Latest / The Road to Accountable AI / Alexandru Voica: Responsible AI Video

Transcript

Alexandru Voica: Responsible AI Video · The Road to Accountable AI

• 0:00
  Alex, thank you so much for joining me on the road to accountable AI.
• 0:05
  Thank you for having me.
• 0:07
  First of all, why don't you tell us a little bit about Synthesia um and the, the AI video Avidar market that you're involved
• 0:17
  Yes, uh we're the world's largest AI video platform for the enterprise.
• 0:22
  uh And what that means in practice is that uh tens of thousands of businesses log into our platform every day and make content with AI.
• 0:32
  And that type of content, the best way to think about it is uh rather than making sort of film or entertainment, businesses are making informational content.
• 0:43
  So they're making
• 0:44
  employee training videos, customer support videos, sales enablement or product marketing.
• 0:49
  It's that type of content.
• 0:50
  And essentially that means that most of the time they want to share some information or some knowledge with their stakeholders, whether that's employees or customers.
• 1:01
  um And yeah, they do that in our platform.
• 1:05
  It's a web-based platform.
• 1:07
  You log in and it's very similar to PowerPoint.
• 1:11
  So just how in PowerPoint you're going to put together a
• 1:14
  slides into a presentation, the interface has uh an editor and you kind put together visuals.
• 1:23
  You can drop in uh these AI avatars, which act as kind of digital presenters.
• 1:28
  You can write in a script and then click generate and a few minutes later you get a video.
• 1:35
  So an obvious uh concern with a tool like that is that people will misuse it, whether that's deep fakes, misinformation, non-consensual images or anything like that.
• 1:47
  How big a problem do you find those kinds of misuses to be?
• 1:52
  So one of the things that's probably important to mention is that from the moment the company was started, which was eight years ago, the founders had a very clear
• 2:07
  understanding of not just the possibilities of this technology, but also the risks.
• 2:13
  And when the company was started, we adopted this, we developed this framework and then kind of adopted internally.
• 2:22
  called the three C's, and that's consent, control, and collaboration.
• 2:26
  So on the consent front, uh the idea was to never create a digital replica of a real person without their consent.
• 2:36
  So even to this day, whenever we clone someone on our platform, so you can go into Synthesia and make a clone of yourself that looks and sounds.
• 2:45
  Just like you saw there's a version of me in Synthesia.
• 2:47
  I you know, I write a little script and it kind of delivers that script on my behalf But every time we create a new uh Clone of a real person we always get their their consent so
• 2:59
  someone has to sit in front of a camera and has to sort of give their explicit consent to us before we begin that process of cloning when it comes to control we've implemented
• 3:13
  content moderation at the point of creation.
• 3:15
  So when someone puts together a video on our platform, as I mentioned earlier, they click generate and that video then gets rendered by our platform.
• 3:27
  But before that process begins, we have a mix of automated and human-based moderation that checks that the video complies with our policies.
• 3:37
  And our policies on our website, we're very transparent about the...
• 3:41
  the type of content that you can and cannot make on our platform.
• 3:44
  And then collaboration is uh essentially what I'm doing right now.
• 3:48
  talk to government, we talk to civil society, we try to ensure that the rules that are being developed and the regulation around these technologies are fair, but also I think
• 4:01
  even more importantly, that they are practical.
• 4:04
  And we've had a very good track record of sort of identifying
• 4:09
  areas where you can actually implement practical regulation.
• 4:12
  So an example of this has been in the UK, where the UK was, I think to my knowledge, one of the first, if not the first country in the world to make non-consensual deep fake
• 4:23
  pornography illegal, um both in terms of creating it and distributing it.
• 4:29
  And that's the type of legislation when it comes to AI, it's probably the type of legislation that will actually make a difference.
• 4:36
  What does it take to make regulation like that practical?
• 4:39
  Because I assume some anti-deepfake regulation might be something that exposes a company like yours to too much liability or is not something you can actually work with.
• 4:49
  So what is it that the UK actually got right?
• 4:52
  So I think what the UK got right is a focus on uh outcomes, like mitigating uh harmful outcomes.
• 4:59
  I think we as a society, um probably in the United States as well as the UK where I am right now, we can agree on many things that we don't want this technology to do.
• 5:09
  And I think if we start at the top of the list, definitely non-consensual deep fake pornography is probably right there, especially the type of content that affects children.
• 5:18
  um So rather than,
• 5:22
  other countries or other regions have tried to approach a regulation from a point of, let's try to influence how the technology is built and hope that that uh action will lead
• 5:35
  to a reduction of harmful outcomes.
• 5:37
  I think that's been a failure.
• 5:39
  And here, an example of that is the UAI Act, where the UAI Act tries to put some guardrails around the size of the models and how you build the technology.
• 5:52
  Like I said, with the hope that downstream from that, you'll uh reduce risk.
• 5:57
  The UK has taken a different approach, which is let's try to identify those kind of harmful outcomes and let's make those things illegal.
• 6:06
  And this is an example where AI-based uh non-consensual pornography is clearly something that as a society we don't want.
• 6:15
  Maybe a few fringe elements of our society want that, but generally speaking, most people do not want that.
• 6:21
  By making that illegal, um it reduces sort of the likelihood that people will engage in this type of behavior.
• 6:28
  And where have you found the greatest challenges with the first two, the control and the consent piece to actually do that effectively?
• 6:37
  So the good news is that at Synthesia, AI governance is something that we've cared about from the beginning.
• 6:44
  So our CEO and his team knew that if we wanted to be a leader in this space, um security and privacy had to be sort of built into the DNA of the company because they knew that if
• 6:58
  you were to sell this type of technology, trust is really important because
• 7:03
  When you talk about large language models and those types of technologies, with text, there are definitely risks there and there are harms.
• 7:11
  But when it comes to video, it's such a personal technology.
• 7:15
  Being able to replicate someone's likeness is quite a personal thing.
• 7:24
  I think the fact that we understood this from the very beginning and we hired...
• 7:28
  One of the examples I always give is...
• 7:30
  We had a head of security when the company was just about 50 employees, right?
• 7:34
  And then we created this framework from the very beginning of um the company.
• 7:40
  And I think that ensured that anything that we built after that, you know, in the last sort of eight years of the company's existence, anything that we built, we always kind of
• 7:51
  considered.
• 7:52
  transparency and security and all these elements at the moment that we were building it.
• 7:57
  A lot of technology companies, what they typically do is they'll build something and then they'll think of some of these issues afterwards.
• 8:04
  What I like about us is we, at the moment that we're building the product, we are thinking proactively about some of these issues.
• 8:13
  I guess though, just to play devil's advocate, someone might say that all sounds good, but there's a cost and you're going to move slower or deployment of the videos may take longer
• 8:24
  because they have to be reviewed.
• 8:25
  How do you factor those competitive concerns in?
• 8:30
  It is always a challenge.
• 8:32
  I'm going to be honest here that, of course, have we lost some potential business along the line because we've taken some of these steps?
• 8:42
  Yes.
• 8:43
  Some of our competitors engage in non-consensual deepfakes.
• 8:48
  Some of our competitors do no content moderation.
• 8:51
  I would argue that as long as we remain very good at
• 8:58
  identifying like the positive use cases and then kind of investing energy in attracting those types of customers.
• 9:08
  I think we're not necessarily gonna be too slow to kind of grow.
• 9:15
  The other factor to think about is that we are an enterprise focused company.
• 9:20
  So we're not building sort of consumer products.
• 9:22
  And when you're selling to that space of the Fortune 500, they really need to
• 9:28
  trust you and trust your brand.
• 9:31
  So having that kind of trust, it's not just morally something that people probably enjoy to have, but from a business point of view, if you are selling in that space, no executive
• 9:48
  of a Fortune 500 company wants to think about what happens if someone steals my avatar and starts making sort of rogue content with it, right?
• 9:58
  And if you're not careful, if you're clumsy, if you don't think about security and privacy, um when it comes to the decision making, um it can affect uh the procurement
• 10:09
  process if you don't have a lot of these things um done professionally.
• 10:17
  As you talked about, this has been an important element for the company from the beginning.
• 10:21
  Are there ways that your mechanisms of preventing these misuses have evolved over time?
• 10:30
  Yeah, so I think a good example of that is when we started to sort of reach scale, it was very clear that we needed to kind of approach content moderation in slightly different
• 10:48
  way.
• 10:49
  And at the beginning, a lot of the ways we did content moderation was...
• 10:54
  because the technology wasn't there, was based on simple things like keywords or these kind of like simple rules-based systems.
• 11:03
  um But as the kind of technology improved, and especially more recently where we've had these large language models that are very good at capturing sort of nuance and being able
• 11:14
  to uh perform some of these content moderation tasks, so uh our uh approach to content moderation has improved and has kind of evolved.
• 11:24
  And we recently did, uh we participated in last year in this um cybersecurity hackathon that was organized by NIST.
• 11:34
  NIST is the US uh standards authority for when it comes to kind of technology, have a very uh good cybersecurity team and they recruited some security researchers and they asked
• 11:48
  them to test our content moderation system as well as our cloning systems.
• 11:53
  And for an entire day, they tried to break them and were unsuccessful.
• 12:02
  And they published this report uh that you can go and read about how exactly they tried to do that.
• 12:09
  And they published some of the results.
• 12:11
  And it was very encouraging to see that we were able to withstand uh that intense testing.
• 12:17
  And we continue to do that.
• 12:19
  we constantly sort of, this process called red teaming, where you're
• 12:23
  trying to break these systems, these kind of security systems.
• 12:27
  We constantly do that, but that was just an example of a public, um we did this in a public forum and the results are public for people who want to read them.
• 12:35
  think uh Wired did an article about it and they kind of published the report.
• 12:42
  So if people want to read that, it's available.
• 12:45
  There was an incident a couple years ago though where some political misinformation videos did slip through and I'm curious what you learned from that experience.
• 12:55
  Yeah, so I think what happened there was, um and I think when you say a few, there were literally a few videos where um they slipped past our content moderation systems and they
• 13:09
  ended up on social media, I think on YouTube and Twitter at the time.
• 13:14
  And I think what we learned from that uh as a practical example is that I would say that what we used to...
• 13:23
  kind of know as traditional, you know, harmful content.
• 13:27
  So when someone is sort of purposely trying to mislead you with something, that is kind of a space that I think, you know, where the threats are very much known.
• 13:40
  What was interesting about that case is that the content itself was not misleading.
• 13:46
  So it didn't necessarily contain misleading information.
• 13:50
  the way it was kind of or the reasons why it was made was to kind of drive polarization.
• 13:56
  So for example, you know, making videos about um the number of school shootings in the US, the content, if you looked at the content, there was nothing misleading there.
• 14:08
  The numbers were accurate.
• 14:10
  They were obviously quite high.
• 14:12
  um
• 14:13
  But the reason why they were making that content was to kind of drive polarization in the population and kind of get people to sort of um start uh talking about this topic and kind
• 14:26
  of get into heated conversations.
• 14:29
  So I think that was kind of an interesting example of where the content itself was not misinformation or disinformation, but the campaign was intended as such.
• 14:41
  um
• 14:43
  And I guess the other lesson that we learned from that was that, you know, when it comes to that type of content, we, we, one measure that we took was to restrict it to enterprise
• 14:57
  accounts.
• 14:57
  So rather than, rather than prevent people from making that type of content completely, if you, if you just restrict it to the kind of organizations that are kind of vetted and
• 15:11
  um We understand who they are as customers.
• 15:14
  um Then you have at least some ability to sort of explain like, this video was made by this organization and then you can sort of go and talk to them.
• 15:26
  If you disagree with what's presented in the video, please go and ask them about why they made that type of content.
• 15:33
  um So I think, yeah, it was definitely an interesting example.
• 15:38
  Let's talk more broadly now beyond content moderation.
• 15:41
  One of the reasons I wanted to talk with you was an essay that you wrote on the transformer site that was titled Audits Not Essays.
• 15:50
  So just for the people listening or watching, can you just describe what that thesis is?
• 15:57
  Yeah, I think the reason why I wrote that article was there's a lot of conversation, um obviously in the US, but mostly in Europe, about the lack of AI adoption.
• 16:12
  mean, there was this uh quite uh popular study from MIT.
• 16:18
  I personally think it was a bit misleading um that showed that 95 % of projects sort of uh don't end up in a good place.
• 16:26
  was terribly misleading.
• 16:28
  did a LinkedIn post on it.
• 16:29
  It doesn't say what people think it said.
• 16:31
  No one should use that stuff.
• 16:33
  But there were others and there's a lot of conversation about an AI bubble.
• 16:38
  And the reason why I wrote that essay was I tried to kind of provide a realistic roadmap for like, this is how you get AI adopted, especially in these kind of large enterprises.
• 16:50
  And I tried to explain in there what enterprises care about, and then how do you address those concerns?
• 16:58
  And you can sort of move faster and get businesses to actually use AI.
• 17:03
  um But because this is particularly a problem in Europe, I think the United States is very good at this.
• 17:08
  So just to give you an example, like Synthesia is used by 90 % of the Fortune 100, which is amazing.
• 17:15
  But there aren't that many companies like us.
• 17:17
  Like there are maybe OpenAI and Google and Anthropic and a few others have that kind of level of penetration.
• 17:25
  But I think we would want more companies to be able to be successful.
• 17:31
  And especially in Europe, was a survey from last year done by Eurostat, which is like the European Commission's statistics body.
• 17:40
  it's their own study that showed that only 13 % of small businesses had used AI.
• 17:50
  And that's a shockingly low number.
• 17:53
  like I said, I just wanted to sort of explain how can we get more businesses to use AI.
• 18:01
  The main thesis I had was that a lot of times uh businesses need trust.
• 18:08
  And there's a lot of people writing about trust, but uh not a lot of people sort of uh implementing kind of a trust stack, if you will.
• 18:19
  And I tried to explain in that essay, what does a, how can you prove trust?
• 18:25
  And I basically um framed it in this kind of like ISO standards.
• 18:31
  because ISO standards are internationally recognized, but they're also operationally um relatively easy to implement.
• 18:38
  So whether you're a large organization or a small organization, it's relatively straightforward.
• 18:45
  It requires some investment, of course, but it is, provided you have good internal governance, it is easy to sort of uh adopt those standards.
• 18:56
  And once you have those standards,
• 18:58
  when you're trying to sell in the enterprise space, uh procurement uh is a lot easier.
• 19:04
  It's not, obviously, it's not going to happen overnight, but you can reduce conversations around trust and security from months to hopefully weeks, essentially.
• 19:16
  And that basically gets sort of AI adopted faster.
• 19:21
  that important proviso you said there provided you have good internal governance complying with something like an ISO standard is not that hard.
• 19:29
  isn't there a challenge for the organizations that are not already at that point where they've made the investment in AI governance to look at these these whether it's know a
• 19:38
  standard or some other kind of audit framework.
• 19:42
  Yeah, I think uh that's a good point, a legitimate point.
• 19:46
  But that's why I tried to explain in that uh essay why or how you can get to good governance.
• 19:57
  And um essentially, I used the example of ISOFOR 2001.
• 20:01
  So ISOFOR 2001 is the standard for AI risk management that was built um both as sort of a
• 20:11
  to try to sort of address some of the product risks that the EUA act had defined, as well as, as I said, give companies on both sides, so the companies that are developing the
• 20:23
  technology and the companies that adopting the technology, uh some sort of common language so that uh they can sort of understand the risks and how to mitigate them.
• 20:35
  to speed up AI adoption.
• 20:37
  So it's specifically designed for AI risk management.
• 20:40
  It's kind of a standard for AI risk management.
• 20:42
  um And I tried to explain like some of the things that we did to kind of, um as we kind of went through that process.
• 20:51
  um But it is, it is, you have to invest to both invest in terms of the, people to kind of build uh those systems.
• 21:05
  as well as as a company to have a culture of uh risk management essentially.
• 21:12
  Do we have all of the standards that we need at this point?
• 21:17
  I think we have a good combination of standards to try to build something around when it comes to AI specifically to build something that could could become if you like a Golden
• 21:32
  standard so I think if you combine like ISO for 2001 with ISO 27,001 with and then with ISO 27,700 and one those kind of three standards taken together
• 21:46
  give you the combination of risk management, data privacy, and data security to build this trust stack.
• 21:55
  And then you could also add SOC 2 to that equation as well.
• 22:00
  And I think you have a very good start for companies, especially large companies who procure AI systems or services, to be able to say, we can begin a conversation because we
• 22:15
  you know, you're ISO certified.
• 22:18
  Do you find that the enterprises you're dealing with are starting to move to that point?
• 22:22
  Because this dynamic obviously works well if the large customers are insisting on it for procurement, which for various other standards and other industries is the case, but it
• 22:33
  requires them to actually want it.
• 22:36
  Yeah, think especially when it comes to these larger organizations, it has helped us speed a lot of um commercial um agreements.
• 22:48
  um And I think what I'm hopeful what will happen is that it will be, they'll have a downstream effect into kind of small and medium sized businesses.
• 23:00
  So as more and more of these large organizations adopt this technology,
• 23:05
  small businesses would feel empowered to do so because a lot of times the small businesses are very risk averse.
• 23:15
  don't have, large organizations will have teams of lawyers, they'll have large security teams.
• 23:22
  If you're a small business with maybe 10, 15 people, you're lucky if you have maybe access to legal advice, if you have access to...
• 23:33
  um you know, security advice, but you're not going to have that resource in house.
• 23:39
  if, if you see that these large organizations are using these technologies, you feel like there's a certain level of comfort that that gives you to be able to make that step.
• 23:52
  But I guess the concern would be the larger organizations that have the resources to do the standards compliance and have all that in house.
• 24:00
  They would just be in a better position even if the smaller companies might need to go down that path that they want to sell to the enterprise.
• 24:08
  So isn't there a danger that the standards and these requirements actually accentuate that competitive differentiation?
• 24:16
  um There is always that risk of a moat, if you like.
• 24:21
  um I think the way to sort of uh make it easier is what I'm kind of advocating for is this kind of um scaling security as you kind of scale in size, if you like.
• 24:37
  So there was this piece of work, there's an organization called in the US called Responsible Innovation Labs, and they did this kind of
• 24:45
  mapping from small company to large organization, how you should think about sort of responsible innovation.
• 24:54
  Like what are the things, if you're just a few people starting a small business to if you're, you know, 5,000 plus employees in size, um how can you think about sort of AI risk
• 25:10
  management?
• 25:10
  um
• 25:13
  because I had the opportunity to sort of provide some feedback into um that process, I think they provided a very good um roadmap for how could this kind of work in practice.
• 25:27
  So what I'm trying to say is that you shouldn't necessarily force ISOFOR 2001 on a small company with like I with 10 people.
• 25:36
  But what I'm saying is that once you reach a certain size and ISO for 2001 becomes a natural sort of common language between AI providers and AI adopters, because that will
• 25:52
  move faster, it will then signal to the small organizations that, hey, if this large company is using this system and or service, they're happy with it, they're comfortable
• 26:01
  with it.
• 26:02
  I can also use it because I have the confidence that the
• 26:06
  the company I'm buying it for has that kind of level of trust.
• 26:09
  um think it's unrealistic to try to impose those kinds of rules down into small business.
• 26:21
  And you suggested that the kind of comprehensive approach in the European AI Act is problematic.
• 26:28
  Are there other things that governments should be doing to encourage the adoption of these more standards-based and bottom-up kinds of governance practices?
• 26:40
  um I think it's um one of the things that um I guess confuses me when I look at something like the EOIA Act is this sort of um obsession, if you like, with controlling how
• 26:55
  technology is built with the aim of influencing how that technology is used.
• 27:02
  And I disagree with this approach because I don't think it works in the real world.
• 27:08
  So...
• 27:08
  oh
• 27:09
  If someone who has never built an AI system uh comes up with these arbitrary rules about how that um system needs to be built, not how it should operate, but how it needs to be
• 27:25
  built, um I think we're going to have this unfortunate situation where you can be fully compliant with the EOIA Act.
• 27:38
  but have a system that can do a lot of damage.
• 27:41
  um Because if you look at the AI Act, kind of um criteria that it uses is, for example, the size of the model.
• 27:49
  So saying it's like, if it goes above a certain compute threshold, em becomes, there's kind of, yeah, systemic, and you know, I've seen in the last, just in the last few months,
• 28:03
  I've seen systems that are just as capable as,
• 28:07
  you know, GPT-5, but they are 20 times smaller.
• 28:12
  uh And so they wouldn't necessarily fall under the incidence of the AI Act because they don't meet that compute threshold, but they have just the same level of capabilities of
• 28:22
  those systems.
• 28:23
  And this is what I mean when I say that rather than try to influence um how technology is used by controlling how it's built, you should try to focus on the outcomes that you want.
• 28:37
  And if you try to apply that type of thinking to adoption as well, so not just kind of risk mitigation, but also adoption, I think you can take some kind of sensible measures to
• 28:51
  accelerate that adoption.
• 28:52
  So the perfect example that I can give is if you are a government or uh if you run kind of public services, um
• 29:06
  At the moment, in the UK, just to take an example of the UK, the UK government spent 570 million pounds last year on AI products and services.
• 29:21
  And if you look at the list, because it's public, if you look at the list of where they bought those services from, I think it's like 90, when I calculated it, was like 90
• 29:32
  something, close to 100%, 90 something percent of...
• 29:35
  of all of that came from companies outside the UK.
• 29:40
  Now, I'm not saying that it should be, know, 90 % of those services should come from companies from the UK, but there's clearly something wrong here.
• 29:48
  If your goal as a UK government is to encourage adoption and development of AI technology in the UK, yet all of your procurement happens from outside the UK, it's very hard to sort
• 30:01
  of...
• 30:02
  uh
• 30:04
  you're sending the message that the UK technology is not good enough and you don't create sort of an internal market for some of these UK AI companies to um thrive because
• 30:15
  obviously at the end of the day, if they don't have revenue, they're not competitive.
• 30:20
  So I think there's this kind of balance, if you like, between the kind of Soviet approach of the state needs to fund everything and everything needs to go through the state.
• 30:33
  versus um the state completely takes the hand off the wheel and doesn't try to at least sort of have some incentives for the companies, like the kind of homegrown companies to at
• 30:47
  least have a shot at procurement.
• 30:50
  um So that's probably kind of, I don't know, I don't have a uh great answer to where that space should be, but it's definitely not at the extremes.
• 30:59
  Like I said, I'm not a fan of this kind of like,
• 31:02
  state controlled, know, the state controls all the means of production, but I'm also not a fan of like, let's just kind of uh sort of pay no attention where we're kind of
• 31:15
  distributing our, as a government, we're distributing our money basically.
• 31:20
  All right, one more thing before we wrap up.
• 31:23
  As you mentioned, Synthesia has been in this world of AI video for a long time.
• 31:28
  Now we're getting into an environment where the fidelity of these tools is increasing, the availability for lots of different use cases is much more widespread.
• 31:37
  What do you see as some of the new kinds of challenges or opportunities people might not be thinking about that you're focusing on with this area in the future?
• 31:49
  um That's a good question.
• 31:50
  The one thing that I'm probably the most excited about is something that is to some extent personal to me.
• 31:58
  So when I was a kid, I'm from Romania, and when I was a kid and I was learning English, um this was like the 1990s.
• 32:08
  Romania was in a kind of a deep depression.
• 32:11
  The economy was really struggling.
• 32:13
  My parents were really struggling financially, but they knew that they wanted,
• 32:18
  they saw that I had the passion to kind of learn English and they kind of searched and they found someone who was in our town and knew a bit of English and they kind of arranged
• 32:30
  a bit of tutoring for me.
• 32:32
  And so I started with that and then I kind of kept going and eventually I learned English to hopefully a very good standard.
• 32:39
  uh But that, if I think about my town and sort of the kids in my town, that was not a...
• 32:47
  an opportunity that was available to um anyone else or maybe very few people, very few kids in my town had that kind of opportunity.
• 32:54
  And when I kind of fast forward to where we are today, I think we're in a place now or we're getting very close to a place where these um models are very good at um personalized
• 33:07
  education.
• 33:08
  And if you look at the, uh think it was OpenAI themselves or maybe someone else, they published this um research as
• 33:16
  how do people use ChatGPT?
• 33:19
  And I think one of the top three, if not the top three, definitely in the top five, was kind of personal tutoring, like education, essentially, like personal education.
• 33:29
  And what I'm excited about is this kind of opportunity for personalized education.
• 33:34
  And when I mean personalized education, I don't mean it just in the sense of, you know, what I described like personal tutoring.
• 33:42
  I mean it also in the sense of...
• 33:44
  that adapts to the way that you like to learn.
• 33:47
  Because the way I like to learn and the way you like to learn is probably very different.
• 33:51
  Like some people are very good, like visual learners.
• 33:54
  Some people are very good at learning through analogies.
• 33:57
  And um like I said, these models are getting to the point where they can help you learn by adapting to you and kind of your personal needs.
• 34:06
  um And if you combine that with...
• 34:09
  um the kind of technology we're building, is video, these AI avatars, kind of AI video, you can generate this kind of like on the fly learning experience that's not just
• 34:22
  personalized to you, but the great thing about it is it's always available.
• 34:26
  So going back to my example, you know, I would meet once a week with this person who was kind of tutoring me on English.
• 34:34
  We would spend about an hour together and he would give me recommendations of like books to read and things like that.
• 34:40
  And then I would have to wait for another week until he kind of was able to meet with me again.
• 34:44
  But with these systems, you know, they're always there.
• 34:47
  And I think what's also interesting is there because they're always available.
• 34:52
  It gives you the incentive to sort of keep going.
• 34:54
  So just, just to kind of show you what that means in practice is we used to have, in our internal, uh, Slack channel, we used to have this,
• 35:04
  chat bot that you could ask it information, but it was in a public channel.
• 35:10
  So everyone could see what you were asking the chat bot.
• 35:15
  And when we kind of transformed that into like an individual uh experience, like anyone could go and interact one-on-one, not in a public setting, but in a one-to-one setting, I
• 35:27
  think the usage of it increased by seven times.
• 35:32
  Because all of a sudden, people are like,
• 35:34
  Okay, uh I can ask dumb questions.
• 35:37
  No one will know.
• 35:38
  It's just me and the chat box.
• 35:39
  uh But also because it's kind of there just for me, I can use it.
• 35:44
  I don't have to worry about other people sort of using it at the same time.
• 35:47
  I can just use it for myself.
• 35:49
  And um I think that's kind of uh a great opportunity that when you combine these kinds of technologies together, we're gonna get, we're gonna see some, hopefully some interesting
• 36:00
  experiences being built.
• 36:02
  Fantastic, Alex.
• 36:03
  Thank you so much.
• 36:04
  Thank you.
• 36:05
  Thank you as well.