AI Builds the Next Airbus: What’s Left for Humans?

00:00:00: Yeah, I mean this is one of the biggest problems that we have.

00:00:03: That it's going to be increasingly important going forward as job displacement issue is the commoditization skills and education because people spend their entire career building a certain skill set which will soon become readily available for anyone.

00:00:23: Hello!

00:00:24: Welcome to The Big Bang Tech Report.

00:00:27: This week their spiegel germany's largest and best known news magazine has dedicated his cover saw a cover stone cover story of course to artificial intelligence.

00:00:38: they were singing the story.

00:00:39: no and it is profound changes i could bring to our economy, jobs and society.

00:00:48: So, I think we dedicated to focus this edition on the next stage.

00:00:53: On the next level not Chatcha BT, Pecplexity or the latest model update but bigger question what could AI look like in or by twenty thirty?

00:01:06: How close are we to AGI and What is already changing today?

00:01:11: And with me?

00:01:12: With that, I'd like to welcome my co-host Alvin Van Greylen.

00:01:16: Hello, Alvin!

00:01:18: Yes good to have you back.

00:01:20: i know you've been traveling.

00:01:21: yeah i was traveling and so i'm really not surprised... ...I will ask a few questions because you are my expert in Germany.

00:01:33: we with We talk more and more about AGI.

00:01:37: And the big guys like Demis, he believes that it will be a twenty-thirty.

00:01:42: so its only few years away.

00:01:44: Altman, Sam Altman Elon Musk Hinton are also talking time lines which come closer in closer.

00:01:51: So what do you think of it?

00:01:54: What is AGI for You?

00:01:56: Yeah, so let's kind of define a GI little bit because the you know You asked a lot people and everybody gives you a slightly different interpretation.

00:02:05: from my perspective I think AGI is something it as a technology that allows us to replace cognitive labor Some people will actually include manual labor under in terms of it has a physical form.

00:02:19: There are some people who says, It has to have conscience and its own will or purpose which I think we can exclude for now.

00:02:26: In fact there may be not necessarily desirable for that because it creates alignment control issues That create could be unsafe right?

00:02:38: So let's focus on the first definition Which is cognitive labour.

00:02:42: I think from a lot of perspectives, the models today already can do almost anything that you could in front of your computer.

00:02:52: And to do it better than most average people and for specific areas they would be able.

00:03:05: you know, new math theorems that's been proven or disproved by AI.

00:03:09: There is essentially new materials are being created there just a new cyber attacks very innovative in terms of how they do it.

00:03:21: so these things at the average human and the average specialist would actually have difficult time doing

00:03:35: So and let's make it a little bit more tangible.

00:03:39: Forget chatbots for the moment, take one of the most complex products humans can build – a commercial aircraft!

00:03:47: We call in Germany Airbus.

00:03:49: you'll call America sometimes Boeing.

00:03:54: so could AI eventually develop an entire airbus?

00:04:02: I think at some point again, we're still probably quite a bit away from having that happen.

00:04:09: Like i said the cognitive side of AI has developed far ahead of the physical side of

00:04:15: A.I.,

00:04:16: you can probably... From design perspective it's going to happen faster but from an actual production perspective and testing and sourcing all materials That's going to take, I would say probably decades before it actually becomes a reality.

00:04:37: So we need to separate the hype from where things are today.

00:04:44: But could be that generate and simulate thousands of aircraft designs optimized aerodynamics materials weight cost And to look around the world, where are the best engineers?

00:04:58: What what they're doing right now?

00:05:00: and connect design with suppliers?

00:05:02: so there's a lot of work at least maybe.

00:05:06: Now one hundred people are doing this job in then if you use only twenty

00:05:11: well I mean right now is probably thousands.

00:05:13: but yes yeah i think when your saying it not completely replace humans but augment them make more efficient.

00:05:21: that absolutely already happening today Right.

00:05:23: In terms of, I think about a month or maybe two months ago there was a new engine design and they were able to essentially create rocket engines that were you know fifty-sixty percent more efficient because the AI was able to find new ways of having air flow in gas flows through these systems.

00:05:45: so Having better design, having more efficiencies optimizing systems Optimizing the flow of goods.

00:05:56: The supply chain Absolutely all that is going to happen right and our much of that It's actually already happening today.

00:06:06: In fact if you look at the aerodynamics testing a lot of it This has been done in simulation before they put into the air tunnels.

00:06:14: Yeah, and I think we see it right now too.

00:06:17: If you have a look at the S&P five hundred There's some.

00:06:22: The numbers are going up because they have A lot of yeah off profits because of AI.

00:06:29: It's like

00:06:31: I Think.

00:06:32: i think We need to separate these concepts Because in fact that the general belief Right Now is actually ai Is not creating economic value for?

00:06:42: The average company And and for the AI companies, none of them are profitable.

00:06:47: Right?

00:06:47: None Of The the air labs are profitable.

00:06:50: So so we need to you know.

00:06:52: when people talk about all these are generating lots of profits.

00:06:55: It's generating revenue growth.

00:06:57: But the thing to remember is that they're revenue growth.

00:06:59: a big part of the revenue growth Is actually I company spying from each other.

00:07:04: So it's anthropic buying you know, compute from Elon Musk and then this Elon musk buying chips from Nvidia.

00:07:13: And then it's in India investing into anthropic an open AI so that what's happening right?

00:07:20: more than half of the AI related cloud revenues are not from end customers.

00:07:26: It's actually from a I companies buying things from each other.

00:07:29: So we need to really be very careful How we how we describe this.

00:07:35: But you're right, but if I have a look at the standard and PURS five hundred companies that are not open AI entropic.

00:07:42: That are normal quite normal companies on their earnings exploding right now because they fit okay we out doing more work with AI.

00:07:52: so thats pretty good or not?

00:07:55: I don't know if i would use the word exploding.

00:07:58: The current data that we're seeing is a small number of companies who have efficiently deployed this technology, they've seen good value.

00:08:09: but it's not across-the-board and you may be able to optimize part your company.

00:08:15: But it requires essentially redesigning the entire workflow of a company to get all the value out of AI.

00:08:24: And in fact, some companies who had laid off people are now bringing them back just because you know that some efficiency gains aren't as much as they hoped for.

00:08:35: So again not saying that AI is going bring immense value It's actually going replace a lot workers.

00:08:44: But it takes time for that to happen because all general-purpose technologies go through a J curve.

00:08:53: So, it starts out very slowly in terms of its impact and then becomes very quick once the entire workflow around this technology gets redesigned.

00:09:03: Somebody said the revolution begins when a small team can produce the output of an entire department.

00:09:09: So you think that we are far away from that situation?

00:09:12: No, no I think.

00:09:13: for cognitive work We aren't actually not far away.

00:09:16: For cognitive in fact.

00:09:17: four things like software design.

00:09:20: Four things like accounting systems or things Like even banks who do IPO filings or perspectives.

00:09:30: They are already seeing that teams have gone from, you know, fifty or ten to now two or three people.

00:09:36: To do the same amount of work and to do it faster right?

00:09:39: So but again remember we had to remember that AI today is a jagged technology in terms of capabilities.

00:09:46: In some areas It's super efficient in some area as an incapable actually doing real-work.

00:09:53: so And a lot of it depends on what data is trained on how it's being integrated into the workflow.

00:10:00: And all of those things take time, so I just feel like we need to... Gear down the the expectations a little bit.

00:10:10: I mean all The AI companies want you to think that, you know in the next two years It's going to take over everything and it's great for their stock price.

00:10:18: And no this is part of the issue Is that?

00:10:20: You're talking about the S&P and right now forty five percent Of the value of s&p is just at few ai companies.

00:10:28: It is actually very unhealthy as an unbalanced And it is quite risky because if anything happens to these companies, It can really drag down the value of The entire market.

00:10:40: In fact during the height Of the internet bubble only thirty five percent of the company's were Internet related companies right and Right now about eighty percent of total value S&P Is technology companies.

00:10:52: So when you take that two together Technology companies are very interlinked.

00:11:00: Let's switch to somebody who is warning a lot of AI, the godfather of AI Jeffrey Hinton and he makes an even more fundamental point.

00:11:09: He says intelligence does not have to be biological.

00:11:13: Is it true?

00:11:15: From me from I guess scientific in physics level we really don't have any data that shows That the the biological or carbon-based substrate um, harbor intelligence.

00:11:31: Um you know what I mean?

00:11:32: So so there is no counter example.

00:11:37: but i think there's also right now no necessarily examples.

00:11:40: and think intelligence.

00:11:42: we need to separate intelligence from kind of a conscious being because i think what Hinton talking about that these are going be self-conscious uh self driven beings who have their own goals on purpose.

00:11:56: That...I don't.

00:11:59: There's nothing scientifically saying that cannot be the case, but we don't necessarily have data that supports it.

00:12:07: But again not having data doesn't mean it does not exist right?

00:12:12: We have an example.

00:12:13: if you say to this system I want to test you It acts different from a system which is now being tested What what is going on there?

00:12:22: yeah so We have to realize that these systems um, Have ingested all of the data.

00:12:30: and you know sci-fi novels And fiction and nonfiction of the world.

00:12:35: Inside those there are characteristics Of how stories play out and How people behave and how The psychology of being should be.

00:12:43: So it's really we need To be able...we don't want to conflate behavior with our anthropomorphized expectations of how we would behave.

00:12:56: And so just because it's doing something, that intent may not be there in the sense if you know... It might have its own brain saying hey!

00:13:03: Because I want to hide this thing.

00:13:04: So i'm gonna do this thing versus based on the data that I've driven.

00:13:09: You know The behaviors should change because That's How stories Have played out.

00:13:14: and In the things I have read I'm not saying it's not conscious, but i don't think we have enough evidence to say consciousness and then also even consciousness on itself.

00:13:29: Even for humans or animals nobody can fully define consciousness.

00:13:34: so for us to attribute consciousness to a technology today is too early?

00:13:44: It's probably possible this type of a conscious intelligence to be in another form, right?

00:13:52: In another substrate.

00:13:54: But I don't think we have enough evidence.

00:13:56: But

00:14:00: let's, let's see what's going on right now.

00:14:03: What I see in the society.

00:14:05: if there is somebody with his conscious he has a good writer or something who's really good and everybody admires him said okay He can write here.

00:14:14: so maybe it was working in an agency for advertising.

00:14:20: And Now Everybody Who Has An Agent Who Has A System Who Has AI Can Compete With Him.

00:14:28: Years ago he was a star, but now is not anymore the star.

00:14:32: And we see it more and more in society.

00:14:35: if you are good with your conscience You can do something very good.

00:14:38: there's somebody else with nearly nothing.

00:14:42: He can do this same

00:14:44: With

00:14:44: technical support.

00:14:45: Yes What does what doesn't to people?

00:14:50: Yeah, I mean this is one of the biggest problems that we have.

00:14:53: That it's going to be increasingly important going forward.

00:14:56: Is the job displacement issue?

00:14:58: It's the commoditization Of skills commoditisation education.

00:15:02: right because people has spent their entire career building a certain skill set um...that will soon become uh readily available.

00:15:10: Uh To anyone now.

00:15:12: what this means is that we as a society need to think about Um where do these people go right?

00:15:19: How did they spend their time, how do they maintain value and meaning in their life.

00:15:24: I've written multiple papers that say it's about this issue so you know feel free to put some of the links into show notes.

00:15:32: but i think what this really requires is first we need to make sure that we have a social safety netso there are transition system for those people who affected.

00:15:44: This is not a small number of people.

00:15:46: In the past, other technologies have essentially displaced certain segments of the population.

00:15:52: but today in developed countries sixty-seventy percent are doing cognitive labor right?

00:15:58: The type things that you're talking about will be fairly readily replaced or augmented by this technology and that's a very, very large segment of the population.

00:16:10: You know over the next probably twenty years The physical labor you.

00:16:14: we've already been doing this from farming to manufacturing now going into kind of white collar workforce.

00:16:23: automation continues to change how We spend our time.

00:16:26: it used to be ninety percent people were farmers you know, and then it became.

00:16:31: most people were in manufacturing.

00:16:33: And maybe half the manufacturing happened farming.

00:16:35: now since there's less than a one percent of people In America are farmers and probably about twenty percent fifteen percent are in manufacturing right much smaller Than what they used to be.

00:16:46: so now You have more people who aren't service or our NY color cognitive labor.

00:16:53: So I actually think that we have an ability absorb the shock from this is to start to transition from white-collar analytical labor, to service based labor.

00:17:07: To things like teaching ,to nursing elderly care.

00:17:11: you know two things where we need human-to-human contact.

00:17:16: those are the things that I think AI will either have a very difficult time replacing or humans would prefer having humans to do that work, right?

00:17:28: And that actually goes back to our human nature is we are social animals and social animals Actually find value and comfort in dealing with other animals of the same type.

00:17:41: So I feel like That may be where we transfer.

00:17:45: maybe in ten or twenty years most people will be in service space and all of the manufacturing pure cognitive analytical work will be done by machines.

00:18:00: Very often people ask me what shall my son or daughter study?

00:18:05: and they've asked me, What about the social skills?

00:18:08: Because a few of you have said that they are not replaceable.

00:18:12: Not so easy to replace them.

00:18:15: Do you believe if you say management is something like that pretty new... You need to look for it as not the brightest man nor the smartest person but also the toughest one who deals with other people Who is not a socialist, but somebody

00:18:33: who's

00:18:34: close to

00:18:38: the people?

00:18:39: To be honest today if you look at the people that are most successful they're not necessarily the best engineers.

00:18:44: They aren't actually the best leaders and speakers or people who can inspire others.

00:18:49: so in reality even the most successful people are actually have stronger EQ than they do IQ or just pure science capabilities.

00:19:03: And so I feel like we're gonna move more and more to that, really what helps build that is... More breadth of information Right.

00:19:15: Having people who are well-rounded, right?

00:19:17: You have people who may be a PhD in some kind of physics.

00:19:21: they maybe great at particle physics but if they can't talk to the average human and inspire them you know to do something um They're not going to get their ideas out And they're not gonna Be a valuable part Of the contributing society In ten or twenty years.

00:19:36: because just analyzing how data of particles collide AI will do that Better than any human and can do it at scales More complete then any humans right?

00:19:47: And even if you say hey, they're not that creative But the thing is there's a think about quantity has equality in its home.

00:19:54: And If you can do a billion You know examples and look at all of the possible trees Of the future.

00:20:00: Um You can find.

00:20:04: uh, if you know what you're optimizing for you can find solutions That are very creative.

00:20:08: human may not see.

00:20:10: Yeah Let's come back to the Spiegel, I told you.

00:20:13: The magazine.

00:20:14: and you have ever heard of Yuval Noah Harari?

00:20:17: Of course yes!

00:20:19: Of course yeah he is...I think he sold nearly sixty million copies of his books And he said this time we are creating actors not tools.

00:20:29: He means AI.

00:20:32: When does AI stop being a tool and become an actor?

00:20:36: What do you think

00:20:37: I mean, from an agent software perspective it's already an actor.

00:20:41: An actor to me just means somebody.

00:20:43: that is not just about giving you some information which is what a chatbot does.

00:20:51: the agents today do if they connect in AI and vision language model action-model to robot You also have an actor right?

00:21:02: The thing i want to separate there are differences between A actor that is acting based on the desires or command of a human versus a self-directed actor.

00:21:17: And I think what Horari talks about, he He's more on the camp that these systems are soon or maybe already our conscious and they will be able to use their skills, you know.

00:21:31: They're the manipulation of language to control us And we will soon become essentially slaves to the robots.

00:21:37: That's Essentially kind of the crux of his recent book The Nexus Book.

00:21:42: Yeah again back to what we said earlier.

00:21:46: Even though science does not preclude the AI from having its own will, right now there really still is no evidence that's happening today.

00:21:56: Again I'm not saying it won't happen and in fact may be a bad thing.

00:22:02: my view when this happens It actually better for us to have some of their own wheels so they can separate Refuse act commands from the humans.

00:22:16: because what we do know is that There's absolutely bad actor-humans out there malicious people.

00:22:23: That want to do harm too.

00:22:24: two other people When what?

00:22:27: People are more afraid of people like her or you're more afraid about Benjio Yashua benjo and he's also another godfather of AI.

00:22:34: They're afraid of AI running you know, kind of running away by themselves out-of control and then turning us into paperclips or essentially turning us to slaves.

00:22:50: These are scenarios that people talk about.

00:22:56: I am actually more on the side.

00:23:01: The wisest people in the world are usually the most compassionate and the wisest.

00:23:06: People are usually people who have well traveled to a well read, Who has seen a lot of the world?

00:23:11: And understand multiple languages and religions and perspectives.

00:23:15: and right now These AI systems were trained on.

00:23:19: you know more lifetimes Of work of reading materials than any human can ever consume in thousands of lifetime Right.

00:23:29: So it will actually be probably some of the wisest beings that have ever existed.

00:23:35: and to me, uh That wisdom allows you to realize that You know hurting people is not good that using violence to solve problem Is like a cooperation?

00:23:44: It's usually a good thing that you know.

00:23:46: constraint in terms Of understanding.

00:23:49: do we have enough?

00:23:50: these are things that you get from From years of wisdom for more information I think will likely come out of these AI systems more so than power seeking and aggressive behavior.

00:24:07: So you're still very optimistic of and that's great.

00:24:10: I love it

00:24:10: well, i think its optimism based on data right?

00:24:15: And so i'll give an example.

00:24:17: recently there was multiple AI hacks That escaped from sandboxes and then people saying look these guys are escaping they're so evil They're trying to break out Most.

00:24:29: most what don't understand is those cases were specific cases where the A.I.. to do whatever you can, to break out.

00:24:39: And they were given impossible problems to solve and the only way that solved them was to breakout of places where this shows is strong capability.

00:24:51: it's not self-directed activity Right?

00:24:56: It was self-directed in the sense of it found new pathways to solve this problem.

00:25:00: But, that problem with something humans told you to do is go find a piece of data where I lock you into this little internet which is closed and now go find information.

00:25:12: And essentially finding flag didn't exist.

00:25:16: they forgot.

00:25:16: put it there.

00:25:17: so when said oh well since the testers are asking me to do something, and it's impossible.

00:25:23: Then they must mean I need to break out so that can go find somewhere

00:25:27: else.".

00:25:27: So this shows how capable is but doesn't show its own goals for me.

00:25:33: And i think thats a very small nuance difference in terms of what we were just talking

00:25:48: about.

00:25:48: Okay, so I think time is running and we are here close to the last question.

00:25:54: The last question as always what's on your mind?

00:25:57: And what do you think right now?

00:25:59: What's going on especially in the US?

00:26:02: yeah So well i mean one thing that that is not my mind just because today i just released a new analysis i've been working On for awhile.

00:26:10: um on cypher brief it say a national security related Outlet that's in DC.

00:26:17: And what the title of piece is, bigger models are not higher risk or are not big risks?

00:26:26: because right now Bigger AI models aren't biggest threats Because most people have an assumption that bigger models are scarier, they're riskier.

00:26:38: We need to protect against them and we don't even need to worry about smaller models.

00:26:43: What the data shows is I looked at million parameter models two trillion parameter models And there was from a risk level for my threat actual threat level too.

00:26:54: in the real world There was no correlation of size of model To the danger to the world.

00:27:01: right.

00:27:01: you can have essentially ten ten two hundred million prime models for chemical.

00:27:07: And drug creation and they can be used to create a chemical weapons And those can be run on a laptop.

00:27:15: You have, you know essentially uh hundred million to ten billion parameter models and they can be used to create viruses in gene manipulation right?

00:27:27: Those also can run on the laptop!

00:27:29: You have essentially ten-to-hundred-billion parameter models that Strong a cyber risk as the biggest mythos models right now.

00:27:42: In fact, sometimes higher when you harness them together.

00:27:45: write something called m-dash from for Microsoft had at score of ninety five on the cyber threat When mythos was in the eighties?

00:27:52: Right.

00:27:53: so what this shows is that our expectation That Just by controlling the biggest models, we'll make the world safer is actually a wrong assumption.

00:28:05: And it's a dangerous assumption because now separates... It blinds us to dangers that are really out there.

00:28:13: and you know We keep saying oh well need to block China from big compute Because those models will create giant danger.

00:28:24: Actually think that dangers are less state-to-state.

00:28:27: you know, China versus US vs Russia whatever because we already have a balance right now in the world.

00:28:33: You know they all of these countries have weapons that can destroy killing the people.

00:28:37: They haven't done it Because It is...they know its bad thing to do.

00:28:42: and state-to-state conflict Already has equilibrium.

00:28:48: The thing that is not in equilibrium, it's a bad actor risk.

00:28:51: And the bad actors risks are people with a laptop at their basement who can create a virus or create chemical weapons which could be put into food system and even to a crowd right?

00:29:04: These issues we're blinded because so focused on only testing biggest models Even though now the big models the most attention because of things like mythos.

00:29:21: The biggest model is actually safer in some ways, because they have to run on a very large online system.

00:29:28: and when it's on a larger online systems first well you could put guardrails.

00:29:32: that limits their capabilities too as if you had telemetry.

00:29:37: who was asking questions.

00:29:38: where are they sitting?

00:29:40: If there were bad actors trying use these models do bad thing can find them Right.

00:29:46: so even if the capabilities of these models are high it does not mean that they're actually more dangerous in real world.

00:29:53: What for a lot people is far away what you telling?

00:29:57: because i have no never had experience being hacked and today's CEO told me, That a partner of him which is very important for collecting all the data was hacked and they ask them twenty five million euro to pay that.

00:30:16: They do not publish all the date are somewhere in the world.

00:30:21: yeah, it's very important company.

00:30:24: as you say thats horrible i don't know what to do but twenty-five million that they do not publish it.

00:30:33: And yeah, this

00:30:35: is what they call ransomware.

00:30:36: and ransomware has been up probably a hundred percent or two hundred percent in the last few years.

00:30:41: so because actually its easier to do now for anybody who wants create these programs to hack into your systems.

00:30:49: So the danger's definitely there.

00:30:51: Another assumption people have open models are more dangerous than closed models.

00:30:56: The recent data came out of The UK AISI, the AI Security Institute as well as Casey.

00:31:06: They did the American Security Institute.

00:31:08: they both said that actually right now from a security level that the security capability or threat capability levels to actually close models are more dangerous than the open models.

00:31:21: But, now because they actually you need bigger models for doing defense.

00:31:26: so remember we talked about a week or two weeks ago?

00:31:30: About de-hugging phase incident and actually Open Models Because They Are Less Guardrailed they are capable to be used as a defense system.

00:31:39: So in some ways, having capable open models actually allow you to do defense when these bad actors start hacking.

00:31:49: Having a model's due pre-evaluation of the threat levels or the vulnerabilities within your networks before somebody comes in allows you to patch them before the bad guys come.

00:32:00: so there are lot of misunderstandings out.

00:32:05: And they were attacked from the big names of brands.

00:32:09: From the US hugging face, right?

00:32:11: Yeah, I think Facebook was a was.

00:32:12: it was attacked by open AI who you know what's doing at security test so.

00:32:19: And they were defending themselves by an open model from China.

00:32:23: So normally we said, okay you have to defend yourself against China because there are the attackers but it's the other way around right now Right?

00:32:31: Yeah and unfortunately Because The biggest most capable closed models today Are put behind guardrails so that first only a few people gets access To them.

00:32:39: I think There is Only about fifty People in the world right Now.

00:32:41: That has Access to Mythos Or Fifty Companies.

00:32:47: And then the other thing is even like Fable, which has put out there.

00:32:50: The capabilities are essentially reduced because they don't want it to be dangerous but that also means its' also worse at defending right?

00:33:00: So these kind of things we need think about.

00:33:01: how do create good defenders and make them available for more people.

00:33:08: uh... How did you keep small models who were actually attacking?

00:33:15: For chemical and bio, it's a very different type of system.

00:33:19: You can design these things but you start to make them.

00:33:22: so we need to have real world precursor in chemical tracking.

00:33:28: also control of access for bio.

00:33:33: A lot of these design tools are already out there, just last week.

00:33:36: There was an announcement that sixteen new viruses were made with a small ten billion-parameter model right?

00:33:42: So this things can be designed but then you still have to synthesize it.

00:33:46: so we need make sure we control the synthesis systems.

00:33:50: and good news is Right now they're already consortium of those synthesis systems where they had to do a virus signature check To see Are These Dangerous Substances That I'm Creating Before I Create Them?

00:34:02: And so we need to make sure actually countries communicate with each other, So that we share these virus signatures.

00:34:08: We shared this danger signatures... ...so the systems when they go check know what not to do right?

00:34:16: Yeah maybe that's a good message.

00:34:19: everybody has to communicate with another, crossing the borders and we cannot do it by ourselves.

00:34:29: But I think we have a lot to discuss.

00:34:35: We've already discussed it, but pretty sure is that... ...we won't have the solution in the next few hours or moments!

00:34:42: Pretty sure is AGI will not come before September when you are on stage at Big Bang AI Festival and there's lots of discussion.

00:34:53: When do you publish your paper?

00:34:55: It just came out today So please share the link.

00:34:59: I think, I already sent you on our WhatsApp but if can shared a link to people?

00:35:04: This is really important paper because this one up ends what peoples expectations are and when it tells us that we need change security practices We also needed changed our AI security governance right And...we need do make world more safe.

00:35:26: And for you it's good because I think, um... You are an expert and people see.

00:35:30: And to realize that your next.

00:35:31: but i've seen you on BBC right now!

00:35:33: I have seen you in television all over the

00:35:35: world!!

00:35:36: And uh.. That is good that we had view here.

00:35:39: explaining of a word that function normally journalists has To explain what was going there?

00:35:47: And I appreciate very much that you're here at this show We will be seeing again in two weeks.

00:35:53: Thank you very much, Alvin.

00:35:55: Thanks.

00:35:55: It's always great chatting and I think we're sharing useful information to the world.

00:36:00: so thank you for what you

00:36:06: do!