Building Matilda

WEBVTT
1
00:00:00.720 --> 00:00:04.679
The Tech Policy Design Institute acknowledges and pays our respects
2
00:00:04.799 --> 00:00:09.759
to all First Nations people. We recognize and celebrate that,
3
00:00:09.880 --> 00:00:15.119
among many things, Indigenous people were Australia's first tech innovators.
4
00:00:18.199 --> 00:00:19.960
You know, we'll keep the good fight going. We're going
5
00:00:20.000 --> 00:00:22.000
to keep building Matilda. And it's two am in the lab.
6
00:00:22.120 --> 00:00:24.879
Let's do the hard thing. Let's write code right and
7
00:00:24.920 --> 00:00:27.280
we built the code and the ship cycle that led
8
00:00:27.320 --> 00:00:29.559
up to Matilda was absolutely the hardest I've ever done
9
00:00:29.600 --> 00:00:33.920
in my life. If we say we're an Australian made
10
00:00:33.960 --> 00:00:36.799
AI company, that means something. We're here to build a
11
00:00:37.159 --> 00:00:41.719
sustainable AI industry for the next generation that's going to
12
00:00:42.039 --> 00:00:44.520
last the test of time, right ten years from now
13
00:00:44.799 --> 00:00:48.479
when America goes Actually, we don't want Australia to be
14
00:00:48.520 --> 00:00:50.880
that smart. What are we going to do? Grovel on
15
00:00:50.920 --> 00:00:53.920
our hands, in our feet and go please. And that's
16
00:00:54.000 --> 00:00:56.679
my biggest fear is I don't want my kids to
17
00:00:56.759 --> 00:01:01.079
be disadvantaged in a world where AI absolutely leaves the differentiated.
18
00:01:16.599 --> 00:01:20.079
Welcome to Tech Mirror. I'm your host Johannah Weaver, and
19
00:01:20.159 --> 00:01:23.200
this is the podcast where we talk about how technology
20
00:01:23.400 --> 00:01:27.280
is shaping our world and how we the humans can
21
00:01:27.319 --> 00:01:31.239
shape technology back now. I'm not sure if maybe the
22
00:01:31.239 --> 00:01:34.640
PM's been listening into the podcast, but on the fifteenth
23
00:01:34.719 --> 00:01:38.159
of July he gave a landmark speech all about how
24
00:01:38.200 --> 00:01:42.040
we can shape AI in Australia's interests and I was
25
00:01:42.120 --> 00:01:45.400
really interested in the way that speech was framed. It
26
00:01:45.480 --> 00:01:47.760
was very familiar in the way that he was talking
27
00:01:47.799 --> 00:01:51.680
about Australia needing to shape our future rather than letting
28
00:01:51.719 --> 00:01:55.200
that future shape us. And in this speech he talked
29
00:01:55.200 --> 00:01:59.719
about how and here I'm quoting him, Australia's power, our
30
00:01:59.760 --> 00:02:04.879
aid agency, our choice lies in embracing change and shaping it.
31
00:02:05.599 --> 00:02:11.159
Not just adopting or accommodating AI, but designing it, making
32
00:02:11.199 --> 00:02:15.280
it and building the capability right here, and today's guest
33
00:02:15.439 --> 00:02:19.120
is doing just that. Dave Lempers is co founder and
34
00:02:19.240 --> 00:02:22.639
CEO of main Code and together with a small but
35
00:02:22.840 --> 00:02:27.479
very mighty team, he has built Matilda, an Australian developed
36
00:02:27.560 --> 00:02:31.120
AI platform that was beata launched just days before the
37
00:02:31.199 --> 00:02:34.400
PM speech. So, Dave, welcome to Tech Mirror.
38
00:02:34.599 --> 00:02:35.439
Thank you for having me.
39
00:02:36.120 --> 00:02:40.919
So, Dave, you've spent years building technology for international companies
40
00:02:41.039 --> 00:02:45.719
like Microsoft or AWS, Amazon web servers, but also for
41
00:02:45.759 --> 00:02:51.800
Australian company Easygo, which spans gaming, crypto and gambling, and
42
00:02:52.199 --> 00:02:56.319
then in twenty twenty four you established main Code. What
43
00:02:56.479 --> 00:02:59.280
made you take the leap and establish an AI company
44
00:02:59.319 --> 00:03:04.639
of your own and to do it in Australia.
45
00:03:03.080 --> 00:03:05.719
So that story goes back a little bit too when
46
00:03:05.759 --> 00:03:11.159
I first started working in Australia. So I graduated from Swinburne,
47
00:03:11.319 --> 00:03:15.800
you know, with like robotics hardware background, and started working
48
00:03:15.800 --> 00:03:18.039
in industry for a couple of years. Did my first
49
00:03:18.080 --> 00:03:20.080
startup when I was very young, in my early twenties,
50
00:03:20.599 --> 00:03:23.400
and found it very frustrating to be in Australia where
51
00:03:23.719 --> 00:03:25.680
people felt, you know I used to say, right like,
52
00:03:25.680 --> 00:03:27.719
they just wanted to get it out of the ground
53
00:03:27.759 --> 00:03:29.879
or office sheeps back. But if you were building something
54
00:03:29.919 --> 00:03:33.439
really cool technically, no one really cared, and so bootstrapping
55
00:03:33.560 --> 00:03:35.759
was the norm. You kind of had to do every
56
00:03:35.800 --> 00:03:40.599
deal paycheck to paycheck. And when the opportunity came up
57
00:03:40.639 --> 00:03:43.240
to go join Windows Azure right in as kind of
58
00:03:43.240 --> 00:03:45.560
a very early engineer on that team, I was like,
59
00:03:45.599 --> 00:03:48.400
I'm super excited to go experience the US. When I
60
00:03:48.439 --> 00:03:51.120
got there, I got to see what real hardcore engineering
61
00:03:51.159 --> 00:03:53.599
looked like and that tip of the spear thinking, but
62
00:03:53.680 --> 00:03:56.159
specifically are on innovation right, and that you've got to
63
00:03:56.159 --> 00:04:00.879
go out into this very uncomfortable cold frontier because that's
64
00:04:00.919 --> 00:04:03.199
where you're going to discover and solve real hard problems.
65
00:04:03.199 --> 00:04:05.400
But you've got to have a thick skin because people
66
00:04:05.439 --> 00:04:08.879
are going to stand on the shore and they're gonna say, oh, like,
67
00:04:09.080 --> 00:04:12.319
look at these idiots going off into the frontier into
68
00:04:12.319 --> 00:04:14.159
certain death. And you've got to have a kind of
69
00:04:14.199 --> 00:04:17.040
conviction about why you're doing it. And I think when
70
00:04:17.040 --> 00:04:19.399
I got back to Australia and met Ed you know,
71
00:04:19.560 --> 00:04:22.360
and was working for it at Easygo, we would talk
72
00:04:22.399 --> 00:04:25.279
a lot about the fact that Australia hadn't really had
73
00:04:25.279 --> 00:04:27.360
its moment yet, you know, you know, you'd have the
74
00:04:27.480 --> 00:04:29.319
ed Lassians and all that, but let's be frank, like,
75
00:04:29.639 --> 00:04:32.959
they're not you know, tech juggernauts who've stimulated the technical
76
00:04:33.000 --> 00:04:36.120
ecosystem in Australia at all. And I think their moves
77
00:04:36.160 --> 00:04:39.160
overseas show you that that's where they ultimately wanted to
78
00:04:39.199 --> 00:04:43.800
take technology in Australia. So we said, what if we
79
00:04:43.920 --> 00:04:48.560
built a silken Valley minded startup and did the hard
80
00:04:48.600 --> 00:04:51.480
thing right? So main codes principles. One of the biggest
81
00:04:51.480 --> 00:04:53.839
principles is do the hard thing, what would that look
82
00:04:53.920 --> 00:04:56.560
like and could we do it? And so my background
83
00:04:56.600 --> 00:04:59.879
in AI, you know, Ed's belief that you know, AI
84
00:05:00.079 --> 00:05:04.160
is absolutely the future for Australia and the world, we
85
00:05:04.240 --> 00:05:06.600
kind of conceive this idea of main Code and that
86
00:05:06.759 --> 00:05:10.160
start of the journey and the goal was always what
87
00:05:10.240 --> 00:05:12.079
are the hard things that people say we shouldn't do
88
00:05:12.199 --> 00:05:14.839
or can't do? Can we do it? And that was
89
00:05:14.879 --> 00:05:16.839
pretty much the start of the journey.
90
00:05:18.240 --> 00:05:21.079
So you're talking about edre Ed Craven who's co founder
91
00:05:21.079 --> 00:05:25.199
and CEO of Easygo but also a majority shareholder in
92
00:05:25.319 --> 00:05:28.120
main Code, you know, and it resonates actually with me.
93
00:05:28.240 --> 00:05:30.040
I think when you're a founder, having to have a
94
00:05:30.079 --> 00:05:31.759
thick skin because people look at you and go, why
95
00:05:31.759 --> 00:05:33.759
are you doing that? But actually that's what happens when
96
00:05:33.759 --> 00:05:36.879
you're doing something for the first time, right. But also
97
00:05:37.240 --> 00:05:39.120
your values, they really stood out to be is be
98
00:05:39.240 --> 00:05:41.560
quite different. So another one of your values that I
99
00:05:41.600 --> 00:05:44.240
really liked is that you say it's two.
100
00:05:44.199 --> 00:05:45.040
Am in the lab.
101
00:05:45.240 --> 00:05:47.519
This is not a nine to five job that you're
102
00:05:47.560 --> 00:05:49.639
doing when you come and work for main Code. And
103
00:05:49.720 --> 00:05:52.480
I think in an era where so much of the
104
00:05:52.560 --> 00:05:55.519
focus is on work life balance and just owning the
105
00:05:55.600 --> 00:05:59.959
fact that, like this is a passion and a lifestyle choice.
106
00:06:00.079 --> 00:06:03.560
Almost it really stood out in terms of the way
107
00:06:03.560 --> 00:06:06.439
that you've positioned it. Can you tell us a little
108
00:06:06.439 --> 00:06:10.319
bit about what Matilda is. So main Code is the company.
109
00:06:10.560 --> 00:06:15.519
Matilda is the application that you have built. Now most
110
00:06:15.519 --> 00:06:20.319
people know of Chatchipt and Claude. Underneath those sit the
111
00:06:20.360 --> 00:06:24.759
particular models that drive the AI assistance or AI chatbots.
112
00:06:25.439 --> 00:06:28.680
What is Matilda is like what have you built? Have
113
00:06:28.879 --> 00:06:31.959
you built the AI assistant or have you also built
114
00:06:31.959 --> 00:06:33.920
the model or both.
115
00:06:34.120 --> 00:06:36.879
Matilda's actually like a product family is probably the best
116
00:06:36.879 --> 00:06:40.800
way to think about Matilda. And so Matilda encompasses Matilda Chat,
117
00:06:40.879 --> 00:06:42.680
which is you can go on to Matilda dot main
118
00:06:42.720 --> 00:06:46.319
Code dot com today, register your account, use it like
119
00:06:46.360 --> 00:06:48.920
you would use Chatchipt or claud.
120
00:06:49.199 --> 00:06:50.720
And I encourage people to do that.
121
00:06:51.279 --> 00:06:54.240
Gala the Apple app. We've got the Android app coming
122
00:06:54.279 --> 00:06:56.480
as soon as Google jumps up and helps us if
123
00:06:56.480 --> 00:06:59.920
they're listening. But you know, the goal is that we
124
00:07:00.040 --> 00:07:02.720
wanted to build a state of the art capability. So
125
00:07:02.800 --> 00:07:07.199
Matilda is a family of things. So it's Matilda where
126
00:07:07.240 --> 00:07:10.360
app the iOS client, it's the software layer. We also
127
00:07:10.399 --> 00:07:13.240
have Matilda code so you can get if you reach
128
00:07:13.279 --> 00:07:16.000
out to us, we'll give you Matilda the CLI coding agent.
129
00:07:16.040 --> 00:07:18.480
We've got a desktop coding agent. All of these are
130
00:07:18.480 --> 00:07:20.759
early in beta right now, we're doing some really exciting
131
00:07:21.040 --> 00:07:24.319
pilot programs with people like ey But so that's kind
132
00:07:24.319 --> 00:07:26.519
of like where people enter right like they're either developers
133
00:07:26.560 --> 00:07:30.079
and they're using Matilda code or their businesses and consumers
134
00:07:30.199 --> 00:07:34.160
using matildamain code dot com. And then underneath that software
135
00:07:34.199 --> 00:07:39.040
layer is a very complex software system, which you know,
136
00:07:39.079 --> 00:07:41.680
there's the layer that deals with your sessions and your
137
00:07:41.759 --> 00:07:44.639
chats and the orchestrator and all that, and then there's
138
00:07:44.879 --> 00:07:47.199
a layer of the model serving. So you'd see in
139
00:07:47.240 --> 00:07:49.360
some of our research we spend a lot of time
140
00:07:49.399 --> 00:07:52.879
focusing on how to improve inference. We've done a lot
141
00:07:52.879 --> 00:07:56.519
of customized inference work. We work very closely with AMD
142
00:07:57.439 --> 00:08:00.120
on how to do that better on our stack. And
143
00:08:00.160 --> 00:08:02.680
then all of that software sits on top of a
144
00:08:02.800 --> 00:08:06.279
custom built cluster that we designed and operate. So that's
145
00:08:06.319 --> 00:08:09.800
everything from the front end network through the application database cloud.
146
00:08:09.879 --> 00:08:12.480
So we have a product called main Cloud which is
147
00:08:12.480 --> 00:08:16.120
what actually runs just like AWS, but it's fully Australian
148
00:08:16.160 --> 00:08:20.360
made and run in Telstra data centers. And then under
149
00:08:20.360 --> 00:08:24.680
that the actual models that we train, so we train
150
00:08:24.759 --> 00:08:26.920
a number of different models. Matilda is not just one
151
00:08:26.959 --> 00:08:29.920
model because obviously it's a mixture of models, but they're
152
00:08:30.120 --> 00:08:33.320
capable of doing coding, interacting with chat. They even do
153
00:08:33.440 --> 00:08:39.879
tool calling and coding support and everything required to satisfy
154
00:08:39.919 --> 00:08:42.000
what I would consider twenty twenty six stead of the
155
00:08:42.080 --> 00:08:46.840
art kind of requirements. All of that is built, operated,
156
00:08:46.919 --> 00:08:49.240
run own by main Code. And you know, we work
157
00:08:49.279 --> 00:08:53.279
with partners like Telstra, Dell amd to as our suppliers
158
00:08:53.279 --> 00:08:56.519
and providers, but we do every single piece of code,
159
00:08:56.840 --> 00:09:00.200
every line of code is written, managed, operated by main Code.
160
00:09:00.360 --> 00:09:02.960
I just want to reflect for a moment. You're a
161
00:09:03.039 --> 00:09:06.279
company that is making what many people would call sovereign
162
00:09:06.440 --> 00:09:11.080
Australian AI, and yet you've said before that you don't
163
00:09:11.159 --> 00:09:16.240
find the term sovereignty or sovereign technology particularly helpful. You've
164
00:09:16.240 --> 00:09:18.159
actually gone so far as to say that you're ditching
165
00:09:18.200 --> 00:09:21.000
the term because it's not useful for builders. Can you
166
00:09:21.279 --> 00:09:24.879
explain what your reversion to the word sovereignty is because
167
00:09:24.919 --> 00:09:27.399
many people would expect someone building a company like yours
168
00:09:27.440 --> 00:09:29.480
to really you know, some people lean into it as
169
00:09:29.480 --> 00:09:30.840
a sales pitch almost.
170
00:09:31.159 --> 00:09:35.240
So Main Code from the very start made a decision
171
00:09:35.279 --> 00:09:38.679
that we weren't going to engage in the dooma rhetoric
172
00:09:39.360 --> 00:09:43.559
and theater and performance that open AI and anthropic do, right,
173
00:09:43.679 --> 00:09:46.559
So I think it's I was just responding to Paul
174
00:09:46.759 --> 00:09:50.080
Smith at thefr just before our call on topics of
175
00:09:50.240 --> 00:09:54.840
like chat GPT, saying that GPT has escaped a sandbox,
176
00:09:55.039 --> 00:09:58.320
you know, and it's just so theatrical and performative, and
177
00:09:58.519 --> 00:10:00.320
there's no way for you and I'd actually to say
178
00:10:00.320 --> 00:10:03.080
if that's true, there's no way to know if that's accurate.
179
00:10:03.159 --> 00:10:06.639
It's just fun, right, and even anthropic. Like every time
180
00:10:06.679 --> 00:10:08.480
I see their CEO in his face and he's like
181
00:10:08.559 --> 00:10:11.360
so pained and he's like whoa, it's like the theatrics
182
00:10:11.399 --> 00:10:14.480
are full tilt. And one of the things we decided
183
00:10:14.519 --> 00:10:16.039
as main Code is we're not going to do that.
184
00:10:16.519 --> 00:10:19.799
We're not going to engage in the rhetoric, the performance, theater,
185
00:10:20.320 --> 00:10:22.440
the clutching off the pearls, the ringing of the hands
186
00:10:22.639 --> 00:10:24.639
are like none of that, right, we were going to
187
00:10:24.679 --> 00:10:27.799
build great Australian made product, and we'll let those products
188
00:10:27.799 --> 00:10:30.519
stand for themselves, right, And we're not going to go
189
00:10:30.559 --> 00:10:33.600
out there and use terms like sovereign right, which is
190
00:10:33.879 --> 00:10:37.440
doesn't mean anything. What like Joanna, what does it mean? Right?
191
00:10:38.000 --> 00:10:40.840
If you say to someone like we're a sovereign AI company,
192
00:10:41.000 --> 00:10:43.960
that doesn't mean anything. If we say we're an Australian
193
00:10:44.000 --> 00:10:46.639
made AI company, that means something. You know you can
194
00:10:46.720 --> 00:10:49.240
understand what that means. And the other thing, too, is
195
00:10:49.279 --> 00:10:52.679
that we're not here to sell a product and you know,
196
00:10:52.759 --> 00:10:56.279
try and you raise bajillion dollars of capital and again
197
00:10:56.360 --> 00:11:00.440
do the US playbook. We're here to build a sustainable
198
00:11:00.840 --> 00:11:04.799
AI industry for the next generation that's going to last
199
00:11:04.840 --> 00:11:06.679
the test of time. Right. We don't want to just
200
00:11:06.759 --> 00:11:09.919
flame out and then drop off like Mistral for example,
201
00:11:10.000 --> 00:11:13.320
having to like sell their soul to Microsoft, right, like
202
00:11:13.399 --> 00:11:14.840
you know, and we don't have to do any of
203
00:11:14.840 --> 00:11:17.360
those things. So we're trying to build a one hundred
204
00:11:17.440 --> 00:11:20.240
year company. And I think you do that by just
205
00:11:20.320 --> 00:11:23.000
working on good product, getting in people's hands, listening to
206
00:11:23.000 --> 00:11:24.919
them when they tell you it didn't work, fixing it,
207
00:11:25.440 --> 00:11:29.840
rinse and repeat. But yeah, I think sovereign as a
208
00:11:29.919 --> 00:11:34.200
term is overloaded, it's been weaponized. It gets thrown around
209
00:11:34.200 --> 00:11:38.559
Canberra by people who are looking to ship policy not product.
210
00:11:38.720 --> 00:11:39.600
That's not our game.
211
00:11:42.480 --> 00:11:46.480
When you're talking day there about building one hundred year company,
212
00:11:47.240 --> 00:11:50.080
What did you What was your reaction when you listened
213
00:11:50.240 --> 00:11:53.960
to the PM speech? Did you listen to it? Was
214
00:11:53.960 --> 00:11:57.559
it significant for you? And what was there anything in
215
00:11:57.600 --> 00:12:00.679
it that surprised you? Did it get you excited? The response?
216
00:12:01.559 --> 00:12:03.679
So I did listen to it because I was very,
217
00:12:04.039 --> 00:12:06.840
very keen and very excited to hear. First of all,
218
00:12:06.840 --> 00:12:08.759
I think it was the feeling of there was a
219
00:12:08.759 --> 00:12:12.639
pivot happening, right. So we've been on the hill regularly
220
00:12:13.039 --> 00:12:17.600
speaking to Senate committees, to you know, senior ministers, Andrew Charlton,
221
00:12:17.759 --> 00:12:19.639
you know, and I sat down and got to talk
222
00:12:19.639 --> 00:12:22.399
about I showed in one of the very early demos
223
00:12:22.399 --> 00:12:24.919
of Matilda. He was one of the first people in
224
00:12:24.960 --> 00:12:28.440
Australia to actually see Matilda. This was over six months ago,
225
00:12:28.919 --> 00:12:31.399
and so I was genuinely excited because I felt like
226
00:12:32.279 --> 00:12:35.440
where the conversation had fallen flat, It felt like there
227
00:12:35.440 --> 00:12:37.399
had been a change, right, and some of the things
228
00:12:37.440 --> 00:12:40.399
that we've been saying had started to take roote. What
229
00:12:40.519 --> 00:12:43.360
was disappointing was that it was just a lot of
230
00:12:43.399 --> 00:12:46.240
words and very little signal. Right, you know, essentially the
231
00:12:46.279 --> 00:12:48.320
two things I took out of it as my comments
232
00:12:48.320 --> 00:12:52.919
today after the speech were focus on data centers boring. Right,
233
00:12:52.960 --> 00:12:55.799
we're talking about an asset class, you know, disguised. You know,
234
00:12:55.840 --> 00:12:58.320
it's a wolf in sheep's clothing. Right, It's not AI
235
00:12:58.879 --> 00:13:01.120
any more than if we were to go set up
236
00:13:01.120 --> 00:13:03.519
an electrical factory we could call that AI, like you know,
237
00:13:03.559 --> 00:13:06.240
electricity for AI. Right, Like it's just come on, right,
238
00:13:06.320 --> 00:13:09.440
Like at some point the basic asset class doesn't matter.
239
00:13:09.799 --> 00:13:13.559
And this this absolute fat you know, infatuation with data
240
00:13:13.600 --> 00:13:16.799
center's data center. Who cares it's a concrete warehouse we
241
00:13:16.840 --> 00:13:20.200
could sell to Bunnings after you know, it doesn't work
242
00:13:20.240 --> 00:13:22.399
out right, And but we got this deal with Nvideo
243
00:13:22.799 --> 00:13:25.559
to what park a bunch of tin in a warehouse
244
00:13:25.600 --> 00:13:30.639
and without the software layer serving tokens as an actual product,
245
00:13:30.679 --> 00:13:33.080
not just rented to Latrobe to run you know, their
246
00:13:33.559 --> 00:13:35.440
data center worklod It's like it's all the it's all
247
00:13:35.440 --> 00:13:38.879
performative theatrics. That frustrates me. So when the Prime Minister
248
00:13:38.960 --> 00:13:41.039
gets up there and goes we're going to talk about
249
00:13:41.159 --> 00:13:46.320
data centers. It's like, okay, welcome, welcome to twenty twenty four.
250
00:13:46.759 --> 00:13:49.679
And then the other part of that was this focus
251
00:13:49.720 --> 00:13:52.759
again on the copyright and it's like, oh my goodness,
252
00:13:52.799 --> 00:13:56.000
that horse is so dead. It's so dead and flogged
253
00:13:56.039 --> 00:14:00.960
to pieces, like that's what you're breathing life into. Meanwhile,
254
00:14:01.159 --> 00:14:05.960
no discussion about procurement preference for Australian made AI, no
255
00:14:06.360 --> 00:14:11.360
clear actionable plan with atomic steps towards moving Australia towards
256
00:14:11.600 --> 00:14:16.279
real meaningful goals in the AI landscape. No exports support
257
00:14:16.360 --> 00:14:19.320
again you know as we speak to companies overseas, No
258
00:14:19.519 --> 00:14:23.440
government support behind main code we've got. We've just launched
259
00:14:23.519 --> 00:14:27.039
the first bad off, the only actual top to bottom
260
00:14:27.080 --> 00:14:29.879
ten to token built in Australia, State of the art
261
00:14:30.360 --> 00:14:34.279
AI platform, right like, we're not talking about two fellas
262
00:14:34.279 --> 00:14:38.120
and a backshed running open code right like this this
263
00:14:38.159 --> 00:14:39.879
has been built over a year and a half and
264
00:14:40.360 --> 00:14:44.519
again just blinders, right, And so it was it was
265
00:14:44.720 --> 00:14:48.440
very disappointing. Part of me felt a little bit like
266
00:14:48.559 --> 00:14:51.639
what did I expect, you know, from the PM and
267
00:14:51.679 --> 00:14:57.120
the government. And I'm sad Joina that Australian made AI
268
00:14:57.360 --> 00:15:03.600
still is a talk point and a political point scoring activity,
269
00:15:04.320 --> 00:15:06.200
and that they don't What that tells me is they
270
00:15:06.200 --> 00:15:11.120
don't actually understand the massive economic and industrial potential for
271
00:15:11.200 --> 00:15:15.279
the kids coming up now who don't want to go
272
00:15:15.759 --> 00:15:19.200
work in Telstra and A and Z and those traditional companies.
273
00:15:19.399 --> 00:15:23.080
They want their shot in an AI industry that's not
274
00:15:23.200 --> 00:15:26.000
being developed here. And we're being told about CGT and
275
00:15:26.039 --> 00:15:28.559
all of this kind of stuff for generations that are
276
00:15:28.679 --> 00:15:32.600
absolutely home and host right, So the kids coming up
277
00:15:32.639 --> 00:15:35.200
who are looking at eight to ten times salary on
278
00:15:35.279 --> 00:15:38.000
buying their first home in a market with no inventory,
279
00:15:38.840 --> 00:15:42.559
and the government's not doing anything about this incredibly powerful
280
00:15:43.159 --> 00:15:46.720
industrial revolution, and so it is frustrating. You know, we'll
281
00:15:46.759 --> 00:15:49.200
keep the good fight going. We're going to keep building matildren,
282
00:15:49.960 --> 00:15:52.080
doing what we can, but it is frustrating when you've
283
00:15:52.080 --> 00:15:56.039
then listen to that presentation and then you see all
284
00:15:56.080 --> 00:15:58.120
the boot licking around that as well, right of like,
285
00:15:58.240 --> 00:16:01.000
oh what a legend, that's the vision we needed, And
286
00:16:01.039 --> 00:16:04.759
it's like, sure, look, I kind.
287
00:16:04.559 --> 00:16:07.480
Of fall halfway between the two things you're saying there.
288
00:16:07.799 --> 00:16:09.240
So I was in the room when the PM was
289
00:16:09.279 --> 00:16:12.360
speaking and there was an amazing energy in the room,
290
00:16:12.399 --> 00:16:14.960
and I kind of I asked myself, why are we
291
00:16:15.000 --> 00:16:17.759
so excited about this? And I do think it's because
292
00:16:17.799 --> 00:16:20.639
so many people, yourself and your team included, have spent
293
00:16:21.240 --> 00:16:24.519
so much shoe lever in trying to get national leadership
294
00:16:24.559 --> 00:16:27.440
and attention on this issue. And so I think we
295
00:16:27.559 --> 00:16:30.200
do need to applaud the fact that the PM stood
296
00:16:30.279 --> 00:16:33.279
up and made this speech and has said I am
297
00:16:33.360 --> 00:16:37.600
prioritizing this, but I'm also well on the record and
298
00:16:37.639 --> 00:16:40.480
saying that it's in the delivery of you know, how
299
00:16:40.480 --> 00:16:44.840
do we actually translate the vision into reality? So on that, Dave,
300
00:16:45.519 --> 00:16:48.720
what do you think of the most important things for
301
00:16:48.759 --> 00:16:53.159
the government to actually do, whether it's policy, whether it's investment,
302
00:16:53.639 --> 00:16:56.399
to deliver on the vision and the ambition.
303
00:16:58.799 --> 00:17:01.799
And I'm going to take this very pragmatically, Joanna, because
304
00:17:01.799 --> 00:17:05.000
I think that's the right approach. We speak to customers
305
00:17:05.119 --> 00:17:10.480
every day now like we're inundated with inquiries. Is one
306
00:17:10.559 --> 00:17:12.720
level of the inundation. The other one is people are
307
00:17:12.759 --> 00:17:14.880
just like, we're ready to sign up, like we've got
308
00:17:14.880 --> 00:17:17.079
customers were like, we want played with it now for
309
00:17:17.119 --> 00:17:18.799
a week, and how do we sign up now and
310
00:17:18.839 --> 00:17:20.839
pay you money and we just want to use this product. Right.
311
00:17:20.839 --> 00:17:24.480
It's fantastic. But I think what's very clear is that
312
00:17:24.599 --> 00:17:27.799
there are now products in market. We're no longer a
313
00:17:27.880 --> 00:17:29.519
year ago when I was going up to the hill
314
00:17:29.799 --> 00:17:31.839
and I was meeting with the PM's office and going, look,
315
00:17:31.839 --> 00:17:34.759
we got this idea. We're ordering a cluster. We're building
316
00:17:34.759 --> 00:17:37.720
a team where we're now cluster is in place. We've
317
00:17:37.720 --> 00:17:40.799
built software that not even a m d as. We've
318
00:17:40.839 --> 00:17:42.920
had to do things that not even the US companies
319
00:17:42.920 --> 00:17:46.559
have done to fully operate the cluster and serve Matilda.
320
00:17:46.720 --> 00:17:51.359
We are the only true full stack NEO AI cloud
321
00:17:51.359 --> 00:17:54.839
in Australia because we don't use any Amazon or Microsoft services.
322
00:17:54.880 --> 00:17:58.400
We're fully hosting ourselves. These are incredible milestones that have
323
00:17:58.519 --> 00:18:01.440
not occurred in the Australian market when you even consider
324
00:18:01.519 --> 00:18:04.680
the large players like at Lassian and Canvas still use
325
00:18:04.799 --> 00:18:07.839
Amazon as their hosting partner. We actually one of the
326
00:18:07.880 --> 00:18:10.720
few real products built in Australia that run our own stack.
327
00:18:11.279 --> 00:18:13.920
So when we speak to companies like ey who are like,
328
00:18:14.000 --> 00:18:17.119
we need built in Australia, run in Australia, like I
329
00:18:17.119 --> 00:18:19.799
can go touch the data center, yea somewhere like that
330
00:18:19.960 --> 00:18:25.119
is crucial. But it's the first off the foundation stones.
331
00:18:25.480 --> 00:18:28.839
And if the government doesn't actually pave the way for us,
332
00:18:28.920 --> 00:18:32.160
just like they did with all the other union backed industries, right,
333
00:18:32.240 --> 00:18:35.640
like you know, they blocked out external competition to allow
334
00:18:36.039 --> 00:18:40.400
the local builders to survive. But the thought that we
335
00:18:40.440 --> 00:18:44.319
can you know, long term defend the position in Australia
336
00:18:44.359 --> 00:18:47.680
against an open AI or an anthropic if the government
337
00:18:47.720 --> 00:18:51.119
doesn't stop buying from Australian built companies, like they'll sign
338
00:18:51.200 --> 00:18:54.319
multimillion dollar deals with open AI and then they'll you know,
339
00:18:54.839 --> 00:18:57.400
sit on their hands when it comes to using Aussie
340
00:18:57.400 --> 00:19:00.440
made stuff. It's not right. And I think that what
341
00:19:00.480 --> 00:19:03.079
I want the government to do is actually get proud
342
00:19:03.119 --> 00:19:07.359
of being Australian, right, like you know, let's bring back
343
00:19:07.400 --> 00:19:10.640
Australian made right. And I always say this to people
344
00:19:10.640 --> 00:19:13.359
like you wouldn't go to the freaking Olympics and barrack
345
00:19:13.440 --> 00:19:16.559
for another country, right like you know you And so
346
00:19:16.640 --> 00:19:20.240
as a country that is trying to build its place
347
00:19:20.400 --> 00:19:24.319
in this industry, to not get behind it everywhere you can.
348
00:19:24.880 --> 00:19:26.240
And I think that's what I want to see the
349
00:19:26.240 --> 00:19:27.960
government do. I want the government to say we want
350
00:19:27.960 --> 00:19:30.920
to see more main codes and more Matilda's springing up.
351
00:19:31.160 --> 00:19:34.279
But we recognize that's really freaking hard. Let's pave the
352
00:19:34.319 --> 00:19:36.960
way for at least the next year or two and
353
00:19:37.039 --> 00:19:38.559
let's block these guys up. But at the end of
354
00:19:38.599 --> 00:19:40.400
the day, our product is state of the art. So
355
00:19:40.440 --> 00:19:42.400
it's not like they can say, oh, it's not as capable.
356
00:19:42.519 --> 00:19:45.119
It's absolutely as capable. But here's the thing, Joiner that
357
00:19:45.119 --> 00:19:49.079
frustrates me. America always protects their own right. They'll put
358
00:19:49.079 --> 00:19:52.359
the embargoes, the restrictions, they'll make sure that America wins.
359
00:19:52.599 --> 00:19:55.319
And of course where he is just like no backbone,
360
00:19:55.559 --> 00:19:58.079
you know, no spine to be able to go. Actually no,
361
00:19:58.559 --> 00:20:01.000
don't need your stuff, thank you. That's ahead of them
362
00:20:01.039 --> 00:20:02.799
potentially not even giving it to us anyway.
363
00:20:03.319 --> 00:20:06.400
Yeah, yeah, I mean, in particularly with the export restrictions
364
00:20:06.400 --> 00:20:08.680
that we're placed out of the Trump administration. So I
365
00:20:08.720 --> 00:20:10.759
just want to be really clear on what you're saying there.
366
00:20:11.119 --> 00:20:13.720
One thing is to say, actually, one of the most
367
00:20:13.720 --> 00:20:16.920
powerful things the government can do is to procure Australian
368
00:20:17.039 --> 00:20:20.359
made AI, so you know, actually use main code. And
369
00:20:20.440 --> 00:20:23.440
now there is to say we should not be using
370
00:20:24.359 --> 00:20:28.200
open AI or anthropic and Claude. Are you calling for both,
371
00:20:28.599 --> 00:20:31.160
because that's quite a dramatic call.
372
00:20:31.480 --> 00:20:35.440
Yeah, I'm calling for both because it's it's not unprecedented.
373
00:20:36.079 --> 00:20:38.680
If you look back, and I was doing research into
374
00:20:38.720 --> 00:20:41.839
this from the very first time we did telco electrical,
375
00:20:42.119 --> 00:20:47.519
early semiconductor, auto manufacturing, Australia has a very very long
376
00:20:47.599 --> 00:20:50.880
line of precedent where we have blocked out foreign or
377
00:20:50.880 --> 00:20:52.680
we've made it we're not blocked up, we've made it
378
00:20:52.880 --> 00:20:56.920
very very hard, or we've increased the barrier to entry
379
00:20:56.920 --> 00:21:00.440
in Australia to protect the local industry. And so that's
380
00:21:00.440 --> 00:21:03.240
all I'm calling for, right And what that means then
381
00:21:03.359 --> 00:21:08.319
is that Australian made has more advantage in the procurement
382
00:21:08.400 --> 00:21:13.000
process in government. That's nothing new, and that we also
383
00:21:13.079 --> 00:21:17.400
make it fair for Australian industry to grow by making
384
00:21:17.400 --> 00:21:21.119
it very hard for larger foreign entities that have way
385
00:21:21.160 --> 00:21:23.880
more resources than we do to come and tip the
386
00:21:23.880 --> 00:21:26.319
boat over like that. That's what we're talking about. And
387
00:21:26.400 --> 00:21:28.720
if you look back as well with the car import tax,
388
00:21:28.759 --> 00:21:30.559
the reason that was done is you've got countries that
389
00:21:30.640 --> 00:21:33.559
are like ten fifteen times the size of Australia. You
390
00:21:33.640 --> 00:21:37.559
have to protect Australian industry. Otherwise are we just renting everything?
391
00:21:38.359 --> 00:21:40.240
Is that all we are from now on? Like all
392
00:21:40.279 --> 00:21:43.359
we do is or sales companies? Right? So here's the
393
00:21:43.400 --> 00:21:45.839
other thing. It's like, oh anthropic and open AI vesting
394
00:21:45.880 --> 00:21:49.640
in Australia. How many engineers work here? None? All sales
395
00:21:49.640 --> 00:21:53.559
support or consulting solution architects all about selling into Australia.
396
00:21:53.640 --> 00:21:56.759
There's actually no core engineering done here, right And so
397
00:21:57.000 --> 00:21:59.440
again this idea that these brands come here and actually
398
00:21:59.480 --> 00:22:03.480
uplift our local generations, it doesn't happen. We actually have
399
00:22:03.480 --> 00:22:06.279
to build some stuff here, you know. And I think
400
00:22:06.279 --> 00:22:07.920
that's what I want to see the government do is
401
00:22:07.960 --> 00:22:11.200
recognize that ten, fifteen, twenty years from now we need
402
00:22:11.240 --> 00:22:15.079
a thriving local AI industry where people are going and
403
00:22:15.119 --> 00:22:18.519
working on first class product, selling to Australian businesses and
404
00:22:18.559 --> 00:22:22.039
exporting that product because we have an incredible export opportunity.
405
00:22:22.319 --> 00:22:25.920
The Australian made brand, especially with tokens, Australian made tokens
406
00:22:25.960 --> 00:22:30.319
are seen as very highly valuable, very sought after. We're safe,
407
00:22:30.680 --> 00:22:33.400
we're regulated, we're secure. A lot of countries want to
408
00:22:33.440 --> 00:22:35.160
do business with us before they want to do business
409
00:22:35.160 --> 00:22:36.960
with America and China.
410
00:22:37.480 --> 00:22:40.200
Yeah, and I think that is something that Australia needs
411
00:22:40.240 --> 00:22:43.200
to lean into more that we have a very specific
412
00:22:43.240 --> 00:22:46.039
competitive advantage there. I just want to flag that the
413
00:22:46.079 --> 00:22:49.920
proposal of restricting access would raise and I'm sorry to
414
00:22:50.000 --> 00:22:52.000
be a total nerd on this day, but you know,
415
00:22:52.039 --> 00:22:55.119
all sorts of challenges under free trade agreements and things
416
00:22:55.200 --> 00:22:57.720
like this, but also frankly the Trump administration, who've made
417
00:22:57.759 --> 00:23:00.400
it really clear that they would retaliate if restricts were
418
00:23:00.400 --> 00:23:03.680
made on US companies. I'm just saying that for balance.
419
00:23:04.720 --> 00:23:08.880
I also think Andrew Charlton has spoken a lot of
420
00:23:08.920 --> 00:23:10.400
you gave a speech saying you know, this is the
421
00:23:10.440 --> 00:23:14.319
sliding doors moment. We need to procure Australian products. But
422
00:23:14.720 --> 00:23:17.559
going back to that point of narrative versus action, we're
423
00:23:17.599 --> 00:23:21.240
not yet seeing the government actually following that through. I
424
00:23:21.279 --> 00:23:24.000
want to pick up on another point that you've touched
425
00:23:24.000 --> 00:23:26.599
on quite a few times there, and this is the
426
00:23:26.680 --> 00:23:29.559
compute clusters that you have. So you've made the point
427
00:23:29.559 --> 00:23:32.759
that you don't use AWS for example, you don't rent
428
00:23:33.039 --> 00:23:37.400
any compute from anybody. You've built the clusters yourselves. You're
429
00:23:37.440 --> 00:23:40.640
working with AMD, which is advanced micro devices who are
430
00:23:40.720 --> 00:23:43.720
chip producers. For those who aren't living in breeding this
431
00:23:43.799 --> 00:23:46.319
can we just start with a simple explanation of what
432
00:23:46.400 --> 00:23:49.039
a compute cluster actually is, and then I'll go into
433
00:23:49.079 --> 00:23:49.920
why that's important.
434
00:23:50.039 --> 00:23:53.279
We never use any of these applications. They're hosted in
435
00:23:54.160 --> 00:23:57.079
a data center somewhere with a set of hardware, and
436
00:23:57.119 --> 00:23:59.960
that hardware is under the control and management often into
437
00:24:00.640 --> 00:24:03.960
and so to deliver Australian AI in the form of Matilda,
438
00:24:04.400 --> 00:24:06.799
we have to build a software stack. We could build
439
00:24:06.839 --> 00:24:09.559
that software stack anywhere and you know you can do that.
440
00:24:09.759 --> 00:24:11.240
You know in your laptop, you can do that in
441
00:24:11.400 --> 00:24:14.640
Amazon and Microsoft are Google. But once you actually deployed
442
00:24:14.640 --> 00:24:16.680
and run it and people are using it, that has
443
00:24:16.720 --> 00:24:20.960
to be running somewhere day to day. And what normally
444
00:24:21.000 --> 00:24:24.440
happens is people will just go to Amazon. They'll they'll
445
00:24:24.680 --> 00:24:27.559
spin up a bunch of resources in Amazon, they'll host
446
00:24:27.599 --> 00:24:30.319
their software there and then that's what you're using. You
447
00:24:30.319 --> 00:24:32.559
know when you do it. But every time you do that,
448
00:24:32.599 --> 00:24:35.559
your data is subject to the US Cloud Act. And
449
00:24:35.599 --> 00:24:38.279
should the US government ever decide that they want a
450
00:24:38.319 --> 00:24:41.480
subpoena that data, Amazon will give it up. Having worked
451
00:24:41.519 --> 00:24:44.200
at Amazon and Microsoft, I know how this process works.
452
00:24:44.200 --> 00:24:47.759
We will cough it up without any questions asked. There's
453
00:24:47.759 --> 00:24:50.640
actually a team that's responsible for like downloading all of
454
00:24:50.640 --> 00:24:53.400
your personal history handing it over to the government. Now
455
00:24:53.440 --> 00:24:56.839
that US cloud subpoena can actually come from a third party,
456
00:24:57.559 --> 00:25:01.240
That's what's really scary, right, So another company can petition
457
00:25:01.319 --> 00:25:05.160
the government to subpoena your data. So it's not just
458
00:25:05.160 --> 00:25:06.759
the White House has to decide that they want to
459
00:25:06.799 --> 00:25:09.680
do something, and like your enemy could ask for that data,
460
00:25:09.720 --> 00:25:11.160
and the White House could go, yeah, we actually prefer
461
00:25:11.200 --> 00:25:13.319
the enemy of you, and we'll cough that data up
462
00:25:13.319 --> 00:25:15.640
through the US cloud AXT. So it was all around
463
00:25:15.680 --> 00:25:18.039
like you know, homeland security and you know, being able
464
00:25:18.079 --> 00:25:18.359
to like.
465
00:25:18.440 --> 00:25:22.000
Response to terrorism originally yeah yeah.
466
00:25:21.240 --> 00:25:24.559
But obviously now in so many of these landmark cases
467
00:25:24.599 --> 00:25:29.519
where commercial organizations have compelled through friends of Trump how
468
00:25:29.519 --> 00:25:31.599
to get data of flex civilians and stuff like that
469
00:25:31.640 --> 00:25:35.480
who are seen as agitators or whatever, that's not good. Right.
470
00:25:36.160 --> 00:25:38.960
So when you sit with your doctor like I did recently,
471
00:25:38.960 --> 00:25:40.640
and he goes, oh, can I record this, and I'm
472
00:25:40.680 --> 00:25:44.599
like no, because I don't want my health transcript going
473
00:25:44.599 --> 00:25:47.839
through Heidi Health and then landing in Amazon's data center,
474
00:25:47.880 --> 00:25:50.720
and now my transcripts are with the US. So these
475
00:25:50.759 --> 00:25:53.559
are the kind of motivating scenarios that we came up with.
476
00:25:53.720 --> 00:25:56.839
The problem is Joanna, and I know this intimately because
477
00:25:56.880 --> 00:25:59.480
I ran the data centertainm at Microsoft for a number
478
00:25:59.519 --> 00:26:03.440
of years. Standing up infrastructure to host a twenty four
479
00:26:03.440 --> 00:26:07.599
to seven app is grueling. So what does that look like?
480
00:26:07.720 --> 00:26:09.680
We first of all have to go out to market
481
00:26:09.759 --> 00:26:13.519
and find a hosting partner. We did that. We selected
482
00:26:13.559 --> 00:26:17.160
Telstra because they're the national backbone. They're incredible partner of
483
00:26:17.160 --> 00:26:19.359
ours as well. I cannot speak highly enough of Telstra
484
00:26:19.720 --> 00:26:22.160
and the support they've given to main Code and Matilda.
485
00:26:22.720 --> 00:26:25.880
So they give us a portion of a data center,
486
00:26:25.920 --> 00:26:29.000
highly secure data center ours happens to be out in Clayton.
487
00:26:29.440 --> 00:26:33.519
They that literally looks like a concrete flow with electrical
488
00:26:33.559 --> 00:26:38.559
cables and cooling. We then go and procure racks of servers.
489
00:26:39.200 --> 00:26:42.079
These are high performance, state of the art, your multimillion
490
00:26:42.079 --> 00:26:46.039
dollar racks. We acquired those through Dell. They're fitted with
491
00:26:46.079 --> 00:26:49.799
AD hardware. Those racks roll into the data center. Our
492
00:26:49.880 --> 00:26:53.880
team wire them up, install the software, bring everything up,
493
00:26:53.920 --> 00:26:58.079
install Matilda, punch holes into the Internet everything's flowing into
494
00:26:58.480 --> 00:27:02.559
Telstra into main code. That is a herculean feat because
495
00:27:02.799 --> 00:27:06.079
again you're dealing with companies like Amazon and Microsoft, multi
496
00:27:06.079 --> 00:27:09.799
billion dollar companies, thousands of people who are monitoring and
497
00:27:10.119 --> 00:27:12.839
managing these clusters, and we're doing it just as a
498
00:27:12.839 --> 00:27:15.720
small team in Melbourne. What's been great though, is that
499
00:27:15.720 --> 00:27:17.680
it's forced us to have to write software that no
500
00:27:17.720 --> 00:27:21.039
one else has had to write. So we have cluster
501
00:27:21.079 --> 00:27:24.440
management software, we have fleet management software that, like I said,
502
00:27:24.440 --> 00:27:27.400
when we're working with AMD and Dell, they reached a
503
00:27:27.440 --> 00:27:28.880
point where they were like, we just can't do some
504
00:27:28.920 --> 00:27:31.160
of this stuff. And we're like, all right, it's two
505
00:27:31.200 --> 00:27:33.480
am in the lab. Let's do the hard thing. Let's
506
00:27:33.680 --> 00:27:36.319
write code, right and we built the code and you know,
507
00:27:36.400 --> 00:27:39.039
the ship cycle that led up to Matilda was absolutely
508
00:27:39.079 --> 00:27:41.839
the hardest I've ever done in my life, but it
509
00:27:41.920 --> 00:27:45.839
was so valuable because the capability hadn't existed in Australia
510
00:27:46.359 --> 00:27:49.160
for decades. We don't know how to do these things,
511
00:27:49.640 --> 00:27:51.680
and I think that was the thing that really struck
512
00:27:51.759 --> 00:27:54.279
me the most was when you're dealing with Dell and
513
00:27:54.319 --> 00:27:56.920
AMD and they're like, we actually do know how to
514
00:27:56.960 --> 00:28:00.119
do this, like locally We've got teams in the US
515
00:28:00.240 --> 00:28:03.279
who help Microsoft and Facebook and stuff, but on the
516
00:28:03.279 --> 00:28:06.240
ground here we don't have these capabilities. And after we
517
00:28:06.240 --> 00:28:09.519
were done, Telstred learned so much about how to do cooling,
518
00:28:09.559 --> 00:28:13.279
you know, liquid cooling. They've built a completely dedicated cooling
519
00:28:13.319 --> 00:28:16.799
infrastructure for main Code and Matilda. We've learned so much
520
00:28:16.880 --> 00:28:19.599
and done things that in Australia hadn't been done because
521
00:28:19.599 --> 00:28:21.799
it didn't need to be done. And I think that's
522
00:28:21.839 --> 00:28:24.279
what's been so amazing. And now as we look to
523
00:28:24.319 --> 00:28:29.160
the future of future clusters, where about generation two, generation three,
524
00:28:29.279 --> 00:28:32.480
we're so far ahead even globally, right, like the way
525
00:28:32.519 --> 00:28:35.319
we think about inference and the rewrite we're doing off
526
00:28:35.599 --> 00:28:39.119
the inference servers with RUST, it's state of the art,
527
00:28:39.119 --> 00:28:42.079
it's first class of our research is you know, we
528
00:28:42.400 --> 00:28:45.839
just received word that a couple of major papers have
529
00:28:45.880 --> 00:28:50.319
been accepted into the industry track for upcoming conferences. So
530
00:28:51.240 --> 00:28:52.799
I think what we're showing is that, yes, on a
531
00:28:52.839 --> 00:28:56.039
small scale, we're still state of the art. You know,
532
00:28:56.079 --> 00:28:59.240
we're absolutely holding our head, you know, with the anthropics
533
00:28:59.240 --> 00:29:01.640
and the open aiy, we're just not as big, well
534
00:29:01.720 --> 00:29:03.240
absolutely as capable, And I.
535
00:29:03.279 --> 00:29:07.200
Think it really is extraordinary what you guys have managed
536
00:29:07.240 --> 00:29:09.079
to build in the period of time that you've managed
537
00:29:09.119 --> 00:29:12.200
to build it, and it needs to be celebrated and applauded,
538
00:29:12.279 --> 00:29:14.359
you know, in the same way we celebrate sports teams
539
00:29:14.400 --> 00:29:16.799
for example. There's a couple of threads I want to
540
00:29:17.079 --> 00:29:21.039
pick out of that. One is you've mentioned their inference compute,
541
00:29:21.079 --> 00:29:23.440
So that's the compute that's required if you're using the
542
00:29:23.480 --> 00:29:26.319
model versus the training of the model. One of the
543
00:29:26.440 --> 00:29:28.839
very live discussions that I've been having here in canber
544
00:29:29.160 --> 00:29:32.000
is do you need a different type of compute, a
545
00:29:32.039 --> 00:29:35.200
different type of data center for the different types of compute?
546
00:29:35.559 --> 00:29:38.640
What's your experience been at main code around that?
547
00:29:39.519 --> 00:29:42.799
So that's actually a really quite a salient point, Joanna.
548
00:29:42.880 --> 00:29:45.960
So when we developed the design for MC two, which
549
00:29:46.000 --> 00:29:48.599
is the cluster class that we run Matilda on now,
550
00:29:49.079 --> 00:29:53.799
we were like, let's do this so we can focus
551
00:29:53.799 --> 00:29:56.720
obviously on the training piece, because the model training is
552
00:29:56.960 --> 00:30:00.880
very resource intensive and requires probably the biggest part of
553
00:30:00.920 --> 00:30:03.079
the cluster. But obviously we need to have an inference
554
00:30:03.160 --> 00:30:05.519
layer too, and we focused on that. I think what
555
00:30:05.559 --> 00:30:09.319
we've learned and there's some decision trade offs you have
556
00:30:09.319 --> 00:30:11.759
to make around like networking for example, right like a
557
00:30:11.880 --> 00:30:15.079
training costers networking is very different because all of the
558
00:30:15.119 --> 00:30:16.680
nodes need to be able to talk to each other.
559
00:30:16.759 --> 00:30:18.960
Right when you're doing a training run. Inference you don't
560
00:30:18.960 --> 00:30:20.880
need that, you know, So unless you're doing some kind
561
00:30:20.920 --> 00:30:24.000
of tensor paralyzation, you actually don't need to have those
562
00:30:24.039 --> 00:30:28.200
talking to each other GP to GPU. So we went
563
00:30:28.240 --> 00:30:30.720
with a very general approach in MC two because we
564
00:30:31.000 --> 00:30:34.160
had to be prepared for anything. Right like that classic
565
00:30:34.240 --> 00:30:36.359
V one, We're like, we don't know what we're gonna hit,
566
00:30:36.440 --> 00:30:38.680
so let's be ready to do everything. The problem is
567
00:30:38.960 --> 00:30:42.200
you end up paying a premium for that because training
568
00:30:42.799 --> 00:30:46.640
configuration is more expensive than inference. So as we now
569
00:30:46.640 --> 00:30:49.640
design MC three, we are very conscious of the fact
570
00:30:49.640 --> 00:30:52.279
that inference is actually a very different problem to solve.
571
00:30:52.759 --> 00:30:54.759
But also we've solved a lot of that in the
572
00:30:54.799 --> 00:30:57.880
software layer. We've tuned and we have a paper recently
573
00:30:58.160 --> 00:31:00.319
about how we were able to achieve a one million
574
00:31:00.319 --> 00:31:04.279
context window inside of VLLM, you know, which was not
575
00:31:04.319 --> 00:31:08.000
supported by AMD in the community. So we've we've done
576
00:31:08.000 --> 00:31:10.640
a lot of software innovation in that space as well.
577
00:31:10.640 --> 00:31:12.319
But at the end of the day, you cannot run
578
00:31:12.640 --> 00:31:16.000
the need for raw horsepower. And I think as we
579
00:31:16.039 --> 00:31:18.920
plan our MC three class data centers over the next
580
00:31:18.960 --> 00:31:22.559
twelve months, those are going to be heavily segregated into
581
00:31:23.039 --> 00:31:24.920
dedicated influence and dedicated training.
582
00:31:28.480 --> 00:31:31.920
So, Dave, a lot of the focus is on public
583
00:31:31.960 --> 00:31:36.319
interest compute, that Australia needs more availability for compute for researchers.
584
00:31:36.640 --> 00:31:39.880
If we had more public interest compute, would that have
585
00:31:40.079 --> 00:31:43.519
changed the way that you went about building compute clusters
586
00:31:43.759 --> 00:31:46.599
or building main code or is that like an entirely
587
00:31:46.640 --> 00:31:47.839
separate conversation for you.
588
00:31:48.680 --> 00:31:51.359
It wouldn't have join us. So when you dig into
589
00:31:51.400 --> 00:31:55.240
the dark murky waters off academic clusters, what you find
590
00:31:55.319 --> 00:31:59.400
is that there it's unfortunate, but it's quite a dark space, right.
591
00:31:59.440 --> 00:32:06.440
It's overrun by nineties era it enterprise resellers, and we
592
00:32:06.519 --> 00:32:09.359
hear this all the time and we experience it actually ourselves.
593
00:32:09.359 --> 00:32:13.039
So when we build MC one, we worked with an integrator.
594
00:32:13.240 --> 00:32:16.000
You know, they didn't really understand AI, but obviously wanted
595
00:32:16.000 --> 00:32:18.240
to sell us stuff and let us starn a path
596
00:32:18.279 --> 00:32:20.920
which was just not usable at all. And so that's
597
00:32:20.960 --> 00:32:22.960
actually what made us go, hang on, we have to
598
00:32:23.039 --> 00:32:25.839
learn this ourselves by scratch, design everything and own everything
599
00:32:25.880 --> 00:32:28.839
ourselves because we can't rely on these folks who have
600
00:32:28.880 --> 00:32:31.559
been around since two thousands. We sell to Latrobe and
601
00:32:31.640 --> 00:32:35.160
Deacon and then you speak like I've had conversations with
602
00:32:35.200 --> 00:32:38.880
like soup you know, so o K and when you
603
00:32:38.960 --> 00:32:41.160
hear about like what happens when they buy these classes
604
00:32:41.200 --> 00:32:44.039
and then they sit there. The Maverick Cluster is a
605
00:32:44.039 --> 00:32:46.480
good example now too, right, Like it's kind of sitting
606
00:32:46.519 --> 00:32:50.000
in this limes state where people aren't able to use it,
607
00:32:50.000 --> 00:32:53.039
this infighting about how they run the cluster, how they
608
00:32:53.079 --> 00:32:56.519
run jobs, like it's just it's just a massive asset
609
00:32:56.599 --> 00:32:59.079
going to waste, right, And I think what you find
610
00:32:59.200 --> 00:33:02.319
is that people think that, oh, I just go and
611
00:33:02.359 --> 00:33:04.440
speak to one of these resellers through it in video.
612
00:33:04.880 --> 00:33:06.880
They'll sell me a bunch of b two hundreds and
613
00:33:06.920 --> 00:33:10.799
now I'm off the races and there It's like like
614
00:33:10.960 --> 00:33:12.960
it would be the equivalent of like when our parents
615
00:33:13.000 --> 00:33:15.440
bought their very first PC and brought it home. Right,
616
00:33:15.480 --> 00:33:18.680
It's like without the software stack, that thing is useless.
617
00:33:19.119 --> 00:33:21.880
You know, It's just it's just a glorified asset that
618
00:33:21.920 --> 00:33:23.599
you get to like a trophy that you get to
619
00:33:23.640 --> 00:33:26.400
show people and go back to Monish and ask them
620
00:33:26.400 --> 00:33:29.400
how many actual jobs that they run on Maverick, what
621
00:33:29.440 --> 00:33:31.960
aray are they actually getting from that spend and you
622
00:33:32.000 --> 00:33:33.799
will hear a very sad story.
623
00:33:34.119 --> 00:33:36.160
So do you think it's not worse investing in public
624
00:33:36.200 --> 00:33:39.079
interest compute? Then if we should focus on prime and sector?
625
00:33:39.880 --> 00:33:42.200
No, What I mean, what I meant to say, Joana,
626
00:33:42.359 --> 00:33:44.480
is that I think what we have to appreciate is
627
00:33:44.480 --> 00:33:46.960
that if we are going to build public infrastructure, we
628
00:33:47.039 --> 00:33:49.000
need to go to experts or are doing it today.
629
00:33:49.680 --> 00:33:53.720
And we and these old procurement pipelines that have paved
630
00:33:53.759 --> 00:33:55.640
the way for these companies that have been off for
631
00:33:55.680 --> 00:33:59.519
twenty years. You're not You're gonna get the same outcome
632
00:33:59.640 --> 00:34:02.039
Sirah and Alders got. So that's why I just want
633
00:34:02.039 --> 00:34:04.160
to be really clear, Right, Let's not keep doing the
634
00:34:04.240 --> 00:34:07.119
same thing expecting a different outcome and then asking why
635
00:34:07.160 --> 00:34:10.480
aren't we succeeding? That's not going to work, right. But Yeah,
636
00:34:10.480 --> 00:34:12.719
if someone was to come to main Code and say
637
00:34:12.960 --> 00:34:15.760
help build an MC two grade or an MC grade
638
00:34:15.880 --> 00:34:20.800
architecture for our university or our institution, we would absolutely
639
00:34:20.800 --> 00:34:23.639
help them. It would just look so differently than what
640
00:34:23.639 --> 00:34:26.000
you would get if you went to the classic Precuman
641
00:34:26.079 --> 00:34:28.280
path out to industry.
642
00:34:28.239 --> 00:34:31.280
Which is essentially a much more technical way of saying
643
00:34:31.400 --> 00:34:34.719
what I say, which is that our existing supercomputing infrastructure
644
00:34:34.760 --> 00:34:36.599
is old and clunky and we need to update it
645
00:34:36.639 --> 00:34:38.880
and bring it into the twenty first century. Yours is
646
00:34:38.920 --> 00:34:41.639
a more sophisticated response, but we get to the same point.
647
00:34:42.239 --> 00:34:44.320
And that's why we built main cloud because one of
648
00:34:44.400 --> 00:34:46.960
the things like our own team, right, our own researchers,
649
00:34:46.960 --> 00:34:50.920
because we have academic research residency researchers, we have our
650
00:34:50.920 --> 00:34:54.280
own full time team. They all come from academia, right like,
651
00:34:54.360 --> 00:34:57.480
and so all of these challenges that we were facing
652
00:34:57.519 --> 00:34:59.400
were like, we need to just reboot this whole thing.
653
00:34:59.400 --> 00:35:03.000
So we built in cloud, which is this incredibly sophisticated
654
00:35:03.320 --> 00:35:05.360
layer that sits on top of the training cluster. And
655
00:35:05.360 --> 00:35:08.719
that's what allows you to take that really expensive asset
656
00:35:08.760 --> 00:35:11.880
and suddenly distribute it evenly and fairly to everyone in
657
00:35:11.920 --> 00:35:14.519
a really high performance, managed way. But you have to
658
00:35:14.639 --> 00:35:18.800
listen to the researcher. If you're letting it and procurement
659
00:35:19.000 --> 00:35:22.559
and finance tell the research organization what they need, you
660
00:35:22.639 --> 00:35:24.960
get no research. But if you listen to them and
661
00:35:25.000 --> 00:35:27.599
you hear what they want and you understand how important
662
00:35:27.800 --> 00:35:31.480
running their own custom jobs. Having that cantenorization using the
663
00:35:31.519 --> 00:35:34.639
latest tools, now you get some incredible And then that's
664
00:35:34.639 --> 00:35:39.000
why a lot of these academic organizations and researchers come
665
00:35:39.000 --> 00:35:41.519
to work on the residency program because they're like, I
666
00:35:41.519 --> 00:35:45.519
can't advance my research at A and you or these places.
667
00:35:45.559 --> 00:35:46.960
I'm going to come to main Code where I can
668
00:35:47.000 --> 00:35:48.760
run it on one of their clusters.
669
00:35:49.239 --> 00:35:52.519
So skills, A lot of people say we don't have
670
00:35:52.679 --> 00:35:56.960
enough skills in Australia. What's your experience from a skills perspective.
671
00:35:57.000 --> 00:35:59.000
Has it been hard to get the right people to
672
00:35:59.039 --> 00:36:00.679
do it is?
673
00:36:00.960 --> 00:36:05.199
Yeah, it's hard to build a team. We've only been
674
00:36:05.239 --> 00:36:09.519
able to build a solid ten engineering team over a
675
00:36:09.599 --> 00:36:12.480
year and a half. A bar is excruciatingly high. But
676
00:36:12.920 --> 00:36:15.760
I also think that we lack some of the cultural
677
00:36:15.760 --> 00:36:20.199
alignment around mission driven work. I think a lot of
678
00:36:20.199 --> 00:36:23.000
people struggle when you say this is not a job.
679
00:36:23.079 --> 00:36:26.119
We're trying to actually deliver AI for Australia, just as
680
00:36:26.199 --> 00:36:29.760
much as we don't have ten thousand Olympians. Yeah, it's
681
00:36:29.800 --> 00:36:31.760
hard to find the kind of people to do this work.
682
00:36:33.599 --> 00:36:37.079
You mentioned copyright was old news. Why do you say
683
00:36:37.119 --> 00:36:40.760
that what approach to copyright does main code take and
684
00:36:41.119 --> 00:36:43.440
do you think that the conversation still needs to be
685
00:36:43.480 --> 00:36:44.239
had more broadly?
686
00:36:44.760 --> 00:36:46.440
So the reason I say it's old news is I
687
00:36:46.440 --> 00:36:50.000
think it's been a solved problem and it's a settled conversation. Right,
688
00:36:50.159 --> 00:36:53.960
just don't steal people's things. It's really that simple. Obviously,
689
00:36:53.960 --> 00:36:57.480
that's quite a simplistic view. How do you not steal
690
00:36:57.519 --> 00:37:02.920
people's things. You optimize your training phases to take as
691
00:37:03.000 --> 00:37:05.719
much of the data that's needed that doesn't require copyright
692
00:37:05.960 --> 00:37:10.119
to inform the model's base capability, language capability, and then
693
00:37:10.280 --> 00:37:12.960
post train on the stuff that you actually need that's less.
694
00:37:13.519 --> 00:37:16.079
So that's one of the things we learned very early
695
00:37:16.199 --> 00:37:18.840
was when you change the mixture and the process of
696
00:37:18.880 --> 00:37:21.320
model training, you can actually get a lot more done
697
00:37:21.320 --> 00:37:23.960
with less. And I think the reason was back in
698
00:37:24.000 --> 00:37:25.880
the day, they didn't think that way, so they just
699
00:37:26.159 --> 00:37:29.119
swoove it up everything they needed. But even today though,
700
00:37:29.199 --> 00:37:31.880
Joanna like, even the big labs aren't using as much
701
00:37:32.000 --> 00:37:35.519
off the copyright data as before it and they're able
702
00:37:35.519 --> 00:37:38.360
to deliver state of the art models without infringing anyone's copyright.
703
00:37:38.840 --> 00:37:42.199
So I think this is again where as the builders
704
00:37:42.440 --> 00:37:45.599
evolve and learn how to do it better, these pundits
705
00:37:45.639 --> 00:37:48.639
need to follow the ball, right because they're still going like, ah,
706
00:37:48.639 --> 00:37:51.280
but back in twenty twenty five, you were still here,
707
00:37:51.360 --> 00:37:53.480
but we don't do that anymore, Like those data sets
708
00:37:53.679 --> 00:37:56.800
aren't there anymore, and we don't need them because we've
709
00:37:57.000 --> 00:38:00.320
advanced the model architectures and the training pro this is
710
00:38:00.360 --> 00:38:03.599
to not need it. So that's kind of why it's
711
00:38:03.639 --> 00:38:06.239
a bit of a no non topic or a noop
712
00:38:06.360 --> 00:38:09.119
at the moment. But when there's bigger issues I think
713
00:38:09.280 --> 00:38:12.559
at play, like I'll tell you for example, like is
714
00:38:12.639 --> 00:38:15.880
copyright as much of an issue to Australia as the
715
00:38:15.880 --> 00:38:18.480
fact that we aren't investing in our own stead of
716
00:38:18.480 --> 00:38:21.599
the art technology and our systems are being hacked by
717
00:38:21.639 --> 00:38:25.480
overseas players. Right that you pick which one you know,
718
00:38:25.519 --> 00:38:27.000
like is Keate, so Brown are going to be okay
719
00:38:27.320 --> 00:38:30.400
that her lyrics are in the model versus freaking you know,
720
00:38:30.440 --> 00:38:33.199
Australia getting taken over like and shall be okay?
721
00:38:33.280 --> 00:38:36.639
Right, yeah, And I think there are lots of We
722
00:38:36.679 --> 00:38:39.519
could continue that conversation for a whole other hour, but
723
00:38:39.599 --> 00:38:42.880
we'll keep moving on one of the other One of
724
00:38:42.880 --> 00:38:47.159
the other claims that you really lean into is that
725
00:38:47.559 --> 00:38:51.440
models from overseas don't reflect Australian values. So if I'm
726
00:38:51.480 --> 00:38:55.760
sitting down with main codes Matilda, will I notice a difference?
727
00:38:55.800 --> 00:38:57.440
And I can answer this question because I actually did
728
00:38:57.440 --> 00:38:59.639
sit down with Matilda, but I'd like to hear it
729
00:38:59.679 --> 00:39:02.360
from you. And why is that difference there? Is it
730
00:39:02.400 --> 00:39:05.000
because you've trained it differently? Is it the data has
731
00:39:05.039 --> 00:39:07.440
been Australian data? What is that point of difference?
732
00:39:07.519 --> 00:39:10.760
Yep. So, and there's research on the way, which is
733
00:39:10.800 --> 00:39:14.119
a research paper we're publishing called Voice of Matilda. So
734
00:39:14.159 --> 00:39:17.880
the Voice of Matilda project started almost a year ago,
735
00:39:18.239 --> 00:39:20.480
and it was very clear that we wanted Matilda to
736
00:39:21.639 --> 00:39:26.840
speak not just lexically, but we wanted Matilda to respond
737
00:39:26.920 --> 00:39:30.280
to people like Australians would respond. So we took away
738
00:39:30.320 --> 00:39:36.159
the sick of fancy, the hyperbole, the blooming off token use, right,
739
00:39:36.199 --> 00:39:39.000
which I think is not good for customers. And when
740
00:39:39.000 --> 00:39:41.280
you talk to Matilda you get a straight answer, right,
741
00:39:41.360 --> 00:39:44.400
like that's what's beautiful about Matilda's Matilda will not waffle on.
742
00:39:44.639 --> 00:39:47.840
Matilda would just be like nah, Like I asked Matilda
743
00:39:48.000 --> 00:39:50.400
something The other day, and Matilda didn't even like, you know,
744
00:39:50.480 --> 00:39:52.639
like the classic open a bullshit right of like, oh,
745
00:39:52.719 --> 00:39:54.639
what a great question the time of our age as Dave,
746
00:39:54.679 --> 00:39:57.559
you're so smart, right, Like Matilda just went, oh nah,
747
00:39:57.920 --> 00:39:59.880
don't need to worry about it, and I went okay,
748
00:40:00.199 --> 00:40:04.800
So I love that. That's like like that we've designed
749
00:40:04.800 --> 00:40:08.599
and yes, that's deep research, deep work that's been done
750
00:40:09.000 --> 00:40:13.119
to train and find tune Matilda to be very conscious
751
00:40:13.119 --> 00:40:17.679
of the voice that Matilda uses and customers absolutely adore it.
752
00:40:17.760 --> 00:40:20.159
And not even Aussie customers, you know, we've got people
753
00:40:20.199 --> 00:40:22.639
using it overseas and they're just like this is so refreshing,
754
00:40:23.159 --> 00:40:25.519
like you know, just such a straightforward, you know, and
755
00:40:25.960 --> 00:40:28.920
again it embraces this Australian made spirit. It's like this
756
00:40:28.920 --> 00:40:32.559
feels Australian right, just fair and you know, good, right
757
00:40:32.599 --> 00:40:33.400
and hard working.
758
00:40:34.119 --> 00:40:38.800
And you're clear that Matilda has been trained on Australian
759
00:40:39.199 --> 00:40:44.400
non copywriter data and that's something that you're particularly proud of.
760
00:40:45.079 --> 00:40:47.639
Yeah. Absolutely, And the thing is as well channel like
761
00:40:47.679 --> 00:40:49.760
you know, our data team has been in existence for
762
00:40:50.159 --> 00:40:53.079
a long time now and when they grade the data
763
00:40:53.159 --> 00:40:56.039
right if from green, orange, red, it's just so clear, right,
764
00:40:56.079 --> 00:40:58.480
that like the stuff that's read you don't need. And
765
00:40:58.559 --> 00:41:00.920
even the tax work that we're doing right now, like
766
00:41:01.079 --> 00:41:04.559
very innovative tax work that we're building into Matilda, all
767
00:41:05.599 --> 00:41:09.440
open data, all non copyright. You know, you can build
768
00:41:09.440 --> 00:41:13.639
a significant capability without ever having an infringe on someone's copyright.
769
00:41:14.000 --> 00:41:16.840
Yeah, And for the paper listening to that saying, but
770
00:41:16.960 --> 00:41:20.599
are you taking advantage of models that have come before
771
00:41:20.880 --> 00:41:24.360
that have used copyrighted material? Just to be really clear,
772
00:41:24.400 --> 00:41:28.440
that isn't the case with Matilda. You've started from scratch, So.
773
00:41:28.559 --> 00:41:31.599
If you mean like distillation and using other models to
774
00:41:31.679 --> 00:41:35.840
cross train, so the base models that get used have
775
00:41:35.920 --> 00:41:38.920
it because again, we're all using the same pile of data, right,
776
00:41:39.519 --> 00:41:43.639
Do we go through every vector checking the provenance of
777
00:41:43.760 --> 00:41:46.440
that input souce No, that would just be absolutely insane
778
00:41:46.679 --> 00:41:50.519
and no one could do that anyway. Do we actively
779
00:41:50.679 --> 00:41:54.320
inject data that we know to be copyright Absolutely not,
780
00:41:54.840 --> 00:41:55.440
absolutely not.
781
00:41:56.639 --> 00:41:58.960
And I think that's an important distinction which maybe we'll
782
00:41:59.039 --> 00:42:03.400
unpack in a future episode. For now, Dave, can we
783
00:42:03.880 --> 00:42:09.079
end by asking two questions of you? One, what to
784
00:42:09.239 --> 00:42:13.920
you is success in one year's time for the Office
785
00:42:14.000 --> 00:42:18.440
of AI that the PM announced in his speech.
786
00:42:20.400 --> 00:42:24.559
Honestly, for me, it's that Australians are using Australia.
787
00:42:24.079 --> 00:42:27.719
Made AI and then looking back in ten years, so
788
00:42:27.880 --> 00:42:31.519
imagine we're recording another podcast, but it's twenty thirty six.
789
00:42:32.320 --> 00:42:37.199
Australia is globally recognized for having Australian made AI, which
790
00:42:37.199 --> 00:42:40.400
we are exporting to the world, and it is the
791
00:42:40.440 --> 00:42:45.199
powerhouse of our economy. The PM's speech on the fifteenth
792
00:42:45.239 --> 00:42:48.039
of July twenty twenty six is considered to be the
793
00:42:48.119 --> 00:42:53.679
turning point in Australia's economic but also social history. What happened,
794
00:42:54.000 --> 00:42:57.519
what did what had to happen in twenty twenty six
795
00:42:58.079 --> 00:42:59.400
to make that future?
796
00:42:59.559 --> 00:43:05.440
In twenty Australians started to believe in ourselves and I
797
00:43:05.440 --> 00:43:08.079
think they they made it clear to the government that
798
00:43:08.119 --> 00:43:12.599
if the government didn't back Australians and support Australian innovation,
799
00:43:13.360 --> 00:43:16.480
they weren't going to be in power. And I think
800
00:43:16.760 --> 00:43:19.239
it's so much more simple. At the end of the day, Jana,
801
00:43:19.599 --> 00:43:22.039
talk to your friend and say, do you feel passionate
802
00:43:22.039 --> 00:43:25.280
about us using our own products anymore? Then you feel
803
00:43:25.320 --> 00:43:28.760
passionate about you're buying Aussie walnuts or you know, buying
804
00:43:28.880 --> 00:43:31.400
an Aussie car, Like how much do you care about
805
00:43:31.440 --> 00:43:35.760
being Australian? And I think if Australians suddenly wake up
806
00:43:35.760 --> 00:43:37.519
and go, I want to use Aussie stuff, I want
807
00:43:37.519 --> 00:43:39.320
to actually, I want this to be here. I want
808
00:43:39.320 --> 00:43:41.599
to be proud of our industry. I want to see
809
00:43:41.599 --> 00:43:44.320
my children work in this industry and do well and
810
00:43:44.360 --> 00:43:48.639
have their shot. And the government goes, damn, this matters.
811
00:43:48.719 --> 00:43:50.440
We've got to get behind these folks, right, and we've
812
00:43:50.440 --> 00:43:52.480
got to do the work that they did for the
813
00:43:52.519 --> 00:43:56.360
tin minds and the smelters and the freaking sheep shearing
814
00:43:56.440 --> 00:43:59.800
stations and all that. Right. I think that's the turning point.
815
00:43:59.800 --> 00:44:03.079
That's what turns Australia into a producer and a maker
816
00:44:03.920 --> 00:44:05.840
because the government goes, man, this is where we need
817
00:44:05.880 --> 00:44:09.639
to put our money. But right now, the flip side
818
00:44:09.639 --> 00:44:12.519
of that is that they don't grow a backbone and
819
00:44:12.559 --> 00:44:15.639
they keep giving the money away for free flights to wherever.
820
00:44:16.239 --> 00:44:21.440
And ten years from now when America goes actually, we
821
00:44:21.559 --> 00:44:24.639
don't want Australia to be that smart. So we're going
822
00:44:24.719 --> 00:44:28.079
to restrict the supply of tokens. And what are we
823
00:44:28.119 --> 00:44:30.840
going to do? Grovel on our hands in our feet
824
00:44:30.880 --> 00:44:34.760
and go please you know, and That's my biggest fear
825
00:44:34.840 --> 00:44:37.920
is I don't want my kids to be disadvantaged in
826
00:44:38.000 --> 00:44:42.360
a world where AI absolutely is the differentiator. If you
827
00:44:42.360 --> 00:44:43.920
have a country that has AI and you have a
828
00:44:43.920 --> 00:44:47.159
country that doesn't have AI, that country with AI is
829
00:44:47.159 --> 00:44:50.280
going to crush the other country, Sublizzer.
830
00:44:50.360 --> 00:44:53.719
And the country that makes the AI as opposed to
831
00:44:54.000 --> 00:44:57.280
a country that adopts the AI fucking sneak in. One
832
00:44:57.440 --> 00:45:00.639
final question for you, Dave, it would be very difficult
833
00:45:00.639 --> 00:45:04.920
for Australia to build only Australian AI, right, just the
834
00:45:05.000 --> 00:45:07.800
cost involved in it. There are some types of AR
835
00:45:07.880 --> 00:45:10.079
that we are probably going to need to use from overseas.
836
00:45:10.519 --> 00:45:12.840
How do you think about that distinction? Do you do
837
00:45:12.920 --> 00:45:15.360
you think do you accept that in some instance as
838
00:45:15.400 --> 00:45:17.599
we will need to have overseas model as well.
839
00:45:17.639 --> 00:45:20.480
Like a year and a half ago, Joanna, like we
840
00:45:20.559 --> 00:45:23.559
stood there going like where do you start? And where
841
00:45:23.559 --> 00:45:25.679
we are? A year and a half later, we're like, actually,
842
00:45:25.679 --> 00:45:28.039
none of this stuff is that hard, right, So then
843
00:45:28.079 --> 00:45:31.039
it really just comes down to like how convicted you know,
844
00:45:31.159 --> 00:45:33.880
like if the government was to go Dave, this is
845
00:45:33.920 --> 00:45:36.760
a national imperative, like the space program, like we need
846
00:45:36.800 --> 00:45:39.000
to be on the map with this thing. Of course,
847
00:45:39.000 --> 00:45:40.840
we're not going to be able to compete with a
848
00:45:40.960 --> 00:45:43.159
China or an American and we don't need to. That's
849
00:45:43.199 --> 00:45:46.159
the beautiful thing, right So it and it's not that
850
00:45:46.239 --> 00:45:48.599
black and white either, right, Like, we can be building
851
00:45:48.719 --> 00:45:52.039
very strong relationships, but we have to we have to
852
00:45:52.079 --> 00:45:54.880
at least be a producer, you know. Yeah, and I
853
00:45:54.880 --> 00:45:56.760
agree with you, Joanna, Like, are we going to produce
854
00:45:56.840 --> 00:46:00.639
every single drop of AI? No? But should we be
855
00:46:00.639 --> 00:46:04.199
producing a fair bit of it just in case? Ah? Yeah,
856
00:46:04.239 --> 00:46:10.519
pretty much, right, just as much as like electricity, telco, milk,
857
00:46:11.039 --> 00:46:13.960
bread like, like, these are primary industries that, yes, we
858
00:46:14.039 --> 00:46:16.559
don't produce all of it, and we probably maybe don't
859
00:46:16.559 --> 00:46:19.360
even produce the very best of it. But for the stuff,
860
00:46:19.800 --> 00:46:22.079
I'll just want to leave one little tidbit that kind
861
00:46:22.079 --> 00:46:25.440
of helps people get their head around this. So not
862
00:46:25.519 --> 00:46:28.679
everyone drives a Bentley and not everyone drives a Ferrari, right,
863
00:46:28.920 --> 00:46:31.800
And when you're stuck on Collins Street at nine am,
864
00:46:31.920 --> 00:46:35.679
the difference between the Toyota, the Bentley the Ferrari actually
865
00:46:35.679 --> 00:46:38.280
doesn't matter because ninety percent of the time you just
866
00:46:38.280 --> 00:46:42.000
need to get to work. Australia needs to own that, right,
867
00:46:42.039 --> 00:46:43.519
we need to be able to make sure that we
868
00:46:43.639 --> 00:46:47.199
have AI to keep us moving. We won't may never
869
00:46:47.239 --> 00:46:49.599
be the Bentley And that's okay, right because the people
870
00:46:49.639 --> 00:46:55.519
who need fable class models are actually fictional. Very few
871
00:46:55.519 --> 00:46:58.960
people actually need that level. We use Matilda code internally
872
00:46:59.280 --> 00:47:03.320
Matilda code. Bill's Matilda, right, So we're definitely state of
873
00:47:03.320 --> 00:47:05.519
the art in Australia then easily tip of the spear.
874
00:47:05.679 --> 00:47:07.880
So this whole thing of like you need a super
875
00:47:07.920 --> 00:47:11.159
weapon every ridiculous, right.
876
00:47:11.239 --> 00:47:14.679
So yeah, and what you're describing there, Dave, is exactly
877
00:47:14.719 --> 00:47:17.199
what we've advocated for in the AI Agency work at
878
00:47:17.280 --> 00:47:19.719
saying yes, we need to have international access to some
879
00:47:19.760 --> 00:47:22.199
world leading, world class models. We need to build the
880
00:47:22.239 --> 00:47:25.480
stuff domestically in Australia that's your sovereign control, so that
881
00:47:25.519 --> 00:47:27.239
we actually, you know, we're not at risk of it
882
00:47:27.320 --> 00:47:30.039
being turned off. We need that choice to be able
883
00:47:30.039 --> 00:47:32.400
to have some optionality between it, and we need to
884
00:47:32.480 --> 00:47:35.079
export leverage. We need the stuff that we're exporting. And
885
00:47:35.159 --> 00:47:38.199
so I think, I really I want to end this
886
00:47:38.440 --> 00:47:42.639
by just acknowledging the courage, the determination and the grit
887
00:47:42.760 --> 00:47:45.800
of you and your team. It's been incredible to watch it,
888
00:47:46.119 --> 00:47:48.519
but also the generosity that you have in terms of
889
00:47:48.599 --> 00:47:51.079
engaging with the research part of it. But also you
890
00:47:51.119 --> 00:47:54.400
know you were midshipping Matilda and your team gave us
891
00:47:54.440 --> 00:47:57.960
a case study. It's extraordinary. So thank you so much,
892
00:47:58.559 --> 00:48:00.840
and I hope this is the first of men conversations
893
00:48:00.840 --> 00:48:03.519
that we have and you know, looking forward to seeing
894
00:48:03.800 --> 00:48:05.239
Matilda taking over the world.
895
00:48:05.480 --> 00:48:11.480
Yeah, thanks so much, Johanna.
896
00:48:13.880 --> 00:48:16.679
Well, that's it for this episode of Tech Mirror, which
897
00:48:16.719 --> 00:48:19.559
is brought to you by the Tech Policy Design Institute.
898
00:48:20.000 --> 00:48:22.239
We're based here in Canberra on the lands of the
899
00:48:22.280 --> 00:48:26.519
Nunberwall Nambri people. If you found today's conversation useful or
900
00:48:26.519 --> 00:48:29.480
thought provoking, please do share it with a friend or
901
00:48:29.480 --> 00:48:32.679
a colleague, or leave a review and subscribe wherever you
902
00:48:32.719 --> 00:48:35.679
get your podcasts. For show notes, you can visit tech
903
00:48:35.719 --> 00:48:39.800
Policy dot Au Forward Slash podcast. And this podcast was
904
00:48:39.800 --> 00:48:43.639
made possible with thanks to the generous contributions from government,
905
00:48:43.800 --> 00:48:47.880
industry and philanthropy to the Tech Policy Design Fund, the
906
00:48:47.920 --> 00:48:51.360
full details of which are available on our website. The
907
00:48:51.440 --> 00:48:54.519
team at Audiocraft produced this pod on the lands of
908
00:48:54.559 --> 00:48:57.719
the Gadigel people of the Eora Nation. Music is by
909
00:48:57.800 --> 00:49:01.039
Talia Scopellis. A big thing. Thank you also to the
910
00:49:01.079 --> 00:49:04.519
team at the Tech Policy Design Institute, without whom this
911
00:49:04.599 --> 00:49:07.960
pod would not be possible. Thank you for joining us,
912
00:49:08.360 --> 00:49:11.480
and as always, get in touch and get involved.