to be like, okay, this is, this was not quite what I wanted.

361
00:23:38.036 --> 00:23:39.216
This is, this is what I wanted.

362
00:23:39.516 --> 00:23:39.656
Right.

363
00:23:39.676 --> 00:23:44.257
There were a bunch of like Kenyans that were hired by open AI.

364
00:23:44.437 --> 00:23:44.597
Yes.

365
00:23:44.857 --> 00:23:45.938
To basically do this.

366
00:23:46.158 --> 00:23:46.478
Yes.

367
00:23:46.578 --> 00:23:49.920
And I don't know how many, but like that was a whole thing back in the day.

368
00:23:50.320 --> 00:23:52.341
And, you know, that's fine for that.

369
00:23:52.441 --> 00:23:52.901
That's fine.

370
00:23:52.921 --> 00:24:00.665
But it's quite expensive in terms of like getting human annotated chain of thought data and so on and so forth.

371
00:24:01.026 --> 00:24:02.947
But on the other hand, there's actually...

372
00:24:04.027 --> 00:24:09.695
a disadvantage to using humans, because then you are limited to human reasoning.

373
00:24:10.255 --> 00:24:10.536
Okay?

374
00:24:11.056 --> 00:24:21.730
There could be other ways to reason about a problem that wouldn't quite occur to a human being, that perhaps if we gave the entire control to the machine itself...

375
00:24:22.451 --> 00:24:24.113
the machine could figure out on its own.

376
00:24:24.493 --> 00:24:24.713
Right.

377
00:24:25.034 --> 00:24:31.240
So there's an inherent disadvantage that I think these guys at DeepSeek actually like latched on to.

378
00:24:31.761 --> 00:24:31.941
Right.

379
00:24:32.001 --> 00:24:38.507
And they figured that what we could do is we could use a pure reinforcement learning strategy where there's a reward signal.

380
00:24:38.527 --> 00:24:41.631
So the way the reinforcement learning works is you've got your model.

381
00:24:41.851 --> 00:24:42.071
Right.

382
00:24:42.451 --> 00:24:44.273
And then it tries to find an answer.

383
00:24:45.154 --> 00:24:48.397
This works for well-defined tasks where there's a right and wrong answer.

384
00:24:48.517 --> 00:24:55.403
And there's a really nice way to compare your answer to the right answer and figure out what a reward should be.

385
00:24:55.724 --> 00:25:01.869
A reward being if my answer is closely aligned to the right answer, then I'm on the right track.

386
00:25:02.270 --> 00:25:05.713
And the strategy that I've been using is sort of on the right track.

387
00:25:06.253 --> 00:25:08.654
And otherwise, maybe I need to try a different strategy.

388
00:25:08.674 --> 00:25:12.676
It's actually very close to how neural systems work, right?

389
00:25:12.776 --> 00:25:23.241
We've studied reinforcement learning, like, at the biological level when it comes to, like, you know, a rat trying to find cheese in a maze, right?

390
00:25:23.321 --> 00:25:27.603
We've literally gone into the brain and found that these... How...

391
00:25:28.283 --> 00:25:31.344
at the nitty-gritty level, at the cellular level, how this stuff works.

392
00:25:31.444 --> 00:25:38.548
And now we're kind of translating that into the AI neural network space, the artificial neural network space.

393
00:25:39.528 --> 00:25:45.551
So if we use pure reinforcement learning, what that does is it bypasses human-labeled reasoning steps.

394
00:25:45.951 --> 00:25:47.092
And then we can actually...

395
00:25:48.553 --> 00:25:56.483
And explore just computationally the entire paradigm of how to solve the problem, which is, I think, very cool.

396
00:25:57.325 --> 00:26:02.432
So that's one of the things that's revealed by this paper, the fact that they used pure reinforcement learning.

397
00:26:04.253 --> 00:26:08.676
It was this autonomous exploration and a development of problem solving strategies.

398
00:26:09.097 --> 00:26:20.664
And what you could see and what they report in the paper is you could see the network sort of use that RL technique, reinforcement learning, to get better at finding the right answer.

399
00:26:20.704 --> 00:26:22.065
So what they noticed is...

400
00:26:23.345 --> 00:26:37.474
For example, the length of the answers would get progressively longer because the neural network wanted to get better and better at finding that right answer in this large language model type of thing.

401
00:26:37.735 --> 00:26:48.522
They also had this thing called the aha moment, which is when I was like, oh, this is like working way better than we thought, which is when, you know, in the chain of thought, you can actually output the thought that this thing is doing.

402
00:26:48.882 --> 00:26:49.062
Right.

403
00:26:49.762 --> 00:26:54.405
And the model would start using the word wait.

404
00:26:57.107 --> 00:27:01.430
That was never explicitly put into the training regime.

405
00:27:01.710 --> 00:27:12.957
But what the model would start doing is as it was thinking, if it was trying to get to an answer and it was like sensing that it was rushing things and just trying to get it done quickly, it would use this thing wait.

406
00:27:14.157 --> 00:27:21.001
And then it would backtrack and try to figure out if there's a better way to rethink what it had just thought and then keep going.

407
00:27:21.521 --> 00:27:27.865
That's like some meta level weird stuff, dude.

408
00:27:27.925 --> 00:27:36.950
Like that it's just using this word wait in the context that the word wait means wait and let's try to figure out what we just said in a nice way.

409
00:27:37.530 --> 00:27:40.932
You know, I think that's kind of crazy to backtrack.

410
00:27:41.312 --> 00:27:41.653
And then...

411
00:27:42.874 --> 00:27:49.604
So it developed these autonomous, sophisticated reasoning patterns to try to, you know, solve the problem.

412
00:27:50.045 --> 00:27:54.451
And it's no longer limited by the human guided approach.

413
00:27:54.712 --> 00:27:54.932
Right.

414
00:27:55.192 --> 00:27:55.413
Right.

415
00:27:55.633 --> 00:27:57.135
Which also has bias issues.

416
00:27:57.856 --> 00:27:58.277
Yeah, exactly.

417
00:27:58.457 --> 00:28:00.200
Related to human reinforcement learning.

418
00:28:00.220 --> 00:28:04.987
Because now, if you have a bunch of Kenyans doing it, there's cultural differences.

419
00:28:05.368 --> 00:28:05.929
Exactly.

420
00:28:06.029 --> 00:28:08.072
There's educational differences.

421
00:28:08.733 --> 00:28:12.318
And so this is where people sort of have concerns about...

422
00:28:12.859 --> 00:28:21.128
the Silicon Valley elite being the ones controlling the parameters and mind space that these models sort of have.

423
00:28:22.390 --> 00:28:28.136
But this is like a very pure play version that kind of abstracts away the need for...

424
00:28:29.978 --> 00:28:34.900
at least in some cases, for the human input to be such a big driver in how it determines answers.

425
00:28:35.040 --> 00:28:35.420
That's right.

426
00:28:35.440 --> 00:28:35.940
Yeah.

427
00:28:36.461 --> 00:28:38.502
So that was the reinforcement learning part, right?

428
00:28:38.562 --> 00:28:40.863
Which is like, okay, that's a training strategy.

429
00:28:41.323 --> 00:28:45.405
They also had this philosophy sort of efficiency over everything, right?

430
00:28:45.525 --> 00:28:53.729
Because of what you talked about where, you know, the powers that be, let's say, have crippled their access to hardware.

431
00:28:54.289 --> 00:28:55.229
So what can we do?

432
00:28:55.289 --> 00:28:56.470
We can innovate hardware,

433
00:28:56.730 --> 00:28:59.652
with algorithms and we can innovate with strategy.

434
00:29:00.132 --> 00:29:05.356
So there's a two-pronged approach that they at least revealed in this nature paper.

435
00:29:05.676 --> 00:29:08.318
The first one is this idea of mixture of experts, right?

436
00:29:08.578 --> 00:29:11.960
This architecture where what you can do is have a giant model.

437
00:29:12.120 --> 00:29:21.387
The DeepSync model itself is 671 billion total parameters, which is on par with some of the big ChatGPTs and things like that.

438
00:29:21.447 --> 00:29:24.629
I think ChatGPT5 has like on the order of trillion or more.

439
00:29:25.090 --> 00:29:27.779
But some of the earlier models have hundreds of billions.

440
00:29:27.799 --> 00:29:28.842
So it's in the same regime.

441
00:29:29.304 --> 00:29:29.484
But...

442
00:29:30.492 --> 00:29:36.176
when it comes to actually inference, the number of active parameters used is only 37 billion.

443
00:29:36.817 --> 00:29:40.660
Okay, which is 20 times less than the total size of this network.

444
00:29:41.040 --> 00:29:48.606
Because what they're doing is, they've got this sort of network on top that comes in in the very beginning, and it's kind of a router.

445
00:29:49.066 --> 00:29:57.212
Okay, what it does is, you've got the 600 billion parameter network, but if you ask it a certain question,

446
00:29:58.013 --> 00:30:02.940
There's only a certain subset of that network that would be very good at answering that question.

447
00:30:03.020 --> 00:30:05.904
And that becomes that 37 billion part.

448
00:30:06.404 --> 00:30:14.635
So they've implemented this router mechanism that takes in the prompt of whatever you put in, and then it sort of routes...

449
00:30:15.516 --> 00:30:35.437
that prompt to a subset of the network and now inference becomes faster it becomes less expensive in terms of maybe you don't need an h100 gpu right maybe you can do with that h 800 or you know that that previous model and you can actually like get a quick answer yes and it'll be just as good right

450
00:30:36.017 --> 00:30:36.177
Right.

451
00:30:36.397 --> 00:30:38.038
So that's one of the strategies that they use.

452
00:30:38.398 --> 00:30:41.159
The other strategy that they used was during training.

453
00:30:41.539 --> 00:30:47.180
They had this idea of using a reward system.

454
00:30:47.220 --> 00:30:47.760
Right.

455
00:30:47.921 --> 00:30:59.844
In reinforcement learning, you have to have a way of sort of quantifying how much the reward is and then steering the network towards getting more of that kind of reward.

456
00:31:00.164 --> 00:31:00.364
Right.

457
00:31:00.805 --> 00:31:08.951
So the way that traditional LLMs have been trained is through this process called proximal policy optimization.

458
00:31:09.191 --> 00:31:09.451
Okay.

459
00:31:09.851 --> 00:31:10.372
PPO.

460
00:31:11.292 --> 00:31:15.415
It's this idea that you have your policy maker, which is your LLM.

461
00:31:15.836 --> 00:31:21.080
That's like actually outputting the language that gets output whenever you do the chat.

462
00:31:21.580 --> 00:31:24.722
But then there's this thing on top, which is kind of like the critic model.

463
00:31:25.363 --> 00:31:28.966
And what that critic model is doing is it's during the training process.

464
00:31:29.946 --> 00:31:40.052
It is taking the output of the policy, which is the output of the model, and then it's trying to guess at what the reward is going to be.

465
00:31:40.572 --> 00:31:40.812
Okay?

466
00:31:41.112 --> 00:31:50.778
And then what you can do is you can use this critic model and the advantage that you get, the advantage being what is my guest advantage at this particular strategy?

467
00:31:51.338 --> 00:31:52.859
That's what this critic model calculates.

468
00:31:53.259 --> 00:32:00.062
And what this critic model does is feedback that advantage to the big LLM to tell it, okay, you're on the right track.

469
00:32:00.302 --> 00:32:02.844
This strategy is working or no, it's not working.

470
00:32:03.484 --> 00:32:08.146
Crucially, this is very expensive because now you've got two models that you're training, right?

471
00:32:08.486 --> 00:32:09.407
You've got two models.

472
00:32:09.467 --> 00:32:15.470
And as the LLM gets bigger and bigger, the critic model has to get bigger and bigger to try to figure out.

473
00:32:15.950 --> 00:32:24.314
how to calculate that advantage in whatever scenario, whatever game you're playing, whether it's text generation, whether it's whatever, right?

474
00:32:24.954 --> 00:32:28.835
So if you've got a bunch of H-100s, that's fine, right?

475
00:32:28.875 --> 00:32:33.217
If you've got billions of dollars of funding from random VC firms, that's fine.

476
00:32:33.637 --> 00:32:34.478
But if you're DeepSeek,

477
00:32:35.258 --> 00:32:36.219
That's no longer tenable.

478
00:32:36.619 --> 00:32:41.764
So what these guys did was they used a different strategy called group relative policy optimization.

479
00:32:41.964 --> 00:32:42.184
Okay.

480
00:32:42.264 --> 00:32:42.524
Okay.

481
00:32:42.945 --> 00:32:46.908
They don't have this second critic model that they're doubly training.

482
00:32:47.188 --> 00:32:47.528
Right.

483
00:32:47.669 --> 00:32:50.511
They don't have this 2x parameter blow up.

484
00:32:51.111 --> 00:32:53.954
What they have is their original LLM.

485
00:32:54.694 --> 00:32:58.298
And what it's going to do is it's going to output a bunch of different answers.

486
00:32:59.018 --> 00:33:02.724
Okay, let's say for the sake of argument, let's say 10 different answers.

487
00:33:03.265 --> 00:33:08.753
And then what it's going to do is it's going to calculate the reward for each which we can get because this is a reinforcement learning strategy.

488
00:33:09.674 --> 00:33:09.854
And...

489
00:33:11.378 --> 00:33:13.399
There's going to be a mean and standard deviation.

490
00:33:13.419 --> 00:33:18.420
There's going to be a distribution that you can calculate about that 10 different strategies.

491
00:33:19.041 --> 00:33:24.222
There's going to be some that are above the mean, which means that that is better than those others.

492
00:33:24.862 --> 00:33:28.484
The whole idea is you're trying to get this idea of advantage.

493
00:33:28.664 --> 00:33:32.725
Which are the strategies that give me advantage and which are the strategies that don't?

494
00:33:33.685 --> 00:33:38.027
In the ones before, in the traditional LLMs, the ones that we're doing in the States and in the West,

495
00:33:38.907 --> 00:33:44.349
We've got this separate neural network that is calculating that advantage and telling the original one where to go.

496
00:33:44.690 --> 00:33:54.054
Here, the original one is coming up with different answers and then creating a distribution out of that and then figuring out, okay, I should go here to this part.

497
00:33:55.010 --> 00:33:56.752
Whatever I did here, that's what's good.

498
00:33:56.972 --> 00:33:58.554
Whatever I did here, that's what's bad.

499
00:33:58.614 --> 00:34:00.877
Because I came up with eight, ten different answers.

500
00:34:01.177 --> 00:34:01.938
These guys were bad.

501
00:34:01.958 --> 00:34:02.699
These guys were good.

502
00:34:02.739 --> 00:34:02.899
Yes.

503
00:34:03.040 --> 00:34:03.220
Right?

504
00:34:03.260 --> 00:34:07.165
So you don't need this second giant neural network to train.

505
00:34:07.986 --> 00:34:08.166
Right?

506
00:34:08.406 --> 00:34:11.129
So it's a really interesting...

507
00:34:11.870 --> 00:34:15.311
strategy where you're choked by the hardware.

508
00:34:15.431 --> 00:34:15.611
Yes.

509
00:34:15.671 --> 00:34:19.071
But you're saying, okay, well, I've got a few tricks up my sleeve.

510
00:34:19.131 --> 00:34:19.331
Yes.

511
00:34:19.531 --> 00:34:19.771
Right?

512
00:34:19.891 --> 00:34:20.772
Yes.

513
00:34:20.812 --> 00:34:21.912
I think it's rather cool.

514
00:34:22.372 --> 00:34:40.595
I think that this is an important, you know, there is always, because of this nexus between nation states trying to accrue and maintain power and the importance of frontier technologies to that goal, we're always going to see policy

515
00:34:41.475 --> 00:34:45.856
impact how these technologies grow and progress.

516
00:34:46.537 --> 00:35:00.381
The Biden administration made a bet that hardware strangulation to China was the best path at creating advantage for the U.S. in the race.

517
00:35:01.061 --> 00:35:02.822
And it was a well-informed bet, right?

518
00:35:02.862 --> 00:35:05.143
Because like everyone who had been doing

519
00:35:06.223 --> 00:35:09.284
AI thought that that was the masterpiece, right?

520
00:35:09.344 --> 00:35:11.505
The hardware was it, right?

521
00:35:11.565 --> 00:35:14.526
Ever since 2012 when AlexNet came out.

522
00:35:14.887 --> 00:35:18.108
And what they could do is actually use GPUs to train in parallel.

523
00:35:18.488 --> 00:35:25.391
And then just all of the mathematics of matrix multiplication and backpropagation could be done on GPUs.

524
00:35:25.711 --> 00:35:30.153
That's when it was, you know, ever since then we've sort of taken for granted that, oh, GPUs are really...

525
00:35:30.973 --> 00:35:31.473
What matter?

526
00:35:31.813 --> 00:35:33.814
But we've created good enough GPUs.

527
00:35:34.034 --> 00:35:34.234
Yes.

528
00:35:34.915 --> 00:35:38.516
That we've taken for granted that like the best GPU is always the best.

529
00:35:38.836 --> 00:35:41.778
But actually the H800 is like a good enough GPU now.

530
00:35:41.898 --> 00:35:42.098
Yes.

531
00:35:42.678 --> 00:35:46.019
That it can do enough of the math.

532
00:35:46.239 --> 00:35:46.460
Yes.

533
00:35:47.020 --> 00:35:49.941
That now these guys can actually go and...

534
00:35:51.546 --> 00:35:54.170
You know, start tweaking the knobs.

535
00:35:54.190 --> 00:35:55.251
Yeah, start tweaking the knobs.

536
00:35:55.992 --> 00:35:59.978
It's a really, you know, the cost difference is crazy.

537
00:36:00.458 --> 00:36:05.305
The equivalence of efficacy of the models, given the cost difference, is even crazier.

538
00:36:06.006 --> 00:36:07.808
And it is going to continue to be...

539
00:36:08.389 --> 00:36:13.073
This hardware-software tension, where do you innovate more?

540
00:36:14.294 --> 00:36:22.360
Where has the most aggregate value in terms of continuing to increase the power and efficacy of these models?

541
00:36:22.701 --> 00:36:24.542
It's going to continue to be a rolling story.

542
00:36:25.102 --> 00:36:30.387
I think, again, to your point at the top of when we started this story, it is very interesting.

543
00:36:31.377 --> 00:36:41.890
That for the amount of time, I remember using GPT-2 in my company trying to see, you know, trying to gauge when is this stuff going to be ready for production years ago.

544
00:36:42.952 --> 00:36:49.600
And it is, the jury is not out yet on...

545
00:36:51.367 --> 00:37:00.450
on this journey to AGI and artificial super intelligence, which of these two levers that we're pulling are going to create those incremental?

546
00:37:00.650 --> 00:37:07.453
Because we've sort of seen, the models are kind of starting to aggregate at a similar level of performance.

547
00:37:07.473 --> 00:37:07.713
Yeah.

548
00:37:08.333 --> 00:37:14.215
And the gains from your walled garden are becoming less and less.

549
00:37:14.635 --> 00:37:19.197
And this is a classic battle in every technology revolution that's ever existed

550
00:37:20.585 --> 00:37:22.797
in terms of the tension between closed source

551
00:37:23.811 --> 00:37:24.931
and open source.

552
00:37:25.351 --> 00:37:28.832
Open source has always been, all servers are on Linux.

553
00:37:28.993 --> 00:37:30.193
Linux is open source.

554
00:37:30.773 --> 00:37:31.553
Why is that?

555
00:37:32.013 --> 00:37:33.074
Everyone uses Python.

556
00:37:33.114 --> 00:37:34.354
Everyone uses Python, open source.

557
00:37:34.374 --> 00:37:40.836
And the reason, the theory of the case is when you open source, there's a more of a distributed incentive to create security.

558
00:37:40.876 --> 00:37:41.336
Exactly.

559
00:37:41.456 --> 00:37:43.117
Because everyone wants it to work.

560
00:37:43.177 --> 00:37:47.118
You're having more viewpoints inputting into the system itself.

561
00:37:47.558 --> 00:37:51.179
And you're limited by your resources if you go closed source by your talent.

562
00:37:51.739 --> 00:37:57.606
The way in which in the U.S. we deal with that is Facebook gives billion-dollar offers to the best AI researchers.

563
00:37:58.107 --> 00:38:02.192
But at some point, this is a forever tension in technology revolutions.

564
00:38:02.732 --> 00:38:07.618
And it's interesting to see it play out in this geopolitical sphere where the –

565
00:38:09.826 --> 00:38:15.013
the winner or the import of the battle is so high.

566
00:38:15.234 --> 00:38:17.457
Yeah, yeah, the stakes are really high here.

567
00:38:17.477 --> 00:38:18.098
It's really high.

568
00:38:18.218 --> 00:38:19.700
Yeah, yeah, yeah.

569
00:38:19.760 --> 00:38:21.082
And so great first story.

570
00:38:21.122 --> 00:38:26.169
I love the AI story is always great because a lot of times people just talk about the end product.

571
00:38:26.890 --> 00:38:28.852
We haven't had a lot of research papers to talk about.

572
00:38:28.872 --> 00:38:35.837
No, but it's kind of cool that we got a sort of look under the box to see how it all worked.

573
00:38:35.917 --> 00:38:40.000
And they were also forthcoming about some of the challenges that they had.

574
00:38:41.081 --> 00:38:50.348
When they first started out with DeepSeek R1-0, which is the one that they didn't release, they were having this trouble where the output would be half Chinese, half English.

575
00:38:51.009 --> 00:38:51.209
Right?

576
00:38:51.469 --> 00:38:54.031
So then what they did was they had cold start data.

577
00:38:54.291 --> 00:38:57.953
So it wasn't completely all blind reinforcement learning.

578
00:38:57.973 --> 00:39:05.818
What they did was they primed the model with human data, okay, to actually sort of give it a running start.

579
00:39:05.978 --> 00:39:06.138
Yes.

580
00:39:06.218 --> 00:39:10.080
And not have it, like, from the very beginning start doing random stuff.

581
00:39:10.360 --> 00:39:14.383
So what you do is you can have a small subset of human-labeled training data.

582
00:39:14.463 --> 00:39:14.663
Yes.

583
00:39:14.783 --> 00:39:16.144
That, like, creates a sort of...

584
00:39:17.144 --> 00:39:34.753
neighborhood of parameters okay where where okay like the the model sort of settles in on this neighborhood of where the numbers should be for all of these weights and then you can fine tune it and fine tuning is a coarse-grained word because obviously like a lot of training was done afterwards but like

585
00:39:35.573 --> 00:39:40.837
After that initial sort of like, okay, don't do like the stupid stuff of like half Chinese, half English.

586
00:39:41.217 --> 00:39:45.060
But then now get into the, you know, how do you problem solve?

587
00:39:45.541 --> 00:39:48.923
How do you actually make a really nice LLM?

588
00:39:49.043 --> 00:39:53.166
So they were forthcoming about some of their challenges, which I think is kind of interesting.

589
00:39:53.366 --> 00:40:01.753
And again, it goes back to the sort of the narrative violation that the current CCP has put ourselves in where...

590
00:40:02.513 --> 00:40:11.637
in the US we're still viewing their operating system as what it was 20 years ago, but they are being nimble and understanding the changing environment and trying to reposition.

591
00:40:11.677 --> 00:40:11.817
Yeah.

592
00:40:12.498 --> 00:40:15.439
Again, you can argue that it's contrived or not.

593
00:40:16.019 --> 00:40:16.259
Yeah.

594
00:40:16.359 --> 00:40:16.580
I mean.

595
00:40:18.281 --> 00:40:22.184
Knowing China, of course, there's stuff that we don't know about, right?

596
00:40:22.264 --> 00:40:26.228
Of course, probably DeepSeek is entrenched with the CCP.

597
00:40:26.308 --> 00:40:27.189
I'm just saying.

598
00:40:28.330 --> 00:40:29.250
I'm just saying.

599
00:40:29.310 --> 00:40:30.411
It's probably true.

600
00:40:30.852 --> 00:40:32.753
But they're doing the work.

601
00:40:32.794 --> 00:40:34.755
They're publishing peer-reviewed papers.

602
00:40:34.815 --> 00:40:43.063
They're coming out with cutting-edge AI, and they're doing it without the H-100s, or at least not with the kind of access...

603
00:40:43.583 --> 00:40:45.805
that companies have in America.

604
00:40:45.925 --> 00:40:50.129
Perhaps they do have H-100s from shell companies in Bhutan or whatever.

605
00:40:50.169 --> 00:40:55.214
Which is one of the arguments is that they've been doing this illicit market trading in order to get access.

606
00:40:55.434 --> 00:40:56.775
But the point is they don't even need it.

607
00:40:57.075 --> 00:40:57.576
Yeah, yeah.

608
00:40:57.596 --> 00:40:58.297
It turns out like...

609
00:40:59.057 --> 00:41:03.358
You know, people are, now that this is out, they're going to try using these strategies now.

610
00:41:03.638 --> 00:41:06.399
And we're going to see if the rubber meets the road.

611
00:41:07.419 --> 00:41:14.361
And, you know, since DeepSeek came out, Meta was already releasing open weight models.

612
00:41:14.822 --> 00:41:19.843
I think they've actually incorporated some of the mixture of experts architecture into the llama models.

613
00:41:19.903 --> 00:41:29.430
I think people are taking this sort of reinforcement learning strategy very seriously, too, where they're saying, OK, maybe I don't need a bunch of Kenyans to just keep annotating my outputs.

614
00:41:29.490 --> 00:41:29.710
Right.

615
00:41:29.770 --> 00:41:30.030
Correct.

616
00:41:30.070 --> 00:41:30.291
Yeah.

617
00:41:30.551 --> 00:41:35.995
And we now just in the last month and a half have an open source, open AI model.

618
00:41:36.555 --> 00:41:42.539
And it is very easy to argue that we would not have an open source.

619
00:41:43.720 --> 00:41:47.402
Weights open AI model without Deep Seagull.

620
00:41:47.422 --> 00:41:47.662
Yeah.

621
00:41:48.043 --> 00:41:48.143
Yeah.

622
00:41:48.163 --> 00:41:49.403
There was not an incentive for.

623
00:41:49.423 --> 00:41:50.184
No, why would they do it?

624
00:41:50.224 --> 00:41:50.904
Open AI to do so.

625
00:41:50.924 --> 00:41:51.144
Yeah.

626
00:41:51.505 --> 00:41:54.606
But they had to answer investors and the competition, et cetera.

627
00:41:55.027 --> 00:41:56.828
Great, great update here.

628
00:41:56.848 --> 00:41:58.549
We'll definitely be touching on AI more.

629
00:41:59.009 --> 00:42:02.791
We're going to go to our next story, which I'm very excited about because I don't know what any of it means.

630
00:42:02.991 --> 00:42:03.711
Yeah.

631
00:42:03.851 --> 00:42:10.495
The headline on our second story, scientists just made the first time crystal that you can see.

632
00:42:11.015 --> 00:42:12.856
We have an ocarina of time.

633
00:42:12.996 --> 00:42:13.296
Yes.

634
00:42:13.357 --> 00:42:24.222
Physicists at the University of Colorado Boulder have created the first time crystal that humans can actually see using liquid crystals that swirl into a never ending pattern when illuminated by light.

635
00:42:24.582 --> 00:42:27.364
They put out their paper in nature materials.

636
00:42:27.724 --> 00:42:29.285
There's been a lot of coverage of this one.

637
00:42:29.745 --> 00:42:33.728
I always see all these TikToks about time crystals and all this.

638
00:42:33.908 --> 00:42:36.229
Yeah, it's very Marvel, very Doctor Who.

639
00:42:36.269 --> 00:42:36.870
Yeah, yeah, yeah.

640
00:42:36.910 --> 00:42:39.411
You know, like, this is how we beat Thanos.

641
00:42:39.532 --> 00:42:39.832
Mr.

642
00:42:39.872 --> 00:42:41.012
Strange, Doctor Strange.

643
00:42:41.032 --> 00:42:42.954
Yeah, yeah, all sorts of, yeah.

644
00:42:43.014 --> 00:42:46.856
So help me understand, you know, what exactly is going on with this time crystals paper?

645
00:42:46.936 --> 00:42:51.419
Yeah, so, well, first, let's talk about what are time crystals.

646
00:42:51.499 --> 00:42:52.260
I would like to know that.

647
00:42:52.300 --> 00:42:52.540
Okay.

648
00:42:52.960 --> 00:42:56.201
So before we even do that, normal crystals.

649
00:42:56.622 --> 00:42:56.842
Okay.

650
00:42:57.382 --> 00:42:58.883
We talked about diamond the other day.

651
00:42:59.043 --> 00:42:59.963
Yes.

652
00:43:00.463 --> 00:43:02.724
The simplest crystal I can think of is table salt.

653
00:43:03.064 --> 00:43:03.384
Okay.

654
00:43:03.404 --> 00:43:05.605
N-A-C-L, sodium chloride.

655
00:43:05.645 --> 00:43:08.387
You can imagine there's a sodium atom and a chlorine atom.

656
00:43:08.467 --> 00:43:09.647
Sodium atom, chlorine atom.

657
00:43:09.707 --> 00:43:10.868
Sodium atom, chlorine atom.

658
00:43:11.168 --> 00:43:16.450
In this regular sort of chessboard-like structure, but a 3D sort of chessboard.

659
00:43:16.470 --> 00:43:16.690
Mm-hmm.

660
00:43:17.030 --> 00:43:18.912
of sodium chlorine, sodium chlorine.

661
00:43:19.353 --> 00:43:21.355
The reason why this is a crystal, right?

662
00:43:21.675 --> 00:43:28.282
What characterizes it as a crystal is you've got a discrete space symmetry breaking, okay?

663
00:43:28.542 --> 00:43:35.549
This is a technical term in physics that just means if I take the crystal and I move it over by one...

664
00:43:36.502 --> 00:43:39.084
spacing, I will get the same thing back.

665
00:43:39.304 --> 00:43:39.524
Okay.

666
00:43:39.745 --> 00:43:39.985
Okay?

667
00:43:40.205 --> 00:43:44.808
You can imagine I've got the NaCl and then there's another NaCl here, right?

668
00:43:45.069 --> 00:43:52.534
I can take that entire crystal and move it by a whole molecule and I'll have the same thing repeating, right?

669
00:43:52.754 --> 00:43:53.975
Just like in a chessboard, right?

670
00:43:54.075 --> 00:43:55.536
Imagine an infinite chessboard

671
00:43:56.417 --> 00:43:58.397
that thing is a crystal in some sense, right?

672
00:43:58.417 --> 00:44:01.338
It's a 2D crystal because there's black, white, black, white.

673
00:44:01.578 --> 00:44:08.919
If I take up the chessboard and I move it over two sort of squares, I'll get the same exact chessboard, right?

674
00:44:09.339 --> 00:44:13.680
So that's what's called symmetry breaking in space, okay?

675
00:44:13.920 --> 00:44:18.241
Where it's no longer continuous because you can't just arbitrarily move this thing around.

676
00:44:18.261 --> 00:44:20.942
You have to move it by the lattice spacing, right?

677
00:44:20.962 --> 00:44:22.582
The lattice constant of the crystal.

678
00:44:22.662 --> 00:44:22.862
Yes.

679
00:44:23.540 --> 00:44:25.201
So that's what makes it a crystal.

680
00:44:25.221 --> 00:44:25.762
Okay.

681
00:44:25.782 --> 00:44:29.064
And it's also resilient to forces and stress, right?

682
00:44:29.084 --> 00:44:31.586
Like a diamond, for example, great way.

683
00:44:32.167 --> 00:44:33.147
I put stress on it.

684
00:44:33.568 --> 00:44:43.155
The crystalline structure of the carbon atoms isn't really going to change because the way that these carbon atoms are bonded to each other, they're resilient to like push and pull.

685
00:44:43.475 --> 00:44:44.776
You know, I heat it up.

686
00:44:45.036 --> 00:44:45.236
That...

687
00:44:46.337 --> 00:44:47.978
that structure is going to be resilient.

688
00:44:48.338 --> 00:44:48.498
Right?

689
00:44:48.858 --> 00:44:52.079
So there was a guy, Frank Wilczek, he was a Nobel laureate.

690
00:44:52.119 --> 00:45:02.684
He won the Nobel Prize in 2004 for discovering something called asymptotic freedom, which has to do with the ways that quarks behave inside the atomic nucleus.

691
00:45:03.404 --> 00:45:04.945
We can get into that some other day.

692
00:45:05.085 --> 00:45:28.434
I think we need some asymptotic freedom in the US right now, but that's a whole nother story That's a whole nother story as well for another day But he won the Nobel Prize in 2004 and he's been very active in physics ever since he's discovered things like called axions and like he's done stuff with cosmology And one of the things that he proposed back in 2012 was this idea of a time crystal Okay

693
00:45:28.794 --> 00:45:32.836
So he figured it's a very interesting, quite simple argument.

694
00:45:32.896 --> 00:45:33.156
Okay.

695
00:45:33.536 --> 00:45:39.298
He said that the laws of physics are symmetric to space translations.

696
00:45:39.618 --> 00:45:39.798
Right.

697
00:45:40.098 --> 00:45:49.402
Meaning if I do a, if I do an experiment here and then I, you know, take up the experiment and I move it somewhere else, the results of the experiment should be the same.

698
00:45:49.562 --> 00:45:49.762
Right.

699
00:45:49.862 --> 00:45:53.844
If, if all the environments are the same and that's because there's no special coordinate

700
00:45:54.404 --> 00:46:18.041
in the universe right there's no like like there's no like oh it's just like here is where where things happen right there's no um there's no special spatial coordinate well similarly the laws of the universe are also symmetric in time right if i do an experiment now and then i do it the next day should be the same yeah so if we have things that break spatial symmetry

701
00:46:18.721 --> 00:46:27.709
By creating crystals in space, right, where you have clearly like carbon making the diamond or NACL making table salt.

702
00:46:27.969 --> 00:46:30.011
So you clearly have spatial crystals.

703
00:46:30.471 --> 00:46:34.995
What if I have a breaking of the time symmetry to create a time crystal?

704
00:46:35.336 --> 00:46:36.217
And what would that look like?

705
00:46:36.257 --> 00:46:42.442
What that would look like is you've got some kind of periodic system that returns to its space.

706
00:46:42.662 --> 00:46:45.563
original form periodically.

707
00:46:45.823 --> 00:46:54.286
So just like how with a space crystal, like the chessboard, I can take it up and I can move it in space and then put it back and it'll be the same thing in time.

708
00:46:54.806 --> 00:46:56.387
For example, a sine wave, right?

709
00:46:56.467 --> 00:47:02.409
Mathematically, a sine wave is symmetric in time to that period, right?

710
00:47:02.489 --> 00:47:07.811
I can pick up the sine wave, I can move it one entire period and put it back and it'll be the same sine wave.

711
00:47:08.651 --> 00:47:10.152
I should be able to have that

712
00:47:11.863 --> 00:47:13.584
in like the physical world.

713
00:47:13.784 --> 00:47:14.045
Okay.

714
00:47:14.065 --> 00:47:14.265
Okay.

715
00:47:14.565 --> 00:47:17.407
Because, you know, it's kind of same, same.

716
00:47:17.447 --> 00:47:17.647
Yeah.

717
00:47:17.687 --> 00:47:17.887
Right.

718
00:47:17.987 --> 00:47:18.768
Theoretically.

719
00:47:18.828 --> 00:47:20.909
Theoretically, it should be possible.

720
00:47:20.929 --> 00:47:21.870
And it kind of makes sense.

721
00:47:22.390 --> 00:47:26.653
If I have the same sort of symmetry breaking in space, I should have the same sort of symmetry breaking in time.

722
00:47:27.174 --> 00:47:28.555
Well, it wasn't quite that simple.

723
00:47:28.975 --> 00:47:29.695
Okay.

724
00:47:29.735 --> 00:47:35.317
So right after he published that, it was, you know, Frank Wilczek is doing this Nobel Prize winner.

725
00:47:35.357 --> 00:47:36.478
He's a very famous physicist.

726
00:47:36.498 --> 00:47:38.119
So everyone's like, okay, is this possible?

727
00:47:38.619 --> 00:47:42.580
And actually it turns out the way that he had envisioned it, it is not.

728
00:47:43.261 --> 00:47:49.123
So people came out and they published something called the no-go theorem for time crystals, which was that.

729
00:47:50.143 --> 00:47:57.648
Frank Wilczek's original idea was that this time crystal is going to be in the ground state, which means it's in its lowest energy state.

730
00:47:58.208 --> 00:47:59.489
There's nothing pumping on it.

731
00:47:59.889 --> 00:48:02.851
It's just, you know, crystals can be in their lower energy state, right?

732
00:48:02.951 --> 00:48:08.054
I can cool a crystal down to like near zero Kelvin and it's going to remain in its crystal form.

733
00:48:08.834 --> 00:48:14.098
And that's what really like that means that it's an inherent property of the material itself, right?

734
00:48:14.338 --> 00:48:17.239
And it's not something that coming from outside creating this order.

735
00:48:17.340 --> 00:48:17.500
Mm-hmm.

736
00:48:18.260 --> 00:48:24.184
Frank Wilczek's idea was that you can have a system that is in its ground state, right?

737
00:48:24.244 --> 00:48:25.524
In its lowest energy state.

738
00:48:25.544 --> 00:48:26.505
There's nothing coming in.

739
00:48:26.825 --> 00:48:29.947
And it's still going to behave in this sort of periodic manner.

740
00:48:29.967 --> 00:48:30.147
Yes.

741
00:48:30.307 --> 00:48:30.548
Okay?

742
00:48:30.748 --> 00:48:32.409
That turns out can't be the case.

743
00:48:32.549 --> 00:48:32.809
Okay.

744
00:48:32.869 --> 00:48:33.109
Okay?

745
00:48:33.349 --> 00:48:39.353
So the no-go theorem said that a true equilibrium ground state cannot have persistent periodic motion.

746
00:48:39.713 --> 00:48:44.176
Meaning, like, if you've cooled something down to its ground state, it's going to be stationary.

747
00:48:44.356 --> 00:48:44.496
Right.

748
00:48:44.616 --> 00:48:45.957
And that kind of makes sense, right?

749
00:48:46.217 --> 00:48:50.940
Because if it's moving around in its ground state, that means there's some kinetic energy that I can extract.

750
00:48:51.240 --> 00:48:54.322
And it's not really the true ground state, right?

751
00:48:54.782 --> 00:48:56.123
Okay, so that's fine.

752
00:48:56.163 --> 00:48:57.924
But then people started wondering, all right, fine.

753
00:48:58.284 --> 00:49:01.386
I don't need something to be in equilibrium at its ground state.

754
00:49:01.806 --> 00:49:03.047
What if I pump it with stuff?

755
00:49:04.568 --> 00:49:05.930
Can I create a crystal that way?

756
00:49:05.970 --> 00:49:07.172
What if I pump it with energy?

757
00:49:07.532 --> 00:49:09.195
So it's now a non-equilibrium state.

758
00:49:09.235 --> 00:49:12.940
Like there's energy coming in and then there's energy going out.

759
00:49:13.340 --> 00:49:19.729
But the way that this energy is transformed inside the system creates this periodic motion within the system.

760
00:49:19.830 --> 00:49:19.950
Right?

761
00:49:19.970 --> 00:49:20.150
Right.

762
00:49:20.971 --> 00:49:22.172
Turns out that was possible.

763
00:49:22.552 --> 00:49:26.795
And for the longest time, people were studying these things called discrete time crystals.

764
00:49:28.076 --> 00:49:31.498
They were systems that were kept out of equilibrium by this outside force.

765
00:49:32.019 --> 00:49:37.743
And it's not as simple as just like you just pump it rhythmically and then this thing goes by that.

766
00:49:37.763 --> 00:49:43.426
Then it's like, oh, like me on a swing, like doing my legs is a time crystal.

767
00:49:43.446 --> 00:49:45.248
Like me at the playground is a time.

768
00:49:45.308 --> 00:49:45.468
No.

769
00:49:45.828 --> 00:49:47.849
No, that's not it's not quite as simple.

770
00:49:47.869 --> 00:49:51.831
What you need is you need you need the pumping to be done at a certain frequency.

771
00:49:51.911 --> 00:49:54.952
And then the crystal itself is doing it at a different frequency.

772
00:49:55.272 --> 00:49:56.373
Maybe it's at a harmonic.

773
00:49:56.753 --> 00:49:56.953
Right.

774
00:49:57.873 --> 00:50:01.295
And there were specialized quantum systems that could do this.

775
00:50:01.755 --> 00:50:01.935
Right.

776
00:50:02.355 --> 00:50:05.737
There were nuclear spins, trapped ions, cold, cold atoms.

777
00:50:06.617 --> 00:50:09.439
at extremely low temperatures, quantum weirdness.

778
00:50:10.040 --> 00:50:13.422
You know, it's like cool, but it's cool.

779
00:50:13.843 --> 00:50:14.803
But that's about it.

780
00:50:15.724 --> 00:50:20.708
Like maybe in 100 years, it can lead to like, you know, quantum devices and things like that.

781
00:50:21.468 --> 00:50:26.873
This paper is very cool because it's doing it in a classical system, which means no quantum weirdness.

782
00:50:27.553 --> 00:50:29.214
And it's doing it at room temperature.

783
00:50:29.495 --> 00:50:30.055
Wow.

784
00:50:30.295 --> 00:50:33.718
And it's doing it at a big enough scale that we can see with our naked eye.

785
00:50:34.318 --> 00:50:36.479
This is giving me LK99 vibes.

786
00:50:37.339 --> 00:50:38.960
Yeah, but this is out of UC Boulder.

787
00:50:39.020 --> 00:50:39.940
Yeah, so it's legit.

788
00:50:40.000 --> 00:50:41.441
And so it's pretty legit.

789
00:50:41.601 --> 00:50:46.182
UC Boulder has probably the best AMO department out of anywhere in the world.

790
00:50:46.202 --> 00:50:46.462
Okay.

791
00:50:46.622 --> 00:50:46.903
Okay?

792
00:50:47.003 --> 00:50:49.644
Atomic Molecular Optical.

793
00:50:50.584 --> 00:50:51.364
Those guys do it.

794
00:50:51.664 --> 00:50:54.285
Let's invest in our university research institutions.

795
00:50:54.325 --> 00:50:54.505
Yeah.

796
00:50:54.585 --> 00:50:58.647
I mean, you know, Boulder has like the most PhDs out of any city.

797
00:50:58.687 --> 00:50:59.347
I heard this recently.

798
00:50:59.447 --> 00:51:01.568
Yeah, and a lot of it has to do with UC Boulder.

799
00:51:01.728 --> 00:51:05.810
Out of UC Boulder came NIST, which is the National Institute for Standards and Technology.

800
00:51:06.051 --> 00:51:08.532
There's a bunch of other science labs just in Boulder.

801
00:51:08.852 --> 00:51:12.774
UC Boulder itself has like four Nobel Prizes in physics just for AMO.

802
00:51:13.294 --> 00:51:16.416
Okay, like not like, okay, there's one for cosmology, one for...

803
00:51:16.676 --> 00:51:20.298
It's just all like in the lab making like materials, right?

804
00:51:20.618 --> 00:51:22.639
They made the first Bose-Einstein condensate.

805
00:51:22.979 --> 00:51:27.742
They're the ones who figured out how to trap ions using lasers and cool them down using lasers...

806
00:51:28.022 --> 00:51:29.122
So they're very good.

807
00:51:29.402 --> 00:51:35.784
So when a paper comes out of UC Boulder and it's an AMO, it's like, okay, these guys probably talked to their department.

808
00:51:35.924 --> 00:51:38.045
And if it didn't pass, like, their own.

809
00:51:38.105 --> 00:51:39.765
Yeah, they have the best in the world there.

810
00:51:39.785 --> 00:51:40.445
Yeah, yeah, yeah.

811
00:51:40.465 --> 00:51:47.487
It's like, so it's a really cool little thing that they've made out of liquid crystals, right?

812
00:51:47.547 --> 00:51:48.628
These are the LCDs.

813
00:51:50.088 --> 00:51:53.230
And they made a macroscopic time crystal.

814
00:51:53.650 --> 00:51:54.671
Okay.

815
00:51:55.131 --> 00:52:04.597
And what you can do is, what's really cool is, first of all, you don't need a laser or oscillating magnetic field like in those old time crystals.

816
00:52:04.677 --> 00:52:10.721
In the old ones, you needed a quantum system and you needed a laser, which means coherent light, single frequency.

817
00:52:11.081 --> 00:52:13.883
Or like a magnetic field, it's oscillating at a single frequency.

818
00:52:13.903 --> 00:52:15.004
And then that sort of...

819
00:52:16.799 --> 00:52:20.021
Drives this quantum system to do this periodic motion.

820
00:52:20.081 --> 00:52:21.601
This is just like a light.

821
00:52:21.621 --> 00:52:26.543
You just shine a light on it, and then it uses the energy from that light to create this periodic motion.

822
00:52:26.743 --> 00:52:27.003
Okay?

823
00:52:27.543 --> 00:52:32.886
And it just uses LCD, like liquid crystal technology, pneumatic liquid crystals.

824
00:52:33.366 --> 00:52:37.788
So the way that liquid crystals work, like, you know, you've used liquid crystals before.

825
00:52:39.288 --> 00:52:40.489
It's a really simple technology.

826
00:52:40.509 --> 00:52:41.709
It actually won the Nobel Prize.

827
00:52:42.649 --> 00:52:43.550
The idea is...

828
00:52:44.230 --> 00:52:46.412
So what does it mean by a liquid crystal?

829
00:52:46.632 --> 00:52:46.892
Okay.

830
00:52:47.272 --> 00:52:49.694
A crystal you usually think of as a solid.

831
00:52:49.994 --> 00:52:50.194
Yes.

832
00:52:50.234 --> 00:52:50.415
Right?

833
00:52:51.475 --> 00:52:56.079
A solid meaning that you've got atoms in a regularly spaced pattern and they don't really move.

834
00:52:56.859 --> 00:53:04.785
In the liquid crystal, you've got molecules that are all sort of... You can think of them as like rods, many tiny rods.

835
00:53:04.825 --> 00:53:06.446
They're molecules that are elongated.

836
00:53:07.147 --> 00:53:10.109
And you can manipulate their direction...

837
00:53:11.590 --> 00:53:12.591
using an electric field.

838
00:53:12.871 --> 00:53:13.031
Okay.

839
00:53:13.271 --> 00:53:13.531
Okay?

840
00:53:14.412 --> 00:53:18.875
So if you put an electric field in this direction, all of the molecules align in that direction.

841
00:53:19.396 --> 00:53:23.619
And you can have an electric field in this direction, which means all the molecules align in this direction.

842
00:53:23.939 --> 00:53:28.963
And what these molecules do is because they're small enough, they can actually influence the propagation of light.

843
00:53:29.964 --> 00:53:30.204
Okay?

844
00:53:30.644 --> 00:53:33.526
And so that's exactly how liquid crystal displays work.

845
00:53:33.827 --> 00:53:35.268
What you do is you have a backlight.

846
00:53:35.288 --> 00:53:35.648
Okay?

847
00:53:35.808 --> 00:53:39.892
OK, that like lets and then you have a polarizer that polarizes all the light.

848
00:53:39.932 --> 00:53:42.715
So all of the electric fields are propagating in one direction.

849
00:53:43.416 --> 00:53:49.362
And what you can do is you can have a liquid crystal interface in between another polarizer.

850
00:53:49.682 --> 00:53:55.408
So now if the light gets to the other polarizer and it's polarized in the same direction, it's going to be let through.

851
00:53:55.428 --> 00:53:55.528
Yeah.

852
00:53:55.908 --> 00:54:00.233
But if the light is polarized in a different direction, in the perpendicular direction, then it's going to be blocked.

853
00:54:00.734 --> 00:54:09.404
And what you can do is use an electric field to manipulate the liquid crystals in the middle to influence the propagation of light.

854
00:54:09.965 --> 00:54:12.047
That's the fundamental technology.

855
00:54:12.248 --> 00:54:12.508
Okay.

856
00:54:12.748 --> 00:54:15.411
So you can have an electric field that sort of rotates the light.

857
00:54:16.172 --> 00:54:20.437
And then if it rotates in exactly the right way, then the light is going to go through and I'm going to see something.

858
00:54:20.998 --> 00:54:28.947
But if there's no electric field or if I make the electric field this way, then all the rods are going to be, you know, sort of in the perpendicular direction.

859
00:54:29.247 --> 00:54:31.470
The light is not going to rotate and I'm going to get blocked.

860
00:54:31.510 --> 00:54:31.670
Yes.

861
00:54:31.930 --> 00:54:33.953
And that's fundamentally how liquid crystals work.

862
00:54:34.373 --> 00:54:39.157
They're liquid in the sense that, you know, just how in a liquid the molecules can move around.

863
00:54:39.457 --> 00:54:45.762
In this case, these liquid molecules, these rod-like molecules can move around and they can be controlled by an electric field.

864
00:54:46.022 --> 00:54:49.845
And they can be controlled by light itself, right?

865
00:54:50.205 --> 00:54:55.169
So they figured what I can do is I can create a liquid crystal.

866
00:54:56.650 --> 00:55:01.474
And this liquid crystal has exactly the right properties where it can...

867
00:55:02.400 --> 00:55:09.390
sort of act like a giant sort of, I guess, okay.

868
00:55:09.650 --> 00:55:17.040
So this liquid crystal can be oriented in such a way that when it's shined on by light,

869
00:55:18.422 --> 00:55:21.986
What you're going to get is deformations in the crystal.

870
00:55:22.186 --> 00:55:22.266
Yes.

871
00:55:22.286 --> 00:55:22.566
Okay?

872
00:55:22.746 --> 00:55:24.068
Tiny little deformations.

873
00:55:24.088 --> 00:55:31.175
So what you can imagine is, let's say all of the liquid crystals, all of these rods are oriented in one direction.

874
00:55:31.255 --> 00:55:31.515
Okay?

875
00:55:32.176 --> 00:55:35.439
And now what I do is I introduce a tiny kink.

876
00:55:36.260 --> 00:55:36.940
in this lattice.

877
00:55:37.321 --> 00:55:43.844
So they're all oriented in one direction, but I've taken like a tiny little neighborhood and I've like bent them in a certain way.

878
00:55:44.184 --> 00:55:44.424
Okay.

879
00:55:44.805 --> 00:55:50.267
What you can do is create a material such that that bend, which becomes this sort of topological defect.

880
00:55:50.287 --> 00:55:50.467
Yes.

881
00:55:50.628 --> 00:55:54.330
This like idea of like, okay, this neighborhood has a tiny little bend or a twist.

882
00:55:54.830 --> 00:56:00.173
That twist can then influence the stuff around it, which can influence the stuff around it, which can influence the stuff around it.

883
00:56:00.453 --> 00:56:04.335
And what you get is this sort of traveling wave of disturbance.

884
00:56:04.455 --> 00:56:04.655
Yes.

885
00:56:04.995 --> 00:56:05.255
Okay.

886
00:56:05.595 --> 00:56:09.437
There are what are called topological solitons in physics.

887
00:56:09.737 --> 00:56:15.719
This idea that like I've got this like disturbance and that disturbance kind of acts like a particle.

888
00:56:16.339 --> 00:56:16.599
Okay.

889
00:56:16.879 --> 00:56:21.021
Because it travels the same way as a particle.

890
00:56:21.481 --> 00:56:26.223
The orientation of that disturbance can have a charge in the sense that if I...

891
00:56:26.943 --> 00:56:29.585
twist it clockwise, then it's a plus one.

892
00:56:29.665 --> 00:56:32.146
And if I twist it counterclockwise, it's a minus one.

893
00:56:32.426 --> 00:56:38.950
And that charge is conserved in the same way that sort of charge is conserved if I were to have electrons moving around, right?

894
00:56:39.150 --> 00:56:43.713
And they can cancel each other out, kind of the same way that charges cancel each other out, right?

895
00:56:44.033 --> 00:56:47.836
So they created this thing and they actually did the numerical calculation.

896
00:56:47.856 --> 00:56:52.700
So they wrote down an equation for how these solitons would interact with each other.

897
00:56:53.000 --> 00:57:01.147
And all of these deformations act like little quasi particles that interact with each other and have a physics to themselves.

898
00:57:01.187 --> 00:57:01.367
Yes.

899
00:57:01.807 --> 00:57:06.649
That is an emergent property of the little tiny rods themselves.

900
00:57:06.669 --> 00:57:09.469
That initial disturbance that has now created a cascade.

901
00:57:09.690 --> 00:57:09.930
Yeah.

902
00:57:09.970 --> 00:57:13.291
That is now independent of the initial source necessarily being.

903
00:57:13.371 --> 00:57:14.771
Yeah.

904
00:57:15.051 --> 00:57:17.572
And it's like it's on top of the rods themselves.

905
00:57:17.752 --> 00:57:20.353
The rods actually aren't the particles anymore.

906
00:57:20.393 --> 00:57:20.553
Yeah.

907
00:57:20.753 --> 00:57:27.797
It's like the way in which the rods are oriented becomes a particle mathematically.

908
00:57:28.137 --> 00:57:29.098
Right.

909
00:57:29.978 --> 00:57:35.501
You can describe the twist itself as this thing that propagates rather than the rods.

910
00:57:35.521 --> 00:57:37.622
Because the rods are just staying where they are.

911
00:57:37.702 --> 00:57:37.883
Right.

912
00:57:38.443 --> 00:57:38.643
Right.

913
00:57:38.683 --> 00:57:39.824
That's the crystal part of it.

914
00:57:41.024 --> 00:57:45.547
But the way that the rods move, that disturbance moves around like a particle.

915
00:57:46.087 --> 00:57:47.088
This is really...

916
00:57:47.128 --> 00:57:48.109
Which is kind of cool, right?

917
00:57:48.129 --> 00:57:50.491
No, it's very... That's, like, very interesting.

918
00:57:50.551 --> 00:57:50.991
Yeah, dude.

919
00:57:51.051 --> 00:57:54.794
Solitons have been around, like, I think in the 1800s, there's this British guy.

920
00:57:54.814 --> 00:57:55.495
Forget his name.

921
00:57:56.355 --> 00:57:57.076
Maybe we'll put it up.

922
00:57:57.176 --> 00:57:57.356
Yes.

923
00:57:57.476 --> 00:58:03.821
But it's a funny story because what he saw was he was, you know, these British physicists back in the day.

924
00:58:03.901 --> 00:58:04.822
Very rich, right?

925
00:58:04.982 --> 00:58:07.104
He's on horseback in the countryside, right?

926
00:58:07.424 --> 00:58:11.087
And he's, like, on horseback, and he sees a canal, and then there's a boat.

927
00:58:11.467 --> 00:58:13.309
And the boat, like, breaks suddenly.

928
00:58:13.669 --> 00:58:13.909
Okay.

929
00:58:14.210 --> 00:58:17.812
And when the boat breaks suddenly, it created this wave.

930
00:58:18.132 --> 00:58:20.274
And he noticed that the wave just kept propagating.

931
00:58:20.874 --> 00:58:22.896
And then he was on horseback and he's got nothing to do.

932
00:58:22.916 --> 00:58:24.617
He's probably, you know, got an estate.

933
00:58:24.757 --> 00:58:24.957
Right.

934
00:58:25.017 --> 00:58:26.338
That like gets him money.

935
00:58:26.598 --> 00:58:26.778
Yes.

936
00:58:26.798 --> 00:58:28.239
Like in Downton Abbey or whatever.

937
00:58:28.560 --> 00:58:29.520
So he's like, fuck it.

938
00:58:29.580 --> 00:58:31.081
I'm going to follow this wave.

939
00:58:31.362 --> 00:58:33.023
And he followed it for two miles.

940
00:58:33.223 --> 00:58:33.423
Hmm.

941
00:58:33.503 --> 00:58:38.205
There was a wave that kept reinforcing itself and it acted kind of like a particle, right?

942
00:58:38.245 --> 00:58:42.588
In the sense that the wave didn't really dissipate until two miles later.

943
00:58:42.928 --> 00:58:46.569
And he could follow this thing and he discovered the first soliton.

944
00:58:46.870 --> 00:58:51.532
There's a little plaque where he lived where it's like this is the guy who discovered the first soliton.

945
00:58:52.012 --> 00:58:57.297
And ever since then, they've been observed in all sorts of sort of material sciences.

946
00:58:57.337 --> 00:58:59.960
We've observed them in biology and quantum systems.

947
00:59:00.300 --> 00:59:05.164
And this is exactly one of those things where it's like there's a wave that's being propagated.

948
00:59:05.264 --> 00:59:11.030
And that wave sort of reinforces itself because of the way that the medium itself works.

949
00:59:11.070 --> 00:59:11.310
Works.

950
00:59:11.570 --> 00:59:11.730
Right.

951
00:59:11.750 --> 00:59:12.011
You know?

952
00:59:12.031 --> 00:59:12.371
Right.

953
00:59:12.491 --> 00:59:15.654
It is a combination of that input and the structure of the medium.

954
00:59:15.714 --> 00:59:15.934
Yes.

955
00:59:16.274 --> 00:59:16.835
Exactly.

956
00:59:16.935 --> 00:59:17.776
The liquid crystal.

957
00:59:18.436 --> 00:59:31.085
And the structure of the liquid crystal enables it to continue the propagation of the wave independent, like as an emergent property of the structure's initial condition.

958
00:59:31.165 --> 00:59:31.685
Exactly.

959
00:59:31.865 --> 00:59:32.406
Yeah, yeah.

960
00:59:32.686 --> 00:59:36.929
And so they figured that they could use this to make a classical...

961
00:59:37.369 --> 01:00:02.782
time crystal right because this is not quantum this is this is just the the rods themselves are like you know multiple many many atoms big right so these guys are interacting through just classical effects yes right and so what you can do is you can have a liquid crystal embedded inside glass right that have these two layers and as one layer sort of moves it influences the light on the other layer and then that thing moves and there's this feedback cycle

962
01:00:03.523 --> 01:00:08.789
And that feedback cycle is what creates this propagation of the time crystal.

963
01:00:08.869 --> 01:00:17.540
And it's really a space-time crystal because what you're seeing, you can actually look at the videos and you can see the colors move.

964
01:00:20.645 --> 01:00:21.486
In real time.

965
01:00:22.207 --> 01:00:27.373
And the wavelength is like, you know, about half a millimeter to a millimeter, which means you can see it with your naked eye.

966
01:00:27.874 --> 01:00:33.961
And it's moving through this liquid crystal in this sort of propagating, you know, traveling wave.

967
01:00:34.442 --> 01:00:34.642
Right?

968
01:00:35.583 --> 01:00:35.784
And...

969
01:00:36.644 --> 01:00:45.489
It was huge because one of the big things that defines a crystal, right, is it has to be, as I said before, crystals have to be resilient.

970
01:00:45.689 --> 01:00:45.869
Yes.

971
01:00:46.269 --> 01:00:46.769
To stress.

972
01:00:46.989 --> 01:00:47.149
Right.

973
01:00:47.189 --> 01:00:47.710
Okay.

974
01:00:47.730 --> 01:00:51.832
If I take a diamond and I press it or I heat it up, the diamond is going to retain its temperature.

975
01:00:52.152 --> 01:00:53.592
So how did they actually test that here?

976
01:00:53.632 --> 01:00:56.154
Well, what you can do is you can like poke it.

977
01:00:56.394 --> 01:00:56.574
Yeah.

978
01:00:57.154 --> 01:01:02.402
And the pattern stays, which means that it is a true time crystal.

979
01:01:02.422 --> 01:01:02.702
Right.

980
01:01:02.842 --> 01:01:06.548
It is resilient to the poking.

981
01:01:07.068 --> 01:01:07.769
It already has.

982
01:01:08.070 --> 01:01:10.193
Yeah, it's got some time crystal going through.

983
01:01:10.493 --> 01:01:12.356
And then when I poke it, the time crystal sort of.

984
01:01:12.676 --> 01:01:14.418
Like it recovers where I poke it.

985
01:01:14.438 --> 01:01:16.901
Of course, it's going to like mess up, but I remove it.

986
01:01:17.201 --> 01:01:20.384
And then the crystal regains its properties.

987
01:01:20.424 --> 01:01:23.648
The other thing they did was they would vary the brightness of the light.

988
01:01:24.048 --> 01:01:28.233
Because it can't be that like, you know, it depends really heavily on the brightness of the light.

989
01:01:28.253 --> 01:01:30.275
So they randomly varied the brightness of the light.

990
01:01:30.535 --> 01:01:32.077
The crystal kept going.

991
01:01:32.357 --> 01:01:39.765
Which means that what this is is really an intrinsic property of the material itself rather than the drive that's coming from the outside.

992
01:01:40.026 --> 01:01:40.246
Yes.

993
01:01:40.306 --> 01:01:40.446
Right?

994
01:01:40.546 --> 01:01:51.198
No, that's a very – that's exactly – And then the final nail in the coffin for me was what they did was they would turn off the light and then they would turn it back on again.

995
01:01:52.059 --> 01:02:06.868
And if this is not a true crystal, if this is simply a response to the light, then when I turned it off, I mean, sorry, when I turned it back on, the crystal would start up at exactly the same phase every time.

996
01:02:07.248 --> 01:02:10.250
Meaning I turn it on five minutes later, the same pattern.

997
01:02:10.390 --> 01:02:14.293
I mean, sorry, five seconds later, the same pattern emergence at exactly the same spot.

998
01:02:14.973 --> 01:02:21.759
But what would happen was they would turn it off and then they turned the light back on and it would randomly order itself.

999
01:02:22.319 --> 01:02:27.103
The same crystal would start, but at different starting points in that wave.

1000
01:02:27.303 --> 01:02:28.064
You see what I'm saying?

1001
01:02:28.104 --> 01:02:30.305
Like the phase offset was completely random.

1002
01:02:30.325 --> 01:02:30.546
Yes.

1003
01:02:30.806 --> 01:02:35.029
Which again tells you this is not a property of the outside drive.

1004
01:02:35.069 --> 01:02:35.289
Right.

1005
01:02:35.389 --> 01:02:38.212
It's a property of the material itself.

1006
01:02:38.232 --> 01:02:38.352
Right.

1007
01:02:38.672 --> 01:02:43.014
So this is truly a time crystal that is macroscopic and it's something we can see.

1008
01:02:43.575 --> 01:02:45.876
This is this is I mean, this is right.

1009
01:02:45.916 --> 01:02:46.996
This is this is really fast.

1010
01:02:47.096 --> 01:02:47.997
It's really cool.

1011
01:02:48.097 --> 01:02:52.019
And and like, OK, you might be thinking, OK, this is like this is cool.

1012
01:02:52.139 --> 01:02:52.359
Right.

1013
01:02:53.560 --> 01:02:56.341
Then I got into reading about the applications.

1014
01:02:56.441 --> 01:02:57.262
OK. OK.

1015
01:02:57.702 --> 01:02:59.523
The applications for this are like.

1016
01:03:01.360 --> 01:03:03.222
like science fiction to me.

1017
01:03:03.482 --> 01:03:03.743
Okay.

1018
01:03:04.644 --> 01:03:08.287
Most of them have to do with like cryptography and security.

1019
01:03:08.768 --> 01:03:09.008
Okay.

1020
01:03:09.529 --> 01:03:10.610
First one I'll tell you about.

1021
01:03:11.711 --> 01:03:14.374
Imagine we make this thing cheap.

1022
01:03:16.055 --> 01:03:17.437
We can stick it on a dollar bill.

1023
01:03:19.101 --> 01:03:19.782
Oh, right.

1024
01:03:19.942 --> 01:03:21.703
As a cryptographic signature.

1025
01:03:21.783 --> 01:03:23.064
As a cryptographic signature.

1026
01:03:23.104 --> 01:03:29.148
Now, if you want to test whether your Benjamin is a real Benjamin, you take your iPhone flashlight, stick it on there.

1027
01:03:29.448 --> 01:03:30.809
It's going to have this pattern, right?

1028
01:03:31.150 --> 01:03:32.811
Because all it needs is an external drive.

1029
01:03:33.151 --> 01:03:37.074
And the pattern is we can make it really hard to duplicate.

1030
01:03:37.094 --> 01:03:37.274
Yes.

1031
01:03:37.534 --> 01:03:37.734
Right?

1032
01:03:37.914 --> 01:03:39.936
So, you know that strip that's on the Benjamin?

1033
01:03:40.056 --> 01:03:40.216
Yes.

1034
01:03:40.276 --> 01:03:44.119
We can have that strip be a time crystal, like literally like…

1035
01:03:44.799 --> 01:03:45.419
propagates.

1036
01:03:45.479 --> 01:03:45.699
Yes.

1037
01:03:45.839 --> 01:03:46.019
Right?

1038
01:03:46.079 --> 01:03:47.500
And then you'll know, okay, that's a real one.

1039
01:03:47.520 --> 01:03:50.920
It has a signature that only... And it's going to be so hard to duplicate.

1040
01:03:50.960 --> 01:03:51.140
Right.

1041
01:03:51.360 --> 01:03:51.540
Right.

1042
01:03:51.840 --> 01:03:52.081
Right?

1043
01:03:52.381 --> 01:03:52.561
Right.

1044
01:03:53.121 --> 01:03:54.201
That's so cool.

1045
01:03:54.221 --> 01:03:56.641
No, that's very... That's actually really... Yeah.

1046
01:03:56.862 --> 01:04:02.583
And the other one is like 2D... Like, so, you know, usually we have 2D barcodes, like a QR code is a 2D.

1047
01:04:03.043 --> 01:04:06.464
Now you can have a 2D plus one barcode.

1048
01:04:06.564 --> 01:04:11.405
You can have a QR code that is 2D.

1049
01:04:11.745 --> 01:04:13.025
And then when you shine a light on it,

1050
01:04:13.625 --> 01:04:18.549
the QR code itself changes into different QR codes.

1051
01:04:18.949 --> 01:04:23.272
You can have multiple QR codes on top that have different periodicities.

1052
01:04:23.813 --> 01:04:30.098
So that like, let's say one is like two seconds and the other one is three seconds and the other one is five seconds.

1053
01:04:30.398 --> 01:04:37.783
Well, then the next time that it's going to become a full QR code is gonna be five times, three times two, which is 30, right?

1054
01:04:37.803 --> 01:04:40.866
So you're gonna have to wait 30 seconds to get the actual QR code.

1055
01:04:41.687 --> 01:04:49.155
You can imagine you've introduced a whole new dimension to security.

1056
01:04:49.255 --> 01:04:51.117
A manipulatable dimension.

1057
01:04:51.458 --> 01:04:51.738
Yeah.

1058
01:04:51.858 --> 01:04:55.282
And the whole point is this is at room temperature.

1059
01:04:55.302 --> 01:04:55.642
Right.

1060
01:04:55.662 --> 01:04:55.843
Right.

1061
01:04:56.503 --> 01:04:58.404
And it's macroscopic.

1062
01:04:58.484 --> 01:04:59.604
It's something I can see.

1063
01:05:00.105 --> 01:05:05.947
I don't need a telescope or like weird ions and lasers to actually, oh, look, it's doing the thing.

1064
01:05:06.387 --> 01:05:09.029
This is why I made the LK99 reference.

1065
01:05:09.089 --> 01:05:11.930
Not because of whether it was real or not real, but because...

1066
01:05:12.250 --> 01:05:14.511
It's room temperature and macroscopic.

1067
01:05:14.551 --> 01:05:14.791
Yeah.

1068
01:05:14.931 --> 01:05:22.212
And the reason why that's material is that means the transition from theoretical to production application.

1069
01:05:22.232 --> 01:05:22.392
Yeah.

1070
01:05:22.652 --> 01:05:27.234
Now we've just got to make it cheap, which is honestly kind of like the easier part of the innovation.

1071
01:05:27.254 --> 01:05:27.454
Yes.

1072
01:05:27.514 --> 01:05:27.734
Right.

1073
01:05:27.794 --> 01:05:28.194
Exactly.

1074
01:05:28.234 --> 01:05:28.474
You know.

1075
01:05:28.494 --> 01:05:29.454
It's an engineering problem.

1076
01:05:29.474 --> 01:05:29.594
Yeah.

1077
01:05:29.654 --> 01:05:32.595
Now it's like, okay, because we can make LCDs real cheap now.

1078
01:05:32.895 --> 01:05:33.055
Right.

1079
01:05:33.075 --> 01:05:33.255
Right.

1080
01:05:33.675 --> 01:05:33.895
Right.

1081
01:05:34.035 --> 01:05:34.435
This is.

1082
01:05:34.975 --> 01:05:38.056
And like this, you know, it doesn't need a power source.

1083
01:05:38.076 --> 01:05:39.396
You just need to put it on there.

1084
01:05:39.416 --> 01:05:42.057
The power source is the light that I'm going to shine on it.

1085
01:05:42.958 --> 01:05:43.759
It's kind of cool.

1086
01:05:43.839 --> 01:05:58.137
I'm seeing from the abstract here, their potential technological utility includes optical devices, photonics, space-time crystal generators, telecommunications, anti-counterfeiting design.

1087
01:05:58.418 --> 01:05:59.960
And again, this is just a first –

1088
01:06:00.600 --> 01:06:05.106
Yeah, I mean, I only talked about the anti-counterfeiting and the security part.

1089
01:06:05.126 --> 01:06:08.090
But yeah, as you said, devices, all sorts of stuff.

1090
01:06:08.150 --> 01:06:13.217
There's all sorts of implications in terms of what we can get out of now this fundamental...

1091
01:06:15.051 --> 01:06:23.758
like implementation experimentally at room temperature in a macroscopic form, which is, yeah, I mean, there's a reason why it was picked up by so many news outlets.

1092
01:06:23.778 --> 01:06:25.320
Yeah, it was covered by a lot.

1093
01:06:25.340 --> 01:06:26.461
This is quite a big deal, yeah.

1094
01:06:26.521 --> 01:06:27.341
This is quite good.

1095
01:06:27.381 --> 01:06:29.623
They didn't get the, they got nature materials.

1096
01:06:29.643 --> 01:06:33.547
They could have put it in nature, but they said, we'll put it in nature materials.

1097
01:06:33.607 --> 01:06:34.988
They probably tried in nature first.

1098
01:06:35.068 --> 01:06:37.310
Usually everyone does.

1099
01:06:37.370 --> 01:06:39.412
First time crystals that you can see.

1100
01:06:40.887 --> 01:06:42.088
Really interesting story.

1101
01:06:42.108 --> 01:06:44.889
And the background is also helpful to understand.

1102
01:06:44.909 --> 01:06:48.170
We live in L.A., so we know a lot of people that like to talk about crystals.

1103
01:06:48.210 --> 01:06:49.771
This is a little bit different than those crystals.

1104
01:06:49.791 --> 01:06:52.252
Yeah, this one actually works.

1105
01:06:52.873 --> 01:06:55.794
Speaking of California, for our third story...

1106
01:06:56.234 --> 01:06:58.276
We're now going to move into the immunology space.

1107
01:06:58.756 --> 01:07:04.240
Headline, hidden viruses in our DNA could be medicine's next big breakthrough.

1108
01:07:04.880 --> 01:07:16.188
Scientists at La Jolla Institute for Immunology, when they're not out surfing on the beautiful sand beaches of San Diego, have decoded the 3D structure of an ancient viral protein hidden...

1109
01:07:17.128 --> 01:07:35.256
Within our DNA our own DNA our own DNA this protein found on cancer and autoimmune cells has a unique shape that could unlock new Diagnostics and therapies this was published in science advances Again, this is another California University.

1110
01:07:35.616 --> 01:07:38.017
We have the most papers the most patents and

1111
01:07:38.597 --> 01:07:39.538
We got it.

1112
01:07:40.038 --> 01:07:42.059
I just always have to make sure people know that we are winning.

1113
01:07:42.279 --> 01:07:42.819
We are winning.

1114
01:07:43.459 --> 01:07:43.959
We are winning.

1115
01:07:43.999 --> 01:07:46.660
This is La Jolla, California, baby.

1116
01:07:46.980 --> 01:07:48.721
It's right next to UC San Diego.

1117
01:07:48.941 --> 01:07:55.604
UC San Diego is an amazing, amazing crucible for biotechnology and molecular biology research.

1118
01:07:55.624 --> 01:07:59.005
It's kind of like the Silicon Valley for biology in some sense.

1119
01:08:00.125 --> 01:08:02.146
And this is a really groundbreaking achievement.

1120
01:08:02.286 --> 01:08:09.948
It's the first time that we have ever published the structure of a human endogenous retrovirus.

1121
01:08:10.248 --> 01:08:10.508
Okay.

1122
01:08:10.788 --> 01:08:14.208
And this is a virus that hides within our own DNA, UNRs.

1123
01:08:14.249 --> 01:08:15.649
I didn't even know that was possible.

1124
01:08:15.669 --> 01:08:16.069
Yeah.

1125
01:08:16.549 --> 01:08:18.129
This is crazy, dude.

1126
01:08:18.229 --> 01:08:22.350
8% of our DNA is viral endogenous DNA.

1127
01:08:29.852 --> 01:08:51.900
infiltrating that have infiltrated dude it's it's they're like they're like sleeper cells dude i was this is when people say the aliens uh manufactured uh created no look this is exactly what i'm talking about little viruses are they're writing our stuff it's crazy it's crazy to think about so so first of all i was like i was like how does this even get in why is this in my dna bro right like

1128
01:08:52.720 --> 01:08:53.561
What's going on?

1129
01:08:53.661 --> 01:08:55.041
Okay, so it's crazy.

1130
01:08:55.121 --> 01:08:57.502
These things are called retroviruses.

1131
01:08:57.543 --> 01:09:04.926
Retrovirus is a virus that has, instead of using DNA as its genetic material, it uses RNA as its genetic material.

1132
01:09:05.426 --> 01:09:10.749
And what it does is when it infects a cell, it uses something called reverse transcriptase.

1133
01:09:11.249 --> 01:09:14.070
Usually transcriptase goes from DNA to RNA.

1134
01:09:14.090 --> 01:09:15.371
This is the reverse process.

1135
01:09:15.651 --> 01:09:18.352
So it's taking RNA and it's making double-stranded DNA.

1136
01:09:18.772 --> 01:09:20.153
So that's reverse transcriptase.

1137
01:09:20.513 --> 01:09:26.836
And then what ends up happening is that reverse transcriptase goes inside the nucleus of our cells.

1138
01:09:26.896 --> 01:09:27.056
Yes.

1139
01:09:27.376 --> 01:09:35.760
And then there's something called integrase, which is another viral protein that takes that DNA, that bit of viral DNA, and sticks it inside our chromosomal DNA.

1140
01:09:36.200 --> 01:09:39.603
Now, if this happens to our skin cells, okay, we got like some viral DNA.

1141
01:09:39.863 --> 01:09:49.911
But if it happens to our sperm and egg, now the next organism that gets that sperm and egg, all of the DNA is going to have it, right?

1142
01:09:50.352 --> 01:09:51.733
These are called germline mutations.

1143
01:09:51.753 --> 01:09:55.016
I was going to say, going back to our CRISPR story, it's the germline.

1144
01:09:55.036 --> 01:09:56.697
It's the germline mutations, right?

1145
01:09:56.757 --> 01:09:56.897
Yes.

1146
01:09:56.937 --> 01:09:59.499
So this is a germline insertion that happens.

1147
01:09:59.820 --> 01:10:02.502
And then now the virus becomes something called a provirus.

1148
01:10:02.942 --> 01:10:05.025
which is something that's just hanging out in our DNA.

1149
01:10:05.085 --> 01:10:05.645
Yes.

1150
01:10:05.826 --> 01:10:08.288
8% of our DNA comes from this.

1151
01:10:08.328 --> 01:10:08.909
Comes from this.

1152
01:10:09.009 --> 01:10:10.251
Dude, that's 1 in 12.

1153
01:10:10.671 --> 01:10:11.532
That's like a lot.

1154
01:10:11.552 --> 01:10:12.073
That's a lot.

1155
01:10:12.173 --> 01:10:13.314
That's a lot.

1156
01:10:13.434 --> 01:10:16.478
I had no idea this was even a thing when we first put the story in the show notes.

1157
01:10:16.498 --> 01:10:17.219
Yeah, this is crazy.

1158
01:10:17.259 --> 01:10:18.460
I was like, this is fascinating.

1159
01:10:18.560 --> 01:10:19.580
Yeah, yeah, yeah.

1160
01:10:19.600 --> 01:10:25.142
So all of that junk DNA, 8% of that is viruses just hanging out.

1161
01:10:25.442 --> 01:10:25.662
Okay.

1162
01:10:25.902 --> 01:10:29.623
Now, most of these viruses, they get mutations over the years and they can become inactive.

1163
01:10:29.943 --> 01:10:30.163
Okay.

1164
01:10:30.384 --> 01:10:34.325
But there's a specific virus called the human endogenous virus K.

1165
01:10:35.385 --> 01:10:36.406
The Herb K virus.

1166
01:10:36.447 --> 01:10:36.827
Okay.

1167
01:10:37.047 --> 01:10:39.170
It's a provirus now because it's within our DNA.

1168
01:10:39.571 --> 01:10:42.374
But this thing is the most recent infiltrator.

1169
01:10:42.775 --> 01:10:43.015
Okay.

1170
01:10:43.276 --> 01:10:46.320
And it's because it's the most recent infiltrator.

1171
01:10:46.900 --> 01:10:48.442
It has the fewest number of mutations.

1172
01:10:48.843 --> 01:10:49.123
Right.

1173
01:10:49.384 --> 01:10:51.446
Which means it's still a problem.

1174
01:10:51.466 --> 01:10:51.707
Mm-hmm.

1175
01:10:52.027 --> 01:10:52.267
Okay.

1176
01:10:52.908 --> 01:10:53.748
And it actually...

1177
01:10:53.989 --> 01:10:58.372
So in most healthy cells, there's epigenetic silencing.

1178
01:10:58.452 --> 01:10:58.733
Okay.

1179
01:10:59.213 --> 01:11:00.674
Our DNA has figured out...

1180
01:11:00.714 --> 01:11:02.896
I don't know who you are.

1181
01:11:02.976 --> 01:11:10.122
I'm going to just put a bunch of random stuff on you so that no one ever transcribes you into RNA and then you become a protein.

1182
01:11:10.182 --> 01:11:10.502
Okay.

1183
01:11:10.642 --> 01:11:10.762
So...

1184
01:11:11.723 --> 01:11:12.784
Our DNA is very good.

1185
01:11:12.884 --> 01:11:14.584
If you're healthy, the cell is healthy.

1186
01:11:15.325 --> 01:11:16.846
It's like, I don't know what this is.

1187
01:11:16.906 --> 01:11:18.066
It's like a bouncer at a club.

1188
01:11:18.186 --> 01:11:18.506
Yeah.

1189
01:11:18.566 --> 01:11:20.987
It's like, you're already in here, but you're cut off.

1190
01:11:21.028 --> 01:11:21.548
Yeah.

1191
01:11:21.568 --> 01:11:21.948
You know?

1192
01:11:22.128 --> 01:11:24.969
It's like, I don't know how you got in, but no more drinks.

1193
01:11:25.550 --> 01:11:25.810
Okay?

1194
01:11:26.650 --> 01:11:32.853
Every once in a while, you get something like a cancer cell or you get autoimmune disorders.

1195
01:11:34.054 --> 01:11:38.556
The cell is losing control and this virus can start acting up.

1196
01:11:39.497 --> 01:11:48.987
So it's actually been detected, this H-Herv-K envelope, the envelope protein that is the protein that encodes for the envelope of the virus.

1197
01:11:49.328 --> 01:11:50.529
That's the thing that snuck in.

1198
01:11:51.470 --> 01:11:53.973
That has been known to be expressed internally.

1199
01:11:54.153 --> 01:12:12.660
In cancers like breast cancer, ovarian cancer, prostate cancer, melanoma, and also in autoimmune disorders like lupus, diabetes type 1, SLE, these kinds of autoimmune disorders, you can actually see the cells that are affected express this HERV-K protein.

1200
01:12:12.740 --> 01:12:12.960
Got it.

1201
01:12:13.020 --> 01:12:14.961
Which means that it is actually a problem.

1202
01:12:15.121 --> 01:12:15.341
Okay.

1203
01:12:15.381 --> 01:12:15.861
Right.

1204
01:12:15.901 --> 01:12:17.302
Which means it needs to be solved.

1205
01:12:17.442 --> 01:12:17.622
Yes.

1206
01:12:17.722 --> 01:12:17.982
Okay.

1207
01:12:18.323 --> 01:12:21.905
And in biology, like half the battle is solving the structure.

1208
01:12:22.045 --> 01:12:22.205
Right.

1209
01:12:22.325 --> 01:12:24.166
Because all of biology is lock and key.

1210
01:12:24.286 --> 01:12:24.486
Right.

1211
01:12:24.786 --> 01:12:26.207
You solve the structure of the thing.

1212
01:12:26.247 --> 01:12:33.131
Then you can figure out what other structures can latch on and like mess with it and like inactivate it and things like that.

1213
01:12:33.211 --> 01:12:33.391
Right.

1214
01:12:33.691 --> 01:12:37.073
So the first first battle is to actually solve the structure.

1215
01:12:37.734 --> 01:12:40.315
Solving the structure for this thing was really hard.

1216
01:12:40.535 --> 01:12:40.736
Okay.

1217
01:12:40.816 --> 01:12:41.056
Okay.

1218
01:12:41.456 --> 01:12:44.638
Because it's this envelope protein.

1219
01:12:44.658 --> 01:12:44.918
Okay.

1220
01:12:45.918 --> 01:12:51.281
It used to be, back in its heyday, before it snuck in and now it's just like an infiltrator.

1221
01:12:52.141 --> 01:12:56.383
Back in its heyday, it was the envelope protein for a retrovirus.

1222
01:12:56.703 --> 01:13:03.286
And these envelope proteins are incredibly finicky by design.

1223
01:13:04.290 --> 01:13:08.494
You want your envelope protein to be finicky because imagine you're a virus, right?

1224
01:13:08.774 --> 01:13:11.956
And you're getting to the outside of a cell that you want to infiltrate.

1225
01:13:12.257 --> 01:13:15.519
You want the protein to first be able to infiltrate through the membrane.

1226
01:13:15.960 --> 01:13:19.723
And then when it's in the membrane, change its conformation to open up.

1227
01:13:20.383 --> 01:13:20.583
Right?

1228
01:13:20.883 --> 01:13:24.364
So you want this thing to have this sort of meta stability is what it's called.

1229
01:13:24.384 --> 01:13:25.724
Where you've got two states.

1230
01:13:26.124 --> 01:13:29.465
There's this pre-fusion state and then a post-fusion state.

1231
01:13:29.625 --> 01:13:29.785
Right?

1232
01:13:30.005 --> 01:13:31.545
Pre-fusion meaning like how do I get in?

1233
01:13:31.805 --> 01:13:33.165
Post-fusion, how do I get in?

1234
01:13:33.285 --> 01:13:37.906
You have the Trojan horse go to the gate in one state and then when it's inside the gate it's open.

1235
01:13:37.926 --> 01:13:38.466
You got to open up.

1236
01:13:38.707 --> 01:13:38.927
Yeah.

1237
01:13:39.227 --> 01:13:39.907
Right?

1238
01:13:40.067 --> 01:13:41.787
So you want this meta stability.

1239
01:13:41.927 --> 01:13:42.067
Yes.

1240
01:13:42.327 --> 01:13:46.048
Now that becomes a real problem when you're trying to get structure.

1241
01:13:46.368 --> 01:13:46.588
Yes.

1242
01:13:46.888 --> 01:13:47.108
Right.

1243
01:13:47.248 --> 01:13:53.413
Because you can imagine you're trying to image this thing, but this this thing is in two different conformations.

1244
01:13:53.613 --> 01:13:55.054
So like you don't know what you're looking at.

1245
01:13:55.134 --> 01:13:55.335
Yes.

1246
01:13:55.555 --> 01:13:56.696
And it's just annoying.

1247
01:13:56.776 --> 01:13:56.976
Yes.

1248
01:13:57.056 --> 01:13:57.256
Right.

1249
01:13:57.476 --> 01:14:00.899
So that's that's the big thing that this group solved.

1250
01:14:01.179 --> 01:14:07.424
OK. What they did, the first thing they did was they took the protein itself and they put in protein.

1251
01:14:09.424 --> 01:14:18.347
they tweaked the protein in a certain way that they could, like they created these things called disulfide bridges, which are basically like staples on the protein.

1252
01:14:18.367 --> 01:14:26.689
What you do is you create, you put in cysteine, which is this particular amino acid that has sulfur atoms, and those sulfur atoms can find each other and sort of bond.

1253
01:14:27.089 --> 01:14:30.410
So you create a more stable protein, which is a protein,

1254
01:14:31.808 --> 01:14:36.230
protein that is still very much like the original protein.

1255
01:14:36.350 --> 01:14:41.831
It still preserves a lot of the structure, but it prevents the protein from doing this finicky stuff.

1256
01:14:41.871 --> 01:14:42.051
Yes.

1257
01:14:42.332 --> 01:14:50.874
And then what they do was they use a groundbreaking technology, the one the 2017 Nobel Prize in Chemistry, called cryo-electron microscopy.

1258
01:14:51.075 --> 01:14:51.235
Okay.

1259
01:14:51.375 --> 01:14:51.615
Okay?

1260
01:14:52.755 --> 01:14:54.357
It used to be called blobology.

1261
01:14:54.437 --> 01:14:55.598
Blobology.

1262
01:14:55.738 --> 01:14:59.981
Yeah, from all the detractors, all the people who would make fun of these cryo-EM people.

1263
01:15:00.361 --> 01:15:07.067
It was called blobology because back in the day, what you do is you get a protein and you like image it with electron microscope.

1264
01:15:07.627 --> 01:15:10.109
But because of this problem of like, you know, the...

1265
01:15:11.350 --> 01:15:15.994
the proteins move around and then they're not stable and all this other stuff.

1266
01:15:16.334 --> 01:15:21.298
And you would get blobs, giant blobs, and you'd be like, it kind of looks like this.

1267
01:15:21.859 --> 01:15:28.924
So all of the X-ray crystallographers and the NMR spectroscopy people, they're like, look at these blobologists, right?

1268
01:15:29.144 --> 01:15:33.508
Because with X-ray crystallography and with NMR, I can get down to atomic resolution.

1269
01:15:33.528 --> 01:15:38.572
I can get to two or three angstroms of resolution where I'm like, that atom is there and it's there.

1270
01:15:38.652 --> 01:15:38.832
Yeah.

1271
01:15:39.052 --> 01:15:44.016
The problem with X-ray crystallography and NMR is it's really good for soluble proteins.

1272
01:15:44.216 --> 01:15:48.319
So proteins that are found within the cell doing their thing.

1273
01:15:48.599 --> 01:15:52.022
But for these transmembrane proteins, it's very hard to crystallize.

1274
01:15:52.242 --> 01:15:54.124
It's very hard to make them behave properly.

1275
01:15:54.884 --> 01:15:57.807
Cryo-Yam was a technique that probably could have been used for that.

1276
01:15:58.347 --> 01:16:00.168
Along came the computing revolution.

1277
01:16:00.188 --> 01:16:00.629
Mm-hmm.

1278
01:16:01.198 --> 01:16:05.879
And now all of a sudden what you could do is you could take thousands of images in cryo-EM.

1279
01:16:06.240 --> 01:16:10.341
What you do is you have a bunch of proteins that are sort of frozen in this state.

1280
01:16:10.501 --> 01:16:11.221
Yes.

1281
01:16:11.741 --> 01:16:14.662
Using this thing called vitreous freezing.

1282
01:16:14.862 --> 01:16:15.022
Yes.

1283
01:16:15.282 --> 01:16:19.364
Where you're not like freezing it with ice because if you freeze it with ice, the protein is going to like misfold, right?

1284
01:16:19.384 --> 01:16:19.784
Maybe, yeah.

1285
01:16:19.824 --> 01:16:25.366
So you do this like weird like sort of vitreous freezing where the protein retains its structure.

1286
01:16:25.686 --> 01:16:29.407
You image it a bunch of times and then you feed that thing to a computer.

1287
01:16:29.467 --> 01:16:29.947
Yeah, yeah.

1288
01:16:30.087 --> 01:16:32.188
What you're getting is 2D shadows, right?

1289
01:16:32.328 --> 01:16:34.149
The proteins are in different orientations.

1290
01:16:34.449 --> 01:16:35.849
You get a bunch of 2D shadows.

1291
01:16:36.170 --> 01:16:44.233
But a computer now can take all of those 2D shadows and say, what is the 3D thing that would create all of these 2D shadows?

1292
01:16:44.313 --> 01:16:46.514
In 2012 was the big one when it came out.

1293
01:16:46.874 --> 01:16:52.396
And there was finally a protein that cryo-YAM had solved that no one else could solve with any other technique.

1294
01:16:52.496 --> 01:16:54.316
And they solved it to three angstrom precision.

1295
01:16:54.676 --> 01:16:56.877
And everyone was like, oh, it's here.

1296
01:16:56.977 --> 01:16:59.217
And then in 2017, they won the Nobel Prize in chemistry.

1297
01:16:59.557 --> 01:17:04.299
It was three guys, Joaquin Frank, Yaques Duboche, and Richard Henderson.

1298
01:17:05.699 --> 01:17:09.080
So blobology now became like a Nobel Prize winning technology.

1299
01:17:09.340 --> 01:17:11.421
Ain't no bunch of blobs now, huh?

1300
01:17:11.441 --> 01:17:11.801
Yeah, exactly.

1301
01:17:11.821 --> 01:17:13.741
Where's your Nobel, huh?

1302
01:17:13.821 --> 01:17:17.502
I think they were very vindicated when they actually got it right.

1303
01:17:17.802 --> 01:17:24.484
So these guys, they actually used the cryo-electron microscopy to get the data from this.

1304
01:17:25.005 --> 01:17:29.226
After having messed with the protein to create these staples and make it sort of…

1305
01:17:30.286 --> 01:17:31.287
stable, right?

1306
01:17:31.867 --> 01:17:36.011
And this was a very big deal because now you've got the 3D structure for this thing.

1307
01:17:36.031 --> 01:17:39.854
You can compare it with other retroviruses that have been solved like HIV, like SIV.

1308
01:17:39.874 --> 01:17:40.414
Yes.

1309
01:17:40.915 --> 01:17:54.746
And this one is very different because the protein that came out is like taller and skinnier than the HIV protein, which is interesting in its own right because now we've got this difference.

1310
01:17:54.806 --> 01:17:58.309
We can now start creating antibodies to actually target this thing.

1311
01:17:58.329 --> 01:17:59.570
This lock and key thing you were just talking about.

1312
01:17:59.630 --> 01:18:00.030
Exactly.

1313
01:18:00.070 --> 01:18:00.751
And that's what they did.

1314
01:18:00.771 --> 01:18:01.311
They weren't done.

1315
01:18:01.351 --> 01:18:04.932
They created monoclonal antibodies to actually target this thing.

1316
01:18:05.313 --> 01:18:13.896
And what they could do now is you can actually have diagnostics where, you know, in autoimmune disorders, for example, right, you have these neutrophils, which are these kinds of immune cells.

1317
01:18:14.276 --> 01:18:21.600
And what they do is they they're expressed in these cells and then in an autoimmune disorder,

1318
01:18:22.741 --> 01:18:26.545
now your immune cells think that, oh, this is a bad cell.

1319
01:18:26.645 --> 01:18:28.327
I'm going to attack my own cell.

1320
01:18:28.387 --> 01:18:28.908
Right?

1321
01:18:29.388 --> 01:18:35.135
So maybe now that we know the structure, we can actually like prevent that from happening.

1322
01:18:35.576 --> 01:18:35.776
Right?

1323
01:18:35.856 --> 01:18:38.660
We can train the immune system or maybe cloud...

1324
01:18:39.620 --> 01:18:43.763
these antibodies in such a way that the immune system no longer attacks its own.

1325
01:18:44.403 --> 01:18:53.148
On the other hand, we can also have ways to early diagnose certain cancers because cancer cells express this protein specifically, right?

1326
01:18:53.469 --> 01:19:02.054
So if we get like a blood sample from somebody, we can have an early diagnostic where you have these monoclonal antibodies that'll go and...

1327
01:19:02.634 --> 01:19:27.362
actually attached to these cancer cells and you'll have a marker now be like oh i see this herb k protein yes that shouldn't be expressed in a healthy cell the epigenetics should totally silence it and it's not but the bouncer is not working right right so now now i can i can i can start therapies earlier they're still in the club getting too drunk they haven't been cut off yeah so let's go ahead and let's go ahead and and try to figure out how to get these guys out

1328
01:19:27.602 --> 01:19:28.642
That's really...

1329
01:19:28.862 --> 01:19:29.682
Right.

1330
01:19:29.822 --> 01:19:31.183
It's the first time that we've done this.

1331
01:19:31.643 --> 01:19:37.544
And the idea is now that we've done it once by confirming that this methodology is a means by which you can do so.

1332
01:19:37.604 --> 01:19:37.804
Yeah.

1333
01:19:38.004 --> 01:19:39.784
This can likely be extended to other ones.

1334
01:19:39.804 --> 01:19:41.465
This is the most recent and all these other things.

1335
01:19:41.485 --> 01:19:41.605
Yeah.

1336
01:19:41.625 --> 01:19:43.305
And this is the one that's like most prevalent.

1337
01:19:43.605 --> 01:19:53.247
But the idea of like sort of changing a protein shape using these editing mechanisms where now I can have this like staple disulfide bridge that sort of...

1338
01:19:53.747 --> 01:19:56.870
Stabilizes it such that I can now do cryo-EM.

1339
01:19:57.070 --> 01:19:57.671
Yes.

1340
01:19:57.871 --> 01:19:59.252
This is a pretty interesting technique.

1341
01:19:59.272 --> 01:19:59.432
Right.

1342
01:19:59.572 --> 01:20:03.055
The same group actually had done it for other proteins.

1343
01:20:03.075 --> 01:20:04.897
And then they tackled this one because this one's hard.

1344
01:20:04.997 --> 01:20:05.397
Yeah.

1345
01:20:05.417 --> 01:20:05.577
Right.

1346
01:20:05.618 --> 01:20:06.338
This one's like.

1347
01:20:06.779 --> 01:20:07.819
If you can do it with this one.

1348
01:20:07.859 --> 01:20:08.200
Yeah.

1349
01:20:08.340 --> 01:20:09.921
Then it's like, okay, now it's like.

1350
01:20:09.961 --> 01:20:10.462
Now we can.

1351
01:20:12.323 --> 01:20:12.824
Fast.

1352
01:20:13.084 --> 01:20:13.404
Yeah.

1353
01:20:13.604 --> 01:20:14.225
It's a cool one.

1354
01:20:14.545 --> 01:20:19.289
I always love our microbiology immunology stories because.

1355
01:20:20.090 --> 01:20:29.214
People don't understand how well we can look at and analyze and understand things that are really, really, really small.

1356
01:20:29.234 --> 01:20:38.078
Yeah, like we know exactly where to put in the cysteine so that the sulfur staples to another sulfur, but it doesn't really change the rest of the structure.

1357
01:20:38.418 --> 01:20:42.840
That's like to have like this thousands of atoms worth of...

1358
01:20:43.600 --> 01:21:00.121
protein and know exactly where to put this thing so that i'm just going to clip this part yes because that's the part that's like getting messed up i would like our next department of health and human services secretary to be able to describe in detail any number of these concepts

1359
01:21:00.421 --> 01:21:08.584
Yeah, because that is usually a good prerequisite to say you're going to actually understand the import of medicine for public health.

1360
01:21:09.085 --> 01:21:09.885
Exactly.

1361
01:21:10.365 --> 01:21:12.386
But but that's a good I like that one.

1362
01:21:12.506 --> 01:21:12.906
I like that.

1363
01:21:12.966 --> 01:21:15.747
And we've touched on CRISPR a lot on this podcast.

1364
01:21:16.087 --> 01:21:16.868
We've touched on a variety.

1365
01:21:16.928 --> 01:21:17.828
And it's again, it's interesting.

1366
01:21:17.848 --> 01:21:27.072
Again, all of these innovations are happening simultaneously in these cross domain areas that have not only independent value, but also like collective value.

1367
01:21:27.092 --> 01:21:27.192
Yeah.

1368
01:21:27.672 --> 01:21:32.914
As we move into what we're seeing is going to be this like biotech revolution.

1369
01:21:32.974 --> 01:21:33.234
It is.

1370
01:21:33.454 --> 01:21:34.055
It really is.

1371
01:21:34.095 --> 01:21:34.455
We're here.

1372
01:21:34.655 --> 01:21:43.458
For AI applying the tools, the fundamental hardware tools, the creativity of how we're problem solving.

1373
01:21:45.119 --> 01:21:45.479
Yeah.

1374
01:21:45.519 --> 01:21:48.060
We're going to go to our last story, which is.

1375
01:21:48.660 --> 01:21:52.904
A lot of times I do from really, really small now to like – Now we're getting really big.

1376
01:21:52.924 --> 01:21:55.126
Now we're getting really big in space.

1377
01:21:55.227 --> 01:21:56.368
California again.

1378
01:21:56.468 --> 01:21:57.629
California forever goodbye.

1379
01:21:57.989 --> 01:22:04.195
Headline, brightest fast radio burst ever detected could help solve an enduring cosmic mystery.

1380
01:22:04.275 --> 01:22:07.459
I love how people with headlines make it so unclear what you're talking about.

1381
01:22:07.479 --> 01:22:07.679
Yeah.

1382
01:22:08.119 --> 01:22:08.499
Summary.

1383
01:22:08.539 --> 01:22:25.769
Researchers at University of California, Santa Cruz, we're going to Northern California, use newly developed chime out trigger telescopes and deep space imaging to challenge long-held assumptions about what causes these mysterious cosmic signals.

1384
01:22:25.889 --> 01:22:31.592
I think we talked about one of these, if I'm not misremembering, the wow signal was one of these...

1385
01:22:32.212 --> 01:22:32.493
Yes.

1386
01:22:32.713 --> 01:22:33.914
Like radio burst.

1387
01:22:33.974 --> 01:22:34.955
It was a radio burst.

1388
01:22:35.095 --> 01:22:35.355
Right.

1389
01:22:35.375 --> 01:22:35.596
Yeah.

1390
01:22:35.616 --> 01:22:38.018
We initially saw and everyone freaked out and was like aliens.

1391
01:22:38.218 --> 01:22:38.418
Yeah.

1392
01:22:39.179 --> 01:22:39.700
It could have been.

1393
01:22:39.820 --> 01:22:44.604
And that's not the we haven't characterized that as a fast radio burst.

1394
01:22:44.664 --> 01:22:44.865
Fair.

1395
01:22:45.065 --> 01:22:46.947
Because we don't really know what the hell that was.

1396
01:22:46.967 --> 01:22:47.407
That's fair.

1397
01:22:47.447 --> 01:22:48.548
That's what I was like.

1398
01:22:48.588 --> 01:22:49.029
What was that?

1399
01:22:49.049 --> 01:22:49.970
Yeah.

1400
01:22:50.010 --> 01:22:50.470
What was that?

1401
01:22:50.550 --> 01:22:50.790
Fair.

1402
01:22:51.091 --> 01:22:51.651
No, that's fair.

1403
01:22:51.811 --> 01:22:52.212
That's fair.

1404
01:22:52.232 --> 01:22:52.452
Fair.

1405
01:22:52.872 --> 01:22:56.374
But it seems like we're now using, again, some of the tools we have.

1406
01:22:56.454 --> 01:22:58.255
Also, this was in the Astrophysical Journal letters.

1407
01:22:58.355 --> 01:22:58.876
Yes.

1408
01:22:58.896 --> 01:23:00.897
In terms of the publication, CNN has covered this.

1409
01:23:01.938 --> 01:23:03.919
UCSC also covered it in their own blog.

1410
01:23:03.979 --> 01:23:04.659
Yeah.

1411
01:23:04.839 --> 01:23:08.942
It's a collaboration between UCSC, some Canadian institutions, Northwestern.

1412
01:23:10.202 --> 01:23:11.563
It was a pretty big collaboration.

1413
01:23:11.663 --> 01:23:13.384
Again, because we need it.

1414
01:23:14.125 --> 01:23:14.825
Okay.

1415
01:23:14.845 --> 01:23:17.867
So what are, because I mischaracterized it.

1416
01:23:18.187 --> 01:23:19.770
What is this fast rate of burst idea?

1417
01:23:19.890 --> 01:23:22.654
What did they do?

1418
01:23:24.056 --> 01:23:25.939
Why is detecting a bright one relevant?

1419
01:23:26.099 --> 01:23:30.025
Yeah, this one's incredibly bright.

1420
01:23:30.205 --> 01:23:32.769
This thing within a fraction of a second output...

1421
01:23:33.669 --> 01:23:36.251
as much energy as the sun does for like four days.

1422
01:23:36.611 --> 01:23:36.972
Okay.

1423
01:23:37.232 --> 01:23:38.713
But this isn't even the brightest.

1424
01:23:38.953 --> 01:23:39.173
Okay.

1425
01:23:39.413 --> 01:23:42.215
This is the brightest we've had because it's the closest one to us.

1426
01:23:42.715 --> 01:23:46.938
But fast radio bursts, the first one was ever discovered in 2007.

1427
01:23:47.879 --> 01:23:49.080
It was called a Lorimer burst.

1428
01:23:49.500 --> 01:23:56.925
It was actually taken from archival data from the Parks Observatory, which is a giant 64-meter radio telescope in Australia.

1429
01:23:57.445 --> 01:24:01.308
It's been called the greatest scientific instrument...

1430
01:24:02.148 --> 01:24:03.549
that Australia has ever produced.

1431
01:24:03.569 --> 01:24:03.989
Okay.

1432
01:24:04.830 --> 01:24:07.532
The greatest scientific instrument that Australia has ever produced.

1433
01:24:08.152 --> 01:24:08.312
Okay.

1434
01:24:08.452 --> 01:24:10.333
And it's...

1435
01:24:10.734 --> 01:24:17.418
The same dish actually was used during the Apollo 11 broadcast to, like, broadcast to the TVs around the world.

1436
01:24:17.518 --> 01:24:20.380
Because NASA had, like, a thing at Goldstone...

1437
01:24:20.520 --> 01:24:20.700
Yes.

1438
01:24:21.040 --> 01:24:22.041
...that they were using.

1439
01:24:22.161 --> 01:24:22.321
Yes.

1440
01:24:22.441 --> 01:24:23.882
But this one was way bigger.

1441
01:24:24.042 --> 01:24:24.202
Yeah.

1442
01:24:24.262 --> 01:24:27.905
And the TV signal that they were getting from this one was just better than the rest.

1443
01:24:28.225 --> 01:24:30.887
And so NASA just used it for the whole two and a half hours...

1444
01:24:31.747 --> 01:24:32.367
Which is kind of cool.

1445
01:24:32.387 --> 01:24:35.168
Which is why the Five Eyes Ally, you know, structure is awful.

1446
01:24:35.188 --> 01:24:35.789
Yeah, yeah, you know.

1447
01:24:35.969 --> 01:24:37.209
Yeah, Australia, yeah.

1448
01:24:37.629 --> 01:24:39.070
They got our back when we need it.

1449
01:24:39.730 --> 01:24:41.451
So, you know, they were on the right side.

1450
01:24:41.471 --> 01:24:42.952
They were facing the moon, so that was nice.

1451
01:24:43.092 --> 01:24:43.672
Yep.

1452
01:24:44.112 --> 01:24:49.894
So this burst was discovered by Duncan Larimar from West Virginia University.

1453
01:24:50.334 --> 01:24:55.997
Basically, what he did was he assigned an undergrad to look at archival data from the Parks Observatory.

1454
01:24:56.317 --> 01:24:56.557
Okay.

1455
01:24:57.037 --> 01:25:00.179
And the Parks Observatory at the time had just released a bunch of data.

1456
01:25:00.459 --> 01:25:02.340
It was a radio telescope that was just beaming up.

1457
01:25:03.080 --> 01:25:05.261
And this undergrad would come in every week.

1458
01:25:05.341 --> 01:25:12.405
And despite all of his classes and stuff, he would look at a bunch of data by hand and try to characterize it and then try to find something interesting.

1459
01:25:12.465 --> 01:25:14.186
And he found something very interesting.

1460
01:25:14.606 --> 01:25:17.788
There was a brief millisecond duration, five milliseconds.

1461
01:25:18.728 --> 01:25:20.650
but it was ridiculously bright.

1462
01:25:21.171 --> 01:25:26.917
On top of that, it was not only bright, it had this very good characterization called dispersion.

1463
01:25:27.398 --> 01:25:27.638
Okay.

1464
01:25:27.938 --> 01:25:29.740
Dispersion is the idea that

1465
01:25:31.369 --> 01:25:39.112
When light moves through a medium, through stuff, different frequencies are going to propagate at different speeds.

1466
01:25:39.692 --> 01:25:42.133
This is fundamentally why we get a rainbow, okay?

1467
01:25:42.613 --> 01:25:49.916
Violet light actually propagates at a slower speed in water than red light does, okay?

1468
01:25:50.276 --> 01:25:54.338
And because of that, when light goes into a water bubble,

1469
01:25:54.878 --> 01:25:57.601
the violet light is going to bend differently than red light.

1470
01:25:57.881 --> 01:26:00.003
And that's what's going to cause the rainbow to come out.

1471
01:26:00.403 --> 01:26:00.643
Okay.

1472
01:26:00.863 --> 01:26:02.965
And then it gets like sort of focused and we get a rainbow.

1473
01:26:03.005 --> 01:26:04.306
But the prism has the same idea.

1474
01:26:04.646 --> 01:26:07.369
When it goes through glass, violet light goes slower.

1475
01:26:07.409 --> 01:26:08.630
And so it's going to bend differently.

1476
01:26:09.210 --> 01:26:15.195
Dispersion is this idea that there's a frequency relation to how fast light moves through medium.

1477
01:26:15.696 --> 01:26:16.737
And in interstellar,

1478
01:26:18.532 --> 01:26:28.755
background right in the interstellar medium of these ionized particles right where it's like atoms without electrons electrons moving around freely You're gonna have that opposite effect.

1479
01:26:28.915 --> 01:26:31.076
You're gonna have slower slower light.

1480
01:26:31.156 --> 01:26:39.439
So slower frequencies moving slower than the faster frequencies just has to do with the different physics that goes on between a glass and a plasma but at the end of the day

1481
01:26:40.443 --> 01:26:49.025
What that means is if there's a transient burst of light, the faster stuff is going to arrive first and the slower frequencies are going to arrive later.

1482
01:26:49.385 --> 01:26:51.406
And that's what's called a dispersion measure.

1483
01:26:51.966 --> 01:26:52.186
Okay.

1484
01:26:53.386 --> 01:27:03.949
This thing that they discovered in the Parkes Observatory data, which is now called the Lorimer burst, had this very specific dispersion measure, which showed that it was of extra galactic origin.

1485
01:27:04.129 --> 01:27:04.349
Okay.

1486
01:27:04.409 --> 01:27:05.789
There was so much dispersion.

1487
01:27:06.009 --> 01:27:06.209
Yeah.

1488
01:27:06.309 --> 01:27:10.030
That meant that there was a lot of stuff that it had to get through.

1489
01:27:10.090 --> 01:27:10.350
Yeah.

1490
01:27:11.190 --> 01:27:14.692
To actually create that delay between the fast and the slow.

1491
01:27:15.032 --> 01:27:24.735
And then when they back calculated how bright this thing must have been to be that far away, they were like, oh, this thing was incredibly bright.

1492
01:27:25.115 --> 01:27:26.476
This was back in 2007.

1493
01:27:26.536 --> 01:27:28.156
Since then, we've discovered multiple.

1494
01:27:28.517 --> 01:27:31.738
There's been some that we've discovered where for that brief delay,

1495
01:27:32.598 --> 01:27:38.424
millisecond duration, this fast radio burst was brighter than the entire galaxy that it came from.

1496
01:27:38.444 --> 01:27:38.965
Wow.

1497
01:27:39.105 --> 01:27:39.385
Okay.

1498
01:27:39.565 --> 01:27:39.846
Okay.

1499
01:27:40.066 --> 01:27:40.366
Yeah.

1500
01:27:40.406 --> 01:27:48.594
So it's been, it's been, it's been quite an insane sort of last few decades of trying to find these fast radio bursts.

1501
01:27:48.675 --> 01:27:48.895
Right.

1502
01:27:49.255 --> 01:27:51.018
The problem is they're really short lived.

1503
01:27:51.378 --> 01:27:51.559
Right.

1504
01:27:51.819 --> 01:27:52.019
Okay.

1505
01:27:52.200 --> 01:27:53.001
Five milliseconds.

1506
01:27:53.101 --> 01:27:55.365
Unlike 3i Atlas, which we're watching for months.

1507
01:27:55.385 --> 01:27:55.645
Yeah.

1508
01:27:55.665 --> 01:27:55.885
Yeah.

1509
01:27:55.905 --> 01:27:58.469
Which we know for months and we can like coordinate this thing.

1510
01:27:58.730 --> 01:28:00.492
You know, it's just, it's there and it's gone.

1511
01:28:00.592 --> 01:28:01.153
And it happens.

1512
01:28:01.274 --> 01:28:02.415
And it just so happens.

1513
01:28:02.475 --> 01:28:02.676
Right.

1514
01:28:03.116 --> 01:28:03.277
So.

1515
01:28:04.878 --> 01:28:06.919
It's been kind of challenging to actually find them.

1516
01:28:07.359 --> 01:28:08.279
It's been kind of rare.

1517
01:28:08.859 --> 01:28:20.163
Along came this detector called Chime in Canada, which is the... Let me just get the thing right because I don't want any Canadians to get mad at me.

1518
01:28:20.923 --> 01:28:25.948
It's called the Canadian Hydrogen Intensity Mapping Experiment, CHIME.

1519
01:28:26.108 --> 01:28:29.812
It's a massive radio telescope that doesn't look like most radio telescopes.

1520
01:28:30.092 --> 01:28:36.159
So, you know, most radio telescopes, you've got a circular dish with a receiver and the light goes that way.

1521
01:28:37.320 --> 01:28:39.002
This thing is a bunch of half pipes.

1522
01:28:39.790 --> 01:28:41.251
that are just laid out on the ground.

1523
01:28:41.291 --> 01:28:41.872
Oh, interesting.

1524
01:28:41.972 --> 01:28:43.594
Giant half pipes, okay?

1525
01:28:43.654 --> 01:28:56.706
100 meters that are laid out on the ground, four of them in a valley in British Columbia, as south as you can possibly get because what they're using, they're using the Earth's rotation to basically scan the sky in this radio frequency.

1526
01:28:56.806 --> 01:28:56.966
Yep.

1527
01:28:57.206 --> 01:28:59.267
And, you know, Canada is pretty far north.

1528
01:28:59.287 --> 01:29:01.247
So they want to get as south as possible.

1529
01:29:01.267 --> 01:29:03.248
So it's like right across the border from America.

1530
01:29:04.908 --> 01:29:05.848
And it's this giant thing.

1531
01:29:06.148 --> 01:29:08.569
And it's made such that it's really sensitive.

1532
01:29:08.949 --> 01:29:11.750
It's got all these capabilities for as soon as a signal comes in.

1533
01:29:12.030 --> 01:29:16.271
It's kind of like the Vera Rubin in terms of they've got online capabilities to analyze the signal.

1534
01:29:16.811 --> 01:29:19.612
And it's also got a really good giant field of view.

1535
01:29:19.812 --> 01:29:23.313
So it's capturing a big part of the sky as it sort of goes around the earth.

1536
01:29:23.853 --> 01:29:30.898
And very recently, so Chime has found thousands of fast radio bursts since 2018.

1537
01:29:31.078 --> 01:29:31.298
Okay.

1538
01:29:31.438 --> 01:29:31.698
Okay.

1539
01:29:31.818 --> 01:29:32.539
It's found a bunch.

1540
01:29:32.919 --> 01:29:39.723
This one is very, very nice because usually when Chime finds fast radio bursts, it can't localize where it came from.

1541
01:29:39.843 --> 01:29:40.043
Right.

1542
01:29:40.123 --> 01:29:43.125
Because it's got this giant field of view, it just sees sort of a signal.

1543
01:29:43.265 --> 01:29:43.405
Yes.

1544
01:29:43.526 --> 01:29:45.627
And it's like, okay, I saw a fast radio burst.

1545
01:29:45.667 --> 01:29:45.747
Yes.

1546
01:29:45.767 --> 01:29:46.587
Don't know where it came from.

1547
01:29:47.608 --> 01:29:54.894
So what they did was, in collaboration with America, they made more chime stations.

1548
01:29:54.934 --> 01:29:56.215
There's one in California.

1549
01:29:56.235 --> 01:29:57.677
There's one in West Virginia.

1550
01:29:57.737 --> 01:29:59.138
There's one also in Canada.

1551
01:29:59.558 --> 01:30:02.501
And now what you can do is triangulate.

1552
01:30:02.701 --> 01:30:03.421
My favorite.

1553
01:30:03.461 --> 01:30:03.682
Right?

1554
01:30:03.862 --> 01:30:08.266
What you can do is, if you get a fast radio burst, you're going to get that signal all over the place.

1555
01:30:08.286 --> 01:30:12.809
And you can triangulate where it came from and actually resolve what part of the sky it came from.

1556
01:30:12.829 --> 01:30:15.912
Because you have multiple receiving points.

1557
01:30:16.152 --> 01:30:17.713
at different places on the planet.

1558
01:30:17.793 --> 01:30:18.214
Exactly.

1559
01:30:18.234 --> 01:30:19.455
And so the angle...

1560
01:30:19.835 --> 01:30:24.179
It's going to be different, and the timing is going to be different, and you can back calculate and see where it came from.

1561
01:30:24.479 --> 01:30:30.784
And what they could do is they could resolve it to a really tiny area in the sky, right around the Big Dipper.

1562
01:30:31.625 --> 01:30:36.089
The equivalent would be to resolve a quarter from 100 kilometers away.

1563
01:30:36.589 --> 01:30:38.651
Very tiny, very tiny, tiny spot.

1564
01:30:39.031 --> 01:30:39.211
Right?

1565
01:30:39.531 --> 01:30:41.613
And that gives you the advantage.

1566
01:30:41.893 --> 01:30:47.737
Because now, if you know that it came from there, you've got all these other telescopes on standby that can point right there.

1567
01:30:48.378 --> 01:30:48.618
Right?

1568
01:30:49.078 --> 01:30:50.059
That's exactly what they did.

1569
01:30:50.099 --> 01:30:51.520
They got the James Webb.

1570
01:30:51.560 --> 01:30:52.501
They called in the big guys.

1571
01:30:52.681 --> 01:30:53.641
They called in the big guns.

1572
01:30:53.661 --> 01:30:56.602
They called in the big guns because this one was really bright.

1573
01:30:56.722 --> 01:30:57.202
Okay.

1574
01:30:57.242 --> 01:31:02.143
In terms of this is the brightest fast radio burst that we've ever seen because of its proximity to Earth.

1575
01:31:02.483 --> 01:31:07.385
Not in terms of intrinsic brightness, but because of its proximity to Earth, now it's really exciting, right?

1576
01:31:07.405 --> 01:31:12.406
Because now we can point optical telescopes to it and we can actually see what part of this galaxy that it came from.

1577
01:31:12.806 --> 01:31:17.107
It came from NGC 4141, which is this barred spiral galaxy.

1578
01:31:17.767 --> 01:31:24.512
And what it does is, and you can localize the neighborhood of that galaxy.

1579
01:31:24.552 --> 01:31:28.515
Not just that it came from that galaxy, but it came from this corner of the galaxy.

1580
01:31:29.035 --> 01:31:29.996
Okay.

1581
01:31:30.536 --> 01:31:33.498
And now you can point James Webb Space Telescope to it.

1582
01:31:33.839 --> 01:31:35.320
And James Webb was pointed to it.

1583
01:31:35.680 --> 01:31:37.621
It found a faint signal in the infrared.

1584
01:31:37.922 --> 01:31:38.402
Okay.

1585
01:31:38.962 --> 01:31:40.143
Right around that same spot.

1586
01:31:40.163 --> 01:31:40.323
Okay.

1587
01:31:40.343 --> 01:31:40.563
Okay.

1588
01:31:41.504 --> 01:31:45.671
And that probably came from some massive stars, like a red giant or something like that.

1589
01:31:45.811 --> 01:31:51.700
Now, it doesn't quite fit that a red giant or a big star would create a fast radio burst.

1590
01:31:51.760 --> 01:31:52.060
Okay.

1591
01:31:52.140 --> 01:31:52.421
Okay?

1592
01:31:52.881 --> 01:31:55.766
Because those things are, like, slow.

1593
01:31:56.006 --> 01:31:56.206
Right.

1594
01:31:56.647 --> 01:31:59.370
They're not five millisecond timescales.

1595
01:31:59.630 --> 01:31:59.810
Right.

1596
01:32:00.151 --> 01:32:04.575
Five millisecond timescales are the stuff of black holes, neutron stars.

1597
01:32:05.115 --> 01:32:07.057
They revolve really fast.

1598
01:32:07.358 --> 01:32:11.081
So there's timescales that sort of correspond to astronomical objects.

1599
01:32:11.181 --> 01:32:11.402
Yes.

1600
01:32:11.482 --> 01:32:12.543
Right?

1601
01:32:14.725 --> 01:32:16.587
So now it leaves us with a conundrum.

1602
01:32:16.607 --> 01:32:18.068
It's like, where did this thing come from?

1603
01:32:18.088 --> 01:32:18.348
Mm-hmm.

1604
01:32:19.104 --> 01:32:20.185
There's hypotheses.

1605
01:32:20.526 --> 01:32:21.887
There's not good data.

1606
01:32:21.987 --> 01:32:23.328
Because again, this is the first time.

1607
01:32:23.589 --> 01:32:30.856
I mean, Chime and the Chime outrigger stations only came online like this year.

1608
01:32:31.097 --> 01:32:31.377
Okay.

1609
01:32:31.497 --> 01:32:33.039
So there's the first thing that it's found.

1610
01:32:33.419 --> 01:32:37.683
And it's very exciting because it's kind of a proof of concept that it's like looking and it can find it.

1611
01:32:37.904 --> 01:32:39.065
And then we can go off with...

1612
01:32:39.385 --> 01:32:46.069
with James Webb Space Telescope, with Keck, with Gemini, and pointed at that and then actually see a follow-up observation, right?

1613
01:32:46.309 --> 01:32:50.112
So we still don't know really where this thing came from.

1614
01:32:50.412 --> 01:32:53.674
What we do know, which is very interesting, is this might be a one-off.

1615
01:32:54.875 --> 01:32:59.457
Most of the fast radio bursts that we've seen, they seem to be repeating fast radio bursts, right?

1616
01:32:59.497 --> 01:33:02.019
Where they've got this rhythmic sort of structure.

1617
01:33:02.319 --> 01:33:03.820
The rhythm can be very long.

1618
01:33:03.840 --> 01:33:05.641
It can be from days to months to years.

1619
01:33:06.041 --> 01:33:06.222
But...

1620
01:33:07.082 --> 01:33:32.325
they've got this rhythmic thing where it's like okay you've got one really bright one but then every once in a while the same source should put off like sort of smaller bursts which i think is why it matters that we have this this multi this multi-sensor data confirmation multi-sensor collaboration to be able to isolate a smaller point in the sky because if we do find these rhythmic sources we can point these things these other tools there

1621
01:33:32.825 --> 01:33:34.325
for longer periods of time.

1622
01:33:34.345 --> 01:33:34.806
Exactly.

1623
01:33:34.846 --> 01:33:36.786
To try to potentially catch what is this.

1624
01:33:37.066 --> 01:33:43.488
It reminds me of the transients conversation we had the other day where it's like there's this rhythmic pattern of the planet, the exoplanet going in front of a star.

1625
01:33:43.528 --> 01:33:43.768
Yeah.

1626
01:33:43.848 --> 01:33:47.149
And so you have to monitor it for certain periods of time to understand that it's not a one-off.

1627
01:33:47.229 --> 01:33:48.329
It's not a one-off, yeah.

1628
01:33:48.489 --> 01:33:54.611
And so what's interesting here is like what we're potentially seeing with this specific really, really bright one may have been a one-off.

1629
01:33:54.831 --> 01:33:55.591
It may have been a one-off.

1630
01:33:55.611 --> 01:33:58.972
I mean, we think it's a one-off because Chime was going around

1631
01:34:00.003 --> 01:34:05.465
That same spot in the sky for the past seven years, it's had hundreds of hours in that region of the sky.

1632
01:34:05.485 --> 01:34:05.946
It's the only one.

1633
01:34:06.006 --> 01:34:08.006
So we've only seen it once, right?

1634
01:34:08.447 --> 01:34:12.448
And it would have found even smaller radio bursts, but it hasn't.

1635
01:34:12.468 --> 01:34:12.829
It hasn't.

1636
01:34:13.149 --> 01:34:15.910
And so it's like, what could be this one-off thing?

1637
01:34:16.270 --> 01:34:22.953
It could be, they're thinking it might be a Magnetar, which is this, a Magnetar is a- Not a Pokemon.

1638
01:34:23.273 --> 01:34:25.254
No, it sounds like one though.

1639
01:34:25.594 --> 01:34:29.136
A Magnetar is sort of a newly born neutron star.

1640
01:34:30.136 --> 01:34:32.137
that's highly magnetized.

1641
01:34:32.277 --> 01:34:40.160
So it's got really powerful magnetic fields and it's in a sort of new stellar neighborhood that has magnetic material around it.

1642
01:34:40.500 --> 01:34:47.302
So it can create these fast radio bursts by sort of, maybe it's like accumulating material and then it blows up or something.

1643
01:34:48.082 --> 01:34:52.764
Maybe it's accumulating material from the red giants that the James Webb Space Telescope saw.

1644
01:34:53.204 --> 01:34:54.685
James Webb saw some signal.

1645
01:34:54.885 --> 01:34:55.506
They just don't know.

1646
01:34:55.566 --> 01:34:56.486
It's like not good enough.

1647
01:34:56.506 --> 01:34:57.047
It's not good enough.

1648
01:34:57.327 --> 01:34:59.789
And because it's transient, it only happened for a little bit.

1649
01:34:59.829 --> 01:35:01.210
And then it's gone.

1650
01:35:01.290 --> 01:35:05.352
It's not like a star that I can just keep looking at and getting data.

1651
01:35:06.653 --> 01:35:07.814
It could be a magnetar.

1652
01:35:08.134 --> 01:35:10.316
It could also be a colliding neutron star.

1653
01:35:10.716 --> 01:35:15.600
But if it was colliding, sort of revolving colliding neutron stars, we would have seen LIGO.

1654
01:35:15.740 --> 01:35:19.062
We would have seen gravitational waves, which we haven't.

1655
01:35:19.442 --> 01:35:20.923
in the hours preceding or since.

1656
01:35:21.303 --> 01:35:31.768
So maybe the collision is gonna happen later, and this is like maybe they're getting close and doing some weird dance, and then that released a giant fast radio burst.

1657
01:35:32.289 --> 01:35:35.070
But it's a new scale of astronomy, right?

1658
01:35:35.470 --> 01:35:42.216
Where before we had like only a few stations that were relying on the Earth's rotation to like maybe catch something.

1659
01:35:42.236 --> 01:35:42.316
Yes.

1660
01:35:42.556 --> 01:35:43.116
Right.

1661
01:35:43.497 --> 01:35:47.900
And now we've got a dedicated sort of instrument that is doing it on the Northern Hemisphere.

1662
01:35:48.160 --> 01:35:53.324
You can imagine we can create a global celestial radio instrument that now does this.

1663
01:35:53.464 --> 01:35:55.766
Actually, the VLA also took a look at it.

1664
01:35:55.886 --> 01:35:56.026
Yes.

1665
01:35:56.046 --> 01:36:01.911
Because now that we can locate it here, the VLA took a look at it for like several days afterwards trying to see if there's going to be a...

1666
01:36:03.632 --> 01:36:04.153
A repeat.

1667
01:36:04.393 --> 01:36:04.573
Yes.

1668
01:36:04.633 --> 01:36:05.293
Didn't see one.

1669
01:36:05.774 --> 01:36:07.495
The ZLA being the Very Large Array.

1670
01:36:07.535 --> 01:36:13.258
Yes, the Very Large Array in New Mexico, which is a giant interferometer of a bunch of radio dishes on train tracks.

1671
01:36:13.858 --> 01:36:20.262
For those who, I think the movie was Contact with Jodie Foster.

1672
01:36:20.362 --> 01:36:21.042
Yeah.

1673
01:36:21.102 --> 01:36:22.103
Written by Carl Sagan.

1674
01:36:22.143 --> 01:36:23.063
Written by Carl Sagan.

1675
01:36:23.103 --> 01:36:24.844
Was at the Very Large Array.

1676
01:36:25.585 --> 01:36:26.025
Yeah.

1677
01:36:26.065 --> 01:36:30.968
But again, this goes back to, I'm glad the outriggers were able to get out before funding out.

1678
01:36:31.128 --> 01:36:32.089
Yeah.

1679
01:36:32.269 --> 01:36:41.974
This is why things like, you know, Chime, the Vera Rubin, which have this very specific use case of SkyService that are not that are just saying we're just going to get this.

1680
01:36:42.234 --> 01:36:43.534
Yeah, this data set.

1681
01:36:44.295 --> 01:36:45.916
No, because we don't know where to look.

1682
01:36:45.936 --> 01:36:46.776
We don't know where to look.

1683
01:36:47.196 --> 01:36:48.277
We're just looking everywhere.

1684
01:36:48.477 --> 01:36:49.678
So we're going to create this map.

1685
01:36:49.698 --> 01:36:49.878
Yeah.

1686
01:36:50.238 --> 01:36:51.919
And then that'll help us know where to look.

1687
01:36:51.939 --> 01:36:52.159
Yeah.

1688
01:36:52.479 --> 01:36:58.302
And then everyone can write their proposals to then use these specific other tools to then look much more deeply.

1689
01:36:58.442 --> 01:36:58.622
Yeah.

1690
01:36:58.943 --> 01:37:05.428
And it's kind of, I mean, fast radio bursts are only 20 years old, you know, in terms of the grand scale of astronomy.

1691
01:37:05.568 --> 01:37:05.768
Right.

1692
01:37:05.829 --> 01:37:08.371
That's hundreds of years old science.

1693
01:37:08.391 --> 01:37:08.551
Yes.

1694
01:37:08.891 --> 01:37:09.091
Right.

1695
01:37:09.151 --> 01:37:11.554
Like only 20 years ago, we found the first one.

1696
01:37:11.954 --> 01:37:14.656
And these things are so ridiculously bright.

1697
01:37:15.337 --> 01:37:21.883
And it's like, what could possibly be making something that is as bright as a galaxy for a fraction of a second?

1698
01:37:21.983 --> 01:37:22.163
Right.

1699
01:37:22.563 --> 01:37:22.743
Right.

1700
01:37:22.763 --> 01:37:23.043
You know?

1701
01:37:24.124 --> 01:37:26.084
Don't get the Dyson Sphere people in on this.

1702
01:37:26.124 --> 01:37:28.465
Because they're like, oh, obviously.

1703
01:37:28.525 --> 01:37:31.286
It's an alien civilization that's harnessed the power of a planet.

1704
01:37:31.346 --> 01:37:37.808
And then when they recycle their generatron, it sends out this massive, I'm being facetious.

1705
01:37:37.868 --> 01:37:38.128
Yeah.

1706
01:37:38.308 --> 01:37:41.109
We should have an episode about Freeman Dyson, though.

1707
01:37:41.169 --> 01:37:42.689
He's one of my favorite physicists.

1708
01:37:42.869 --> 01:37:43.089
Same.

1709
01:37:43.129 --> 01:37:45.050
He's one of my favorite physicists of all time.

1710
01:37:45.070 --> 01:37:46.010
I love Dyson.

1711
01:37:47.030 --> 01:37:47.791
Very clever.

1712
01:37:48.411 --> 01:37:48.971
Interesting.

1713
01:37:49.031 --> 01:37:49.891
Really funny, too.

1714
01:37:49.931 --> 01:37:50.832
Yeah, yeah.

1715
01:37:50.872 --> 01:37:51.652
Really funny guy.

1716
01:37:52.252 --> 01:37:55.054
Brightest fast radio bursts ever detected.

1717
01:37:55.334 --> 01:37:55.554
Yeah.

1718
01:37:55.834 --> 01:37:57.836
Because of some of the tools that we have.

1719
01:37:57.936 --> 01:37:59.277
This is such a new area.

1720
01:37:59.817 --> 01:38:00.978
Such a new area.

1721
01:38:01.458 --> 01:38:03.619
And we haven't figured it all out yet.

1722
01:38:03.899 --> 01:38:04.120
No.

1723
01:38:04.200 --> 01:38:05.200
There are really interesting areas.

1724
01:38:05.240 --> 01:38:07.982
No, there are mysteries out there in the universe.

1725
01:38:08.162 --> 01:38:10.884
And we are doing good work to...

1726
01:38:11.504 --> 01:38:16.108
Try to narrow the path towards discovery on some of these things.

1727
01:38:16.128 --> 01:38:16.289
Yeah.

1728
01:38:16.609 --> 01:38:27.899
And it's like, you know, when the Vera Rubin, I was listening to the press conference for the Vera Rubin the other day, and they were talking about, they used Donald Rumsfeld's Unknown Unknown.

1729
01:38:28.099 --> 01:38:28.660
Classic.

1730
01:38:29.040 --> 01:38:37.244
But it actually made sense here because there are the known unknowns, which is like what is causing the expansion of the universe?

1731
01:38:37.965 --> 01:38:38.825
What is dark energy?

1732
01:38:38.885 --> 01:38:39.786
What is dark matter?

1733
01:38:40.206 --> 01:38:41.907
What are these fast radio bursts?

1734
01:38:42.207 --> 01:38:46.309
We know that they're there, but they're unknown in what they are.

1735
01:38:46.709 --> 01:38:51.592
And then he said, what I'm really excited about with the Vera Rubin is the unknown unknowns.

1736
01:38:51.852 --> 01:38:56.354
Because before 2007, fast radio bursts were an unknown unknown, right?

1737
01:38:56.474 --> 01:38:56.654
Yeah.

1738
01:38:56.974 --> 01:38:59.355
People couldn't even imagine that something like this would exist.

1739
01:38:59.395 --> 01:38:59.555
Right.

1740
01:38:59.635 --> 01:39:03.316
Where in five milliseconds, it dumps like a ridiculous amount of energy.

1741
01:39:03.516 --> 01:39:03.676
Yes.

1742
01:39:03.916 --> 01:39:07.277
You know, that timescale for that amount of energy is insane.

1743
01:39:07.537 --> 01:39:10.178
And who knows how many, what number of unknown unknowns.

1744
01:39:10.398 --> 01:39:11.479
The Verirubin is going to find.

1745
01:39:11.499 --> 01:39:12.519
As we have these better tools.

1746
01:39:12.539 --> 01:39:12.779
Yeah.

1747
01:39:12.939 --> 01:39:14.240
And what the Chime is going to find too.

1748
01:39:14.280 --> 01:39:15.220
Maybe it'll find like...

1749
01:39:15.900 --> 01:39:16.621
Super fast.

1750
01:39:16.861 --> 01:39:17.081
Yes.

1751
01:39:17.141 --> 01:39:18.602
Radio bursts, you know, that are like.

1752
01:39:18.963 --> 01:39:19.183
Right.

1753
01:39:19.643 --> 01:39:20.664
At the microsecond level.

1754
01:39:20.824 --> 01:39:21.004
Right.

1755
01:39:21.044 --> 01:39:21.645
You never know.

1756
01:39:21.685 --> 01:39:22.165
Never know.

1757
01:39:22.845 --> 01:39:23.506
You never know.

1758
01:39:23.966 --> 01:39:29.030
We always cover a really fascinating array of stories each week.

1759
01:39:29.531 --> 01:39:32.373
We started with a deep dive on Deep Seek.

1760
01:39:32.793 --> 01:39:33.614
No pun intended.

1761
01:39:34.334 --> 01:39:40.640
And the geopolitical power war that is currently happening between the U.S. and China as leaders.

1762
01:39:40.780 --> 01:39:42.542
Europe, you know, eat your heart out.

1763
01:39:43.102 --> 01:39:45.925
You know, Mistral is not really in the same category.

1764
01:39:46.825 --> 01:39:52.771
We went then to time crystals, the first time crystals that are room temperature and macroscopic.

1765
01:39:53.291 --> 01:39:53.952
That I can see.

1766
01:39:53.992 --> 01:39:54.872
That we can see.

1767
01:39:56.092 --> 01:39:58.414
Out of University of Colorado Boulder.

1768
01:39:59.295 --> 01:40:07.321
Implications for any number of use cases, like security features, technology devices.

1769
01:40:08.202 --> 01:40:11.965
The third story was really fascinating about hidden viruses in our DNA.

1770
01:40:12.746 --> 01:40:14.247
Could be big for medicine breakthroughs.

1771
01:40:14.307 --> 01:40:15.067
Yeah.

1772
01:40:15.127 --> 01:40:16.749
The fact that viruses are 8%

1773
01:40:18.410 --> 01:40:18.830
Of our DNA.

1774
01:40:18.850 --> 01:40:19.671
Of our DNA.

1775
01:40:21.052 --> 01:40:24.555
And that there's these indicators that drive cancer and autoimmune diseases.

1776
01:40:24.835 --> 01:40:29.398
And understanding the structure, understanding the lock, now gives us the ability to build a key.

1777
01:40:30.239 --> 01:40:31.500
And then fast radio bursts.

1778
01:40:32.201 --> 01:40:33.482
Still an enduring mystery.

1779
01:40:33.542 --> 01:40:33.782
Yeah.

1780
01:40:35.163 --> 01:40:42.747
about where is all of this massive scale of light in such a short time scale?

1781
01:40:43.427 --> 01:40:48.370
What cosmological event could create something like this?

1782
01:40:51.292 --> 01:40:52.173
Another great week.

1783
01:40:52.253 --> 01:40:53.153
Again, the feedback.

1784
01:40:53.433 --> 01:40:57.436
We're still on the top charts on Apple for Science Podcasts.

1785
01:40:57.676 --> 01:40:58.557
Keep telling your friends.

1786
01:40:58.857 --> 01:41:00.318
Please keep sharing with your friends.

1787
01:41:00.798 --> 01:41:02.059
The comments are hilarious.

1788
01:41:02.899 --> 01:41:04.140
All of you are comedians.

1789
01:41:04.560 --> 01:41:05.801
Some really, really funny comments.

1790
01:41:06.161 --> 01:41:06.822
I do talk.

1791
01:41:07.402 --> 01:41:08.844
I don't just sit here and nod my head.

1792
01:41:09.264 --> 01:41:11.787
Also, my hairline is not that crazy.

1793
01:41:11.847 --> 01:41:13.129
This is not Stephen A. Smith.

1794
01:41:13.609 --> 01:41:14.691
I've always had a big forehead.

1795
01:41:14.771 --> 01:41:16.853
And he's not the black guy from New Girl.

1796
01:41:17.354 --> 01:41:20.177
I'm not Lamorne Morris, although I appreciate it.

1797
01:41:20.217 --> 01:41:24.482
Lamorne, if you want to have us on the Lamorning podcast, we're literally down the street.

1798
01:41:24.903 --> 01:41:26.284
We'd love to talk to you about science.

1799
01:41:27.406 --> 01:41:33.780
My name is Lester Nare, joined as always by my co-host and our resident PhD, Krishna Chowdhury.

1800
01:41:34.041 --> 01:41:35.684
This is From First Principles.

1801
01:41:36.085 --> 01:41:37.027
We'll see you all next week.

1802
01:41:37.368 --> 01:41:37.629
Peace.
