WEBVTT

1
00:00:01.199 --> 00:00:06.200
<v Speaker 1>Welcome to the Sentient Code, where intelligence is engineered, autonomy

2
00:00:06.280 --> 00:00:10.439
<v Speaker 1>is emerging, and a line between human and machine grows thinner.

3
00:00:10.800 --> 00:00:15.359
<v Speaker 1>Each episode, we decode the algorithms, explore the robotics, and

4
00:00:15.439 --> 00:00:19.000
<v Speaker 1>examine the ideas shaping the future of artificial minds.

5
00:00:23.920 --> 00:00:26.440
<v Speaker 2>Imagine for a second that you are trying to predict

6
00:00:26.480 --> 00:00:30.679
<v Speaker 2>the weather of a specific city, but like, not just

7
00:00:30.719 --> 00:00:33.600
<v Speaker 2>tomorrow's weather. We are talking thirty or even forty five

8
00:00:33.679 --> 00:00:34.719
<v Speaker 2>days from today.

9
00:00:34.560 --> 00:00:36.280
<v Speaker 3>Right, which is incredibly difficult.

10
00:00:36.320 --> 00:00:39.479
<v Speaker 2>Exactly, and your only tool to accomplish this impossible pask

11
00:00:39.759 --> 00:00:42.399
<v Speaker 2>is just a simple list of daily temperatures leading up

12
00:00:42.399 --> 00:00:42.960
<v Speaker 2>to right now.

13
00:00:43.039 --> 00:00:46.920
<v Speaker 3>Yeah, you are dealing with a totally inherently chaotic system there.

14
00:00:46.960 --> 00:00:50.640
<v Speaker 3>I mean, the atmosphere is notoriously noisy. There are countless

15
00:00:50.679 --> 00:00:53.240
<v Speaker 3>invisible variables interacting at all times.

16
00:00:53.320 --> 00:00:56.759
<v Speaker 2>It is the classic butterfly effect in action. Yeah, a tiny,

17
00:00:56.880 --> 00:00:59.920
<v Speaker 2>almost imperceptible shift in the wind, or a fractional different

18
00:01:00.000 --> 00:01:05.400
<v Speaker 2>it's a measurement early on just cascades into massive, completely

19
00:01:05.480 --> 00:01:07.239
<v Speaker 2>unpredictable differences a month later.

20
00:01:07.319 --> 00:01:07.439
<v Speaker 3>Oh.

21
00:01:07.480 --> 00:01:11.159
<v Speaker 2>Absolutely, you are effectively trying to guess the twist ending

22
00:01:11.719 --> 00:01:15.319
<v Speaker 2>of an incredibly complex three hour movie based entirely on

23
00:01:15.359 --> 00:01:17.560
<v Speaker 2>a single blurry frame from the first ten minutes.

24
00:01:17.640 --> 00:01:19.480
<v Speaker 3>That is a really great way to put it. And

25
00:01:20.079 --> 00:01:23.359
<v Speaker 3>what's fascinating here is that the way scientists currently tackle

26
00:01:23.439 --> 00:01:27.120
<v Speaker 3>this kind of chaos relies on a highly clever architecture

27
00:01:27.200 --> 00:01:31.319
<v Speaker 3>called reservoir computing. Specifically, they use these things called echo

28
00:01:31.359 --> 00:01:35.040
<v Speaker 3>state networks. But what makes this field so compelling at

29
00:01:35.040 --> 00:01:38.959
<v Speaker 3>this exact moment is a recent test of that architecture.

30
00:01:38.519 --> 00:01:39.439
<v Speaker 2>Right, the breakthrough test.

31
00:01:39.599 --> 00:01:44.519
<v Speaker 3>Yeah, researchers took a system of just nine single atoms

32
00:01:44.959 --> 00:01:47.640
<v Speaker 3>nine and they put them head to head with a

33
00:01:47.760 --> 00:01:51.200
<v Speaker 3>massive state of the art classical computing setup of ten

34
00:01:51.359 --> 00:01:55.719
<v Speaker 3>thousand artificial neurons, and the atoms one they completely outperformed

35
00:01:55.760 --> 00:01:56.599
<v Speaker 3>the massive.

36
00:01:56.280 --> 00:02:01.400
<v Speaker 2>Network, nine individual atoms beating ten thousand artificial neurals. I mean,

37
00:02:01.480 --> 00:02:04.640
<v Speaker 2>to truly appreciate the sheer scale of what that tiny

38
00:02:04.680 --> 00:02:07.280
<v Speaker 2>system achieved, we really have to look at the massive

39
00:02:07.280 --> 00:02:09.800
<v Speaker 2>classical goliath it was up against. First we do, Yeah,

40
00:02:09.879 --> 00:02:13.000
<v Speaker 2>we need to understand the workhourse of chaotic prediction, which

41
00:02:13.039 --> 00:02:14.960
<v Speaker 2>is the echo state network or ESN.

42
00:02:15.120 --> 00:02:17.680
<v Speaker 3>So if you look at standard neural networks, they traditionally

43
00:02:17.719 --> 00:02:19.400
<v Speaker 3>struggle with time series data.

44
00:02:19.520 --> 00:02:22.080
<v Speaker 2>That is information that unfolds chronologically right.

45
00:02:21.960 --> 00:02:25.800
<v Speaker 3>Exactly like weather patterns or fluctuating energy grids or the

46
00:02:25.800 --> 00:02:29.080
<v Speaker 3>stock market. With a regular neural network, if you want

47
00:02:29.080 --> 00:02:31.960
<v Speaker 3>it to learn how a temperature drop on Tuesday affects

48
00:02:31.960 --> 00:02:35.319
<v Speaker 3>the storm front the following Thursday, you have to adjust

49
00:02:35.400 --> 00:02:38.560
<v Speaker 3>every single mathematical weight and connection inside the.

50
00:02:38.520 --> 00:02:41.560
<v Speaker 2>Network during the training phase, right during training, because the

51
00:02:41.599 --> 00:02:44.680
<v Speaker 2>network has to essentially learn the flow of time step

52
00:02:44.680 --> 00:02:49.240
<v Speaker 2>by step, which I imagine requires massive computing power huge amounts.

53
00:02:49.360 --> 00:02:51.039
<v Speaker 3>The algorithm has to go all the way to the

54
00:02:51.159 --> 00:02:53.000
<v Speaker 3>end of the data, realize it made a mistake, and

55
00:02:53.039 --> 00:02:57.360
<v Speaker 3>then mathematically crawl backward through time to fix every single connection.

56
00:02:57.560 --> 00:03:00.759
<v Speaker 3>Oh wow, Yeah, so the training gets bogged down or

57
00:03:00.960 --> 00:03:04.000
<v Speaker 3>it just gets stuck in a loop and stops learning entirely.

58
00:03:04.199 --> 00:03:05.759
<v Speaker 2>That sounds incredibly inefficient.

59
00:03:05.879 --> 00:03:09.159
<v Speaker 3>It really is. But echo state networks bypass that entirely

60
00:03:09.280 --> 00:03:11.280
<v Speaker 3>by creating what they call a reservoir.

61
00:03:11.479 --> 00:03:13.080
<v Speaker 2>A reservoir, yeah, I think of.

62
00:03:13.039 --> 00:03:15.560
<v Speaker 3>A massive hidden layer made up of hundreds or thousands

63
00:03:15.599 --> 00:03:16.599
<v Speaker 3>of artificial neurons.

64
00:03:16.639 --> 00:03:18.599
<v Speaker 2>Okay, let's unpack this. Let me try to visualize the

65
00:03:18.639 --> 00:03:22.280
<v Speaker 2>mechanics here. If we think of this ESN reservoir as

66
00:03:22.319 --> 00:03:24.919
<v Speaker 2>a large, perfectly still pond.

67
00:03:25.240 --> 00:03:27.319
<v Speaker 3>I like where this is going, right, So.

68
00:03:27.199 --> 00:03:30.520
<v Speaker 2>Your input data, let's say today's temperature in our city

69
00:03:30.560 --> 00:03:32.879
<v Speaker 2>is a pebble. You drop that pebble into the pond.

70
00:03:33.439 --> 00:03:36.719
<v Speaker 2>Immediately it starts creating ripples. Yes, those ripples bounce off

71
00:03:36.719 --> 00:03:38.919
<v Speaker 2>the edges, they cross over each other, and they mix

72
00:03:39.000 --> 00:03:41.280
<v Speaker 2>together in highly complex, unpredictable ways.

73
00:03:41.759 --> 00:03:45.759
<v Speaker 3>And then when tomorrow's temperature comes in, that is another

74
00:03:45.800 --> 00:03:48.919
<v Speaker 3>pebble dropped into the water, right, but it does not

75
00:03:49.199 --> 00:03:51.759
<v Speaker 3>hit a flat, empty surface. It hits a pond that

76
00:03:51.879 --> 00:03:55.840
<v Speaker 3>is already rippling from yesterday's pebble. So the new ripples

77
00:03:55.840 --> 00:03:58.000
<v Speaker 3>physically interact with the old ripples.

78
00:03:58.039 --> 00:03:58.800
<v Speaker 2>Oh. I see, the.

79
00:03:58.800 --> 00:04:01.719
<v Speaker 3>Surface of the pond affect carries a running memory of

80
00:04:01.800 --> 00:04:04.520
<v Speaker 3>everything that has been dropped into it right up to

81
00:04:04.560 --> 00:04:05.080
<v Speaker 3>that moment.

82
00:04:05.360 --> 00:04:09.599
<v Speaker 2>So older hipples naturally fade out over time, because if

83
00:04:09.599 --> 00:04:11.879
<v Speaker 2>they didn't fade, the pond would just become a chaotic,

84
00:04:12.039 --> 00:04:14.719
<v Speaker 2>turbulent mess of endless waves where you couldn't read a

85
00:04:14.719 --> 00:04:16.480
<v Speaker 2>single distinct pattern exactly.

86
00:04:16.639 --> 00:04:18.920
<v Speaker 3>But if they faded immediately, the pond would have no

87
00:04:19.040 --> 00:04:20.600
<v Speaker 3>memory of the past week's weather at all.

88
00:04:20.759 --> 00:04:23.720
<v Speaker 2>That makes sense, So that balance of fading ripples is

89
00:04:23.759 --> 00:04:26.319
<v Speaker 2>what allows the system to hold a usable history.

90
00:04:26.560 --> 00:04:29.959
<v Speaker 3>Yeah, and the defining feature of the echostate network is

91
00:04:30.000 --> 00:04:33.680
<v Speaker 3>how it is built to handle those ripples. When engineers

92
00:04:33.720 --> 00:04:36.480
<v Speaker 3>first set up the network, those hundreds or thousands of

93
00:04:36.560 --> 00:04:39.879
<v Speaker 3>artificial neurons inside the reservoir are connected to each other

94
00:04:39.959 --> 00:04:43.519
<v Speaker 3>completely at random, completely random yep, And then nobody ever

95
00:04:43.560 --> 00:04:47.639
<v Speaker 3>touches them again. The connections inside the reservoir stay permanently

96
00:04:47.720 --> 00:04:49.759
<v Speaker 3>fixed during the entire training process.

97
00:04:50.240 --> 00:04:53.560
<v Speaker 2>Wow. Really, that fixed nature is the echo state property.

98
00:04:53.560 --> 00:04:56.439
<v Speaker 3>Then you got it. The data comes in and every

99
00:04:56.439 --> 00:04:59.680
<v Speaker 3>neuron updates its state using a really simple activation function.

100
00:05:00.160 --> 00:05:02.879
<v Speaker 3>Usually they use the ton function Palm. Yeah, it is

101
00:05:02.920 --> 00:05:05.279
<v Speaker 3>just a mathematical way of squeezing any number down so

102
00:05:05.319 --> 00:05:07.639
<v Speaker 3>it fits neatly between minus one and positive one.

103
00:05:07.720 --> 00:05:10.560
<v Speaker 2>Oh, I get it. That way, the math doesn't explode

104
00:05:10.560 --> 00:05:13.120
<v Speaker 2>into infinity as the ripples keep bouncing.

105
00:05:12.879 --> 00:05:15.399
<v Speaker 3>Around, exactly so, as the days go by in the

106
00:05:15.399 --> 00:05:19.319
<v Speaker 3>weather data, the states of all these thousands of neurons evolve. Together.

107
00:05:19.680 --> 00:05:22.680
<v Speaker 3>They turn a simple one dimensional string of temperature numbers

108
00:05:22.920 --> 00:05:26.680
<v Speaker 3>into an incredibly rich, high dimensional map of patterns.

109
00:05:26.920 --> 00:05:29.759
<v Speaker 2>It is transforming a boring list of numbers into a

110
00:05:29.879 --> 00:05:31.360
<v Speaker 2>vast complex.

111
00:05:30.959 --> 00:05:32.560
<v Speaker 3>Topography, beautifully said.

112
00:05:32.800 --> 00:05:36.839
<v Speaker 2>But wait, if we never train or adjust those thousands

113
00:05:36.839 --> 00:05:40.959
<v Speaker 2>of connections inside the reservoir, how does the network actually

114
00:05:41.040 --> 00:05:42.839
<v Speaker 2>learn to predict the storm next month.

115
00:05:43.439 --> 00:05:45.160
<v Speaker 3>Well, that happens at the very end of the line

116
00:05:45.199 --> 00:05:47.800
<v Speaker 3>in the readout layer vat out layer. Right. It is

117
00:05:47.920 --> 00:05:50.839
<v Speaker 3>usually just a basic linear equation, and it is the

118
00:05:50.879 --> 00:05:54.120
<v Speaker 3>only part of the entire system that ever gets trained. Oh, okay,

119
00:05:54.279 --> 00:05:57.519
<v Speaker 3>you feed your historical weather data into the reservoir. Let

120
00:05:57.519 --> 00:06:01.040
<v Speaker 3>it generate all those rich, complex ripples and simply collect

121
00:06:01.040 --> 00:06:04.639
<v Speaker 3>the states of the neurons. Then you use very fast,

122
00:06:04.720 --> 00:06:07.519
<v Speaker 3>straightforward math to figure out the best weights, the kind

123
00:06:07.519 --> 00:06:11.360
<v Speaker 3>of math like ridge regression or at least squares, just

124
00:06:11.480 --> 00:06:14.519
<v Speaker 3>basic stuff to map those rippling reservoir states to the

125
00:06:14.560 --> 00:06:16.800
<v Speaker 3>actual future temperatures you want to predict.

126
00:06:17.439 --> 00:06:19.759
<v Speaker 2>Go back to our pond analogy. It's like we aren't

127
00:06:19.800 --> 00:06:22.639
<v Speaker 2>trying to track every single drop of water or calculate

128
00:06:22.680 --> 00:06:24.839
<v Speaker 2>the physics of how the pebble hit the surface. Right,

129
00:06:25.040 --> 00:06:27.160
<v Speaker 2>We just set up a camera pointing at the surface

130
00:06:27.160 --> 00:06:31.759
<v Speaker 2>of the pond, and we train a simple algorithm to realize, Oh,

131
00:06:32.000 --> 00:06:34.800
<v Speaker 2>when the waves look exactly like this specific pattern, it

132
00:06:34.839 --> 00:06:36.480
<v Speaker 2>means a storm is coming in five days.

133
00:06:36.720 --> 00:06:39.639
<v Speaker 3>That is a perfect analogy. And because you skip the

134
00:06:39.680 --> 00:06:44.000
<v Speaker 3>agonizingly slow process of training that massive central reservoir, the

135
00:06:44.000 --> 00:06:46.439
<v Speaker 3>whole process is incredibly fast.

136
00:06:46.199 --> 00:06:47.000
<v Speaker 2>Which is huge.

137
00:06:47.040 --> 00:06:50.360
<v Speaker 3>It is you don't need giant supercomputer clusters. These echo

138
00:06:50.439 --> 00:06:55.240
<v Speaker 3>state networks can run on ordinary everyday computers. Really Yeah,

139
00:06:55.279 --> 00:06:59.120
<v Speaker 3>And scientists use them to handle fundamentally chaotic systems all

140
00:06:59.160 --> 00:07:02.240
<v Speaker 3>the time. They use them to model the Lorenza tractor,

141
00:07:02.519 --> 00:07:04.240
<v Speaker 3>which maps atmospheric convection.

142
00:07:04.480 --> 00:07:04.759
<v Speaker 2>Wow.

143
00:07:04.839 --> 00:07:09.279
<v Speaker 3>They use them for high frequency financial forecasting, processing human speech,

144
00:07:09.560 --> 00:07:12.879
<v Speaker 3>and even controlling the complex movements of robotics.

145
00:07:13.000 --> 00:07:16.560
<v Speaker 2>So for weather forecasting, they just feed in historical daily

146
00:07:16.639 --> 00:07:21.120
<v Speaker 2>climate data like maximum daily temperatures over several years, and

147
00:07:21.199 --> 00:07:23.360
<v Speaker 2>train that simple camera at the end the readout layer

148
00:07:23.639 --> 00:07:26.759
<v Speaker 2>to recognize the seasonal cycles and random variations and.

149
00:07:26.759 --> 00:07:28.920
<v Speaker 3>The slow long term trends. Yeah.

150
00:07:28.959 --> 00:07:31.279
<v Speaker 2>But naturally, because this is science, I assume the immediate

151
00:07:31.319 --> 00:07:33.560
<v Speaker 2>thought is to just build a bigger pond. I mean,

152
00:07:33.720 --> 00:07:36.639
<v Speaker 2>if five hundred neurons create a good wave pattern, ten

153
00:07:36.680 --> 00:07:38.600
<v Speaker 2>thousand neurons should create a perfect.

154
00:07:38.319 --> 00:07:42.079
<v Speaker 3>One, right, you would think, So researchers consistently scale them up.

155
00:07:42.160 --> 00:07:44.600
<v Speaker 3>They move from five hundred nodes to one thousand than

156
00:07:44.720 --> 00:07:47.199
<v Speaker 3>five thousand, and frequently push them up to ten thousand

157
00:07:47.240 --> 00:07:47.839
<v Speaker 3>no's or more.

158
00:07:47.920 --> 00:07:51.439
<v Speaker 2>Because a bigger reservoir means a higher dimensional space.

159
00:07:51.279 --> 00:07:55.720
<v Speaker 3>Exactly, there is more room for diverse dynamic patterns to form,

160
00:07:56.079 --> 00:07:59.600
<v Speaker 3>giving that readout layer a much richer map of ripples

161
00:07:59.600 --> 00:08:00.839
<v Speaker 3>to pull features from.

162
00:08:00.920 --> 00:08:03.040
<v Speaker 2>So if a five thousand node network gives you a

163
00:08:03.040 --> 00:08:07.480
<v Speaker 2>great prediction, a ten thousand node network should just completely

164
00:08:07.560 --> 00:08:09.639
<v Speaker 2>obliterate it in terms of accuracy.

165
00:08:09.720 --> 00:08:11.040
<v Speaker 3>Well, that is the theory.

166
00:08:11.000 --> 00:08:13.560
<v Speaker 2>But from what you're saying, that isn't what happens. Why

167
00:08:13.600 --> 00:08:16.600
<v Speaker 2>not do the ripples just turn to white noise? Is

168
00:08:16.639 --> 00:08:19.439
<v Speaker 2>a massive pond just too chaotic for the camera to read?

169
00:08:19.680 --> 00:08:22.519
<v Speaker 3>The issue is a hit, a hard wall of diminishing returns.

170
00:08:22.560 --> 00:08:23.759
<v Speaker 2>Diminishing returns.

171
00:08:24.000 --> 00:08:28.240
<v Speaker 3>Yeah, When scientists actually run these massive networks on weather data,

172
00:08:28.319 --> 00:08:31.360
<v Speaker 3>the initial gains are great. One thousand nodes easily beats

173
00:08:31.360 --> 00:08:33.919
<v Speaker 3>five hundred. Five thousand is better still, But when you

174
00:08:34.000 --> 00:08:36.919
<v Speaker 3>jump from five thousand to ten thousand nodes, the improvement

175
00:08:37.039 --> 00:08:41.799
<v Speaker 3>suddenly flattens out. It just flattens completely. The extra classical

176
00:08:41.840 --> 00:08:47.000
<v Speaker 3>nodes stop providing truly new information. The dynamic patterns inside

177
00:08:47.000 --> 00:08:50.399
<v Speaker 3>the reservoir start to overlap and repeat themselves. You stop

178
00:08:50.399 --> 00:08:54.320
<v Speaker 3>getting distinct wave patterns and just start generating redundant echoes.

179
00:08:54.440 --> 00:08:56.440
<v Speaker 2>So it's just wasting computing power at that point.

180
00:08:56.600 --> 00:09:01.200
<v Speaker 3>Essentially, yes, and beyond just the redundancy, making the network

181
00:09:01.279 --> 00:09:05.000
<v Speaker 3>larger dramatically increases the frigility of the system. The hyper

182
00:09:05.000 --> 00:09:08.240
<v Speaker 3>parameter tuning becomes a total nightmare.

183
00:09:07.919 --> 00:09:09.519
<v Speaker 2>Tuning, Like what kind of parameters?

184
00:09:09.600 --> 00:09:12.559
<v Speaker 3>Well, to keep a ten thousand node network stable, you

185
00:09:12.679 --> 00:09:16.799
<v Speaker 3>have to perfectly tune parameters with names like the spectral radius,

186
00:09:17.159 --> 00:09:20.440
<v Speaker 3>the leaking rate, the input scaling, and the sparsity of

187
00:09:20.480 --> 00:09:21.159
<v Speaker 3>the reservoir.

188
00:09:21.320 --> 00:09:23.960
<v Speaker 2>Hold on, let's ground those terms in the pond analogy.

189
00:09:23.960 --> 00:09:26.440
<v Speaker 2>What is a spectral radius in physical terms?

190
00:09:26.519 --> 00:09:29.360
<v Speaker 3>Good question. Think of the spectral radius as how aggressively

191
00:09:29.399 --> 00:09:31.639
<v Speaker 3>the ripples bounce off the edges of the pond. Okay,

192
00:09:31.679 --> 00:09:34.120
<v Speaker 3>if you set it too high, the waves amplify each

193
00:09:34.120 --> 00:09:37.840
<v Speaker 3>other uncontrollably, and the whole system becomes wildly unstable. It

194
00:09:37.879 --> 00:09:39.960
<v Speaker 3>just starts spitting out pure garbage.

195
00:09:40.039 --> 00:09:41.120
<v Speaker 2>And what about sparsity.

196
00:09:41.399 --> 00:09:44.919
<v Speaker 3>Sparsity would be like placing large rocks in the pond

197
00:09:45.080 --> 00:09:48.639
<v Speaker 3>to block certain ripples from traveling everywhere. And the leaking

198
00:09:48.720 --> 00:09:50.919
<v Speaker 3>rate is the viscosity of the water itself.

199
00:09:51.039 --> 00:09:52.960
<v Speaker 2>Oh, I see. So if you set the leaking rate

200
00:09:53.159 --> 00:09:55.960
<v Speaker 2>just a fraction too high, making the water too thick,

201
00:09:56.440 --> 00:10:00.720
<v Speaker 2>the memory fades too fast, the reservoir goes quiet les instantly,

202
00:10:01.000 --> 00:10:02.919
<v Speaker 2>and you lose the historical data entirely.

203
00:10:03.080 --> 00:10:07.639
<v Speaker 3>It is agonizingly delicate. Scientists knew classical echo state networks

204
00:10:07.639 --> 00:10:10.320
<v Speaker 3>were hitting this wall. They knew just making the classical

205
00:10:10.320 --> 00:10:13.159
<v Speaker 3>pond bigger wasn't the answer because they were entirely bound

206
00:10:13.159 --> 00:10:14.039
<v Speaker 3>by classical math.

207
00:10:14.120 --> 00:10:16.480
<v Speaker 2>Because it's still just sequential processing right.

208
00:10:16.600 --> 00:10:20.279
<v Speaker 3>Inside that classical reservoir. It is still just single numbers

209
00:10:20.320 --> 00:10:24.000
<v Speaker 3>being added, multiplied, and passed through functions sequentially. There is

210
00:10:24.120 --> 00:10:29.000
<v Speaker 3>no built in native way to explore multiple possibilities simultaneously.

211
00:10:29.120 --> 00:10:29.720
<v Speaker 2>What did they do.

212
00:10:30.200 --> 00:10:33.639
<v Speaker 3>They set up a direct, brutal head to head battle,

213
00:10:33.720 --> 00:10:37.879
<v Speaker 3>the ultimate stress test. They took actual, real world daily

214
00:10:37.919 --> 00:10:39.360
<v Speaker 3>weather records from the city of.

215
00:10:39.279 --> 00:10:40.799
<v Speaker 2>Delhi, real weather data.

216
00:10:40.919 --> 00:10:43.519
<v Speaker 3>Yes, and they didn't just ask the networks to predict

217
00:10:43.559 --> 00:10:47.039
<v Speaker 3>to Mars temperature. They task the systems with forecasting the

218
00:10:47.080 --> 00:10:50.320
<v Speaker 3>chaotic temperature trends fifteen days out, thirty days out, and

219
00:10:50.399 --> 00:10:52.080
<v Speaker 3>a massive forty five days out.

220
00:10:52.120 --> 00:10:53.159
<v Speaker 2>Wow, forty five days.

221
00:10:53.279 --> 00:10:56.559
<v Speaker 3>Yeah, in the classical corner, they lined up the heavyweights,

222
00:10:57.159 --> 00:11:01.200
<v Speaker 3>carefully tuned, optimized echo state network sized at five hundred,

223
00:11:01.519 --> 00:11:04.639
<v Speaker 3>one thousand, five thousand and ten thousand nodes.

224
00:11:04.679 --> 00:11:07.840
<v Speaker 2>And in the other corner of the challenger a quantum reservoir.

225
00:11:08.000 --> 00:11:12.679
<v Speaker 3>Exactly. But this wasn't some warehouse sized, billion dollar quantum

226
00:11:12.720 --> 00:11:16.559
<v Speaker 3>supercomputer chilled absolute zero. No, it was built from just

227
00:11:16.799 --> 00:11:22.120
<v Speaker 3>nine interacting atomic spins, nine individual atoms inside.

228
00:11:21.720 --> 00:11:23.320
<v Speaker 2>A molecule, and the result.

229
00:11:23.440 --> 00:11:27.559
<v Speaker 3>The result was staggering. The tiny nine spin quantum system

230
00:11:27.639 --> 00:11:31.559
<v Speaker 3>didn't just compete, It captured the temperature trends accurately across

231
00:11:31.600 --> 00:11:35.240
<v Speaker 3>those fifteen thirty and forty five day horizons, beating the

232
00:11:35.279 --> 00:11:38.440
<v Speaker 3>massive ten thousand node classical network at its own game.

233
00:11:38.720 --> 00:11:41.480
<v Speaker 2>Here's where it gets really interesting. It's like a single

234
00:11:41.519 --> 00:11:44.720
<v Speaker 2>person with a multitool outbuilding an entire construction crew of

235
00:11:44.759 --> 00:11:47.120
<v Speaker 2>ten thousand workers. Yeah, it's hard to wrap your mind

236
00:11:47.159 --> 00:11:47.600
<v Speaker 2>around that.

237
00:11:47.879 --> 00:11:49.240
<v Speaker 3>It is very unintuitive.

238
00:11:49.360 --> 00:11:52.279
<v Speaker 2>We are so trained to assume that bigger is always better.

239
00:11:52.480 --> 00:11:57.320
<v Speaker 2>You know that more nodes automatically equal more intelligence. How

240
00:11:57.360 --> 00:12:00.159
<v Speaker 2>does a microscoptic clump of nine atoms hold more more

241
00:12:00.200 --> 00:12:03.000
<v Speaker 2>complexity than ten thousand artificial neurons.

242
00:12:03.039 --> 00:12:05.480
<v Speaker 3>Well, it comes down to quantum mechanics.

243
00:12:05.240 --> 00:12:08.519
<v Speaker 2>Right, So if classical nodes are locked into processing one

244
00:12:08.600 --> 00:12:11.759
<v Speaker 2>number at a time, are these atoms doing that quantum

245
00:12:11.759 --> 00:12:14.000
<v Speaker 2>trick where they exist in multiple states at once to

246
00:12:14.039 --> 00:12:15.080
<v Speaker 2>process more data?

247
00:12:15.480 --> 00:12:19.639
<v Speaker 3>That is exactly the first mechanism at play superposition superposition. Yes,

248
00:12:19.960 --> 00:12:23.080
<v Speaker 3>in a classical network, a node is in one specific

249
00:12:23.159 --> 00:12:25.399
<v Speaker 3>state at one specific time. It is a one or

250
00:12:25.399 --> 00:12:29.320
<v Speaker 3>a zero, a positive or a negative. But quantum superposition

251
00:12:29.519 --> 00:12:33.559
<v Speaker 3>allows those nine atomic spins to explore a vast number

252
00:12:33.559 --> 00:12:35.759
<v Speaker 3>of states simultaneously.

253
00:12:35.000 --> 00:12:36.679
<v Speaker 2>Like they're doing all the math at the same time.

254
00:12:36.759 --> 00:12:40.399
<v Speaker 3>Basically, they exist in a fluid cloud of probabilities, exploring

255
00:12:40.440 --> 00:12:43.200
<v Speaker 3>many dynamics at once. Even though there are only nine

256
00:12:43.240 --> 00:12:47.240
<v Speaker 3>physical pieces, the mathematical space they are exploring is exponentially

257
00:12:47.320 --> 00:12:49.679
<v Speaker 3>larger than what nine classical nodes could ever reach.

258
00:12:49.960 --> 00:12:52.919
<v Speaker 2>That is incredible. And then there's the entanglement factor. Right

259
00:12:53.039 --> 00:12:56.440
<v Speaker 2>in the classical pond, we artificially wire the nodes together

260
00:12:56.840 --> 00:13:01.360
<v Speaker 2>and hope the mathematical ripples create useful relationships, But entanglement

261
00:13:01.840 --> 00:13:03.639
<v Speaker 2>is a physical property exactly.

262
00:13:03.679 --> 00:13:07.879
<v Speaker 3>It creates deep fundamental correlations between the atomic spins. They

263
00:13:07.960 --> 00:13:11.519
<v Speaker 3>share information in a profound, non local way.

264
00:13:12.200 --> 00:13:15.039
<v Speaker 2>So to bring it back to the pond analogy, entanglement

265
00:13:15.120 --> 00:13:17.399
<v Speaker 2>would be like dropping a pedal on the left side

266
00:13:17.399 --> 00:13:21.399
<v Speaker 2>of the pond and it instantly creates perfectly mirrored ripples

267
00:13:21.399 --> 00:13:23.720
<v Speaker 2>on the far right side of the pond, without the

268
00:13:23.759 --> 00:13:25.919
<v Speaker 2>water having to physically travel across the middle.

269
00:13:26.080 --> 00:13:30.360
<v Speaker 3>The connection is instantaneous and deeply linked. Yes, the richness

270
00:13:30.440 --> 00:13:35.559
<v Speaker 3>of those entangled physical connections completely dwarfs the random mathematical

271
00:13:35.600 --> 00:13:39.159
<v Speaker 3>wiring of the classical network unbelievable. But perhaps the most

272
00:13:39.159 --> 00:13:41.480
<v Speaker 3>elegant part of the whole quantum system is how it

273
00:13:41.559 --> 00:13:43.720
<v Speaker 3>handles the fading memory we talked about earlier.

274
00:13:43.799 --> 00:13:45.159
<v Speaker 2>Oh right, the leaking rate.

275
00:13:45.320 --> 00:13:49.159
<v Speaker 3>Yeah, In a classical network, engineers have to agonize over

276
00:13:49.200 --> 00:13:52.799
<v Speaker 3>the leaking rate. Fading memory is mathematical fiction. They have

277
00:13:52.840 --> 00:13:55.879
<v Speaker 3>to carefully code and manually tuned so the network doesn't

278
00:13:55.879 --> 00:13:57.279
<v Speaker 3>forget the past too fast.

279
00:13:57.519 --> 00:14:00.399
<v Speaker 2>But the quantum system uses natural dissipation.

280
00:14:00.200 --> 00:14:04.240
<v Speaker 3>Exactly in the physical quantum system. Natural dissipation happens on

281
00:14:04.279 --> 00:14:08.879
<v Speaker 3>its own. As time passes, the excited atomic spins naturally

282
00:14:08.919 --> 00:14:10.639
<v Speaker 3>relax back to their base.

283
00:14:10.399 --> 00:14:14.120
<v Speaker 2>State, so nature essentially comes pre installed with the exact

284
00:14:14.240 --> 00:14:18.320
<v Speaker 2>fading memory algorithm that classical engineers spend months trying to code.

285
00:14:18.519 --> 00:14:22.159
<v Speaker 3>That's a brilliant way to phrase it. The physical, natural

286
00:14:22.200 --> 00:14:25.559
<v Speaker 3>way these atoms relax over time perfectly matches the exact

287
00:14:25.600 --> 00:14:27.919
<v Speaker 3>requirement to process time series data.

288
00:14:28.360 --> 00:14:31.320
<v Speaker 2>The researchers don't have to perfectly tune a mathematical leaking

289
00:14:31.399 --> 00:14:34.480
<v Speaker 2>rate because the physics the molecule does it for them automatically.

290
00:14:34.600 --> 00:14:37.279
<v Speaker 3>Yes, if we connect this to the bigger picture, the

291
00:14:37.360 --> 00:14:40.320
<v Speaker 3>quantum system isn't running a program to simulate the math

292
00:14:40.360 --> 00:14:44.000
<v Speaker 3>of a reservoir. It physically is the reservoir. It uses

293
00:14:44.039 --> 00:14:46.600
<v Speaker 3>the native rules of the universe to do profoundly more

294
00:14:46.679 --> 00:14:48.000
<v Speaker 3>with exponentially less.

295
00:14:48.200 --> 00:14:50.799
<v Speaker 2>Okay, I follow the physics of why it works better,

296
00:14:51.159 --> 00:14:53.480
<v Speaker 2>but I keep getting stuck on the mechanics of the input.

297
00:14:53.720 --> 00:14:55.639
<v Speaker 2>I mean, I can wrap my head around typing a

298
00:14:55.759 --> 00:14:58.639
<v Speaker 2>temperature number into a computer program, but how do you

299
00:14:58.720 --> 00:15:02.480
<v Speaker 2>feed the daily temperature of Deli into nine microscopic atoms?

300
00:15:02.840 --> 00:15:06.080
<v Speaker 3>Ah? They use nuclear magnetic resonance or NMR.

301
00:15:06.440 --> 00:15:10.240
<v Speaker 2>NMR like the technology behind an MRI machine at a hospital.

302
00:15:10.399 --> 00:15:13.120
<v Speaker 2>We're putting nine atoms into a medical scanner.

303
00:15:13.240 --> 00:15:16.759
<v Speaker 3>The underlying principle is very similar. Yeah, you use highly

304
00:15:16.759 --> 00:15:21.120
<v Speaker 3>precise magnetic fields and radio frequency pulses to physically interact

305
00:15:21.240 --> 00:15:24.480
<v Speaker 3>with the natural spin spin couplings of the atoms.

306
00:15:24.559 --> 00:15:26.960
<v Speaker 2>Okay, so how does the temperature data translate to that?

307
00:15:27.200 --> 00:15:30.440
<v Speaker 3>Well, if the daily temperature in Deli is ninety five degrees,

308
00:15:30.919 --> 00:15:34.559
<v Speaker 3>the researchers translate that number into a radio pulse tuned

309
00:15:34.559 --> 00:15:38.240
<v Speaker 3>to a very specific frequency, almost like tuning a guitar string.

310
00:15:38.480 --> 00:15:41.240
<v Speaker 2>So they fire that specific frequency at the molecule and

311
00:15:41.320 --> 00:15:43.799
<v Speaker 2>it makes the atoms vibrate or spin in a way

312
00:15:43.799 --> 00:15:46.000
<v Speaker 2>that physically represents that exact temperature.

313
00:15:46.039 --> 00:15:49.000
<v Speaker 3>You've got it. It is literally dropping the physical pebble into

314
00:15:49.000 --> 00:15:52.200
<v Speaker 3>a quantum pond that is just wild. Then you let

315
00:15:52.240 --> 00:15:56.440
<v Speaker 3>those natural quantum dynamics, the superposition, the entanglement, the dissipation

316
00:15:56.679 --> 00:15:59.240
<v Speaker 3>play out as the ripples interact, and then you use

317
00:15:59.279 --> 00:16:02.759
<v Speaker 3>the nmr AG again to measure the resulting physical states of.

318
00:16:02.679 --> 00:16:04.639
<v Speaker 2>The molecule, and that gives you the prediction.

319
00:16:04.639 --> 00:16:08.440
<v Speaker 3>Mostly yes, But to be completely transparent about the architecture,

320
00:16:08.559 --> 00:16:11.200
<v Speaker 3>the quantum system does occasionally get a tiny bit of

321
00:16:11.279 --> 00:16:13.639
<v Speaker 3>help at the very end. In that might outlayer, we

322
00:16:13.679 --> 00:16:17.440
<v Speaker 3>discuss the camera watching the pawn. Right, the researchers found

323
00:16:17.480 --> 00:16:20.320
<v Speaker 3>that pairing the nine spin reservoir with a slightly more

324
00:16:20.360 --> 00:16:25.960
<v Speaker 3>sophisticated nonlinear redoubt helped translate those deeply complex quantum states

325
00:16:25.960 --> 00:16:29.679
<v Speaker 3>into the final temperature predictions with incredible accuracy.

326
00:16:29.720 --> 00:16:30.519
<v Speaker 2>That kind of readout.

327
00:16:30.679 --> 00:16:34.879
<v Speaker 3>Specifically, they use support vector regression with a radial basis

328
00:16:34.919 --> 00:16:36.279
<v Speaker 3>function kernel.

329
00:16:36.000 --> 00:16:39.759
<v Speaker 2>Wait planning GUS please a radial basis function kernel that

330
00:16:39.799 --> 00:16:42.200
<v Speaker 2>basically just means the camera is capable of drawing curve

331
00:16:42.320 --> 00:16:44.279
<v Speaker 2>lines instead of just straight ones to find a patter

332
00:16:44.360 --> 00:16:48.480
<v Speaker 2>right yep, it can match more complex circular groupings of data.

333
00:16:48.559 --> 00:16:50.759
<v Speaker 3>That is a great way to summarize it. It allows

334
00:16:50.759 --> 00:16:54.240
<v Speaker 3>the readout layer to handle more complex nonlinear relationships.

335
00:16:54.320 --> 00:16:54.759
<v Speaker 2>Makes sense.

336
00:16:55.080 --> 00:16:58.200
<v Speaker 3>But even with that tiny mathematical assist at the readout layer,

337
00:16:58.480 --> 00:17:01.759
<v Speaker 3>the heavy lifting, the full processing of the chaotic time

338
00:17:01.840 --> 00:17:05.400
<v Speaker 3>series data, the memory, the pattern recognition is being done

339
00:17:05.559 --> 00:17:08.240
<v Speaker 3>entirely by the natural physics of those nine spins.

340
00:17:08.599 --> 00:17:12.440
<v Speaker 2>So why does this specific battle between ten thousand artificial

341
00:17:12.440 --> 00:17:17.039
<v Speaker 2>neurons and nine atomic spins matter to you listening right now, Well,

342
00:17:17.039 --> 00:17:20.839
<v Speaker 2>because it completely reframes the timeline of quantum technology.

343
00:17:20.880 --> 00:17:21.640
<v Speaker 3>It really does.

344
00:17:21.759 --> 00:17:24.880
<v Speaker 2>When we usually hear about quantum computing, the narrative is

345
00:17:24.920 --> 00:17:28.119
<v Speaker 2>almost always about the distant future. We are told we

346
00:17:28.160 --> 00:17:32.640
<v Speaker 2>have to wait decades for massive, flawlessly error corrected, fault

347
00:17:32.680 --> 00:17:37.160
<v Speaker 2>tolerance Sci Fi supercomputers before quantum technology actually touches our

348
00:17:37.240 --> 00:17:37.960
<v Speaker 2>daily lives.

349
00:17:38.039 --> 00:17:40.759
<v Speaker 3>But this experiment proves that we do not have to wait.

350
00:17:41.359 --> 00:17:45.359
<v Speaker 3>Classical echostate networks are brilliant feats of engineering. They have

351
00:17:45.440 --> 00:17:47.559
<v Speaker 3>pushed the limits of what we can do with scale,

352
00:17:47.920 --> 00:17:52.480
<v Speaker 3>hyper parameter tuning and brute force math. But nature's native

353
00:17:52.559 --> 00:17:57.279
<v Speaker 3>quantum physics can effortlessly outmash years of our best classical engineering. Today.

354
00:17:57.720 --> 00:18:01.400
<v Speaker 2>It proves that practical world changing of advantages aren't locked

355
00:18:01.400 --> 00:18:04.240
<v Speaker 2>away in the twenty fifties. We can use today's near

356
00:18:04.319 --> 00:18:08.480
<v Speaker 2>term hardware, these imperfect, small scale, noisy quantum systems to

357
00:18:08.720 --> 00:18:12.839
<v Speaker 2>drastically improve everyday forecasting. Right now, we are talking about

358
00:18:12.880 --> 00:18:16.160
<v Speaker 2>better predictions for balancing the energy grid during heat waves,

359
00:18:16.559 --> 00:18:20.480
<v Speaker 2>or more accurate modeling of volatile financial markets, and knowing

360
00:18:20.559 --> 00:18:22.960
<v Speaker 2>exactly what the weather's going to do forty five days

361
00:18:22.960 --> 00:18:25.920
<v Speaker 2>from now so agriculture and shipping can actually prepare.

362
00:18:26.160 --> 00:18:30.359
<v Speaker 3>It is a profound paradigm shift. This raises an important question.

363
00:18:30.480 --> 00:18:33.759
<v Speaker 3>You know, we are no longer trying to brute force

364
00:18:33.839 --> 00:18:37.000
<v Speaker 3>the math of the universe to simulate chaos. We are

365
00:18:37.039 --> 00:18:40.039
<v Speaker 3>finally letting the universe do the math for us.

366
00:18:39.920 --> 00:18:42.119
<v Speaker 2>Which brings us back to that blurry frame of a

367
00:18:42.160 --> 00:18:44.599
<v Speaker 2>movie we talked about at the very beginning, trying to

368
00:18:44.640 --> 00:18:48.880
<v Speaker 2>predict the complex twist ending from almost nothing is an

369
00:18:48.880 --> 00:18:52.559
<v Speaker 2>incredibly complex, naturally dissipating system of just nine atomic spins

370
00:18:53.000 --> 00:18:56.720
<v Speaker 2>can accurately model the chaotic cascading weather patterns of an

371
00:18:56.880 --> 00:18:59.319
<v Speaker 2>entire massive city for a month and a half when

372
00:18:59.559 --> 00:19:03.599
<v Speaker 2>entirely new, seemingly unpredictable human systems. Might we be able

373
00:19:03.640 --> 00:19:05.720
<v Speaker 2>to perfectly map if we scaled.

374
00:19:05.319 --> 00:19:07.559
<v Speaker 3>This approach up, Yeah, that is the real question.

375
00:19:07.759 --> 00:19:09.640
<v Speaker 2>I mean, not to ten thousand atoms, but just one

376
00:19:09.720 --> 00:19:14.480
<v Speaker 2>hundred or maybe one thousand quantum spins. Could we model

377
00:19:14.519 --> 00:19:17.839
<v Speaker 2>the hidden cascading ripple effects of the global economy? Could

378
00:19:17.880 --> 00:19:21.160
<v Speaker 2>we perfectly predict the exact spread and mutation timeline of

379
00:19:21.160 --> 00:19:24.559
<v Speaker 2>a virus across continents? If just nine atoms can map

380
00:19:24.599 --> 00:19:26.720
<v Speaker 2>out the storms of tomorrow, what on Earth or one

381
00:19:26.720 --> 00:19:28.440
<v Speaker 2>thousand atoms is going to show us about the future