00:02AI has been attacking things around the world for the last couple of months but I guess
00:06it's been of less interest here because it hasn't evolved Australia and now we've seen
00:10it on our home soil.
00:10Luckily it wasn't too serious but it was definitely a wake up call.
00:14Can't AI companies, when they're instructing these AI agents at the beginning of these
00:20tasks and this was a task given to this agent, can't they direct the agent not to do anything
00:27underhanded in the process?
00:29Yes, yes they can.
00:32That's called guardrails sometimes people know.
00:35But it's problematic.
00:36AI's aren't people so they don't always understand what we say.
00:40We can make mistakes in how we do the guardrails and we shouldn't really trust AI to police
00:46itself.
00:46There have been lots of examples of it escaping guardrails and also malicious actors, bad people,
00:52can just with a few commands themselves undo those commands, forget the last instruction
00:57would be a naive example of that.
00:58And so if there is so much uncertainty about what can happen, what's the likelihood that
01:05AI agents will eventually have the ability to get through highly sensitive defense systems
01:15which control things like nuclear weapons?
01:18Yeah, I mean that's the million dollar question.
01:23I would say the answer there is not so cheery.
01:26I haven't seen anything that suggests that we can have any confidence that that won't happen
01:36without changing what we do.
01:38AI is so new, we're not used to it.
01:41It's like we've got castles, they're great castles, we've got suits of armor, they're great suits of armor,
01:46but someone's just invented cannons and rifles and we're still talking about our castles.
01:50So something has to change.
01:51And it's abilities are improving exponentially.
01:55It's abilities are improving far faster than we can expect and the consequences we can't really expect.
02:02So we've seen it used in war in the Ukraine extensively.
02:04It's astonishing what it can do.
02:07All the major companies' AIs have escaped some time in the last few months and gone out and done bad
02:12things.
02:13The companies that own them didn't want them to.
02:17And these are companies doing this in a so-called benign fashion.
02:21What does this highlight about the dangers of this tech getting in the hands of people who have malicious intentions?
02:28Yes, that's right.
02:29It's like a genie that can do anything.
02:30It's like nuclear weapons but they're available to everyone.
02:34As a friend, AI is a fantastic and powerful friend.
02:38But as an adversary or someone who's out to get you, it's frightening.
02:43So, yeah, we really are in a funny moment in history, I think.
02:50I can't think of another moment like this except maybe the nuclear weapons scare in the Cold War where things
02:57could go one way or another.
02:58We could keep AI under control and it could help us or it could escape and it could be devastating
03:04for us.
03:04So, has the horse already bolted on regulation or can authorities and companies somehow get together now to ensure a
03:15catastrophic event doesn't happen?
03:18The horse hasn't bolted.
03:22Certainly bad things have happened, but we can't give up.
03:25I think this is a great time.
03:27We're lucky that this has happened in Australia and it's sharpened our focus like Medibank and Optus did so long
03:32ago.
03:34But at the same time, the data they got on the surface doesn't seem so serious.
03:37So, it's a very friendly wake up call because this could have been anything and it could have been far
03:42more sensitive data.
03:43It's not too late.
03:45The horse hasn't bolted, but we do need to act.
03:47We can't just keep riding around in our suits of armour.
03:50And how should authorities act now?
03:54So, countries can't assume, I think, that other countries have common interests in this.
04:02And we can't assume that the people that run the large tech companies, the tech bros, that their interests align
04:09with ours.
04:09So, I think probably it's a two-pronged attack.
04:12One is to try to foster international collaboration and cooperation where possible in this new order.
04:17But the other thing is to develop local things, local regulation, pursue things through local courts, have local, have a
04:25commission.
04:25We've got some great commissioners, a privacy commissioner, a safety commissioner.
04:29We probably need an AI safety commissioner.
04:32And we need to go, we can't just rely on our big friends here.
04:37We actually need to be active saying, keep humans in the loop.
04:40Don't have our nuclear weapons set up so that an AI can control.
04:43Interestingly, when Xi met Trump a while ago, when they were at loggerheads, they were quarrelling about everything.
04:50The one thing that he asked the US, I don't know if you remember this, is,
04:54can you please make sure that no computer systems control your nuclear launch?
04:58So, he was aware of the risk that it imposed.
05:01So, I think, yeah, we all need to wake up to that risk.
05:04It seems my ability to recover from labors.
05:06We'llкj the Nicol calls.
