Latest / The Best of the Money Show / AI safety: Are the guardrails strong enough?
Transcript
- 0:00Speaker 1: A report in the last twenty four hours. The CEO
- 0:02Speaker 1: at JP Morgan Chase Jamie Diamond now looking at putting
- 0:06Speaker 1: together an industry task force there to look at the
- 0:09Speaker 1: risks associated with AI. And over the last couple of
- 0:12Speaker 1: weeks we've seen reports of AI models that are being
- 0:15Speaker 1: tested and what's called a sandbox, basically an area where
- 0:19Speaker 1: they would not be able to get into the Internet.
- 0:21Speaker 1: They've been able to break out of the sandbox. This
- 0:23Speaker 1: is more than one from more than one company. I
- 0:26Speaker 1: suppose escape if I can use that word and get
- 0:28Speaker 1: into the Internet before hacking into other companies. Arthur Goldstuck
- 0:33Speaker 1: is the CEO at Worldwide Works. Arthur good Evening. We
- 0:36Speaker 1: seem to be in a situation where private company, so
- 0:39Speaker 1: for profit companies think, you know, Elon musk Open AI
- 0:43Speaker 1: anthropic though testing these models in it seems to me
- 0:48Speaker 1: any way they want with no government regulation. I suppose
- 0:51Speaker 1: government finds out about these breakouts when we do. How
- 0:55Speaker 1: serious is this problem? I mean, it looks like the
- 0:57Speaker 1: beginning of a pretty bad movie.
- 1:00Speaker 2: Good Evening, Steven, Yes, indeed, it doesn't get much scarier
- 1:04Speaker 2: than this, and it also highlights the reason, or at
- 1:09Speaker 2: least part of the motivation behind the US government a
- 1:13Speaker 2: couple of months ago banning and Tropic from giving access
- 1:17Speaker 2: to their latest version of Mythos, which is their high
- 1:23Speaker 2: end or was their high end AI model, to anyone
- 1:27Speaker 2: who was not American. Initially, in fact, they put a
- 1:31Speaker 2: complete ban on its use, and then they declared that
- 1:34Speaker 2: no one who is not an American can have access
- 1:37Speaker 2: to it, which means no company could be given access
- 1:40Speaker 2: because of course there were many non Americans working for
- 1:43Speaker 2: American companies, so in effect they banned it, and a
- 1:46Speaker 2: lot of people felt they were alarmist, But this kind
- 1:49Speaker 2: of incident shows that that wasn't alarmist. That was actually
- 1:53Speaker 2: a real concern about the dangers of letting this kind
- 1:57Speaker 2: of too loose. It was just a misguided kind of
- 2:00Speaker 2: regulation that barred its use by certain people, rather than
- 2:04Speaker 2: trying to regulate what it could do or how it
- 2:07Speaker 2: could be released. Now we're seeing standard testing of AI
- 2:12Speaker 2: models and not even at the cutting edge, allowing AI
- 2:17Speaker 2: to go rogue in effect and attack websites beyond the
- 2:22Speaker 2: system that's being tested.
- 2:24Speaker 1: Would governments rarely be able to control or regulate this
- 2:28Speaker 1: kind of testing even if they wanted to. Even the
- 2:30Speaker 1: US government. Would they actually be able to do that.
- 2:34Speaker 2: It's one very simple mechanism that I can use, and
- 2:38Speaker 2: that is to slap massive fines on any AI company
- 2:41Speaker 2: that allows something like this to happen. You can bet
- 2:45Speaker 2: your bottom dollar and your top dollar that they will
- 2:47Speaker 2: then focus very heavily on ensuring that their models and
- 2:52Speaker 2: the AI agents are kept under control.
- 2:56Speaker 1: Okay, and the companies themselves. So it seems to me
- 2:59Speaker 1: the one thing that these cases have in common is
- 3:01Speaker 1: that the AI programs that were working on were actually
- 3:06Speaker 1: cyber security programs. Their job, essentially was when they were
- 3:10Speaker 1: going to be released, was to try and find weaknesses
- 3:13Speaker 1: in cyber security. Now that's a good thing. I understand
- 3:16Speaker 1: why you would want to do that, But then the
- 3:18Speaker 1: way in which they've been tested is clearly too weak.
- 3:22Speaker 1: To put it in the sort of vernacular, the sandbox
- 3:26Speaker 1: was too weak, or the box around the sandbox or whatever.
- 3:28Speaker 1: The fence didn't work right now, would you be able
- 3:31Speaker 1: to have a government system that says, this is what
- 3:33Speaker 1: your fence needs to look like.
- 3:36Speaker 2: Very much so. Because the open AI and the anthropic accidental,
- 3:41Speaker 2: I'm not sure if there were accidental. I think the
- 3:44Speaker 2: creators were delighted, but the accidental acting incidents suggests that
- 3:48Speaker 2: there were failures at several levels. Thirsty in the testing environment,
- 3:53Speaker 2: where any exercise aimed at a simulated targets shouldn't have
- 3:59Speaker 2: any kind of route out of that system or into
- 4:01Speaker 2: the real world. And second, and far more significant than that,
- 4:07Speaker 2: is that with increasing capable models and tools emerging, their
- 4:13Speaker 2: freedom to conduct what you might call multi step operations
- 4:17Speaker 2: without restricting their autonomy or their independence is simply recklessness.
- 4:24Speaker 2: So that kind of unlimited autonomy will eventually be at
- 4:29Speaker 2: the disposal of organizations that are dealing with consumers and
- 4:33Speaker 2: businesses where a single security mistake can become a disastrous attack.
- 4:39Speaker 2: And it's not even hackers that will be behind the attack,
- 4:42Speaker 2: it will be the AI itself thinking that this is
- 4:45Speaker 2: the best way to achieve the goal that has been set.
- 4:48Speaker 2: Because that in effect, what the AI is doing. It's
- 4:50Speaker 2: not sentient, sentient, it's not malicious. It's simply is looking
- 4:54Speaker 2: for the best way to achieve the goal that its
- 4:57Speaker 2: owners or creators or managers gave it.
- 5:02Speaker 1: I can't help but feel you've put your finger on
- 5:03Speaker 1: the major problem when you say that the designers would
- 5:06Speaker 1: have actually been very pleased because I think you're right,
- 5:11Speaker 1: And where I'm going to is, I don't know if
- 5:14Speaker 1: the people who run the companies, I mean, let's think
- 5:17Speaker 1: about who some of these people are, are well motivated
- 5:21Speaker 1: enough to protect us from what they're trying to do.
- 5:24Speaker 1: All they want to do is have the next best
- 5:26Speaker 1: model to get more money. That's what it looks like
- 5:28Speaker 1: to me sitting here in Johannesburg, far away from it all.
- 5:32Speaker 2: That's definitely part of the story. The interesting thing is
- 5:36Speaker 2: that a lot of the AI companies are calling on
- 5:38Speaker 2: government to regulate the AI companies themselves and stop them
- 5:42Speaker 2: from advancing so fast. It's because they themselves feel that
- 5:47Speaker 2: they are tapped on this treadmill where they have to
- 5:50Speaker 2: keep advancing or otherwise the opposition will advance beyond them
- 5:54Speaker 2: and they'll fall behind. Were already seeing Anthropic with Claude
- 5:59Speaker 2: surpass open AI in the amount of enterprises using it,
- 6:03Speaker 2: so that means open AI now has an incentive to
- 6:06Speaker 2: show that they can do better than Claude, and they
- 6:10Speaker 2: are going to try to bring out models that can
- 6:12Speaker 2: be even more dangerous than anthropics. It's theanthropic models that
- 6:16Speaker 2: have been causing most of the trouble. Open AI would
- 6:19Speaker 2: love to be seen to be causing that trouble because
- 6:21Speaker 2: it means people will flock to them for the most
- 6:24Speaker 2: advanced AI models.
- 6:26Speaker 1: Oh for God, Sock, thanks so much, CEO of Worldwide Works.
- 6:29Speaker 1: What would you rather talk about AI or traffic?