1
00:00:00,080 --> 00:00:03,400
For over 60 years, we've been 
promised robots that can do 

2
00:00:03,400 --> 00:00:05,600
almost anything. 
I think robotics has given 

3
00:00:05,600 --> 00:00:09,520
itself a reputation of being one
of the biggest technology over 

4
00:00:09,520 --> 00:00:12,200
promises ever. 
But right now, something's 

5
00:00:12,200 --> 00:00:15,720
different. 
In this month, October 2025 / 6 

6
00:00:15,800 --> 00:00:18,800
billion dollars have been poured
into robotics by some of the 

7
00:00:18,800 --> 00:00:22,560
biggest names in investing, and 
they've claimed that their 

8
00:00:22,560 --> 00:00:26,120
future is physical AI. 
By the end of today's episode, 

9
00:00:26,440 --> 00:00:29,720
you'll start to understand 
what's actually changed, why 

10
00:00:29,720 --> 00:00:32,320
people are putting so much money
into robotics all of a sudden, 

11
00:00:32,840 --> 00:00:35,800
and whether we're at the 
beginning of an iPhone era 

12
00:00:35,800 --> 00:00:39,080
moment or just the start of 
another period of 

13
00:00:39,080 --> 00:00:41,520
disappointment. 
This is another loop with Jack 

14
00:00:41,520 --> 00:00:43,160
Horton. 
I hope you enjoy the show. 

15
00:00:55,950 --> 00:00:59,470
So let's start by, I guess, 
setting the scene a little. 

16
00:01:00,190 --> 00:01:05,790
I think ever since the Jetsons 
in I think the 1960s or 70s had 

17
00:01:05,790 --> 00:01:10,520
this like robot made, it's been 
promised by I guess 

18
00:01:10,520 --> 00:01:13,520
technologists and robot makers 
that these are the future of 

19
00:01:13,520 --> 00:01:16,760
household helpers. 
And in fact, I think experts, 

20
00:01:16,760 --> 00:01:20,400
you know, ever since the 1950s 
have been pretty confident in 

21
00:01:20,400 --> 00:01:23,840
predicting that, you know, by 
the year 2000, most American 

22
00:01:23,840 --> 00:01:27,400
homes will have some form of 
automation to make coffee or 

23
00:01:27,400 --> 00:01:29,520
close windows. 
But I don't think we've ever 

24
00:01:29,520 --> 00:01:33,120
reached that cool and exciting 
future because none of this has 

25
00:01:33,120 --> 00:01:35,200
happened, really, other than the
ability, I guess, to use a 

26
00:01:35,200 --> 00:01:38,040
mobile phone app to shut your 
windows in certain hotels. 

27
00:01:38,880 --> 00:01:43,040
And also, you can't really 
forget the awful, like, robot 

28
00:01:43,040 --> 00:01:45,600
hoovers and lawn mowers that you
get in some places. 

29
00:01:45,720 --> 00:01:50,040
So then why all of a sudden is 
someone pouring $6 billion into 

30
00:01:50,040 --> 00:01:52,720
robotics ventures? 
Because that's what's happened 

31
00:01:52,720 --> 00:02:07,560
this month. 
SoftBank pretty famous for 

32
00:02:07,720 --> 00:02:10,680
throwing away money at times, 
but they've made the headlines 

33
00:02:10,680 --> 00:02:15,920
recently because they've put 
$5.4 billion alone to acquire a 

34
00:02:15,960 --> 00:02:19,840
massive robot division with the 
CEO of SoftBank saying that 

35
00:02:19,840 --> 00:02:22,520
Softbank's next frontier is 
physical AI. 

36
00:02:22,800 --> 00:02:26,720
And Elon Musk in the last month 
for XAI, such as AI startup 

37
00:02:26,720 --> 00:02:31,560
that's inside of X and Twitter 
has actually started hiring top 

38
00:02:31,560 --> 00:02:34,920
NVIDIA researchers to develop 
what is described as world 

39
00:02:34,920 --> 00:02:37,280
models. 
So these are like AI language 

40
00:02:37,280 --> 00:02:41,520
models and systems that I guess 
give robots an intuitive 

41
00:02:41,520 --> 00:02:45,040
understanding of physical world 
objects and gravity and 

42
00:02:45,400 --> 00:02:47,040
basically how the world just 
works. 

43
00:02:47,240 --> 00:02:50,000
There's also been big humanoid 
robots making strides in the 

44
00:02:50,000 --> 00:02:51,480
world. 
For anyone that's been on 

45
00:02:51,480 --> 00:02:55,240
Instagram and is connected to 
technology accounts, really 

46
00:02:55,480 --> 00:02:57,640
you've just seen the robot 
called Figure. 

47
00:02:58,200 --> 00:03:01,920
So Figure O3 is a humanoid robot
that's basically aimed at 

48
00:03:02,480 --> 00:03:06,600
households to do general chores.
And in their videos, you can go 

49
00:03:06,600 --> 00:03:08,840
YouTube. 
This Figures robot has been, you

50
00:03:08,840 --> 00:03:11,640
know, picking up clutter or 
taking clothes out the washer 

51
00:03:12,160 --> 00:03:15,360
and loading dishwashers. 
And obviously as a result, a lot

52
00:03:15,360 --> 00:03:18,960
of the ecosystem around robotics
is racing to support these 

53
00:03:18,960 --> 00:03:22,480
smarter robots. 
So NVIDIA have announced Project

54
00:03:22,480 --> 00:03:24,920
Groot. 
So Groups is basically platform 

55
00:03:24,920 --> 00:03:28,320
for training robots, they call 
it for robot intelligence. 

56
00:03:28,680 --> 00:03:32,720
And CEO of NVIDIA, Jensen Huang 
actually called this the 

57
00:03:32,720 --> 00:03:35,760
foundation for making robots. 
One of the most exciting 

58
00:03:35,760 --> 00:03:39,680
problems to solve in AI today. 
And there are lots of new 

59
00:03:39,680 --> 00:03:43,200
breakthroughs coming out to 
allow robotics to take giant 

60
00:03:43,200 --> 00:03:46,960
leaps forward because of AI. 
So as you can imagine, Nvidia's 

61
00:03:46,960 --> 00:03:49,520
teamed up with basically every 
single robotics firm in the 

62
00:03:49,520 --> 00:03:53,880
world from Agility to Boston 
Dynamics to start-ups like 

63
00:03:53,880 --> 00:03:56,840
Figure and Sanctuary. 
So obviously there's a, there's 

64
00:03:56,840 --> 00:03:59,160
a renewed focus on robotics 
right now. 

65
00:04:00,040 --> 00:04:02,440
And there's a good question like
why, why is this happening? 

66
00:04:02,840 --> 00:04:05,200
You know, why is this time 
different than all the others? 

67
00:04:05,760 --> 00:04:07,760
And I think it's a really 
interesting question to ask. 

68
00:04:07,960 --> 00:04:09,720
And in one word, it's because of
AI. 

69
00:04:09,920 --> 00:04:13,280
Not the term AI that used to be 
used over the last few decades, 

70
00:04:13,480 --> 00:04:16,320
but the specific recent 
breakthroughs in large language 

71
00:04:16,320 --> 00:04:19,680
models because they've 
completely transformed the way 

72
00:04:19,680 --> 00:04:23,600
that vision and language is 
interpreted by machines. 

73
00:04:24,400 --> 00:04:27,880
So the big bet right now is that
these same breakthroughs can 

74
00:04:27,880 --> 00:04:32,720
give robotics the adaptability 
and I guess St. smarts that 

75
00:04:32,920 --> 00:04:35,880
they've historically lacked. 
So as soon as a human child has 

76
00:04:36,120 --> 00:04:39,400
been birthed into the world, 
they can learn things by just 

77
00:04:39,400 --> 00:04:42,640
picking up a toy or observing 
people through trial and error. 

78
00:04:43,720 --> 00:04:46,720
But yet pre twenty 20s robots 
just cannot do that. 

79
00:04:46,880 --> 00:04:50,800
They're either explicitly coded 
step by step or trained with 

80
00:04:50,800 --> 00:04:53,080
very narrow data sets. 
Which means soon as it 

81
00:04:53,080 --> 00:04:55,360
encounters something new IT just
gets confused. 

82
00:04:55,680 --> 00:04:58,320
Whereas new large language 
models, as most of us have 

83
00:04:58,640 --> 00:05:02,120
experienced, are incredibly 
adaptable and have a huge amount

84
00:05:02,120 --> 00:05:04,920
of training data. 
So they've really turned this 

85
00:05:05,200 --> 00:05:10,000
old world robotics, I guess idea
on its head into something that 

86
00:05:10,000 --> 00:05:12,080
now is potentially very 
plausible and exciting. 

87
00:05:12,200 --> 00:05:16,160
For example, Google's DeepMind 
robotic centre model wasn't 

88
00:05:16,160 --> 00:05:19,560
trained to do 1 specific task. 
It was trained to integrate all 

89
00:05:19,560 --> 00:05:21,680
of its vision and language 
capabilities. 

90
00:05:22,040 --> 00:05:25,760
So this robot is able to 
interpret commands and map them 

91
00:05:25,760 --> 00:05:27,040
to what actions they should 
take. 

92
00:05:27,160 --> 00:05:31,080
So in a demo, when the robot was
asked to hand me a green 

93
00:05:31,120 --> 00:05:34,120
vegetable from the fridge, a 
Gemini robot arm was able to 

94
00:05:34,120 --> 00:05:36,960
understand this is a vegetable 
that's green, this is a fridge 

95
00:05:37,320 --> 00:05:39,080
and hand the person the 
vegetable. 

96
00:05:39,280 --> 00:05:42,840
And coming back to that Xai and 
Elon Musk's idea of a world 

97
00:05:42,840 --> 00:05:45,000
model. 
This is where models suddenly 

98
00:05:45,000 --> 00:05:48,840
can teach robots physics. 
So a world model is essentially 

99
00:05:48,840 --> 00:05:52,800
an internal simulation where AI 
can start to predict outcomes of

100
00:05:52,800 --> 00:05:55,600
actions before taking them 
because they just intuitively 

101
00:05:55,600 --> 00:05:58,120
understand the world. 
It just adds a massive layer of 

102
00:05:58,120 --> 00:06:02,760
reasoning instead of just reflex
or hard coded instructions. 

103
00:06:02,920 --> 00:06:06,360
Another massive breakthrough has
been training data and 

104
00:06:06,720 --> 00:06:10,400
exponential data at scale. 
Obviously companies are now 

105
00:06:10,400 --> 00:06:12,680
using techniques like 
reinforcement learning in 

106
00:06:12,680 --> 00:06:16,520
virtual environments. 
So imagine virtual robots in a 

107
00:06:16,520 --> 00:06:19,480
video game that are all 
essentially characters trying 

108
00:06:19,480 --> 00:06:23,600
tasks in a simulator so they can
learn from failures and share 

109
00:06:23,600 --> 00:06:27,160
knowledge, but without actually 
suffering a negative 

110
00:06:27,160 --> 00:06:29,960
consequence. 
A good example of this might be 

111
00:06:29,960 --> 00:06:32,240
learning and teaching them how 
to stack dishes. 

112
00:06:32,640 --> 00:06:36,120
And if they fail, they haven't 
broken edition real life, it's 

113
00:06:36,120 --> 00:06:38,400
just a video game. 
But all of that data that 

114
00:06:38,400 --> 00:06:42,000
they're capturing through that 
trial and error process is 

115
00:06:42,000 --> 00:06:44,880
maintained in its memory for 
future interactions. 

116
00:06:45,000 --> 00:06:47,160
And this is really exciting 
because, you know, consider hand

117
00:06:47,160 --> 00:06:51,040
eye coordination and just 
general object recognition tasks

118
00:06:51,040 --> 00:06:56,000
like distinguishing a open door 
from an open window used to be 

119
00:06:56,000 --> 00:06:57,560
something to be quite difficult 
for a robot. 

120
00:06:57,560 --> 00:07:01,640
But modern vision models and all
this innovations in training 

121
00:07:01,640 --> 00:07:04,840
data makes this type of 
recognition something that's 

122
00:07:04,840 --> 00:07:07,040
very plausible. 
And a good example of this is 

123
00:07:07,040 --> 00:07:11,600
that Figures CEO Brett Adcock 
actually claimed that their 

124
00:07:11,600 --> 00:07:15,480
internal AI robot learnt how to 
fold a towel from just 80 hours 

125
00:07:15,480 --> 00:07:19,600
of training data, which is 
something that was unimaginable 

126
00:07:19,600 --> 00:07:21,360
10 years ago. 
And we also have things like 

127
00:07:21,360 --> 00:07:24,560
integrations and connectivity 
being 10 times better than say 

128
00:07:24,560 --> 00:07:27,880
1015 years ago. 
We have cloud, Internet things 

129
00:07:27,880 --> 00:07:29,760
and just frameworks that didn't 
exist before. 

130
00:07:30,440 --> 00:07:33,760
A good example is a robot could 
offload really heavy 

131
00:07:33,760 --> 00:07:37,600
computations to the cloud and it
could get over the air updates 

132
00:07:37,600 --> 00:07:41,560
like software updates. 
So when robots learn something 

133
00:07:41,560 --> 00:07:44,320
in the world, they could then 
force a software update to all 

134
00:07:44,320 --> 00:07:46,400
of them, then learn something. 
So yeah, there's been a lot of 

135
00:07:46,400 --> 00:07:49,760
interesting breakthroughs and 
and partially explains the rise 

136
00:07:49,760 --> 00:07:52,600
in excitement and investment 
levels into the field. 

137
00:07:52,680 --> 00:07:55,000
However, and it is a big 
however, it's just never that 

138
00:07:55,000 --> 00:08:07,920
simple. 
Because I think before we get 

139
00:08:07,920 --> 00:08:11,320
carried away with all these 
exciting demos, it's important 

140
00:08:11,320 --> 00:08:15,160
to note some of the real big 
challenges because even today's 

141
00:08:15,160 --> 00:08:18,880
most advanced robots are still 
struggling with basic tasks in 

142
00:08:18,880 --> 00:08:23,040
very controlled environments. 
A Time reporter note that during

143
00:08:23,080 --> 00:08:26,440
figures, that robot figure that 
I mentioned earlier during their

144
00:08:26,440 --> 00:08:30,640
towel folding test, that robots 
were caught frequently dropping 

145
00:08:30,640 --> 00:08:33,760
things, being unable to pick the
towers up, and having to have an

146
00:08:33,760 --> 00:08:36,360
engineer come sort things out. 
I'm not just like criticizing 

147
00:08:36,360 --> 00:08:38,760
them for the sake of it or 
trying to find gotcha moments. 

148
00:08:39,080 --> 00:08:41,960
These are just real challenges 
that they really need to 

149
00:08:41,960 --> 00:08:44,520
overcome if this is something 
that's going to generate revenue

150
00:08:44,520 --> 00:08:47,840
and deliver value in a house. 
I think this really speaks to 

151
00:08:47,840 --> 00:08:51,440
hype cycles because historically
there's been lots of AI and 

152
00:08:51,440 --> 00:08:54,600
robotics hype cycles. 
Hype builds big expectations 

153
00:08:54,600 --> 00:08:58,920
form and then obviously reality 
triggers a big crash to Earth 

154
00:08:59,000 --> 00:09:02,640
and a collapse in funding. 
We saw this in the late 1880s 

155
00:09:02,640 --> 00:09:06,400
and then again in the 2000s when
consumer robots with those, you 

156
00:09:06,400 --> 00:09:09,240
know, rubbish Hoover robots 
popped up everywhere and 

157
00:09:09,480 --> 00:09:11,600
everyone thinking that was going
to be the next big thing. 

158
00:09:11,800 --> 00:09:16,200
But after 2015, when lots of 
high profile robotics companies 

159
00:09:16,200 --> 00:09:20,000
were shut down, then investors 
grew quite rare, wary and and 

160
00:09:20,000 --> 00:09:21,960
just refused to put money into 
the sector. 

161
00:09:22,160 --> 00:09:24,440
Even some of the really exciting
companies that you've probably 

162
00:09:24,440 --> 00:09:27,600
seen all over YouTube, Boston 
Dynamics, you know, that robot 

163
00:09:27,600 --> 00:09:30,480
that's kind of painted yellow, 
the dog that runs around, the 

164
00:09:30,480 --> 00:09:32,960
robot that does backflips that 
you can't push over. 

165
00:09:33,560 --> 00:09:36,360
I mean, their videos have been 
super viral, but they've still 

166
00:09:36,360 --> 00:09:41,560
not made any money. 
Google bought it in 2013, but 

167
00:09:41,760 --> 00:09:44,520
finding no near term product 
market fit, they sold it to 

168
00:09:44,520 --> 00:09:48,200
SoftBank in 2017. 
And SoftBank then sold it to 

169
00:09:48,200 --> 00:09:51,000
Hyundai in 2020 for about $1.1 
billion. 

170
00:09:51,720 --> 00:09:54,560
And the real issue is that 
Boston Dynamics never found any 

171
00:09:54,560 --> 00:09:58,160
mass market applications. 
They started selling Spot the 

172
00:09:58,160 --> 00:10:04,440
Robot Dog in 2019, but it cost 
$75,000 each and really it was 

173
00:10:04,440 --> 00:10:06,440
just doing inspections around 
building sites. 

174
00:10:06,560 --> 00:10:08,960
Atlas, the humanoid robot was 
the exact same. 

175
00:10:08,960 --> 00:10:12,520
It was a very expensive 
prototype with no rear clear 

176
00:10:12,720 --> 00:10:15,760
path to being usable by the mass
market. 

177
00:10:16,000 --> 00:10:18,280
It's just too early. 
And so yeah, I think we can sum 

178
00:10:18,280 --> 00:10:20,920
up robotics challenges into 3 
core buckets. 

179
00:10:21,400 --> 00:10:24,480
First is real world performance 
is just completely unproven. 

180
00:10:24,920 --> 00:10:27,720
You know, robotics demos are 
notorious for being very 

181
00:10:27,720 --> 00:10:32,200
controlled with mock kitchens, 
engineers, even controllers. 

182
00:10:32,760 --> 00:10:36,720
And a task like unloading a 
dishwasher that might take us 2 

183
00:10:36,720 --> 00:10:41,920
to 5 minutes might take a robot 
20 minutes or even longer to 

184
00:10:41,920 --> 00:10:44,560
train to do it. 
Assume it doesn't get stuck or 

185
00:10:44,560 --> 00:10:46,840
break all the glass. 
Second robot is just very 

186
00:10:46,840 --> 00:10:49,040
expensive. 
You know Agility Robotics has a 

187
00:10:49,040 --> 00:10:52,120
humanoid robot and it costs 
about 150 K. 

188
00:10:52,920 --> 00:10:57,680
Tesla's Optimus has tried to 
target A20K price tag, but 

189
00:10:57,680 --> 00:11:01,120
that's going to be years away 
and it's current prices. 

190
00:11:01,400 --> 00:11:04,320
The return on investment for 
this general purpose robot is 

191
00:11:04,320 --> 00:11:07,480
completely unproven. 
We could be decades away from 

192
00:11:07,480 --> 00:11:10,160
this being a reality. 
And 3rd, the struggle with real 

193
00:11:10,160 --> 00:11:14,720
edge cases and generalizations. 
A robot might do a really good 

194
00:11:14,720 --> 00:11:20,160
job of folding some towels, but 
when that towel is wet or if 

195
00:11:20,160 --> 00:11:23,360
it's got lots of tassels on it, 
then often a robot will get 

196
00:11:23,360 --> 00:11:26,640
quite confused. 
So it's got a lot of training, 

197
00:11:26,640 --> 00:11:31,720
data and innovation still to 
take place before a robot can 

198
00:11:31,720 --> 00:11:33,760
really adapt to real world 
scenarios. 

199
00:11:45,150 --> 00:11:48,030
So yeah, to conclude, if I ask 
the question of do I think AI 

200
00:11:48,030 --> 00:11:50,950
powered robots are here to stay 
and ready to empty the 

201
00:11:50,950 --> 00:11:52,950
dishwasher? 
No. 

202
00:11:53,350 --> 00:11:54,950
After digging through all the 
evidence, the answer is 

203
00:11:54,950 --> 00:11:57,150
definitely no. 
Or maybe I'm being too harsh. 

204
00:11:57,150 --> 00:12:00,440
Maybe it's not quite yet, but 
what I would say is that this 

205
00:12:00,440 --> 00:12:03,240
moment starts to feel a little 
bit different because there are 

206
00:12:03,240 --> 00:12:06,240
real big technology 
breakthroughs which overcome 

207
00:12:06,240 --> 00:12:08,960
some of the major challenges 
robotics has faced over the last

208
00:12:09,520 --> 00:12:12,280
50-60 years. 
There's also huge investment 

209
00:12:12,280 --> 00:12:15,160
going into the sector and many 
companies that were badly burned

210
00:12:15,160 --> 00:12:17,560
before are starting to come back
to the table because I think 

211
00:12:17,560 --> 00:12:21,400
most people all know, 
acknowledge, recognise that it 

212
00:12:21,400 --> 00:12:23,360
might take five years, it might 
take 50. 

213
00:12:23,360 --> 00:12:27,680
But robotics is going to be a 
major leap and one of the most 

214
00:12:28,080 --> 00:12:31,080
influential piece of technology 
in human society. 

215
00:12:32,240 --> 00:12:34,360
And really, if you think about 
how technology is typically 

216
00:12:34,360 --> 00:12:38,920
spread, I expect industrial 
adoption first, then gradual 

217
00:12:38,920 --> 00:12:41,920
home integration over the next 6
to 15 years. 

218
00:12:41,920 --> 00:12:44,600
And I know that's a big window, 
but it's so hard to tell. 

219
00:12:45,280 --> 00:12:47,040
History then gives us a lot of 
clues. 

220
00:12:47,040 --> 00:12:51,320
Smartphones went from about 0 to
50% global adoption in about 10 

221
00:12:51,320 --> 00:12:53,680
years. 
Personal computers took 15 to 20

222
00:12:53,680 --> 00:12:56,120
years. 
Even robot vacuums launched in 

223
00:12:56,120 --> 00:13:01,880
2002 took 20 years to reach 14% 
of US households by 2023. 

224
00:13:02,040 --> 00:13:04,360
So yeah, robots and the 
technology surrounding them are 

225
00:13:04,360 --> 00:13:07,560
getting bigger and more well 
funded, with a lot of clever 

226
00:13:07,560 --> 00:13:09,040
people being attracted to the 
sector. 

227
00:13:09,200 --> 00:13:11,840
But it's still far off. 
And in many ways, the real 

228
00:13:11,840 --> 00:13:14,520
questions we should start to ask
ourselves throughout this next 

229
00:13:14,520 --> 00:13:17,760
period is, are we as a society 
ready for them? 

230
00:13:18,520 --> 00:13:20,680
I think that's an even bigger 
and more important question. 

231
00:13:20,880 --> 00:13:23,480
One for another day. 
But anyway, that's it for today.

232
00:13:23,480 --> 00:13:26,320
I hope you enjoyed the 
discussion and I'll see you next

233
00:13:26,320 --> 00:13:26,440
week.