TRANSKRİPT 🔗
WEBVTT
1
00:00:02.110 --> 00:00:07.652
Hi everyone, this is a quick announcement to let you know that I'm running a new series of episodes on the podcast.
2
00:00:07.772 --> 00:00:16.715
The series is called Compute in Australia, and I'll be exploring the question of whether and how we should scale up a large domestic compute industry here.
3
00:00:16.935 --> 00:00:29.540
Think many gigawatts of new AI data centers, some potentially drawing a full nuclear reactor's worth of power or more, which is an order of magnitude larger than any clusters currently operating on the continent.
4
00:00:30.398 --> 00:00:38.222
Compute simply refers to the processing power produced by special chips sitting in racks of servers, which are stacked inside those data centers.
5
00:00:38.702 --> 00:00:43.525
It's become the key input for both the training and running of modern AI systems.
6
00:00:44.326 --> 00:00:48.508
AI will be at least as big as the internet, maybe as big as the industrial revolution.
7
00:00:49.088 --> 00:00:54.416
And like past tech transitions, it's likely to shake up the global distribution of wealth and power.
8
00:00:55.177 --> 00:00:57.561
So what is a middle power like Australia to do?
9
00:00:58.102 --> 00:01:04.011
What would it look like if we were extremely proactive about ensuring that we come out of this transition
10
00:01:04.691 --> 00:01:05.872
better than we entered it.
11
00:01:06.612 --> 00:01:10.434
Where in the AI value chain can we become a significant player?
12
00:01:11.315 --> 00:01:15.357
At least to begin with, an obvious answer seems to be compute.
13
00:01:16.037 --> 00:01:17.979
I've been ruminating on this for a while now.
14
00:01:18.279 --> 00:01:29.105
Early last year, I started thinking about the economic opportunity, whether or not compute could become our next great export industry, and asking a bunch of economist friends what the answer to that question might be.
15
00:01:29.825 --> 00:01:35.131
But this year I started thinking a lot more about the strategic leverage compute could afford us.
16
00:01:35.892 --> 00:01:41.899
That new thinking was primarily motivated by my growing anxiety about our access to frontier AI.
17
00:01:42.760 --> 00:01:45.704
If the connection here isn't immediately obvious, let me explain.
18
00:01:46.905 --> 00:01:54.267
Pretty soon, AI will be so deeply woven into the substrate of society that it becomes a utility like water or the internet.
19
00:01:54.867 --> 00:02:00.129
It's hard to picture this possibility today, especially if you view AI as nothing more than frivolous chatbots.
20
00:02:00.909 --> 00:02:15.634
But imagine a not-too-distant future where almost all economic and social activity, from how businesses negotiate and transact to how you access hospitals and government services, is in some way mediated through AI agents.
21
00:02:16.814 --> 00:02:21.797
At that point, uninterrupted access to the AI substrate becomes critical.
22
00:02:22.457 --> 00:02:30.962
To take
Load full transcript
the internet as an analogy, Australia, more precisely the University of Melbourne, first gained access to the global internet in 1989.
23
00:02:31.582 --> 00:02:40.868
If you'd switched off our access in the early 90s, you would have pissed off some university researchers and early adopters, but you wouldn't have triggered societal upheaval.
24
00:02:41.708 --> 00:02:45.129
If you switched it off today, the story would be quite different.
25
00:02:45.829 --> 00:02:50.190
Of course, today's internet can't actually be switched off like that because it's decentralized.
26
00:02:50.730 --> 00:02:53.831
And this is where the AI slash internet analogy stops.
27
00:02:54.791 --> 00:03:01.953
Assuming frontier models continue to meaningfully outperform open source alternatives, access to the frontier remains crucial.
28
00:03:02.533 --> 00:03:13.221
But control over this access may be concentrated in the small number of companies building frontier AI and in the country or countries in which those companies are based.
29
00:03:14.122 --> 00:03:18.465
At least in the near term, Australia is unlikely to build its own frontier AI.
30
00:03:19.006 --> 00:03:23.930
We don't have the density of talent or the colossal amounts of capital that appear to be necessary.
31
00:03:24.710 --> 00:03:28.334
Only two countries are running that race, the United States and China.
32
00:03:29.095 --> 00:03:35.543
Australia will generally choose American models, but American AI companies exist under American law.
33
00:03:35.843 --> 00:03:38.706
They can be controlled and coerced by American governments.
34
00:03:39.487 --> 00:03:44.332
We caught a glimpse of this in March, when the Pentagon declared Anthropic a supply chain risk.
35
00:03:45.013 --> 00:03:56.880
And again in April, when Anthropic gave early access to a model with unprecedented cyber capabilities, Claude Mythos Preview, to selected organizations, all of whom were US domiciled or headquartered.
36
00:03:57.280 --> 00:04:00.562
No Australian business or government agency was on the initial list.
37
00:04:01.362 --> 00:04:15.129
And we saw yet another glimpse in June when the US government imposed export controls that denied foreign users access to Fable 5 and Mythos 5, though in practice this forced Anthropic to suspend access for all users.
38
00:04:16.210 --> 00:04:22.193
So the problem is, if AI becomes like water, a foreign government will have its hand on the tap.
39
00:04:23.033 --> 00:04:30.096
If middle powers like Australia are bound to import frontier AI, how do we ensure continued access?
40
00:04:30.976 --> 00:04:32.277
This brings us back to compute.
41
00:04:32.797 --> 00:04:43.161
In a loosely analogous way to how Australia gave up on acquiring nuclear weapons, yet maintained some influence over nuclear politics through its preponderance of uranium deposits,
42
00:04:43.921 --> 00:04:51.732
We may not build our own frontier AI, but we can host the vast amounts of compute needed to train and or run these models.
43
00:04:52.453 --> 00:04:59.723
We could use this as leverage for frontier access, an idea that's been called compute for access in AI policy circles.
44
00:05:00.681 --> 00:05:07.283
Very, very roughly, this could look like the Australian government striking deals with hyperscalers and major AI labs.
45
00:05:07.903 --> 00:05:19.027
We fast track approvals and facilitate grid access in return for the labs guaranteeing Australians access to frontier AI on equal terms to the US market, if not the US government.
46
00:05:19.607 --> 00:05:27.210
Of course, such an agreement doesn't bind US administrations, but once billions of dollars of capex are sunk into the Australian continent.
47
00:05:27.850 --> 00:05:40.452
And if we are contractually permitted to throttle data centers power supply or even step in and repurpose scarce compute entirely if we lose frontier access, it does incentivize the labs to advocate for our interests in Washington.
48
00:05:41.472 --> 00:05:46.835
Now, the purpose of this podcast series is not to advocate for Compute for Access specifically.
49
00:05:47.136 --> 00:05:50.638
That's just one example of the strategic benefits Compute might offer.
50
00:05:51.258 --> 00:05:52.619
There are credible alternatives.
51
00:05:53.119 --> 00:05:55.621
And as always, the devil will be in the details.
52
00:05:56.261 --> 00:06:01.945
But that's why I'm doing this as a series of interviews, so I can explore the issue from different perspectives.
53
00:06:02.425 --> 00:06:06.108
And so we can see where the guests answers line up and where they don't.
54
00:06:07.001 --> 00:06:15.207
The compute question, in my opinion, is the most important question in Australia at the moment, and I don't think it's being covered with the rigour it deserves.
55
00:06:16.088 --> 00:06:18.590
So for the next month, this is pretty much all I'm doing.
56
00:06:19.050 --> 00:06:24.835
I'll release episodes across the month as I record them, rather than all at once, because things are moving so quickly.
57
00:06:25.515 --> 00:06:33.701
First up is an interview with Andrew Charlton, who is arguably the key member of the Australian government shaping our AI strategy.
58
00:06:34.302 --> 00:06:35.423
More guests will follow.
59
00:06:36.183 --> 00:06:41.308
Subscribe to the podcast wherever you get your podcasts so that you don't miss the episodes as they drop.
60
00:06:41.669 --> 00:06:43.771
Thanks, and see you for the first episode.