I saw so many failed experiments of take a speech engine connected to an LLM and hope for the best. And I always say hope is not a strategy when it comes to generative AI. Every single piece of working life is going to be accelerated by this and I like to take the view of positive augmentation. I know there's a lot of uncertainty about the the threat of AI to roles and so on, but the threat becomes a certainty if you ignore it. Think it's actually going to be one of the most important soft skills for people is asking the questions.
Christoph Magnussen:I'm sitting down with Keith Griffin. He is Cisco Fellow, one of the highest ranking technical positions at Cisco. He's at the same time VP for the collaboration at Cisco. So stuff you know from Webex for example. He's also one of the key figures writing the AI framework, the safety framework that Cisco guides in their AI products.
Christoph Magnussen:So also part of the AI strategy for Cisco for a very long time since I think ranking back to 2016. So this is what we will do today. My name is Christoph Magnussen. This is AI to the DNA and we live and breathe AI and technology with my company Blackboat every day. This podcast is made for you to also be able to bring AI to your company DNA.
Christoph Magnussen:This is why I share that. I'm very happy that you're here and we do this for the reason that I really want to learn for myself or understand the nature of this technology in a better way. I do it for many years already, but having conversations like these really really helps me to dive deeper into these questions that we need to ask ourselves and learn it. This is what we do today. So leave a comment afterwards, ask the questions that we need to ask in the next sessions, and welcome to the session with Keith Griffin.
Christoph Magnussen:Today, I'm sitting down with Keith Griffin, Cisco fellow. For the ones of you listening, if you don't know what a Cisco fellow is, can tell you Keith is one of the people influencing millions of people collaborating every day because Cisco Fellow is the highest ranking technical position you can have at Cisco. Very few people have this title and I actually targeted you right away when we came to this conference and like I gotta talk to Keith because this is one of the few persons who are able to connect actually in-depth what technology means. And you also wrote as I think since you do AI since 2016 for Cisco already, so long time. Yeah.
Christoph Magnussen:You wrote the AI responsible framework as well. So a lot to talk about. Thank you for making it possible.
Keith Griffin:Oh, sure. I really appreciate you reaching out and and for it to to happen and, yeah, I was one of the coauthors of the the framework and, actually coming up on ten years since we first started experimenting with with initially supervised machine learning, who knew it would turn into such a huge area?
Christoph Magnussen:Quite a while. And I mean, you are a VP at Cisco in the collaboration area and also leading the site in Ireland as I learned. And so you have many different roles, of course, but at the same time, there must be something with this job that excites you because you're with Cisco for quite a while. Yeah. So what is that?
Keith Griffin:It's eighteen years and that's yeah. There's actually a lot with the job. Almost all of it excites me. It's a it's a really unique role. As you mentioned, yeah, the Cisco fellow is a VP equivalent type of of role.
Keith Griffin:It's just amazing to achieve that. I never thought I would actually, and it's a real honor to be in the role. On the technical side of it, I focus on our AI technology strategy and direction. We have an amazing engineering team that that works on on that with me. I also focus on mergers, acquisitions and investments, mostly in the AI space, but generally across collaboration where we place our bets and where we may need to acquire.
Keith Griffin:As you know, Cisco is a highly acquisitive company. We have a lot of acquisition activity, including a very recent one. And then I also run our patent strategy and research strategy that kind of lines up with that type of role really well. So those are really the the key pieces on the the technology piece. And being a fellow, it also spans not just collaboration, but like you mentioned, the responsible AI framework that's across the company, AI governance also across the company and participating in in groups that have broader reach across Cisco and also the industry in some cases.
Keith Griffin:But, yeah, it's really, really fascinating role. I I really enjoy it.
Christoph Magnussen:One personal question I always have in the very beginning for everyone who comes here because spending something in that technological depth means you also try stuff that doesn't work. Is there something that comes to you in your in your mind when you think of these eighteen years in your professional career at Cisco where you say, like, this I I totally I totally miscalculated how this could turn on. Like, where you really stumbled upon something that shows that makes you human and it didn't didn't work out that well.
Keith Griffin:Yeah. There's some there's there's a few examples like that and when I look back on it though, often it's it's not that things maybe didn't work out, but we had the timing wrong. Like Cisco is very very good at identifying market adjacencies and trying to be early in certain areas. And where we find we're not maybe early, but something has happened, can often acquire and then keep pace. We think of ourselves as the world's largest startup from that point of view, and we keep going with that type of mentality.
Keith Griffin:But there were several areas like that. I mean, one area that I got wrong is the current area, AI. I mean, I had no idea it would be this big, you know, so I got it wrong from that point of view, even though it's getting it wrong while getting it right. It's it's been like just phenomenal, the amount of adoption, the rate of change and the pace of it, you can barely keep up. In fact, if you're not augmenting yourself with AI, you're not going to keep up.
Keith Griffin:That's that's been that's that's one that maybe fits to the question. There's also areas like maybe ten plus years ago, we experimented a bit and created product in the social software space for enterprises. That didn't quite work out. It didn't really take off in the enterprise as as such, even though the technology was was really amazing. And I felt we were a little early with that, although in hindsight, it just really didn't take off in the same way as it did in the consumer space.
Keith Griffin:So that's maybe one example of that, although it was a phenomenal area to work in at the time.
Christoph Magnussen:When we talk here at this time, we have a global outage of Cloudflare at the moment happening. That means services like ChatGPT, Claude, I guess Spotify, Facebook, some of these are down. And if people experience that, they experience how they already are attached to this technology. And what they sometimes get wrong is that there is a totally different approach between enterprise and the normal everyday world. If you if you hear that, it was while you were talking, so I was looking at that and I'm like, okay, there is something happening here.
Christoph Magnussen:I thought that could be exciting for for for you and me to to dive into that. When you hear that, how is what you built internally different and how did you start to get there compared to the consumer side of AI?
Keith Griffin:Yeah. And I would say in general, like, it's it's good to to know that for first of all, I I wasn't aware, but it's You were on stage. That's okay. But it's I would say it also that's like, you know, no vendor is fully immune from these types of outages. It happens and it happens frequently.
Keith Griffin:What you can try to do is make the systems as robust as they possibly can be. One of the things that I I really take away from early career actually before Cisco, I worked for Nortel, Canadian telecommunication company and that started with PBX type of design. What was fascinating about that was the level of over engineering, you know, telephone switches
Christoph Magnussen:Yeah.
Keith Griffin:Do not go out. You over engineer every single aspect of that as much as you possibly can. That stuck with me all the way through. Now that's just a personal view, but then as we look at this from an industry perspective, as we get into the enterprise environment, we have to have so much failover built in. And what complicates that currently actually is recent legislative initiatives.
Keith Griffin:So if you have regions where now the data must stay in the region, doesn't give you any more the option to fail over to another region. So you have to get to a new level of granularity in your design. It's kind of an interesting it's an interesting side effect, maybe an unintended consequence of if you have like, you know, data must reside in a certain area, you no longer have the neighboring area to fail over to. But I think in general, yeah, from an enterprise point of view, we have to make sure we have very robust services, we have good level of failover within region, which is the solution to the last problem I just brought up. And that we're also working with the best in class vendors and partners because no one organization can do all of this in the enterprise software space by themselves.
Keith Griffin:So, you know, it it really trickles all the way down. Everybody has to hold themselves accountable to each other and to themselves.
Christoph Magnussen:Do you remember the first step if you take that perspective where you are now and how you build it? But do you remember the first step that you did personally in the organization to, like, get there back in the day. We said, like, okay, what do we need to first step making an overview? I don't know. Like, where did you start?
Keith Griffin:You you mean, like Back in the really starting from, like, college from Meaning,
Christoph Magnussen:taking because AI is totally different beast if you want to tackle it in the in the IT infrastructure.
Keith Griffin:Mhmm.
Christoph Magnussen:And Cisco is around for forty years now. Mhmm. Forty years. I mean, that's a long time. And building change is hard.
Keith Griffin:Yes. Doing that
Christoph Magnussen:in your core field harder. Yeah. Was literally the first thing you did? Did you write something to top management? Did you work on it?
Christoph Magnussen:Did you research? Did you start stumble upon something? Was it?
Keith Griffin:From an AI perspective, when we looked so we had a brief basically from our CTO at the time, which was to investigate supervised machine learning to see, can it make a positive difference to the collaboration user experience? And the sub brief is, can it solve a problem that was previously unsolvable or expensive to solve? So that's what set us about looking at a supervised machine learning model for noise detection. So Right. That the bill.
Keith Griffin:Right? You could do that, but you need maybe an expensive DSP to to solve that problem. Right. Which was solvable, but in an expensive way. Would it benefit the user experience?
Keith Griffin:Yes, it would. And we've very talented data scientists, the very first data scientist Derek Chan that was hired into the team to go to look at that. And almost immediately, he had a solution, you know, he took noises, recordings of dogs barking and keyboard clicks and trained the model. And amazingly it worked, like every single time if there was a noise, it would just activate and say there's a noise, there's a noise. It was amazing.
Keith Griffin:At the same time, it was very very CPU intensive to the level that you couldn't run it. But here's where the the, you know, the magic comes in of the data science scientist. He was able to remove appropriate number of layers of the neural network to maintain the accuracy just losing one or two points of accuracy, but dropping the CPU from say 30% to less than 1%. So that was a really, really big breakthrough moment for us that gave us the belief that this is not just something we could play around with and prototype, but this is something that can really have an impact in collaboration. And it definitely met the brief that we had.
Keith Griffin:But the second thing, back a little bit to your question on enterprise AI, immediately we realized, wait a minute, we are going to have to do this in a way that the user is and the company is very certain about its data privacy. So immediately, we started thinking about how do you do this in a way that preserves data privacy and focuses on that. And the the single line I said to the team, I still say it today is, if you don't do that, if you don't focus on data privacy with AI, if you don't focus on responsible AI framework with AI, all you ever make is a good demo. But you can't really have something deployed unless you can really assure users and an organization that it's appropriate to use.
Christoph Magnussen:I I was quite impressed you showed it on stage. I didn't see it before. You showed a deepfake video of yourself and your AI agent could detect it while you were like while this video was running, which is quite helpful if you have a video conference you don't know or helpful if you see a video and you're not sure. And the second thing where you where you talked about was when you have guardrails built in when it comes to certain types of like hate speech or like stuff that's not appropriate and so on. Tell us a little bit and and tell me a little bit about how you look on that because this is not only a technical topic but you need to make certain decisions.
Christoph Magnussen:I think I I showed on stage certain system prompts and that words really make up this stuff. So there are like people coming together now with technicians talking about how to handle that technology.
Keith Griffin:Yeah. Yeah. There are and it's it's a fascinating area. You know, the deepfake one in particular, in fact, the demo you mentioned, I wasn't planning to show that until I saw your deepfake virtual agent demo and I added it back in because I thought this is a really good way to bring it to life, you know, what the threat is and also what some of the solutions are. And what we've been doing there is looking all and, you know, internally from an engineering point of view of what can we build to detect that.
Keith Griffin:But also there's lots of emerging startups and really solid companies in in that area. So we made an investment in a company called GetReal out of Berkeley. We also partnered with Pindrop. The one I showed on stage was from Karnā as an AI startup. And there's there's many more, Reality Defender.
Keith Griffin:There's there's lots of lots of organizations doing really amazing work in detecting this. And it's almost going to be like an arms race, almost like the antivirus space. Like as soon as you, you know, can detect something, somebody's gonna find out a way to get around the detection and then you have to detect more. And I actually think that the way we have to think about this is maybe not from the point of view of just detecting deepfakes, but ensuring identity. Looking at it from a different perspective where those becomes tools of protecting your identity and validating it.
Keith Griffin:And if you look across the company at Cisco, we have a big practice in security outside of collaboration. So things like Duo and multi factor authentication, there's many ways. I think that's what we have to do is take this very holistic approach. If we get drawn into the romantic notion of AI, fighting AI for deepfake, it's it's quite narrow and it's easy to fail. And we have a president in organization, Jeetu Patel.
Keith Griffin:He has this line that I they landed on me and I'd use it over and over which is, you know, we have to be we have to be right all of the time when it comes to security. The bad guys only have to get it right once. And you know, we really have to we really have to focus on making sure we have a holistic approach, not just rely on one narrow set of technologies. But it is easy to get drawn into the deep fake piece. I see it as definitely an emerging threat and something the threat really is is if it goes to scale as opposed to these kind of very targeted attacks that happen right now.
Keith Griffin:But if it goes to scale, then yeah, it's it's dangerous. And the second part of your question was around the guardrails. I think that's a very, very fascinating area because just like prompting, you can have the system guardrail, but you're also going to have application or business specific or feature specific guardrails. So I think it's going to be necessary for us to get beyond where we are right now with system level guardrails, which as you mentioned, be hate speech, toxicity, jailbreak, trying to force an LLM to do something that it's not intended for. I think we have to get beyond that and maybe also look at things like brand based guardrails.
Keith Griffin:You know, if you're if you create a virtual agent for a booking system and you ask it to book something for you, whether it's a flight or hotel or whatever it might be, you probably don't want the virtual agent to recommend a competitor's hotel or flight or something like that. How do you get it to stay on brand is dangerous to your business as, you know, so something more sinister. So
Christoph Magnussen:we found very typically up 10% off and Exactly. Yeah.
Keith Griffin:So it's and it's fascinating because they are, you know, in an agentic system that can happen unless you somehow try to to control it or keep it on some guardrails to not do that.
Christoph Magnussen:I had a very good conversation with Richard Socher who invented prompt engineering as a term and and he's very not critical, but like skeptical about agents scaling fast. And and lately, Andrej Karpathy said something similar that we more had to indeed the decade of agents instead of, like, the next months. You talked about agents a lot quite a lot and also your colleagues today. And this line that you just mentioned also landed on me like you have to get it right every single time and it's not easy and you need your your clients with you to test it, to try it, to work on it. This is a super complex task.
Christoph Magnussen:Yeah. How do you handle that field? Because I know that from your talks, you're not the hype person. You're more the person who really makes it work and happen. What's your view on agents specifically when it comes to God rates and to get it right every time?
Keith Griffin:Yeah. I think I think it's it needs it needs two there's two aspects to it. The first is is building trust and that's why my main guidance is get single agent right. I agree with that sentiment of maybe the the decade of agency, a good way to look at it. It's about getting started now, but you have to build trust.
Keith Griffin:You have to start with a use case that's that's relatively narrow. I talked to a customer in Australia last week while while down at Cisco Live in Melbourne and they talked about creating an agent that could just handle cancellations. They found they were using our topic analytics and found that the topic that was coming up most was cancellations, which was not the intent of that particular customer service area. So number one, that's a good AI finding right away. The topic analytics gave them an outcome.
Keith Griffin:Then they set about resolving that by putting a virtual agent in in order to be able to deal with a cancellation. And they found naturally enough when you think about this from a human point of view, what follows from a cancellation is a reschedule or a rebooking. But they started only with the cancellation. When they got success, then they expand to rebooking and then more and more. And then at some point, you might say, maybe it's the it's the cancellation agent and separately you have a rebooking agent.
Keith Griffin:Now you get into multi agent and maybe agentic after that. But I really really liked the approach they were taking. They weren't trying to boil the ocean. They were looking at it problem by problem, use case by use case and my guess is they'll be very successful because of that approach that they're they're taking. They were also the second part is after you build trust, think is you must focus on user experience in a relentless way.
Keith Griffin:I saw so many failed experiments of take a speech engine connected to an LLM and hope for the best. And I always say hope is not a strategy when it comes to generative AI. So it it is, you know, that's prone to latency along think times in the LLM. Our team has had this really relentless focus on human like user experience, which is very challenging. I mean, we have a good system, but I think there's still an awful long way to go.
Keith Griffin:The team has been focusing on things like turn detection. If you think about this traditionally, how does turn detection work? You wait for silence. But if I wait for silence, there's that much of a pause and you're tempted to jump in and ask it something else. So voice activity detection, silence detection, looking at also the semantics of what's being said in order to figure out when the agent might want to take its turn to talk next, because you might have hints as to when when it's your turn.
Keith Griffin:For example, if the agent asks you to call out your credit card number, well, it knows it can talk after it gets the sixteenth digit because that's how long a credit card number is. So building that level of sophistication into the agent so that it's somewhat human like, it's still it's still like fallible, you can still have, you know, faults with with that. But that's why I say it's a start and you have to have this relentless focus on user experience. We've great user experience team that have really focused on speech and there's been such a resurgence of speech as an interface and this I think agentic world will will actually bring speech back into being a practical way to interact. It is our natural way after all, but you know, we don't tend to associate it with speaking to machines.
Keith Griffin:But now at this point, yeah, I mean, who knows? It it looks like speech will make somewhat of a comeback. That would would have been one
Christoph Magnussen:of my questions anyways. What's your what's your thought on on when it comes to GUI and and interface design? I mean, there are some thoughts and experiments. And when I arrived last evening, I'm like, oh, interesting. You of course, you guys still have hardware.
Christoph Magnussen:And I was always like, why would you do hardware? And then in in the end, I thought, you know what? There are companies, consumer AI companies investing billions in hardware companies.
Keith Griffin:Yes.
Christoph Magnussen:There is something about touching and so on. So you said audio you think could be one of the interfaces. Let's dive a little bit more into that because audio is tricky in such a different interface when it comes to anything we're used to. Like how how do you approach that? What people do you have in your team to work on
Keith Griffin:that? Yeah. We have a lot of investment in this area. When we put our AI strategy together initially, we focused on three key areas, audio and speech, computer vision and video AI, and then language AI. And language AI somewhat backs speech AI.
Keith Griffin:But we separate also speech and audio. So speech for us are workloads like transcription, closed captions where you have maybe speech in text out type of workload. And audio which is core in the audio pipeline like noise removal or you know acting on the media as it's happening. And just on Friday, we announced our intent to acquire a San Francisco based company called EzDubs which is a speech to speech model for translation. I can Live translation.
Keith Griffin:Live translation. Yes. And it's and it's there's still it's that remains, as you say, it's still a challenging area. That's something that works today on approximately one second plus type of latency. But this team is really amazing and I think we can get that down into the hundreds of milliseconds area over time.
Keith Griffin:They're going to keep as a joint Cisco, we're to keep working on on that. But it's it's really fascinating if you think about it, it will mean that, you know, I can speak here in English, you could hear me in German, you can speak in German and hear me in I can hear that in my language and and the application of that broadly across multi person meetings. You think about contact centers and calling. I think we've even yet to discover where where that could could go. So yeah, we we think about it across those different levels of, you know, pure speech and then the audio piece and they're coming together and we also have a design approach to do as much as possible on the edge or near the edge.
Keith Griffin:So if you take our noise removal model, it runs locally on the device or on the client because well for a number of reasons. One from an engineering point of view, that's where the noise happens. So it makes sense to remove it there. But also it removes and reduces the the the the sort of uncertainty for our customers because from a trust point of view, if you're removing the noise on a closed system at the microphone where it's happening, it's a lot easier to see the data privacy aspect and transparency. Whereas if you're if you're intercepting that audio in the cloud and removing the noise there, well, what else could you be doing?
Keith Griffin:Is it quite a natural question and a good question to to ask. So we try to do as much of that processing as possible on the edge and you mentioned, you know, the the hardware link as well. So when we designed our devices, I mean, over ten years, we have NVIDIA GPUs in those devices. You know, it's it was a very visionary move by the organization to put that in, not just for the video quality, but also for machine learning workloads like face detection on the video side, but also on the audio pipeline. That gives us a really strong advantage in being able to carry out machine learning workloads on the audio pipeline and speech and then when we get to more generic hardware like laptops and various other platforms, we have to use different techniques.
Keith Griffin:But I think that, yeah, that that marriage of hardware and software is core to our design beliefs. We we can't always manage it, but where we have our full stack, then, yeah, we can This do a
Christoph Magnussen:is a question mark for many people look at that. I saw that in the audience earlier today that I think everything in AI has to run-in the cloud and at the vendor, but a lot of things can happen at on device on your local machine. What did you do in order to, like, help your team to to understand better these possible, especially the ones that are not technically, we say like everyone in the team needs to understand that this is a new type of technology. What did you do there?
Keith Griffin:Well, it's we we always, if we can try to focus on solving the problem as close to the problem as possible. So if we can do that on the edge on the device, whether it's video device or phone, we will always take that approach. But we also have a secure development lifecycle that we use inside Cisco as an engineering process and all of our engineering teams are trained on that. And that includes our responsible AI framework also. So I think, know, we tried to create awareness of using the best design patterns that are most appropriate for the problem that we're solving and to your point, not just throw everything to the cloud.
Keith Griffin:I mean, you can and it's good that there's certain workloads that could only work there. When you think about large language models and so on, you just won't really run them locally. But we've also seen an emergence of, you know, customers that either have a preference to or simply can't use cloud, but they want to be part of this AI race that's going And a different part of our business has created something called AI PODs, which are AI ready data centers. It's basically one of the key strategic elements of Cisco. Have AI ready data centers, we have future proof workplaces where collaboration is and all underpinned by digital resilience or security business.
Keith Griffin:But what's what that has opened up for us is we can take things like our collaboration workloads and run them on an AI pod that's capable of running open source smaller smaller large language models. Like you take something like Mistral 7B or Llama type of models, run them on the device that can run-in house, train them what a conversation looks like, and suddenly you can have an in house AI powered meeting, translation, contact center or a calling solution. So this is a it's a very interesting trend. It's interesting. I find it fascinating in multiple ways because, you know, if you really are an organization that can't go to cloud, you need some sort of alternative.
Keith Griffin:But out of that, you drive additional types of costs of running that versus a, you know, a SaaS type of model. You get into now owning equipment again. You also have very interesting power requirements like from running these like GPU based effectively small data center type solutions. But it's a fascinating area and I think it's still very very much emerging. But it you know, back to your question, it gives us a whole range of different design patterns that we can choose to use from cloud to AI PODs to write on the agent on the device Mhmm.
Keith Griffin:Within our collaboration devices.
Christoph Magnussen:If you if you take that one step further and I mean, you mentioned it in in the collaboration space, there are a lot of problems you can tackle and you you see right away. What what I was thinking about and we actually talked about in the car yesterday because I I I think about that a lot and I don't have answer yet. If you take the advantages of what AI models can do and what they can do better and worse, and then you take the artifacts of work itself, how would a I mean, this podcast is called AI to the DNA, so we also go really deep to that. What would an AI native app or application or workplace look like that is not invented yet? We we say Yeah.
Keith Griffin:It's a it's a really that's that's that's something we're challenging ourselves with at at the moment. And I think it goes back to two things you you talk about as well. We use exactly the same language which is automation and augmentation. It's like I said earlier, if people don't adopt AI, especially in professional knowledge based roles, it's so easy to get left behind. You know, it's totally changed the way I work every single day.
Keith Griffin:Mhmm. Even across our organization, I see, you know, AI productivity tools. If you're a user experience designer, you're using AI assisted design tools. If you're generating code sorry, if you're writing code, you're using AI generated code and trying to move it in that way. And that's augmenting humans to be better at their at their job.
Keith Griffin:Then when you get into the workplace itself, looking at, some of the agents that we launched recently like AI note taking, like having a video device you can walk into and you don't have to have a remote part to that meeting. You don't have to be in a Webex meeting or Zoom or Microsoft Teams or Google Meet or platform of your choice. You can just say, want to I want to take notes between the two of us right now. And we just use the video device to activate that because it's capable of of doing that. You know, they're they're small examples, but then you have this agentic flow from it of whatever we agree in the notes, maybe it can take actions on those and follow-up for us and it's kind of supercharging the way we get things done.
Keith Griffin:I know for me, but I personally, I'm brutal at follow-up. I'm normally just moving on to the next thing, the next thing, the next thing. So if I say I'm gonna send you a document like good luck getting it, but like I just very very excited about my follow-up. But if I had an agentic flow that was going to say, hey, you promised this document or even surface it. Here's the document you promised.
Keith Griffin:You wanna go just hit send, that's all you have to do. If it made it that easy for me, maybe I get better at that thing that I'm not so good at. You know, I'm really anybody that knows me will know that I've just because I have to keep moving to the next thing, I I very rarely get to do that follow-up. So I think in a workplace like that, these it's it's likely to be the small incremental things at first that will will make a difference. And and also back to the point about building trust will help us buy into the technologies more.
Keith Griffin:But, you know, there's when you think about the physical workplace itself, you think about how things have changed since COVID in terms of returning to office. There's so much revamping of office spaces and, know, putting these like video and collaboration systems in. It's it's on us to make them work well. It's on us to make them work well with the network as well, which is something again we have an as a company have an advantage in. And it's also really, I think, what's often left behind in this is the IT worker.
Keith Griffin:There are people that have to make this work. They need AI too. The you know, whether it's troubleshooting, being able to figure out and get ahead of issues, outages, whatever it might be. AI is typically good at, you know, pattern matching, like anomaly detection, maybe like outages can be prevented by, you know, looking ahead and assisting the the IT hero in this as well, not just the end end user. So I think it has to really span the whole piece, but I think it's gonna be very interesting for the physical workplace.
Keith Griffin:You know, one of the things that our user experience team is constantly focusing on is how do you make it equitable for people that are in an office and those that are remote, but to participate equally in a in a meeting. So recently with our AI AI director, it is something that it it is able to frame up everybody that's in the room and the active speaker. Before we used to just jump to the active speaker, but now you if you're remote, I can't read the room anymore. I'm just seeing that person. So as we learn and evolve in our user experience thinking, we start like looking at these different types of layouts and techniques and it would be foolish of me to say that's the answer, it's now solved.
Keith Griffin:There you know, it will evolve and it will will change. But there's yeah, there's all sorts of interesting challenges and of course, I'm only describing offices, not everybody works in an office, know, so how we automate and use AI in, you know, in manufacturing and healthcare and all of the broad ways that the people work. It's pretty it's pretty fascinating opportunity.
Christoph Magnussen:Which which of these artifacts you work with today would you want to see get rid? I mean, talking about like PowerPoint presentations, meetings, Excel files, recordings, emails. What would be when you say this could be something personal that you take personally and say like, I really want to get rid of that work artifact with I I
Keith Griffin:don't know what it that it's it's necessary for me to have almost all of the things you mentioned, but I would like to get faster and better and more complete with some of those. So again, with the way I work, you know, as I said, I'm pretty bad with with follow-up, but I'm also bad with producing. I have more going on in here than I can get down on paper or I get
Christoph Magnussen:have the same And
Keith Griffin:it's and so I would love the concept of, you know, if I don't have to spend like five hours designing a presentation, but I can describe what I want and it creates like we saw the AI generated vidcasts like just making So that now it's it's done. And that's that's so that's something and that's what I'm actually experimenting with quite a bit. I run a lot of experiments that I've done manually and therefore have a reference point I run with AI to see how was it better and how close was it to the result that we got. So an example would be, you know, when when we did the acquisition announcement last week, that was a relatively manual process to to get to that. So generally in a process like that, you're going to run a technology landscape, you know, what technologies are out there, you can create a shortlist and and and then you start, you know, working with a smaller number of companies and then you get to result.
Keith Griffin:The same with the deep fake analysis that we did, which was over a year ago. I had a manual process and I compared it then with using our internal AI tools that provided by Cisco IT to run the same process. And what took what took me eight weeks to analyze and the numbers are interesting in it, eight weeks to analyze, but not full time, just whenever I could get to it. It took eight minutes, same number, but minutes rather than weeks to get something that was 80% accurate. I had to spend maybe two or three more hours than tweaking it, you know, but but that was and that was a while ago, but that was a really fascinating breakthrough moment of, okay, I have the results and I know what it should look like for from a manual process.
Keith Griffin:I'm going to run this in an automated way and it didn't get it fully right, but gosh, was it an accelerator. So next time out I went the other way around, I'm not going to run the manual process. I'm going to start with the AI queries and a skill, like you mentioned with the prompts, a skill is knowing to ask the right question and then knowing how to iterate on the context of what you get back to get to where you want to be. I think it's actually going to be one of the most important soft skills for people is asking the questions. Very
Christoph Magnussen:difficult thing and and it brings me to a question when you say that how do you I mean, I totally understand when you share the process because I'm I'm working very very similar, but taking an organization along an enterprise is much different. Much different. What are typical mistakes that you see at the moment when people see the pace on the one side and saying this is insane and then think how their organization can do it or not do it or like what what is what is a very very typical mistake?
Keith Griffin:Well, I I think the most fundamental mistake and it's it's probably not the answer you're looking for. The most fundamental mistake is not using it at all. People that think I'm good.
Christoph Magnussen:Oh, I get that. Yeah. A good point.
Keith Griffin:Yeah. Yeah. I think that people there's certain people that are convinced that it's either a trend and it will go away. There's people that even think, well, I'm far enough along in career, I'll see my career out without it. I I don't think so.
Keith Griffin:I think this is coming like a train. It's very real. It's very useful. That's why it's very real. But I you know, I think that that is one one big mistake.
Keith Griffin:And then the second is, especially in enterprises is different groups not being aligned at the kind of the same pace, right? So organizations have to work together to get things done. If one is very aggressively adopting AI and moving very quickly, you can only move as quickly as the ecosystem around you. So you have to have somewhat of a coordinated approach. Even back on a responsible AI framework, I mean our procurement team and legal team became my best friends.
Keith Griffin:You know, we would work very transactionally before that in engineering and product to those functions like, you know, we need a thing, let's source it and go get us. Are we, you know, do we have all of the right legal framework in place, we're good and it's very standard process. And we had to actually create, you know, this this group that was all working at the same pace because the pace of this is so fast that, you know, whether it's your procurement team, legal team, governance, making everybody lock step and move quickly. Move quickly, but but safely, you can cut corners, with it. But I think that's that's something I observed from, just dealing with, other, organizations that they maybe have one team, it might be IT or engineering because it's technical and they lean in very heavily and the rest of the organization just, you know, hasn't really moved at that pace yet.
Keith Griffin:So, yeah, that's at, you know, at a high level, think that's that's one.
Christoph Magnussen:Yeah. That resonates with me and and my question would be like, what how what did you saw solve internally then to to make that work? Like, really concrete saying, hey, this is a format that we established or this is my way in updating and keeping on track. I mean, I I showed on stage how I do it. I play around usually quite a lot to experiment, which brings me quite often also into trouble.
Christoph Magnussen:Mhmm. I'm saying like, oh, word is maybe a bit too far. But I'm pretty much know what I'm doing and and testing and and and trying it out with fun stuff that is not critical, but that's my way. Mhmm. What did you establish to to to scale it?
Keith Griffin:I I think the key thing was establishing an AI governance function, not a not a department, but just like an overlay in the organization that represented all those groups I mentioned earlier, right, having product and IT and legal and procurement. Every organization will be different. They might not have some of those functions, they might have different functions. But making sure that there's representative from each that can do two things, represent the needs of your group for AI, but also represent what's happening in AI into those groups so that everyone is is matching that pace. So how
Christoph Magnussen:technical are they in the in the governance position?
Keith Griffin:They don't have to be because I mean, they could be from entirely non technical Okay. Functions, but it's about everybody being on the same page and same pace. That's that's what really made us really made it work. And then those that are most motivated will end up driving some of of that. But I think it's very useful function and it's it also, you know, it puts like new life into areas like compliance, governance, know, these are now really, in my view, really critical functions.
Christoph Magnussen:If you take one step back and you look at the 12 year old Keith and you have to explain him this kind of new feel here, and I'm asking for a friend with kids. What's your story to, like, teach it to the young Keith or even to kids in order to say, hey, this is this is something you should look at. And I mean, at the same time, they need to be curious, play with it, but also know, okay, there's not only good stuff happening. What do you do here?
Keith Griffin:Yeah. I I would say, first of all, like, don't ignore it, especially at that, you know, at that age and going into education and it's it's it's embrace it. And it doesn't matter in in what field. It doesn't matter even if you're doing something practical and hands on, there's going to be some form of AI augmentation and automation to it. It's it's it's absolutely inevitable.
Keith Griffin:So the first piece of advice would be, you know, embrace it, understand it and know how to make it an advantage for yourself. You know, it's I think every single piece of working life is going to be accelerated by this. I think the more practical hands on craft trade based roles are maybe the later to be, you know, positively impacted by this, but will be I think in in that way. And the more professional knowledge based even software development are going to be the more immediately affected by this. And and again, I like to take the view of positive augmentation.
Keith Griffin:I know there's a lot of uncertainty about the the threat of AI to roles and so on. But but my view would be, you know, the the threat becomes a certainty if you ignore it. You have to take this on. There's there's research that I I use a lot in talks, that came from, Stanford AI and Stanford Medical, I think it was 2016. And, they asked themselves a research question, could AI replace a radiologist?
Keith Griffin:And it's a it was a very fascinating research where there were certain tasks that AI was really good at. Naturally, it's good at image recognition, but that's not all that radiologists does. They have far broader role. And at that point in time, the output and outcome of the research was that, it was a professor Langlotz, think his output was that AI would not replace a radiologist, but a radiologist using AI would replace a radiologist not using AI. Now take radiologist replace it with lawyer, accountant, software engineer, you know, any knowledge based profession.
Keith Griffin:I think it's a true statement and that's almost that's nine years ago, almost ten years ago now when they published that. Yes, you know, the field of radiology is still thriving. Know, so
Christoph Magnussen:More than that. I mean, Geoffrey Geoffrey Hinton had to roll back Yes.
Keith Griffin:Correct. To
Christoph Magnussen:have got the Nobel Prize because there is this paradox that you create actually more work when you innovate in a certain area. And and people always ask me, like, will it replace the jobs? I think we made a YouTube video about it seven years ago or something. K. And we also said, like, no, it won't because what usually happens with this paradox is that the the work explodes and and gets more Yes.
Christoph Magnussen:You solve it in a different way and that's why I'm so curious and really really understanding in which area will it will it change in effect first. We we talked about system prompts between our talks and it it fascinates me that so few people have an understanding that words direct the system. And you have so many different products. How do you govern the different layers of system prompts? And let's let's get a little bit technical.
Christoph Magnussen:I mean, is Sure. Also something like how do you manage it to to collaborate on system prompts?
Keith Griffin:Yeah. And there's there's a certain amount of like go to prompts that are good for just the overall like safety of a system like basic stuff like only answer if you're 100% certain, you know, like managing the the the core output. But then there's certain things that have to be they have to be domain specific or application specific. So we focus on having some core system prompts, but then also empowering engineers to create prompts that are relevant for that workload. Like a, you know, a contact center interaction is very different from a multi person meeting, very different from maybe, you know, a professional business call between on a calling type of system.
Keith Griffin:It may be different also by modality. So messaging type of prompt may be different from a call summary. And there's slight variations, some of them then are very different. And others also bring in other features into the prompt engineering like in a contact center environment very very likely you're going to need to do PII redaction whereas in like a meeting, well you don't necessarily want that, you don't want the transcript of the meeting saying person said to person I will send you the content. You need the names for the actions.
Keith Griffin:So it's so you're less likely to have something like PII redaction in that whereas in contact center where that's required, you've credit card details and other personal information that becomes very very mandatory. So in these core services like prompting, guardrails, PII redaction, we find that it's important to empower the teams but then every single team on every feature to do with AI has to do responsible AI impact assessment. And it's in that assessment that the checks and balances come into it of Have you implemented that enough, maybe too much, but getting the balance right in that. So from an engineering process point of view and technical point of view, that's how we manage it. And we also have a lot of knowledge sharing in the AI space across the teams.
Keith Griffin:So they'll they'll if somebody is stuck on how do I, you know, ensure safety in a certain type of output, they will ask and maybe somebody else has solved it in a different different area like just good good old like team working together to to get things done. So but it's it's a really important area to have not just the system prompts but domain specific and application maybe even feature specific and also to wrap them in a safe way. I mean a safe it's it's safer to have a system that's behind a UI than like an open text box.
Christoph Magnussen:Can get
Keith Griffin:all sorts of you you covered very well prompt injection and and things like that. And that's actually that's quite a risky area. When you think about something like Agentic AI, then each agent in the Agentic flow becomes a new prompt injection Prompt injection Attack surface. Yeah. So you you really proliferate that problem, you know.
Keith Griffin:But it's it's something that it's it's something that's also an emerging area. You know, this is when you look at the next wave of like where what we need to focus on next. It's it's actually a lot of it's a lot of that, you know, it's like known problems that are maybe expanding in terms of their their their opportunity and threats that we have to go solve.
Christoph Magnussen:Which is a very good closing statement because I already got the sign that you need to head to the next meeting. One last very quick one. Is there a question people should ask you about AI but they never ask you?
Keith Griffin:That's a that's a good one. Well, because I work a lot on the acquisition stuff, the number one question people don't ask me is what are you working on right now? Because that would be first that would would give it away. No. I don't I don't know that's think that generally I deal a lot with very knowledgeable people in the AI space and they're normally spot on with what they ask about.
Keith Griffin:But I really, you know, a question that's maybe not asked enough is the where next and I think that's almost fading out because it's moving so quickly. Yeah. You know, I mean, if you asked me a year ago, I would have said multimodal AI is the future. Now I say agentic AI, but in a multimodal way and in three months time it's going to be like something else. How how do we know?
Keith Griffin:So we just have to we have to stay with it and it's, you know, there's never been a more fascinating time to to work in this space.
Christoph Magnussen:So we will see for a second round.
Keith Griffin:Yes. I look forward
Christoph Magnussen:to that. Thank you so much, Keith.
Keith Griffin:Thanks a lot.