1
00:00:00,121 --> 00:00:04,787
It's been almost a decade since Tesla debuted the Roadster, their flagship car,

2
00:00:04,787 --> 00:00:08,937
and last week on Thursday, Elon Musk finally showed us the car.

3
00:00:09,277 --> 00:00:15,609
I watched a Tesla Roadster car drive up a ramp, turn upside down while stationary,

4
00:00:15,881 --> 00:00:19,980
staying stuck to the ramp, launch itself off said ramp,

5
00:00:20,328 --> 00:00:26,597
fly through the air, and then land perfectly into cruise control without a single

6
00:00:26,597 --> 00:00:28,907
dent or bit of damage on the car.

7
00:00:28,907 --> 00:00:32,217
Now, I understand what I just described to you sounds like AI,

8
00:00:32,217 --> 00:00:34,177
but it is anything but AI.

9
00:00:34,177 --> 00:00:37,727
And it uses some really neat physics tricks that I'm going to get into.

10
00:00:37,727 --> 00:00:43,670
But I realized it is the perfect combination of SpaceX and Tesla engineering

11
00:00:43,867 --> 00:00:49,467
that is going to propel, pun intended, cars, manufacturing and automobiles into the future.

12
00:00:49,467 --> 00:00:55,307
We are actually getting flying cars. But it also told me another story and the rumor mill is going.

13
00:00:55,307 --> 00:00:58,777
That Tesla and SpaceX might actually merge by the end of the year,

14
00:00:58,777 --> 00:01:03,557
forming what is going to be the fifth largest company in the entire world.

15
00:01:03,557 --> 00:01:06,367
I'm still in Japan. It's my final day. I've got my kimono on,

16
00:01:06,367 --> 00:01:10,277
but I'm excited to get into it and back to the regular schedule starting on

17
00:01:10,277 --> 00:01:14,817
tomorrow's episode. So all of this and more on today's episode of In The Loop

18
00:01:14,817 --> 00:01:15,838
brought to you by Qualcomm.

19
00:01:19,615 --> 00:01:25,072
Okay, so first we have to watch this video. There's jets of air propelling itself out of the car.

20
00:01:25,385 --> 00:01:28,926
It's spinning the car onto its front. It's landing it perfectly.

21
00:01:29,333 --> 00:01:33,327
There's a loud whirring noise when you listen to this video of air suction.

22
00:01:33,327 --> 00:01:35,068
You're wondering what on earth is going on.

23
00:01:35,364 --> 00:01:38,237
And then you're thinking, is this video sped up or slowed down?

24
00:01:38,237 --> 00:01:41,245
How on earth is a car able to stay stuck to the wall?

25
00:01:41,680 --> 00:01:44,457
Well, there's actually three neat physics tricks happening here,

26
00:01:44,457 --> 00:01:48,077
which combines SpaceX's engineering with that of Tesla's, which has produced

27
00:01:48,356 --> 00:01:52,747
a new type of car which can only be classified as a flying car and this wasn't

28
00:01:52,747 --> 00:01:57,841
the only neat trick that they showed us at the pre-launch event for the roadster i watched this car

29
00:01:58,108 --> 00:02:04,047
fly over trucks and cars i watched it go from 0 to 60 mile per hour in 0.81

30
00:02:04,047 --> 00:02:08,795
seconds which is the fastest of any road car ever and it basically got me questioning

31
00:02:08,929 --> 00:02:12,597
everything that i know about cars and what they're about to become

32
00:02:12,917 --> 00:02:15,384
so i wanted to understand the first trick which is

33
00:02:15,726 --> 00:02:21,251
how on earth did they get a car to stick to the wall because what you're seeing on your screen.

34
00:02:21,726 --> 00:02:25,972
Is in real time. And so no way could a normal car stay stuck to the wall,

35
00:02:25,972 --> 00:02:27,711
it would have fallen and crashed to the ground.

36
00:02:28,141 --> 00:02:33,142
Well, this car is equipped with an immense fan suction system.

37
00:02:33,142 --> 00:02:39,302
So it's embedded on the underbelly of the car, and it is sucking out all the

38
00:02:39,302 --> 00:02:40,922
air from underneath the car.

39
00:02:40,922 --> 00:02:45,910
And what this does is it creates an immense downwards force or downwards pressure,

40
00:02:46,206 --> 00:02:49,445
pushing the car onto any surface that it is driving on.

41
00:02:49,776 --> 00:02:57,259
Now the power of this force is up to 4,700 kgs, which is around 4.7 tons.

42
00:02:57,590 --> 00:03:00,591
That is a lot of force, 2 g's worth of force.

43
00:03:00,910 --> 00:03:04,550
And so therefore when the car is upside down or perpendicular to the ground,

44
00:03:04,852 --> 00:03:10,064
it's still enough to stick it to whatever surface that it's riding on, in this case, the ramp.

45
00:03:10,169 --> 00:03:14,592
And so that's what effectively is happening underneath the car that causes it to maintain traction.

46
00:03:14,592 --> 00:03:18,222
Now you might be wondering what is the physical or practical reality of this?

47
00:03:18,222 --> 00:03:21,422
Well, you can imagine if the car is moving at high speeds and you don't want

48
00:03:21,422 --> 00:03:25,482
it to barrel roll, or if you want it to have the best traction or safety when

49
00:03:25,482 --> 00:03:29,771
you are turning the steering wheel, this gives you that added traction.

50
00:03:30,003 --> 00:03:34,192
Now, the second trick that you saw were these jet streams that were shooting

51
00:03:34,192 --> 00:03:38,512
out of the car at weird times to try and stabilize its direction and help it

52
00:03:38,512 --> 00:03:42,031
land beautifully after it propelled itself off of the ramp.

53
00:03:42,525 --> 00:03:47,947
What I'm showing you on your screen now is something that Tesla's calling their thruster.

54
00:03:48,063 --> 00:03:49,802
But this thruster isn't like any

55
00:03:49,802 --> 00:03:54,332
typical thruster that you would see on a rocket or that of any other car.

56
00:03:54,675 --> 00:04:00,632
It contains highly pressurized compressed gas in this spherical shape that you're

57
00:04:00,632 --> 00:04:02,047
seeing on your screen right now.

58
00:04:02,414 --> 00:04:07,783
And this gas is basically released and it expands super aggressively or very

59
00:04:07,783 --> 00:04:12,893
quickly in milliseconds, such that it creates this immense force that can push

60
00:04:12,893 --> 00:04:15,537
the car in multiple different directions.

61
00:04:15,845 --> 00:04:19,573
Now, the car itself is fitted out to the max with these thrusters.

62
00:04:19,573 --> 00:04:20,832
They have about 10 of them.

63
00:04:20,942 --> 00:04:26,083
And coincidentally, these are the same thrusters that SpaceX used on the Falcon 9.

64
00:04:26,083 --> 00:04:30,561
So the spherical storage tank that you're seeing right at the end of this thruster

65
00:04:30,857 --> 00:04:33,383
basically is the container for

66
00:04:33,383 --> 00:04:37,370
this highly compressed gas. It's usually a mixture of nitrogen and oxygen.

67
00:04:37,643 --> 00:04:43,018
The cool thing about this is there's zero fuel or petroleum involved anywhere. It is just gas.

68
00:04:43,320 --> 00:04:46,943
And then you have an electric valve, which basically opens at a given point

69
00:04:46,943 --> 00:04:52,213
or time due to various different sensors and allows the gas to release and expand.

70
00:04:52,213 --> 00:04:56,823
And then Newton's law of motion applies here, where if it pushes a car in a

71
00:04:56,823 --> 00:05:01,793
certain direction or forces enough gas in one particular direction it can push

72
00:05:01,793 --> 00:05:04,555
the car stabilize the car in a particular different way

73
00:05:04,862 --> 00:05:08,963
and so as you're going back to the original demo that i showed you when it propels

74
00:05:08,963 --> 00:05:10,998
or launches itself off of the ramp

75
00:05:11,323 --> 00:05:15,543
these gas streams from these different thrusters help stabilize it now the scene

76
00:05:15,543 --> 00:05:21,150
i immediately think of when i watched this was that scene when matthew mcconaughey in interstellar.

77
00:05:21,481 --> 00:05:26,203
Docks the ship and he uses his robot tars to control this pretty precarious

78
00:05:26,203 --> 00:05:32,163
situation using these highly pressurized jet streams of gas to make sure or ensure.

79
00:05:32,503 --> 00:05:35,993
That the spaceship lands in a very specific way.

80
00:05:36,138 --> 00:05:40,633
That's exactly what this Tesla car does. And it's worth pointing out that SpaceX's

81
00:05:40,633 --> 00:05:45,843
supply chain is probably the only one belonging to any space company or any

82
00:05:45,843 --> 00:05:48,821
company in the world that could do this at scale.

83
00:05:49,084 --> 00:05:54,801
Now, Elon Musk did go on to say that the release of this car is probably going to be sometime in 2028.

84
00:05:55,005 --> 00:05:57,963
But the point is, if you are any other space company that wants to try and recreate

85
00:05:57,963 --> 00:06:01,863
this, there's no way that you could get it down to a cost optimal path.

86
00:06:01,863 --> 00:06:04,193
I'll get into prices later. But the idea here is,

87
00:06:04,536 --> 00:06:09,806
this is a first of a kind. But the third and final trick is something that is a lot more relatable.

88
00:06:10,171 --> 00:06:14,678
It is wings on the back of the car that you're seeing on the screen here fold

89
00:06:14,678 --> 00:06:17,298
out from the trunk like area of the car.

90
00:06:17,298 --> 00:06:21,828
And it does so in a very sleek and subtle way. Now, if you're a fan of Fast

91
00:06:21,828 --> 00:06:25,718
and Furious or any racing movie, you typically see this kind of architecture

92
00:06:25,718 --> 00:06:28,908
on the back of a car that goes super fast because there's less drag.

93
00:06:29,129 --> 00:06:30,511
It's way more aerodynamic.

94
00:06:30,766 --> 00:06:35,259
And it creates, most importantly a downward pressure as the car is moving fast.

95
00:06:35,491 --> 00:06:39,348
Now think about it if you have a car that's moving incredibly fast but doesn't

96
00:06:39,348 --> 00:06:42,898
have any downward pressure as soon as you steer the car it goes out of control

97
00:06:42,898 --> 00:06:46,823
and it starts to barrel roll but if you are able to generate downwards force

98
00:06:47,008 --> 00:06:48,622
then we have a different situation.

99
00:06:48,738 --> 00:06:53,178
Now this version or this form of the car is known as track mode and I believe

100
00:06:53,178 --> 00:06:57,728
it's towards a model that doesn't have the expensive fan suction system that

101
00:06:57,728 --> 00:07:00,342
you're seeing in the earlier demos that I showed you.

102
00:07:00,592 --> 00:07:03,971
But it's still cool nonetheless, and it achieves downwards force.

103
00:07:04,528 --> 00:07:10,157
Bigger than any other car or road car or land car that is accessible to an average user.

104
00:07:10,337 --> 00:07:14,758
And it's almost to the level and point of an F1 car. And you're achieving something

105
00:07:14,758 --> 00:07:16,880
similar at a regular consumer car.

106
00:07:17,060 --> 00:07:22,128
Pretty insane. And honestly, I have to give it to Elon and SpaceX and Tesla

107
00:07:22,128 --> 00:07:27,158
in general, because they managed to keep this a secret for so long.

108
00:07:27,158 --> 00:07:31,439
And on this video here, you'll see the footage that they released afterwards

109
00:07:31,555 --> 00:07:36,274
of all the test runs that they did to try and create the thrusters that would

110
00:07:36,530 --> 00:07:39,058
be repurposed from SpaceX rockets to make sense

111
00:07:39,418 --> 00:07:44,208
for a regular consumer car to show the jet streams propelling the car up and

112
00:07:44,208 --> 00:07:46,028
making sure it lands in a safe-like manner.

113
00:07:46,098 --> 00:07:49,225
I'm guessing they went through so many different types of crashes.

114
00:07:49,539 --> 00:07:53,638
The formation of the fan suction system to make sure that it generates enough

115
00:07:53,638 --> 00:07:57,698
downward pressure so that you can literally drive your car upside down and to

116
00:07:57,698 --> 00:08:01,688
create the wing system as well to do this for an average regular car that doesn't

117
00:08:01,688 --> 00:08:04,498
have the fan suction system. Now, in terms of price.

118
00:08:04,952 --> 00:08:08,753
They didn't really release any at all. But they did say that if you want the

119
00:08:08,753 --> 00:08:11,487
performance version, it's a $50,000 deposit.

120
00:08:11,841 --> 00:08:16,033
If you want the thrust version, which was the Jetstreams that you saw on your

121
00:08:16,033 --> 00:08:21,953
screen, it's a $250,000 deposit, a quarter of a million dollars.

122
00:08:21,953 --> 00:08:25,713
And then you have the rocket version, which is, as far as I'm concerned,

123
00:08:25,713 --> 00:08:28,043
undefined, which is invite only.

124
00:08:28,043 --> 00:08:31,423
So it's been, what, almost a decade since they announced the Roadster,

125
00:08:31,423 --> 00:08:35,993
and it's gonna be another two years until we actually get access to this thing.

126
00:08:35,993 --> 00:08:40,813
So it's still TBD whether this actually gets mass produced, but it's such a

127
00:08:40,813 --> 00:08:45,053
cool car demo. But stepping away from the sci-fi of it all, there is a more

128
00:08:45,053 --> 00:08:49,370
strategic reason why I think Elon Musk is creating this car.

129
00:08:49,707 --> 00:08:53,343
He's demonstrating that he can merge some of the most powerful capabilities

130
00:08:53,343 --> 00:08:58,623
and technologies from both companies to create something net new that the world has not seen.

131
00:08:58,623 --> 00:09:02,453
Now, it's no secret that Elon Musk has been incredibly bullish on space.

132
00:09:02,453 --> 00:09:05,673
He's been bullish for various different reasons. He believes getting humans

133
00:09:05,673 --> 00:09:09,223
to extraterrestrial planets is one of the most important things.

134
00:09:09,223 --> 00:09:12,723
Kardashev's law of scaling to kind of harness the energy of the Sun,

135
00:09:12,723 --> 00:09:16,993
in this case, to power AI models and generate compute, an abundance of compute

136
00:09:16,993 --> 00:09:20,315
that he thinks will help power the world and generate tons and tons of wealth

137
00:09:20,536 --> 00:09:22,144
for anyone and everyone in the future.

138
00:09:22,539 --> 00:09:26,993
But also to create as many robots, and robots come in the form of cars or humanoid

139
00:09:26,993 --> 00:09:28,814
robots in Tesla Optimus.

140
00:09:29,278 --> 00:09:34,436
As we can so that you can help humanity propel as AI and robotics merge.

141
00:09:34,436 --> 00:09:38,016
So then it would only make sense if Tesla and SpaceX would merge.

142
00:09:38,016 --> 00:09:41,986
Now, if you watched one of our early episodes almost a year,

143
00:09:42,346 --> 00:09:46,763
maybe a year and a half ago, you knew that we predicted it on this show.

144
00:09:46,926 --> 00:09:53,126
We predicted well before SpaceX actually IPO'd that SpaceX should end up merging

145
00:09:53,126 --> 00:09:56,167
with Tesla because there's so much synergies between the two.

146
00:09:56,446 --> 00:09:59,466
Tesla has all these different battery packs that would be required to power

147
00:09:59,466 --> 00:10:03,386
the data centers that SpaceX and its AI efforts were trying to push out.

148
00:10:03,386 --> 00:10:06,966
And simultaneously, SpaceX could benefit from all the humanoid robots that they

149
00:10:06,966 --> 00:10:10,856
were building to help autonomously build spaceships, to refuel rockets,

150
00:10:10,856 --> 00:10:14,376
and help achieve Elon's cadence of launching a rocket every hour,

151
00:10:14,376 --> 00:10:16,706
whatever that is, by 2032.

152
00:10:16,706 --> 00:10:21,396
And the rumor mill says that we may get an announcement as early as December

153
00:10:21,396 --> 00:10:26,786
or first week of December from Tesla announcing that they have plans to merge.

154
00:10:27,082 --> 00:10:32,249
Now, Elon has been a little vague about this, but he has been suggestive,

155
00:10:32,614 --> 00:10:35,569
and it wouldn't surprise me if this deal ends up happening.

156
00:10:35,790 --> 00:10:40,915
Now, the rumor also says that the most likely structure is SpaceX acquires Tesla

157
00:10:41,107 --> 00:10:46,134
in an all-stock deal, where Tesla shareholders get about 0.7%.

158
00:10:46,528 --> 00:10:50,384
SpaceX shares per Tesla share, plus a likely premium.

159
00:10:50,384 --> 00:10:55,654
And so Barclays says that Musk would own about 37% of the combined company with

160
00:10:55,654 --> 00:10:58,074
roughly 56% voting power.

161
00:10:58,074 --> 00:11:02,434
Now, the combined valuation of both of these companies would put it at the fifth

162
00:11:02,434 --> 00:11:06,674
largest company in the world. And listen, I have thoughts about all of this.

163
00:11:07,034 --> 00:11:12,094
I get the camp of critics that say that Elon is just doing all these different

164
00:11:12,094 --> 00:11:19,424
fancy demos to raise as much money as he can so that he can kind of keep the wheel going.

165
00:11:19,424 --> 00:11:23,644
And there's a lot of validity behind it, because although these concepts are

166
00:11:23,644 --> 00:11:27,924
very cool to watch and talk about and show because it's pretty viral, blah, blah, blah.

167
00:11:27,924 --> 00:11:31,864
The question about whether they can actually scale this to a mass audience is

168
00:11:31,864 --> 00:11:37,750
still uncertain, and it's still probabilistically a harder feat to actually pull off.

169
00:11:38,127 --> 00:11:41,194
We have to figure out the manufacturing of all of these things.

170
00:11:41,194 --> 00:11:45,064
We have to figure out getting the cost down so people can actually afford and buy it.

171
00:11:45,291 --> 00:11:49,848
You need to figure out the safety laws behind this. There has never been a flying car.

172
00:11:50,016 --> 00:11:54,126
There needs to be a lot of regulation and laws formed around these different kinds of things.

173
00:11:54,265 --> 00:11:58,774
And of course, Elon is very infamous for having delayed production times and

174
00:11:58,774 --> 00:12:03,227
the Roadster is actually the most delayed product that has come from one of Elon's companies.

175
00:12:03,871 --> 00:12:09,214
But on the other side, Elon is the only entrepreneur and innovator that is doing

176
00:12:09,214 --> 00:12:13,536
stuff like this. He's the guy that had the wacky wild idea of launching GPUs into space.

177
00:12:13,873 --> 00:12:17,321
And now it looks like it is very likely over the next couple of years.

178
00:12:17,600 --> 00:12:22,981
He was the first one to say that electric cars will become way more popular than petrol or gas cars.

179
00:12:23,295 --> 00:12:27,201
And back then, I remember, everyone thought he was crazy and he was able to pull it off.

180
00:12:27,561 --> 00:12:31,224
Everyone thought he was crazy when he said launching rockets and reusing them.

181
00:12:31,224 --> 00:12:36,170
So catching them and reusing them was impossible and he's been able to demonstrate it. Now of course.

182
00:12:36,733 --> 00:12:40,483
Demonstrating it and mass scale, as I said earlier, is a very different thing.

183
00:12:40,820 --> 00:12:45,140
But consolidating all these companies is going to help him achieve his goal

184
00:12:45,140 --> 00:12:49,380
versus keeping them as separate entities. So I'm glad to see that some of this

185
00:12:49,380 --> 00:12:50,764
consolidation might be happening.

186
00:12:51,066 --> 00:12:54,850
And it's a huge win for SpaceX, especially after their IPO and for Tesla,

187
00:12:54,850 --> 00:12:58,338
whose stock has been fairly flat over the last year.

188
00:12:58,443 --> 00:13:04,640
Now, speaking of IPOs, Anthropic is set to have a blockbuster IPO at $2.2 trillion valuation.

189
00:13:04,640 --> 00:13:10,060
And they're set to have their IPO roadshow start in about two weeks' time.

190
00:13:10,060 --> 00:13:14,450
And the idea here is they'll release their S1 prospectus, which will basically

191
00:13:14,450 --> 00:13:18,487
reveal all the details of their finances over the last year and a half to two years,

192
00:13:18,870 --> 00:13:22,320
and also their pitch as to why you should invest in their IPO,

193
00:13:22,640 --> 00:13:26,950
except the tide seemed to have turned recently on Anthropic.

194
00:13:26,950 --> 00:13:30,132
I'm showing this tweet by Andrew Curran here, where he basically goes,

195
00:13:30,527 --> 00:13:34,750
there was a meeting that happened between Anthropic and investors where they revealed some numbers.

196
00:13:34,750 --> 00:13:42,520
And the headline number is Anthropic's third quarter revenue rose 45% to $16.8 billion.

197
00:13:43,379 --> 00:13:47,560
Now, to put that into context, that is humongous. In the first quarter of this

198
00:13:47,560 --> 00:13:52,590
year, they earned about $4.3 billion, and the second quarter was $11.5 billion.

199
00:13:52,590 --> 00:13:56,086
So between the first and the second quarter, it jumped up over 100%.

200
00:13:56,439 --> 00:14:00,240
And then in this third quarter, it jumped up another 45%, except investors are

201
00:14:00,240 --> 00:14:05,060
reading that as things are slowing down, which is, I guess, accurate,

202
00:14:05,060 --> 00:14:08,537
but still a crazy take because that is a heck ton of money.

203
00:14:08,670 --> 00:14:14,640
And they are on track to hit in Q4, an ARR of about $116 billion.

204
00:14:15,200 --> 00:14:19,690
Now, that is a far cry from what Dario initially predicted at the start of the

205
00:14:19,690 --> 00:14:24,976
year of $10 billion. So Anthropik is easily one of the fastest growing companies in the world.

206
00:14:25,116 --> 00:14:30,758
And yet the market is still reading it as bearish. That Anthropica promised to keep.

207
00:14:31,554 --> 00:14:34,875
Growing their revenue, and they're currently showing a slowdown.

208
00:14:34,875 --> 00:14:39,585
And people are attributing the reasoning to losing enterprise market share to

209
00:14:39,585 --> 00:14:43,025
their competitor, OpenAI, who have launched, in their defense,

210
00:14:43,025 --> 00:14:47,685
a really good coding model that has improved drastically over the last six months.

211
00:14:47,685 --> 00:14:49,775
Codex, people have been shifting over to it.

212
00:14:49,898 --> 00:14:54,265
I think this is a bit of a nothing burger. I'm showing this tweet from John

213
00:14:54,265 --> 00:14:58,245
Dunnigan on my screen, where he basically talks about Anthropic agreeing to

214
00:14:58,245 --> 00:15:02,135
license this technology from a company called Opaque Systems.

215
00:15:02,135 --> 00:15:07,415
Now, Opaque Systems basically helps you or your models train on proprietary

216
00:15:07,415 --> 00:15:11,676
data, like from an enterprise, in a completely encrypted and secretive manner,

217
00:15:11,937 --> 00:15:14,955
which should give enterprises the confidence that they can use cloud models

218
00:15:14,955 --> 00:15:16,815
without knowing that they're going to steal from them.

219
00:15:16,815 --> 00:15:20,345
So I think this is a nothing burger. I think that Anthropics revenue is still

220
00:15:20,345 --> 00:15:24,675
soaring. And I think the fact that their revenue is down due to adjusted basis

221
00:15:24,675 --> 00:15:28,145
performance or whatever that might be, is because they have to spend so much

222
00:15:28,145 --> 00:15:29,590
money to acquire compute.

223
00:15:29,886 --> 00:15:33,935
If you've noticed, the cards or the tables have sort of changed between OpenAI

224
00:15:33,935 --> 00:15:37,565
and Anthropic. Typically, OpenAI is the one that has most of the compute and

225
00:15:37,565 --> 00:15:41,105
Anthropic has to hold back on releasing their models or rolling out to anyone

226
00:15:41,105 --> 00:15:41,665
and everyone specifically.

227
00:15:42,065 --> 00:15:46,260
But that has flipped over the last couple of weeks when OpenAI has suddenly

228
00:15:46,260 --> 00:15:50,060
become constrained and raising the prices of their subscription whilst halving

229
00:15:50,060 --> 00:15:54,220
usage capacity. And Anthropic has actually had more abundance of compute rolling

230
00:15:54,220 --> 00:15:56,548
out to as many users as they can.

231
00:15:56,815 --> 00:16:00,356
So I think that after Anthropic announced, I believe a few weeks ago,

232
00:16:00,547 --> 00:16:05,410
that they wanted to or they plan to spend over half a trillion dollars over

233
00:16:05,410 --> 00:16:10,080
the next 10 years on compute, that this is probably where the money is going

234
00:16:10,080 --> 00:16:11,410
to. But there's also another reason.

235
00:16:11,730 --> 00:16:16,050
I think that these AI labs, not just Anthropic, are very close to achieving

236
00:16:16,050 --> 00:16:18,630
something called recursive self-intelligence, where basically

237
00:16:18,961 --> 00:16:22,298
you have an AI model that can build and design the next version of itself,

238
00:16:22,560 --> 00:16:26,536
which means that this whole acceleration of AI models, the improvement of them,

239
00:16:26,826 --> 00:16:30,611
the power of them will be reached much, much quicker.

240
00:16:31,310 --> 00:16:33,490
As these models work 24-7 overnight.

241
00:16:33,490 --> 00:16:38,140
And Dwakesh on an interview this week actually confirmed, or he said,

242
00:16:38,140 --> 00:16:41,800
at least from himself, his own claim that Anthropic employees have been talking

243
00:16:41,800 --> 00:16:47,040
to him saying that RSI will be achieved as close as within the next nine months.

244
00:16:47,040 --> 00:16:50,730
OpenAI has said something pretty similar, but it's not without evidence.

245
00:16:50,730 --> 00:16:54,410
Anthropic's code changes that they've been pushing autonomously via their models,

246
00:16:54,695 --> 00:16:58,330
via their agents. And by the way, I think it's 50% of all code is currently

247
00:16:58,330 --> 00:17:01,928
being pushed autonomously or generated autonomously by their models.

248
00:17:02,294 --> 00:17:08,000
And almost 90% of all code generated has had AI models involved.

249
00:17:08,000 --> 00:17:11,600
So we're not too far off AI models just taking over the whole process,

250
00:17:11,600 --> 00:17:16,667
but where the gap is, where the delta is, is taste. These AI models...

251
00:17:17,224 --> 00:17:21,103
Whilst they're good at executing, aren't really good at figuring out what tasks

252
00:17:21,103 --> 00:17:24,771
they should do or what the design of the research is that they should go after.

253
00:17:25,316 --> 00:17:26,933
And Anthropic researchers are

254
00:17:26,933 --> 00:17:29,513
basically saying, we're going to get there within the next nine months.

255
00:17:29,751 --> 00:17:33,954
And if you look at the data, changes merged per day on Anthropic's monorepo,

256
00:17:34,204 --> 00:17:35,973
it has absolutely shut up.

257
00:17:36,453 --> 00:17:41,153
It has 10x'd since May. As this tweet says, Claude is able to autonomously iterate

258
00:17:41,153 --> 00:17:45,783
on research loops, micro-optimize complex programs, and build new services from

259
00:17:45,783 --> 00:17:49,313
the ground up at a speed and scale we haven't seen before.

260
00:17:49,313 --> 00:17:52,733
Now, speaking of Frontier Labs, one which is relatively new to the scene,

261
00:17:52,733 --> 00:17:55,813
Meta, has been doing some pretty awesome stuff in the open source world.

262
00:17:55,813 --> 00:17:58,699
So let's get into that on today's episode brought to you by Qualcomm.

263
00:18:02,421 --> 00:18:07,744
Mark Zuckerberg started his whole AI effort at Meta, pushing open source forward.

264
00:18:07,744 --> 00:18:11,184
It started with the Llama series of models and ended with Llama 4,

265
00:18:11,184 --> 00:18:13,674
which was at the time a huge disappointment.

266
00:18:13,674 --> 00:18:18,484
Since then, Meta has basically fired their entire AI team, spent upwards of

267
00:18:18,484 --> 00:18:25,004
$50 billion to acquire around 150 to 200 people and form Meta Super Intelligence

268
00:18:25,004 --> 00:18:28,429
Labs. And they've pushed out a bunch of AI products, which I've absolutely hated.

269
00:18:28,707 --> 00:18:33,164
But recently, a bunch that I absolutely love. And part of this is their series

270
00:18:33,164 --> 00:18:37,548
of models called MuseSpark. You've got 1.1, 1.2, and 1.3.

271
00:18:37,878 --> 00:18:43,284
And 1.3 is a really good model. It's a coding model and an agentic reasoning model.

272
00:18:43,800 --> 00:18:48,124
It's not necessarily frontier, but it works really well for specific niche use

273
00:18:48,124 --> 00:18:54,524
cases or specific work that you might want to kind of point and shoot an AI agent or AI model at.

274
00:18:54,524 --> 00:18:58,544
And this week, Mark Zuckerberg decided to open source that model.

275
00:18:58,544 --> 00:19:02,234
Now, think about this. This is a model that they have spent billions of dollars

276
00:19:02,234 --> 00:19:07,144
pre-training, post-training, proprietary designing, and then now releasing to

277
00:19:07,484 --> 00:19:09,050
the wide world to actually use.

278
00:19:09,363 --> 00:19:13,384
Now, there was a lot of critics, including myself, that basically thought that

279
00:19:13,764 --> 00:19:18,264
once Meta had fired and hired a new AI team, once they had failed on their initial

280
00:19:18,264 --> 00:19:20,894
open source LLAMA series, that they would not open source models.

281
00:19:20,894 --> 00:19:25,214
So it's pretty awesome to see that the American frontier of open source is so

282
00:19:25,214 --> 00:19:27,951
back and it's led by none other than Meta.

283
00:19:28,456 --> 00:19:31,695
But the coolest part about this story is that this new model,

284
00:19:31,858 --> 00:19:37,284
MuseSpark Open, can actually run on the latest Mac Studio, and they launched

285
00:19:37,284 --> 00:19:39,184
it in partnership with Apple.

286
00:19:39,184 --> 00:19:43,494
John Turnus, who is the new CEO and actually a hardware guru at Apple,

287
00:19:43,494 --> 00:19:46,714
he's been there for 18 years building some of the most popular devices that

288
00:19:46,714 --> 00:19:48,895
you and I use in our daily lives, and.

289
00:19:49,435 --> 00:19:53,383
Was the man that kind of helped engineer or make sure that the design of the

290
00:19:53,383 --> 00:19:59,693
Mac Studio was something that could befit a model as large and as big as Muse Spark Open.

291
00:19:59,693 --> 00:20:03,223
Now the stats for this thing is pretty cool. Alexander Wang basically says that

292
00:20:03,223 --> 00:20:07,493
this is an 816 billion parameter model, which is fairly big.

293
00:20:07,493 --> 00:20:10,993
And let me tell you, the American frontier for open source was way,

294
00:20:10,993 --> 00:20:14,333
way, way behind. So seeing something like this become available to anyone and

295
00:20:14,333 --> 00:20:16,563
everyone in the West is pretty awesome to see.

296
00:20:16,903 --> 00:20:19,513
Also, it's just good to see competition in general. And when you look at the

297
00:20:19,513 --> 00:20:25,423
benchmarks, it is great at GDPVAL, which is essentially general common knowledge work.

298
00:20:25,423 --> 00:20:28,316
So if you want to give it access to a tool to help you do your work,

299
00:20:28,583 --> 00:20:29,423
it's really good at that.

300
00:20:29,423 --> 00:20:32,893
But it's also good at agentic reasoning and long horizon tasks,

301
00:20:32,893 --> 00:20:36,963
which basically means if you give it a task that is meant to typically take

302
00:20:36,963 --> 00:20:41,063
days, it can actually keep working on it and have between a 40 to 50% chance

303
00:20:41,063 --> 00:20:42,573
of actually getting it right.

304
00:20:42,573 --> 00:20:44,726
And I expect to see that stat increase in.

305
00:20:45,185 --> 00:20:50,764
Many times over as these models start to improve, which begs the question, how soon will that be?

306
00:20:50,764 --> 00:20:55,184
Well, technically, Meta has released seven new models over the last three months.

307
00:20:55,184 --> 00:21:00,689
So that's a new model every, I think it's 11 or 13 days, which is absolutely insane.

308
00:21:00,979 --> 00:21:05,364
Because just a year ago, we wouldn't get a new model until we waited for three months.

309
00:21:05,364 --> 00:21:09,924
And that was from the Frontier Lab. So for Meta to be so far behind and catch

310
00:21:09,924 --> 00:21:12,064
up to where they are today is really awesome.

311
00:21:12,064 --> 00:21:15,214
And I'm glad to see open source winning. And then the final item on the docket

312
00:21:15,214 --> 00:21:20,464
is Google generally releasing their new Gemini 4 model, Argon.

313
00:21:20,464 --> 00:21:25,124
Now, I spoke about this last week when I had a conversation around whether Google

314
00:21:25,124 --> 00:21:26,614
can actually pull off a comeback.

315
00:21:26,614 --> 00:21:31,304
They had the lead or they were near having the lead at the end of last year,

316
00:21:31,304 --> 00:21:33,114
which was not that long ago.

317
00:21:33,114 --> 00:21:36,454
But since then, they underwent a lot of problems at the company.

318
00:21:36,454 --> 00:21:41,174
The DNA of the company decided to spend resources pretty haphazardly.

319
00:21:41,174 --> 00:21:46,294
They decided to build models that weren't really engineered around what customers

320
00:21:46,294 --> 00:21:49,184
or the market actually desired. They didn't focus on enterprise.

321
00:21:49,184 --> 00:21:50,496
They didn't really focus on consumer.

322
00:21:50,705 --> 00:21:53,754
They built many different models that worked in different subscription packages.

323
00:21:53,754 --> 00:21:58,004
It was a whole palaver and mess, which ended up in Demis Sissibis stepping away

324
00:21:58,004 --> 00:22:01,025
from the company or stepping away from his AI role at the company.

325
00:22:01,379 --> 00:22:06,882
And also a lot of key staff like Noam Shazir leaving to OpenAI and others leaving to Anthropic.

326
00:22:07,114 --> 00:22:10,696
So Google was in a bit of a precarious situation. But they have since released,

327
00:22:10,882 --> 00:22:14,534
finally, after delaying their model, Gemini Argon 4. And the reviews.

328
00:22:14,894 --> 00:22:16,774
I have to say, are actually quite good.

329
00:22:17,017 --> 00:22:21,404
My main concern for this model was that it was benchmarked. Because when I looked

330
00:22:21,404 --> 00:22:25,696
at the benchmarks last week, it looked pretty damn impressive.

331
00:22:25,905 --> 00:22:29,074
But I had to question how on earth they were able to pull this off.

332
00:22:29,074 --> 00:22:32,074
Like, what did it actually feel like to use this product? Now.

333
00:22:32,514 --> 00:22:33,660
I'll be the first to put my hands up.

334
00:22:34,014 --> 00:22:37,852
I don't know yet. I'm going to spend the weekend actually using the product.

335
00:22:38,136 --> 00:22:41,612
And if it's good or if it's bad, I'm going to give you a review next week.

336
00:22:41,612 --> 00:22:45,742
So stay tuned for that episode. But from the feedback that has been given from

337
00:22:45,742 --> 00:22:49,862
people, Ethan Mollick goes, I spent a decent amount of time with Gemini 4 before

338
00:22:49,862 --> 00:22:53,778
launch. Overall, Gemini 4 is a big jump over previous Gemini models.

339
00:22:53,999 --> 00:22:59,082
And it is the first non open air anthropic model to pass the fable Mythos Astrobar.

340
00:22:59,082 --> 00:23:01,325
So he's basically saying this thing is pretty damn good.

341
00:23:01,818 --> 00:23:05,541
It is also much more pleasant and more agentic than previous Gemini models.

342
00:23:05,541 --> 00:23:08,355
Basically, the taste of the model, it's much more easier to deal with.

343
00:23:08,703 --> 00:23:12,641
In fact, a big advantage is that it has a very different personality and sense

344
00:23:12,641 --> 00:23:15,381
of taste than Claude or Astra, which is a big deal.

345
00:23:15,381 --> 00:23:18,521
So when I'm using Claude, it is very verbose.

346
00:23:18,521 --> 00:23:24,401
It says way too many words, and I have to read an essay every time I have to get an answer from it.

347
00:23:24,401 --> 00:23:28,551
Now, with Astra, I don't have to read an essay, but it is quite sycophantic

348
00:23:28,551 --> 00:23:31,271
when I don't want it to be. It's quite agreeable when I don't want it to be.

349
00:23:31,271 --> 00:23:35,325
I want it to push back more. Ethan basically says that Gemini is not like that.

350
00:23:35,592 --> 00:23:40,071
But he also says that it's still not the best model to use when it comes to

351
00:23:40,071 --> 00:23:42,561
coding, when it comes to the majority of his work.

352
00:23:42,561 --> 00:23:47,171
But it's good to sort of bring into the fold and kind of get it to acclimatize

353
00:23:47,171 --> 00:23:48,151
to whatever work you're doing.

354
00:23:48,151 --> 00:23:52,031
And maybe in some cases, it actually benefits or excels. Justine Moore from

355
00:23:52,041 --> 00:23:56,651
A16Z said that it has really good visual intelligence.

356
00:23:56,651 --> 00:24:00,171
And what you're seeing on your screen here is she asked it to create an interactive

357
00:24:00,171 --> 00:24:05,341
museum for Apple computers. And it is pretty damn impressive to see. Am I making sense?

358
00:24:05,661 --> 00:24:09,161
It's just a crazy world. But Gemini is really pushing the frontier forwards.

359
00:24:09,161 --> 00:24:12,761
And, you know, we'll see once I get my hands on it, whether we can do something cool here.

360
00:24:12,761 --> 00:24:16,631
And that is it for the roundup. I will be back in the hot seat next week,

361
00:24:16,631 --> 00:24:19,271
maybe drop the kimono and wear something a little more normal.

362
00:24:19,271 --> 00:24:22,941
But thank you so much for listening. I love interacting with you guys.

363
00:24:22,941 --> 00:24:25,728
I love reading all your comments, whether it's good, bad.

364
00:24:26,036 --> 00:24:28,921
Criticism is always welcome. I'm always trying to figure out different ways

365
00:24:28,921 --> 00:24:32,761
of improving. So, you know, send me a DM, just even say hello or suggestions

366
00:24:32,761 --> 00:24:34,081
on any other topics that you have.

367
00:24:34,534 --> 00:24:37,501
And yeah, I guess that is it. I will see you on the next one.

368
00:24:37,501 --> 00:24:40,478
Thank you for listening to today's episode of In The Loop brought to you by Qualcomm.