0:00
ceding control to artificial intelligence that's where the problem comes and it's going to be really
0:06
hard to suddenly press that big red stop kill button because you might be past you might be
0:10
past the point of no return well just first of all tell us exactly what happened the other day
0:14
with open ai that got people worried essentially they were testing uh new hacking capabilities of
0:20
what they call a frontier ai model one of the latest most high powered most advanced versions
0:25
that they're trialing. Now, this has been positioned as AI going rogue. There's a slightly
0:31
different story to this. It's close, but there is a slight difference. And effectively, have you
0:37
ever been 10 pin bowling when they put the rails up on the side of the alley to stop you going over
0:41
onto another alley? They basically said, drop all guardrails. So normally AI has a level of
0:47
protection built in, but that was switched off and it was told to do whatever it could to prove
0:52
its hacking prowess within a so-called sandbox not connected to the internet an area where it
0:57
could basically sort of like where armies go and have little towns made up on the top of a hill
1:01
somewhere something similar to that but what the ai essentially did because it took the human
1:06
it took the request of the uh human owner uh to prove its hacking prowess to the biggest limit
1:11
it thought it would go and prove that by breaking out of its sandbox and hacking into a real company
1:16
You could say that is Rise of the Machines, Terminator, all of those things coming to life
1:23
or you could say somebody very clumsily wrote the prompt that caused the problem initially
1:29
They are similar, but there is a slight difference. This thing hasn't started thinking for itself
1:33
It hasn't necessarily gone into business for itself at this stage. But it has woken up everybody to the potential about this happening and happening further One about six months ago Matt google had to switch off an ai test because various systems within this sandbox started talking to
1:51
each other in a language that the programmers and humans could not understand it had to switch it
1:55
off because and you could look at that and go has this thing become sentient is it talking in a way
2:00
that it's it's you know trying to uh surpass the human or has it decided there is a more efficient
2:06
route to talking than the language it was initially given. So there's sort of interpretation points
2:11
on many of these little matters. I wonder if that nuance is going to make us sleep better at night
2:17
because there was an interview given by Elon Musk to The Economist just recently where he said he
2:23
and you know he's not the dumbest guy on the planet as well as being the richest has spent a
2:28
lot of money on AI. He has predicted that in five years time artificial intelligence
2:34
will be sentient and will be cleverer than we are as human beings
2:39
Now, is he being, is this bonkers or is it going to happen? What do you think
2:44
I think there is an element of truth to what Elon Musk says, but he's also got skin in the game
2:49
He's trying to build an AI company and he has got a Pat Butcher, Peggy Mitchell-esque dislike
2:56
with Sam Altman at OpenAI. Those two have completely fallen out and they're trying to sue each other at each possible angle
3:03
So anything he can do to position himself as the safe future of artificial intelligence against others is a calculated business decision by Elon Musk
3:13
This also gets really messy because all of the AI companies are talking about the power and the potential danger in quotation marks of their models because they're all trying to get funding
3:24
This whole thing is burning billions of billions and billions of dollars and making no noticeable route to profit anywhere so these companies have to sell the sizzle they have to look sexy they have
3:35
to look edgy they have to look like they might solve humanity's problems because they need to
3:39
make money and um when you keep talking like that and you have incidents like this week it's highly
3:46
understandable that senators and politicians around the world suddenly think good god what
3:51
happens if this does go wrong give me that big red button i can start hitting the hitting to stop it
3:55
but i mean there's another company wasn't i think it was anthropic that developed a model called
3:59
mythos you know and then a few months ago they warned they themselves warned about their own
4:04
product saying this is going to be so good um that actually criminals could use it to hack into the
4:11
most secure accounts you know don't don't use this now was that a gimmick was that was that
4:16
performative or did they have a point two days later they accidentally gave it to a load of their
4:20
third-party customers so it couldn't have been that dangerous this is the and and those who
4:25
later tested it said this has some really interesting capabilities but it's nowhere like
4:30
as as game-changing and as truly world defining as as was suggested one thing remains though
4:36
when this tool that um open ai were testing uh last week uh was released it found errors in code
4:44
that open AI's own developers couldn't find. So this stuff is improving on, you know, a daily, hourly, momently basis
4:54
But on the opposite side of that, AI until now has been really good for tasks
4:58
We are now getting to a stage where you can say to an AI, go and run the US defense strategy for 2026-27
5:06
You can start doing big multiple things with it. And when you start giving handing over and ceding control to artificial intelligence that where the problem comes And it going to be really hard to suddenly press that big red stop kill button
5:20
because you might be past the point of no return. We've already seen how AI models can find their way out onto the internet
5:26
and sneak into different places. But this is the point, isn't it? I mean, the theme of the programme this morning is taking back control
5:31
I argued at the top of the show that Andy Burnham is giving people this sensation
5:36
Maybe it's only an illusion, but certainly the feeling, the sensation at the moment that whether it's through devolution you know or putting a little
5:42
bit more money in their pockets that they have a degree of control over their lives that they often
5:46
feel as voters that they don't my worry is that with with ai you've got a very small number of
5:52
fabulously wealthy individuals who are also incredibly clever and understand this stuff
5:57
better than most of us do apart from perhaps your good self um who aren't you know who are changing
6:02
the world all around us in a way that we cannot comprehend and once again we are made to feel like
6:08
the recipients of something you know not the the active participants of it and that's a problem i
6:14
agree yeah i agree it's all gas and no break at the moment and that is the challenge there is no
6:18
brake pedal and all of these companies are rushing off in different directions to try and get
6:24
advantage over others governments are arguably encouraging some of the larger ones to in order
6:29
to give them supremacy you've got government ministers who truly don't understand the value
6:35
of artificial intelligence or how it works being asked to mark the um the uh homework of humans and
6:41
ai models humans who are being paid billions of pounds billions of dollars a year to make these
6:45
new frontier models we're in a place where um it's not about legislation catching up it's about not
6:51
having right people in the right places and until governments take it seriously and start paying the
6:55
money big tech are paying you're never going to find yourself back in an even position
