Latest / Best of Business / Jack Tame: AI is breaking out

Transcript

Jack Tame: AI is breaking out · Best of Business

• 0:09
  Speaker 1: You're listening to a podcast from Newstalk st B. Follow
• 0:12
  Speaker 1: this and our wide range of podcasts now on iHeartRadio.
• 0:16
  Speaker 2: It has been yet another week of yet more unsettling news.
• 0:23
  Speaker 2: The conflict in the Middle East appears to be escalating again.
• 0:26
  Speaker 2: As of this morning, we face even higher tariffs sending
• 0:30
  Speaker 2: products to the US than under President Donald Trump's previous measure.
• 0:35
  Speaker 2: But if there is one piece of unsettling news I reckon,
• 0:38
  Speaker 2: we probably haven't yet grappled with sufficiently. It is the
• 0:42
  Speaker 2: latest humbling episode with one of the world's biggest AI companies.
• 0:48
  Speaker 2: So OpenAI admitted yesterday that one of its new models
• 0:53
  Speaker 2: has autonomously broken free of its shackles, escaped from the
• 0:58
  Speaker 2: walled off environment where it was supposed to be safely tested,
• 1:02
  Speaker 2: gained access to the wider Internet, and successfully hacked another company. Legally,
• 1:11
  Speaker 2: it is a gray area. There is a pretty strong
• 1:14
  Speaker 2: argument that this was a crime. If a person did it,
• 1:18
  Speaker 2: theoretically they could be charged. But of course a person
• 1:21
  Speaker 2: didn't do it. This wasn't a person. It was an
• 1:24
  Speaker 2: autonomous AI model going to fairly extreme lengths to try
• 1:29
  Speaker 2: and solve a problem that it had been posed by
• 1:32
  Speaker 2: its creators. Again, it was supposed to be contained. Those
• 1:38
  Speaker 2: same masters who posed it a challenge have thought that
• 1:43
  Speaker 2: this model was confined in what they call a sandbox,
• 1:46
  Speaker 2: which is a It's kind of like a digital playpen
• 1:50
  Speaker 2: of sorts, where theoretically, at least you can wall in
• 1:54
  Speaker 2: a model and stop it from escaping to the outside world.
• 1:59
  Speaker 2: But like a toddler who pulls himself up and climbs
• 2:02
  Speaker 2: out of a cot, this model got free. AI models
• 2:07
  Speaker 2: are now so good at exploiting cyber vulnerabilities they can
• 2:11
  Speaker 2: exploit the vulnerabilities of their creators. Ultimately, the model solved
• 2:17
  Speaker 2: the problem that it had been set, but at some cost.
• 2:23
  Speaker 2: There are all sorts of fascinating little details that make
• 2:27
  Speaker 2: this whole episode particularly interesting. So for starters, it looks
• 2:30
  Speaker 2: like open AI didn't really understand what their own model
• 2:35
  Speaker 2: was actually doing, so they couldn't tell in real time
• 2:39
  Speaker 2: that this thing had escaped and was roaming the Internet
• 2:42
  Speaker 2: at large. They couldn't pull a metaphorical power cable out
• 2:47
  Speaker 2: of the wall because they only realized what it had
• 2:49
  Speaker 2: done after the fact. If there's one thing to be
• 2:52
  Speaker 2: grateful for in this instant, it's that the target of
• 2:56
  Speaker 2: the AI hack wasn't a bank, or a cryptocurrency exchange
• 3:03
  Speaker 2: or a weapon system. It was a startup called hugging Face,
• 3:07
  Speaker 2: which is a kind of interface for open source AI.
• 3:11
  Speaker 2: So if anything, the boss of Hugging Face honestly sounded
• 3:14
  Speaker 2: kind of stoked. He was going to chaffed that his
• 3:18
  Speaker 2: business had been targeted. But of course, this is the
• 3:21
  Speaker 2: sort of thing that experts have been warning about four ages,
• 3:27
  Speaker 2: at the most extreme end of the AI dumerism scale.
• 3:33
  Speaker 2: They've suggested that these systems could ultimately decide humans are
• 3:37
  Speaker 2: an inconvenience and that the world would be a much
• 3:40
  Speaker 2: better place if computer systems were in charge instead. And
• 3:44
  Speaker 2: arguably one very important step on the road to AI armageddon.
• 3:49
  Speaker 2: Is this a model breaking free. If you can't be
• 3:55
  Speaker 2: bothered engaging with AI, or even if you think AI
• 3:59
  Speaker 2: is massively overhyped, this week marks a significant milestone. In
• 4:04
  Speaker 2: one sense, it is incredibly impressive that an autonomous system,
• 4:10
  Speaker 2: something completely free of human control, operating totally by itself,
• 4:16
  Speaker 2: is able to pull off something with this complexity. In
• 4:20
  Speaker 2: another sense, it is profoundly unsettling. If this is where
• 4:26
  Speaker 2: the technology is today, who can honestly say where it
• 4:31
  Speaker 2: might be in five or ten years time.
• 4:34
  Speaker 1: For more from News Talks at b Listen live on
• 4:37
  Speaker 1: air or online, and keep our shows with you wherever
• 4:40
  Speaker 1: you go with our podcasts on iHeartRadio