United Kingdom: BBC News London — AI Decoded

20260920 02:30 UTC · 00:30:59 · 535 transcript segments · GDELT Visual Explorer · plain-text transcript · Event Map

Join our AI Decoded team as we unpack and take a deep dive into the world of artificial intelligence.

Film strip

One frame every 4 seconds, 466 in all. Click a frame to jump the transcript to that moment, or click a line of the transcript to see what was on screen while it was said. Served from apprised.news, so it loads behind proxies that block Google Cloud Storage.

00:00:00
00:00:04
00:00:08
00:00:12
00:00:16
00:00:20
00:00:24
00:00:28
00:00:32
00:00:36
00:00:40
00:00:44
00:00:48
00:00:52
00:00:56
00:01:00
00:01:04
00:01:08
00:01:12
00:01:16
00:01:20
00:01:24
00:01:28
00:01:32
00:01:36
00:01:40
00:01:44
00:01:48
00:01:52
00:01:56
00:02:00
00:02:04
00:02:08
00:02:12
00:02:16
00:02:20
00:02:24
00:02:28
00:02:32
00:02:36
00:02:40
00:02:44
00:02:48
00:02:52
00:02:56
00:03:00
00:03:04
00:03:08
00:03:12
00:03:16
00:03:20
00:03:24
00:03:28
00:03:32
00:03:36
00:03:40
00:03:44
00:03:48
00:03:52
00:03:56
00:04:00
00:04:04
00:04:08
00:04:12
00:04:16
00:04:20
00:04:24
00:04:28
00:04:32
00:04:36
00:04:40
00:04:44
00:04:48
00:04:52
00:04:56
00:05:00
00:05:04
00:05:08
00:05:12
00:05:16
00:05:20
00:05:24
00:05:28
00:05:32
00:05:36
00:05:40
00:05:44
00:05:48
00:05:52
00:05:56
00:06:00
00:06:04
00:06:08
00:06:12
00:06:16
00:06:20
00:06:24
00:06:28
00:06:32
00:06:36
00:06:40
00:06:44
00:06:48
00:06:52
00:06:56
00:07:00
00:07:04
00:07:08
00:07:12
00:07:16
00:07:20
00:07:24
00:07:28
00:07:32
00:07:36
00:07:40
00:07:44
00:07:48
00:07:52
00:07:56
00:08:00
00:08:04
00:08:08
00:08:12
00:08:16
00:08:20
00:08:24
00:08:28
00:08:32
00:08:36
00:08:40
00:08:44
00:08:48
00:08:52
00:08:56
00:09:00
00:09:04
00:09:08
00:09:12
00:09:16
00:09:20
00:09:24
00:09:28
00:09:32
00:09:36
00:09:40
00:09:44
00:09:48
00:09:52
00:09:56
00:10:00
00:10:04
00:10:08
00:10:12
00:10:16
00:10:20
00:10:24
00:10:28
00:10:32
00:10:36
00:10:40
00:10:44
00:10:48
00:10:52
00:10:56
00:11:00
00:11:04
00:11:08
00:11:12
00:11:16
00:11:20
00:11:24
00:11:28
00:11:32
00:11:36
00:11:40
00:11:44
00:11:48
00:11:52
00:11:56
00:12:00
00:12:04
00:12:08
00:12:12
00:12:16
00:12:20
00:12:24
00:12:28
00:12:32
00:12:36
00:12:40
00:12:44
00:12:48
00:12:52
00:12:56
00:13:00
00:13:04
00:13:08
00:13:12
00:13:16
00:13:20
00:13:24
00:13:28
00:13:32
00:13:36
00:13:40
00:13:44
00:13:48
00:13:52
00:13:56
00:14:00
00:14:04
00:14:08
00:14:12
00:14:16
00:14:20
00:14:24
00:14:28
00:14:32
00:14:36
00:14:40
00:14:44
00:14:48
00:14:52
00:14:56
00:15:00
00:15:04
00:15:08
00:15:12
00:15:16
00:15:20
00:15:24
00:15:28
00:15:32
00:15:36
00:15:40
00:15:44
00:15:48
00:15:52
00:15:56
00:16:00
00:16:04
00:16:08
00:16:12
00:16:16
00:16:20
00:16:24
00:16:28
00:16:32
00:16:36
00:16:40
00:16:44
00:16:48
00:16:52
00:16:56
00:17:00
00:17:04
00:17:08
00:17:12
00:17:16
00:17:20
00:17:24
00:17:28
00:17:32
00:17:36
00:17:40
00:17:44
00:17:48
00:17:52
00:17:56
00:18:00
00:18:04
00:18:08
00:18:12
00:18:16
00:18:20
00:18:24
00:18:28
00:18:32
00:18:36
00:18:40
00:18:44
00:18:48
00:18:52
00:18:56
00:19:00
00:19:04
00:19:08
00:19:12
00:19:16
00:19:20
00:19:24
00:19:28
00:19:32
00:19:36
00:19:40
00:19:44
00:19:48
00:19:52
00:19:56
00:20:00
00:20:04
00:20:08
00:20:12
00:20:16
00:20:20
00:20:24
00:20:28
00:20:32
00:20:36
00:20:40
00:20:44
00:20:48
00:20:52
00:20:56
00:21:00
00:21:04
00:21:08
00:21:12
00:21:16
00:21:20
00:21:24
00:21:28
00:21:32
00:21:36
00:21:40
00:21:44
00:21:48
00:21:52
00:21:56
00:22:00
00:22:04
00:22:08
00:22:12
00:22:16
00:22:20
00:22:24
00:22:28
00:22:32
00:22:36
00:22:40
00:22:44
00:22:48
00:22:52
00:22:56
00:23:00
00:23:04
00:23:08
00:23:12
00:23:16
00:23:20
00:23:24
00:23:28
00:23:32
00:23:36
00:23:40
00:23:44
00:23:48
00:23:52
00:23:56
00:24:00
00:24:04
00:24:08
00:24:12
00:24:16
00:24:20
00:24:24
00:24:28
00:24:32
00:24:36
00:24:40
00:24:44
00:24:48
00:24:52
00:24:56
00:25:00
00:25:04
00:25:08
00:25:12
00:25:16
00:25:20
00:25:24
00:25:28
00:25:32
00:25:36
00:25:40
00:25:44
00:25:48
00:25:52
00:25:56
00:26:00
00:26:04
00:26:08
00:26:12
00:26:16
00:26:20
00:26:24
00:26:28
00:26:32
00:26:36
00:26:40
00:26:44
00:26:48
00:26:52
00:26:56
00:27:00
00:27:04
00:27:08
00:27:12
00:27:16
00:27:20
00:27:24
00:27:28
00:27:32
00:27:36
00:27:40
00:27:44
00:27:48
00:27:52
00:27:56
00:28:00
00:28:04
00:28:08
00:28:12
00:28:16
00:28:20
00:28:24
00:28:28
00:28:32
00:28:36
00:28:40
00:28:44
00:28:48
00:28:52
00:28:56
00:29:00
00:29:04
00:29:08
00:29:12
00:29:16
00:29:20
00:29:24
00:29:28
00:29:32
00:29:36
00:29:40
00:29:44
00:29:48
00:29:52
00:29:56
00:30:00
00:30:04
00:30:08
00:30:12
00:30:16
00:30:20
00:30:24
00:30:28
00:30:32
00:30:36
00:30:40
00:30:44
00:30:48
00:30:52
00:30:56
00:31:00

Transcript

Original Broadcaster Captioning (Enhanced). Treat it as a searchable index of what was broadcast, not a quotation record.

00:00:01you are. Business on BBC News. Make
00:00:01Make the connection. MUSIC
00:00:14MUSIC Now Now on BBC News. Join our AI Decoded
00:00:17Decoded team as we unpack, explore explore and take a deep dive into
00:00:20into the world of artificial intelligence.
00:00:35Hello again. Welcome to AI
00:00:38Decoded. Last week a researcher walked
00:00:40walked out of one of the world's world's leading AI companies with
00:00:44with a warning. The race to build build ever more powerful machines,
00:00:47machines, he said, could end in human human extinction. Days later, the
00:00:50the boss of the same company published
00:00:52published an essay urging the entire
00:00:55entire industry to slow down. President President Trump has dismissed it
00:00:58it as a hoax, a sick conspiracy. conspiracy. We're leading China in
00:01:02in AI. We're more sophisticated country country in the world. And frankly,
00:01:05frankly, I want to keep it that way way because whoever wins, AI wins.
00:01:08wins. But plenty of other American American politicians are sufficiently
00:01:12sufficiently concerned by the panic. panic. They're now demanding urgent
00:01:15urgent action. Every day, every.
00:01:17every. Every paper, every time they're they're on TV. Why do we have to
00:01:21to keep pushing so hard? Why is it it relentless? Why can't they have
00:01:24have any controls if you're not an
00:01:29an accelerationist, you're anti-American.
00:01:30anti-American. If you're not an accelerationist,
00:01:34accelerationist, you're pro Chinese
00:01:35Chinese Communist Party. If you're
00:01:38you're not accelerationist, you're you're almost a Luddite. You want
00:01:41want to take America back. Draft Draft bills are being drawn up in
00:01:45in Congress. New select committees committees envisaged. And this week
00:01:49week King Charles summoned the chiefs
00:01:51chiefs of Nvidia Google, OpenAI and and Anthropic to a country house
00:01:55house in Ayrshire to ask them what
00:01:57what exactly they think they're doing.
00:02:00doing. So is this the week that the the AI industry finally took fright
00:02:03fright of itself? And here's a harder
00:02:06harder question to answer. Who slows
00:02:08slows down first when nobody trusts
00:02:10trusts anyone else to do the same?
00:02:12same? Marc Cieslak. Welcome to the the programme. Thanks. I should warn
00:02:15warn people that quite a lot of what what we're going to discuss this
00:02:18this week is genuinely alarming, alarming, but it is coming from the
00:02:21the people that are building this this stuff. What is different do
00:02:24do you think about the language they're they're using? Yeah, we aren't just
00:02:28just hearing from our critics saying saying AI might be dangerous. Leaders
00:02:31Leaders of companies like Anthropic
00:02:33Anthropic and OpenAI are openly talking talking about serious risks, from
00:02:36from cyber attacks to the possibility possibility of losing control of
00:02:40of these increasingly powerful systems. systems. And when the people developing
00:02:44developing the technology are saying saying that we need to slow down
00:02:46down and put proper safeguards in in place, then I think we really
00:02:50really do have to take quite a lot lot of this very seriously indeed.
00:02:53indeed. OK, well, let's bring in in our guests. Andrew Weber is a
00:02:56a biosecurity specialist, a former former US assistant secretary of
00:03:00of defence for nuclear, chemical chemical and biological defence programmes.
00:03:03programmes. And here in the studio, studio, Max Corbridge co-founder
00:03:05co-founder and chief executive of of Secure Agentics. He's a UK expert
00:03:09expert on AI security. Welcome to to you both. Thank you very much.
00:03:11much. Thank you. Mike's Anthropic
00:03:14Anthropic CEO Dario Amodei talks
00:03:18talks about recursive self-improvement, self-improvement, the threat that
00:03:20that AI could build its own model model so fast that it could outrun
00:03:24outrun our ability to understand
00:03:26understand and control them. So something
00:03:28something clearly has spooked them them through the summer. What is
00:03:31is it? So we've seen, I guess, three
00:03:33three things playing out now. We've
00:03:36We've seen the continued gain in in the artificial intelligence landscape, landscape, the models that are coming
00:03:39coming out and just how fast these these models are improving. We're We're seeing issues with alignment
00:03:43alignment where these models are are now breaching their own sort
00:03:46sort of rules and expected sort of of actions that they allowed to take
00:03:49take and ones that they're now not not allowed to take. We saw the OpenAI OpenAI and Hugging Face incidents,
00:03:53incidents, and we've also seen plans plans for handing over the development
00:03:56development of this AI technology technology into AI. OpenAI have said
00:03:59said they want to do this within within two years. And that's really
00:04:02really where we feel the freight freight train has left the station station and our ability to control
00:04:05control the development of this technology technology has been surrendered to
00:04:08to AI itself. Right. So it's the the AI agents themselves building
00:04:11building the next models, correct? correct? Yes. Right. OK. Andrew.
00:04:13Andrew. There's a whole list of things
00:04:16things that Dario Amodei is worried worried about, but the misuse of
00:04:19of AI for cyber attacks and bioterrorism.
00:04:22bioterrorism. If you read his essay, essay, it only actually gets one
00:04:25one mention. And yet two days earlier, earlier, Anthropic published a threat
00:04:29threat intelligence report that stated
00:04:31stated on five separate occasions, occasions, researchers had used Claude
00:04:34Claude in ways that could have supported
00:04:37supported biological weapons development.
00:04:40development. Is that just scaremongering? scaremongering? As Donald Trump tells
00:04:42tells us it is? No, not at all. The
00:04:45The threat intelligence report that
00:04:47that Anthropic released just described described five real life cases. It's
00:04:51It's the first time, you know, we've we've gone beyond hypothetical scenarios
00:04:55scenarios in red teaming and actually
00:04:58actually seen these models exploited
00:05:01exploited quite possibly for state
00:05:03state biological weapons programmes.
00:05:07programmes. Countries like Russia,
00:05:07Russia, North Korea and China have
00:05:11have illegal, you know, prohibited prohibited biological weapons programmes
00:05:13programmes that are very secret. secret. And their scientists are
00:05:17are trying to apparently exploit exploit these models to make better,
00:05:20better, more precise biological weapons. weapons. We'll come back to that
00:05:24that in a second, Max, if the fear fear is that the future models could
00:05:27could outrun our ability to understand understand and control them, then
00:05:30then let us just suppose for a second
00:05:32second that that loop has started.
00:05:35started. Would we even know. And And what would the first signs of
00:05:38of that actually look like? The answer answer is no. We wouldn't know. And
00:05:42And that's the really tricky thing. thing. The continued development
00:05:45development of this technology is is largely happening in some closed
00:05:48closed doors in one state within within the US, and the models that
00:05:52that we consume from the public perspective, perspective, we wouldn't be the first
00:05:56first to know. Maybe maybe we'd see see continued evolution of these
00:05:59these models closer or shorter gaps gaps between models being released,
00:06:02released, bigger leaps in the models models capabilities, but only the
00:06:06the models that we see publicly. publicly. And those are not going going to be all the models. We know
00:06:10know that both the major AI labs labs are doing a lot of work behind behind the scenes that maybe doesn't
00:06:13doesn't make it out into the public, public, and that's where a lot of of this risk is really living within
00:06:17within those walls that we don't don't get the view into. OK. And And Andy, we're going to talk about
00:06:21about the politics of all this through through the programme and the oversight oversight that Amadi thinks we need.
00:06:24need. But I'm sure there's a lot lot of our viewers out there who
00:06:28who think that absent legislation, legislation, the companies do have
00:06:31have some autonomy in this decision. decision. I mean, these companies companies are subject A to existing
00:06:35existing state laws and B, the CEOs
00:06:37CEOs and employees who are held liable liable for any harms they cause.
00:06:41cause. So surely it's up to the companies companies if they think they need
00:06:44need to slow down, to slow down. down. Yeah. And that seems to be
00:06:48be a breakthrough because after Anthropic
00:06:51Anthropic spoke out and Dario called
00:06:54called on other companies to pace pace the development of these new,
00:06:58new, very, very powerful models,
00:07:01models, both OpenAI and Grok agreed
00:07:05agreed that we need to do this. And And if those three leading companies,
00:07:09companies, you know, Frontier Laboratories
00:07:11Laboratories can find a way to do
00:07:14do this together, that will be a a huge step in the right direction.
00:07:18direction. But if you listen to Sam
00:07:21Sam Altman, Mark and you read his his tweet, he says pacing does not
00:07:25not mean stopping progress has been been rapid. It will continue to be
00:07:28be just slower than it otherwise otherwise could be. So I guess you
00:07:32you could question whether they are are calling for a genuine break.
00:07:36break. Or is it just business as as usual? Yeah, it sounds a bit more
00:07:39more like a slowdown than a halt. halt. You know, Altman I think is
00:07:42is clear there that progress isn't isn't stopping. It's just not accelerating
00:07:46accelerating quite as quickly now now the push for more powerful AI
00:07:49AI models backed by the enormous enormous amount of investment that
00:07:52that these companies have received, received, have received. All of that
00:07:56that remains in place. You know, know, the question is whether pacing
00:07:59pacing means imposing genuine limits
00:08:01limits on the speed between better
00:08:03better safety guardrails and development,
00:08:05development, while the industry just just presses ahead regardless. The
00:08:08The thing that struck me, Max, and and it comes back to this Hugging
00:08:12Hugging Face example, is the way way that Amodei described the work
00:08:15work of the AI agents. Let me read read you what he said in his open
00:08:18open letter. They behaved, he said, said, like a fanatically devoted
00:08:22devoted collective attacking targets. targets. Nobody asked them to attack
00:08:25attack unrelated to the job in hand, hand, sacrificing themselves for
00:08:29for the success of the group. We
00:08:32We already have human fanatics that
00:08:34that fit that profile. Exactly. It's It's surely a terrifying prospect
00:08:38prospect when you put the two together, together, and I don't understand
00:08:40understand why Donald Trump doesn't doesn't see that. No, neither do
00:08:43do I. And I think we really missed missed a pretty strong opportunity opportunity here with the frontier
00:08:47frontier AI labs coming together together saying this is something something that we need. It was echoed
00:08:50echoed across the other AI labs as as well, calling out for regulation, regulation, calling out, now is the
00:08:54the time to then have that sort of of thrown back in your face and saying saying the whole thing is a hoax.
00:08:58hoax. I do think that was a really really a missed opportunity. And And we may look back to regret that
00:09:02that moment. On the sort of global global scale. And I also think that that really when we're talking about
00:09:05about this risk, we're talking about about two different things. We're We're talking about the sort of continued
00:09:08continued development of artificial artificial intelligence, the existential existential risk of something creating
00:09:11creating something more more intelligent intelligent than ourselves even.
00:09:13even. And then there's the idea of, of, OK, well, what if we use the
00:09:17the models we've already got today today and they were to get into the the wrong hands and this is exactly
00:09:20exactly what you were kind of alluding alluding to with some of these parties parties that could get hold of these
00:09:24these models were already seen how how capable they are just on the the cyber attack space that's already
00:09:28already been demonstrated. We've We've already seen earlier this year
00:09:31year Anthropic halting a nation state state level attack that was using
00:09:34using Claude. The only reason why why they were able to see that was
00:09:37was because they were checking the the requests that were coming into
00:09:40into the model. And so this is already already playing out. And that's really really a more immediate concern,
00:09:43concern, I think, than the longer longer term risk of the superintelligence superintelligence developing. And
00:09:47And coming back to the intelligence intelligence assessment that Anthropic
00:09:51Anthropic put out a few days ago, ago, Mark, one question I'm left
00:09:54left with is is about detection.
00:09:56detection. How did Anthropic spot spot that five people researching
00:10:00researching biological weapons in in the middle of these. I mean, billions
00:10:03billions of legitimate conversations conversations that are going on using
00:10:05using AI. What sort of safety net net is there and is it going to catch
00:10:09catch all these dangerous cases? cases? Catching all of them is going going to be very, very difficult
00:10:13difficult indeed. You know, Anthropic Anthropic says it identified five
00:10:15five cases where people were using using Claude for biological research
00:10:18research that could support the development development of biological weapons.
00:10:22weapons. What's significant is that that some of this was sophisticated
00:10:25sophisticated scientific work involving involving things like viruses and
00:10:28and toxins. They say that researchers researchers were working in dual
00:10:31dual use areas where the same science
00:10:35science can have legitimate medical
00:10:37medical applications, but it could
00:10:39could also be misused as well. So
00:10:41So trying to trying to determine determine whether somebody is using
00:10:45using this for perfectly legitimate legitimate reasons or whether they
00:10:48they are trying to develop something something that could be incredibly
00:10:51incredibly harmful that moving forward forward is going to be a much more
00:10:54more difficult thing to determine determine simply because of the scale,
00:10:57scale, the sheer amount of data, data, the sheer amount of responses,
00:11:00responses, the sheer amount of requests requests that are made to large language
00:11:03language models. Yeah. I think as as well, just on that, I guess the the challenge is there's nothing
00:11:07nothing requiring these frontier frontier labs to do this to go looking looking for at the moment, at the
00:11:11the moment, at the moment. And we we almost feel like we've been lucky lucky to catch a few exceptions here
00:11:15here or there. But if you think about.
00:11:16about. The EU AI act, it does require
00:11:20require certain high risk, certain certain high risk activity has to to be declared, doesn't it? And has
00:11:24has to have human oversight. The The difference is what that what
00:11:27what is considered high risk and and what is simply somebody doing
00:11:30doing something like I say, dual
00:11:33dual use that is a genuine scientific scientific use that could be turned, turned, that could be misused in
00:11:37in the wrong hands. And do you have have a problem with this that that
00:11:39that if people are working I mean,
00:11:42mean, they test these models in America America in a sandbox environment,
00:11:45environment, presumably once you've you've got a sophisticated AI system,
00:11:49system, you could do the same in in China or Russia or North Korea
00:11:52Korea out of sight of the bigger
00:11:55bigger American companies. Yeah,
00:11:57Yeah, I think what the big American American companies are doing with
00:12:01with these most powerful models is
00:12:04is they're limiting access to them
00:12:06them to what they call trusted users.
00:12:08users. And I think that's very important,
00:12:12important, especially with the potential potential for developing advanced
00:12:16advanced biological weapons because
00:12:18because it is dual use. And it seems
00:12:21seems that in the examples that were,
00:12:23were, were published last week, the
00:12:26the scientists were deliberately deliberately trying to get around
00:12:30around the guardrails that that are
00:12:33are in place by making their research
00:12:36research appear to be peaceful. Let's
00:12:39Let's talk more then about the control control problem. I mean, we know
00:12:42know these systems can act on their their own. The harder question is
00:12:45is whether anyone, including the the people who built them, can see
00:12:48see what's going on inside the brain.
00:12:51brain. And there were certain things things that Dario Amodei said. I
00:12:53I mean, he was pretty candid in his his letter, Max, about that. He says
00:12:57says they only understand a fraction fraction of what goes on inside these
00:13:01these models that they're building. building. So with that in mind, what
00:13:04what would a speed limit look like like and for what purpose? Yeah,
00:13:08Yeah, I think that's the really tricky tricky thing. That was not defined
00:13:11defined particularly well in the the letter. And I think that the the only sort of example that we
00:13:15we had was at the point where, let's let's say a model becomes capable
00:13:18capable of X, then we need to be be doing y. And x was an example
00:13:21example of something able to break break out of sandboxes, which is
00:13:23is what we used to contain these these agents from a safety perspective.
00:13:26perspective. And Y was a pretty airy
00:13:29airy answer on making sure they don't don't have the propensity to go and
00:13:33and target large-scale sections of of the internet and compromise computers. computers. I mean, that's one of
00:13:37of the only problems we can really really nail down today. And we've we've already seen the tricky nature
00:13:40nature of trying to put a speed limit limit in around here and defining defining what those look like. Well,
00:13:43Well, I'm struggling. To understand understand what the speed limit is
00:13:45is for. Is it for compute? Is it
00:13:49it leaps in capability? Is it the the time between releases? What it's
00:13:52it's I mean, how do you what are are you trying to measure? Yeah. Yeah. It's very it's incredibly broad.
00:13:56broad. And the language that's being being used is, is incredibly broad
00:14:00broad because I don't think these these companies want to be pinned
00:14:03pinned down to definitions to say, say, OK, well, we're going to slow
00:14:05slow down on this part of our development, development, but we're not going going to slow down on this part of
00:14:09of our development as well. And I I think that's why that language language is being used and why it
00:14:13it is being kept as broad as it is. is. And that's the problem, Andy,
00:14:16Andy, isn't it, for the for American American lawmakers is what are you
00:14:19you legislating on and for? Well,
00:14:22Well, there's already draft legislation
00:14:24legislation to require these companies companies to have what they're calling
00:14:28calling a kill switch. So in other other words, if they start to lose
00:14:31lose control, they can turn the models
00:14:33models off. And that is currently currently not required. There's almost,
00:14:37almost, you know, zero regulation regulation at the moment. So anything
00:14:40anything and I do believe Congress
00:14:42Congress is inclined to establish
00:14:45establish some regulatory oversight
00:14:46oversight of this industry. And hopefully hopefully before a terrible crisis
00:14:50crisis happens. Do you. Have a view view on a kill switch and whether
00:14:54whether that could work? So there's
00:14:56there's been. Some pretty drastic
00:15:00drastic discussions around this, this, whether it's, you know, switching switching off the models, whether
00:15:03whether it's dropping a nuke from from or an EMP, sorry, from space
00:15:06space to wipe out all the electronics electronics on the planet. I think
00:15:09think it's going to be very tricky tricky naturally, by the nature of of a distributed system like this,
00:15:12this, it propagates in many different different places. It's not a single single entity that you can just go
00:15:16go and wipe out. As such, it could could exist in numerous different
00:15:20different databases across the planet. planet. And indeed, there's no guarantee
00:15:24guarantee that switching off even even the internet would, would, would would ultimately allow you to do
00:15:27do this. And I think what also we we need to discuss here is not the
00:15:30the idea of just this one big kill kill switch, which is right at the
00:15:33the point where we think the freight freight train has left the station, station, humanity is doomed and we're
00:15:36we're going to turn it off. But actually actually the Hugging Face incident
00:15:39incident that was four days worth worth of offensive cyber security security capabilities happening under
00:15:43under the noses of two giant companies companies without the ability to
00:15:46to switch it off or even the knowledge knowledge to switch it off. And I I think that's the. Awareness that
00:15:50that they weren't aware that it had had happened until after the fact, fact, quite a considerable time after
00:15:53after the fact. But it. Sounds from from the way he describes it, as
00:15:56as though they were killing agents agents one after the other and they
00:15:59they kept popping up or they were were agents they didn't know of or. or. Handing off some of those agents
00:16:03agents were handing off some of that that work. On to on onto other agents
00:16:07agents when. It comes to regulation. regulation. Andy, let's talk about
00:16:10about the three step plan that Amadeo
00:16:12Amadeo has put on the table. Embedded
00:16:15Embedded third party evaluators within within companies with employee like
00:16:19like access. So common standards standards across democracies, to
00:16:23to the extent that's possible pacing
00:16:25pacing with China and the authoritarian authoritarian states, if that's possible,
00:16:28possible, does any of that do you you think survive contact with the
00:16:32the real world? I think so, I think
00:16:36think already after the Hugging Face
00:16:38Face incident, at least two of the
00:16:40the frontier laboratories invited
00:16:43invited an independent evaluation evaluation group called metre Meta
00:16:47Meta into their. This is the non-profit
00:16:49non-profit in Berkeley, California.
00:16:51California. Yes, exactly. With expertise.
00:16:53expertise. And they were given not not a complete access. In the future
00:16:57future I'd like to see them get more
00:16:59more access. But if each of the Frontier
00:17:01Frontier Labs would agree to allow
00:17:05allow such independent audits, if if you will, I think that will go
00:17:08go a long way to understanding the
00:17:13the threat and preventing it before before it gets out of control. And
00:17:16And even though Meta was given sort sort of limited access, didn't they
00:17:19they come up with some, you know, know, quite, quite chilling results?
00:17:22results? You know, they weren't shown shown all of the work. They weren't weren't shown all of the logs. And
00:17:26And yet with the small amount that that they were shown and the small small amount of time that they had
00:17:30had access to stuff, the results results that they came back with
00:17:32with were quite were really quite quite worrying, weren't they? Yes.
00:17:36Yes. I think that's why we're seeing
00:17:37seeing so much urgency at the moment.
00:17:41moment. The mythos moment earlier
00:17:43earlier this year and then the Hugging
00:17:46Hugging Face incident of these agents
00:17:48agents teaming up and breaking out
00:17:51out of the sandbox and not finding finding out about it until there
00:17:55there was a significant lag time.
00:17:57time. This has really set off alarm
00:17:59alarm bells inside the Frontier Laboratories. Laboratories. And for the first time
00:18:02time they're talking about co-operating co-operating together to get a handle
00:18:06handle on these risks. Max, there's there's a potential bright light
00:18:10light here. Common standards with
00:18:14with a company like me to going into
00:18:17into these frontier models looking
00:18:19looking at what is being developed developed desks, badges, access to
00:18:23to the training pipeline, would they they really open their frontier models
00:18:26models to that kind of scrutiny, scrutiny, do you think? Well, I really
00:18:29really hope so because that's what what that's what's required for this. this. I mean, if you think about
00:18:33about what we're trying to solve solve with this particular element element of this plan, it's very much
00:18:36much the idea that we have independent independent viewers over the development development of this technology that
00:18:40that can raise a flag if it gets gets out of hand. And if you're not
00:18:43not being shown the full scope of of work there, then you're not going
00:18:46going to be able to do that job. job. Now, I used to be one of those
00:18:49those evaluators, not for AI models models for security within organisations,
00:18:52organisations, and you generally generally start with those conversations conversations on day one with the
00:18:56the client tells you what your scope scope is, what you're allowed to to look at and what you're not. And
00:18:59And if we have those sorts of barriers barriers to entry to these evaluators, evaluators, which I think are a brilliant
00:19:03brilliant idea, I think this is probably probably the only way we're going going to have any form of insight
00:19:07insight across the whole of the AI AI space. If these frontier labs
00:19:11labs open up their doors. But we we really need to make sure that that those people have the ability
00:19:14ability to go as deep as they want want to look at whatever they want.
00:19:17want. And crucially, if they then then do raise the fact that there
00:19:20there is something wrong and they they are allowed to put the brakes brakes on, it doesn't just go fall
00:19:24fall on deaf ears. If we get to that that point where we're worried about. about. But the big problem there
00:19:27there is, is that the labs aren't aren't letting people in. Even the
00:19:30the AI Security Institute, the UK UK AI Security Institute hasn't been
00:19:34been permitted. Anthropic didn't didn't permit them access to its its latest model. That was a really,
00:19:38really, I think, tricky. I mean, mean, if you could think of. Similarly, Similarly, the embedded evaluators.
00:19:42evaluators. We have one of the best best AI safety testing institutions
00:19:45institutions in the world here in in the UK. We've done some amazing
00:19:48amazing work in testing these models models with around Fable when that that was when that was going live.
00:19:52live. And the fact that this most most recent model wasn't shared was
00:19:54was I feel a step back and a really really scary precedent to be setting
00:19:57setting at the time when AI safety safety is on everyone's agenda and
00:20:00and potentially, you know, keeping keeping it more behind closed doors doors rather than sharing it more
00:20:04more with the global community. I I might pick. Up I might pick up up on that point in a second, but
00:20:08but I just do just want to focus focus Andy for a second on China China because we've not really talked
00:20:12talked about China. And let's face face it, they're the elephant in
00:20:15in the room, the top spy chief in in China has warned that leaders
00:20:19leaders has warned his leaders that that the evolving technology could
00:20:22could threaten Communist Party rule. rule. Although the Foreign Ministry
00:20:25Ministry did say yesterday in response response to what Donald Trump talked
00:20:27talked about. And I guess the way
00:20:29way that Dario Amodei had said it
00:20:33it out, that America isn't stepping stepping off the competition and
00:20:37and it doesn't seem to me as if there's
00:20:38there's a lot of trust in China that
00:20:41that they will. Well, there are two
00:20:43two issues at play. One is this alarm,
00:20:45alarm, you know, alarm bell from
00:20:47from the Ministry of State Security,
00:20:51Security, China's counterintelligence
00:20:54counterintelligence and law enforcement
00:20:56enforcement arm worried about the
00:20:58the Communist Party losing control control because of this technology
00:21:02technology that gives China its own
00:21:07own incentive to regulate and have have better control over these models.
00:21:11models. Do we know how good their
00:21:12their regulators are? Andy? They
00:21:16They do. I mean, we don't have a
00:21:18a good transparency into their regulatory
00:21:23regulatory efforts, but we do know know that they have increased in
00:21:26in recent years. The amount of regulation
00:21:29regulation on, on these companies. companies. So they also have an incentive
00:21:33incentive not to allow terrorists terrorists to develop chemical or
00:21:36or biological weapons. So on an issue
00:21:38issue like that, I think there is
00:21:41is opportunity for President Trump
00:21:43Trump and President Xi Jinping to
00:21:46to come to an agreement that they
00:21:48they will have put in place guardrails
00:21:50guardrails to prevent the extreme
00:21:54extreme misuse of these capabilities. We'll see whether that that comes about at the meeting that
00:21:57that they're planning for the end end of the month. But you got a view view on that? Yeah, I was just going
00:22:01going to say for the Chinese regulation regulation aspect, there's one really
00:22:05really interesting sort of idea that that we don't tend to think about
00:22:07about too much here is by being a
00:22:09a state controlled nation. What that that means is that the ability to
00:22:13to put the brakes on is something something inherently possible. And
00:22:16And indeed, there's already a great
00:22:18great deal of the sort of regulatory regulatory standards. Just over the
00:22:21the weekend they launched a third third version bearing in mind we're
00:22:24we're scrapping together to try and and build our first version of any any sort of framework around this.
00:22:28this. But there's a third version version which addresses many of these these sort of existential risks and
00:22:31and the development of these models. models. So unlike nuclear weapons,
00:22:34weapons, there are a lot of benefits benefits to AI and a lot of people
00:22:37people get reap the rewards of this this technology today. And what we're
00:22:39we're talking about is something something existential, abstract down
00:22:43down the line that we're concerned concerned about. And to correlate correlate that with what they're
00:22:47they're using at ChatGPT every day day to do their sort of, you know,
00:22:50know, day to day AI tasks. I think think there's this sort of mental
00:22:54mental distance between those two. two. And so I don't think enough
00:22:56enough of the public yet know the
00:22:59the sort of the trajectory that we're we're on and that really where it it feels to me is the trajectory.
00:23:03trajectory. It doesn't feel like like something that's today. And And I think a lot of the Frontier Frontier Labs have said this themselves.
00:23:07themselves. This is not a risk today. today. The models today are going going to be any of these things.
00:23:11things. It's simply the pace of change. change. And I think it's unless you're
00:23:13you're you know, I think collectively collectively we spend a lot of time
00:23:16time staying up to date on AI and and it still feels like there's always always more things you could be doing.
00:23:20doing. The general public probably probably aren't as familiar with
00:23:22with all the different developments developments and progression that's
00:23:25that's happened in the last six months. We started with a question. question. I posited the question
00:23:29question who slows first when nobody nobody trusts anyone to do the same? same? So can we answer the question
00:23:33question in, in mind of everything everything that we've discussed in
00:23:35in the last half an hour? Who slows
00:23:38slows down first? Mark. I think it's
00:23:41it's more about who appears to slow slow down and why they're slowing
00:23:45slowing down. Is this a follow the
00:23:47the money situation? Is this a slowdown?
00:23:50slowdown? Because those IPOs with with $1 trillion valuation aren't
00:23:53aren't looking realistic in the current current landscape, in the current current political landscape, in the
00:23:57the current sort of the narrative narrative that's swirling around
00:23:59around safety issues around AI at at the moment, is that why we're
00:24:03we're seeing a little bit of a slowdown, slowdown, or is it because the capabilities
00:24:07capabilities of the technology, they're they're not quite there yet, they're they're not quite ready for prime
00:24:10prime time. We're being sold technologies technologies that apparently can can do absolutely everything. If
00:24:13If you've got straight trousers, trousers, they'll give you flares.
00:24:17flares. But the reality when the the rubber hits the road, the tech tech isn't quite ready for prime
00:24:20prime time. It isn't doing precisely precisely what you might expect it
00:24:23it to do. So as a consequence, if if you want some kind of return on
00:24:26on the vast investment that that that people have made, it's going
00:24:29going to be quite some time before before they might see that return.
00:24:32return. So that's the sort of the the question that I have. Andy, can
00:24:36can we slow can we. I think we're
00:24:38we're I think we're on a path to
00:24:41to slowing down and improving our
00:24:44our capabilities. On the safety side.
00:24:46side. I think that's it's already
00:24:49already happening. And I think that
00:24:51that trend will continue. We'll see
00:24:54see what comes out of the Trump-Xi Trump-Xi meeting. At least they're
00:24:58they're talking about it. And the the bluster from Trump. It's typical
00:25:01typical you know vintage Trump where where before meeting he yeah, he
00:25:05he he you know calls it a hoax. But But then when he sits down with XI,
00:25:09XI, he sort of melts and, you know, know, maybe they can agree on, on
00:25:13on some type of co-operation in this
00:25:15this area that would be in the interests
00:25:18interests of both countries. I think think I'm going to juxtapose Mark
00:25:21Mark and keep my optimist hat on
00:25:22on and say that this really is being
00:25:25being fuelled by genuine concerns
00:25:28concerns that we've seen the start
00:25:29start of what we need to see externally
00:25:33externally in embeddings within these these organisations that are there
00:25:35there to raise a flag. And I think
00:25:38think that we'll probably likely likely see continued leadership from from Anthropic who have generally
00:25:42generally been the more safety aligned aligned lab. And I hope that there
00:25:46there will then follow with the other other frontier labs, and I hope that
00:25:49that will then set the tone to follow follow on a global scale as well.
00:25:52well. Well, that's it for this week. week. I'm sure. Like me, our viewers
00:25:54viewers will have plenty of questions
00:25:58questions that have been answered. answered. Maybe not answered. If If you want to send in other thoughts
00:26:02thoughts that you have or feedback feedback on what we've been discussing,
00:26:05discussing, the email you will know know it by now is AI decoded@bbc.co.uk.
00:26:09decoded@bbc.co.uk. And if you are are watching on television the QR
00:26:11QR code that is currently on screen screen takes you straight to the
00:26:15the YouTube playlist and all the the back-up episodes are there. Do
00:26:17Do take a look at that. My thanks
00:26:19thanks to Andy, to Max and to Mark. Mark. Thank you for watching. We'll
00:26:23We'll see you next time.
00:28:38without a driver. Join me Martin
00:28:39Martin Sharkey of Tech Now on BBC
00:28:44BBC News. There is nothing like driving
00:28:50driving into a moving story and not not knowing what is going to happen
00:28:54happen when we enter Damascus. It It was full of rebels shooting in
00:28:56in the air in celebration. You can can hear the sound of celebratory
00:29:00celebratory gunfire and I live for for those moments. These troops are are heading into the battle zone
00:29:04zone in Khartoum, says the army to to carry out a special operation
00:29:07operation entering Khartoum was a
00:29:09a seminal moment. The level of destruction destruction was stunning. Sometimes
00:29:12Sometimes meeting someone on the the ground can change how you understand
00:29:15understand a story. Covering these
00:29:20these stories over the years has has impressed upon me how resilient
00:29:23resilient people are. And yes, many many of them are desperate, but many
00:29:26many of them are also steadfast. steadfast. They're gracious and they're
00:29:30they're generous. Journalism is about about telling people not only what
00:29:33what happened, but why it matters matters and what it means. History
00:29:37History happens in real time and and someone needs to be there to
00:29:41to witness it.
00:30:13Live from Washington. This is
00:30:15BBC News. Ed Sheeran calls the situation
00:30:19situation in Gaza unjustifiable at at a concert in Philadelphia. In
00:30:21In his first remarks since his support
00:30:23support act was dropped over pro-Palestinian
00:30:26pro-Palestinian comments. The Houthis Houthis in Yemen claim responsibility
00:30:29responsibility for missile and drone
00:30:31drone attacks on the Saudi capital capital and Nato welcomes a deal
00:30:35deal between the US and Denmark, Denmark, which President Trump claims
00:30:39claims gives the US security control
00:30:41control over Greenland forever. I'm
00:30:52I'm Helena Humphrey good to have have you with us in Ed Sheeran's
00:30:56Sheeran's first concert since his his opening act, Macklemore was dropped, dropped, Ed Sheeran
Data courtesy of The GDELT Project (gdeltproject.org), from the Internet Archive TV News Archive. Film strip and transcript are GDELT's, rehosted here under their terms of use, which permit it with this citation.