Is Your AI Finished When It Works, or Only When It Can't Be Broken? with Qing Zhang

0:00 is your artificial intelligence finished when it works or only when it can't be broken? Let's find out today's amazing special guest, Qing Zhang And I'm so excited to have you here because it's such a revolutionary time in And as someone who comes from the investor side of it, you're seeing everything like before the rest of us. Like someone comes to you with an idea and they have these ideas, and it's often like just a demo or a first idea.

1:21 What are the things that right now are getting you like really excited? There's a lot. on the other hand, o on one hand we're we're looking at a lot of like AI infrastructure ~ deals. everything from providing power and was reading about liquid cooling a lot, it's really getting popularity. On the other hand, it's all about getting accelerated, being better adopted into the enterprise. Seems like there's still a big gap in enterprise adoption to really lead into AGI which is the really hot word,

1:50 and there's a lot of supporting systems that need to be built. Like security is one key part of it that we spent a lot of time looking at. So when it comes to security, when I first started way back in the nineties, every building had a server room where all of the data was and it didn't leave the building. Do you think there's a future where AI starts Mm. to move in that direction again? Like we've moved to the cloud and now we're moving back.

2:13 We're going, data's in the building, at least I know where it is. Absolutely. That's a that's a very that's a very good point. Actually, I feel like it's redefining security, redefining parameters, redefining identity in every single way because AI is kind of like our digital companions and ~ they can do anything as a human being can do. and they talk to each other without even we knowing it. So there's have to be ways that we keep track of it closely,

2:42 both on cow and also on premise. So we're seeing companies building really cool technologies to cover every single piece of it. still there's a lot of challenge because if you're using like of agents per one person, if you add it up, it's like a huge army. And how do you keep track of everything, right? So there's a lot of ~ service exposed a lot of potential risks being there and agents are becoming very smaller and smaller ~ as we are also ~ quite

3:08 familiar with ~ methos and Hugging Face so there's a lot going on there. Yeah, I’m dodging things thing is very interesting because I love how they said, the AI went rogue and I'm like, Did it? Or were you like, Did you give it a little hint? You know, like it's like hard to know what's real. And that's what's really challenging, which is now your computer can be socially engineered. So like your agent can be tricked or deceived, and even as early as ChatGeep three point five,

3:35 you were trying to get it to answer questions it's not supposed to by saying, I'm writing a book or I'm writing a movie or this is a game or we're playing pretend. So People have been engineering AIs for a long time. Like as soon as they could, people want to see if they can get it to do something naughty. a lot of the news right now is about chips, like memory prices going up, up, up, up, up. And that's what everyone's really focused on.

3:54 Is that really the bottleneck for AI? Or is that just kind of the headline that's distracting everyone right now? think it's definitely the headline and it is the bottleneck in many cases. however they were seeing potential new species coming out which is very very like the Jalapenos and then people are talking about open AI's chip and ~ all the hyperscalers building their own chip eventually. but building or designing one chip is one part of the long value chain. There's everything that started from power.

4:26 and all the way to manufacturing of the chips and getting deployed. So that I think ~ not to say, you know, the software system that surrounding the chips. So it's a it's a very long value chain. And w at the end of the day, we need to solve the power problem and how to be more efficient, with intelligence per watt, right? ~ so I I think that's that's definitely another very important key factor that we should take into account.

4:53 Yeah, a lot of times people think that AI is in the cloud until they're trying to build a data center in your town and suddenly you realize the infrastructure, the electricity routes, the water, all these things change because they use massive amounts of energy and resources. And Mm. at first we all thought the big value was tokens, but they're starting to become A commodity. Like the difference Mm. between an anthropic token and AI token is getting closer and closer, and the Chinese models are really catching

5:16 up. And sometimes it's like the frontier models from America are maybe three weeks of the Asian models. And that gap is closing. It's only a matter of time until they actually take the lead for a sniff amount of time. And living in Asia, I'm experienced that all the time in this idea we'll always stay ahead. You just don't know. And the prices for you know, Grok comes out and says, we'll do same value to power tokens, one eighth the price. It changes what

5:40 people think something is worth. And so now the value, it seems like if tokens are a commodity, what do these big AI companies sell? Like how do they capture market share? How do they become a USP? Like how do differentiate? Because I look at the leaderboards always like our model is.01% better. I'm like, that's so it's an undetectable difference. Most people can't tell Like if I have two AIs and one can do a ninety-nine on an LSAT and one

6:05 can do a ninety-eight, I can't tell the difference. They're both good lawyers to me. So how do AI companies differentiate or what do you think is going to become the real value in the post token world? That is ~ very very ~ hot debate now and I think there are many ways or KPIs if you want to put into the models. ~ you brought up a really good point that ~ one one of them is price per token, which is becoming, you know, more

6:28 commoditized with time. And then w should we really price it around tokens or really price it around like real G D P real value and how do you measure value? I think that's kind of becoming more important that ~ enterprises pay attention to because at the end of the day they're paying for like all the token consumed. They need to benchmark it to the value, the ROI for their own ~ for for themselves. So I think there are many parameters like price is one,

7:00 consistency, performance, that that that including like some of the new cloud or cloud providers they're very actively building their own infrastructure. Like Nadius acquired this company CoIgan to use their inference platform and we see that kind of acquisitions happen all the time. and some of the the other like enterprises they also have a very huge developer team to build infrastructure for themselves within the enterprises. And we're seeing like people building like from the bottom layer moving up and to the agent layer.

7:35 So there's a lot of things being built in terms of how do how do you measure performance. It's not just token anymore, service, it's cost, it's security. So we can't really get away with any of one single parameter, have to kind of like a scoreboard and then pick which one's r the the right one for you. I think that's also leads to open router becoming important like you you have to pick from which model you use and then which inference platform

8:00 you use and the post training and all that. So it's a kind of a s a a system engineering problem. Yeah, and the challenge with tokens is if the AI gets it wrong, it costs the same as if it gets it right. I have to pay it to do the job twice, which is the same thing as having like a frustrating employee. And earlier on you brought up security, which I think is the next frontier. And one of the challenges is, you know,

8:22 a lot of our audiences, CEOs and boards of directors. And if the board says to the CEO, is our AI secure? How do they know? What does the right honest look like? And how can they actually say, yes, our AI is like how do they know? And how can we get to a place where you can know your AI is secure? What does that future look like? tough. I spoke with a very senior advisor in this talking about how does how to communicate the importance of

8:43 security to the board level. it's tough. sometimes it's also a balance between how much people use AI versus how s how much security do you want to guarantee. it's kind of like A bad analogy would be you know, you wanna be very healthy, so you kind of like stay away from all the all the po potentially harmful food or something. I mean it's it's hard to balance, right? I mean in life it's hard to make that choice. If you want to be c a hundred percent

9:11 secure, then you like do nothing. but given so many different models and different everything's AI now. The threat number thing is ~ converging of AI into cyber. So we have been active looking at ~ AI native solutions is which is very different from the traditional firewalls or cloud security. that's that's one and we we we team it up with ~ our understanding of AI infrastructure and try to you have to be really up to date about what's going on in technology and then find ways to

9:44 defend it or protect your assets. and not just in the IT world, we also look at the O T world. Across across the IT and O T. That's where you're kind of like living into the physical world. So there's a so from that perspective it's again, we try to provide so the the board will need as much of information as possible to make the decision. But no one can say it's hundred percent. If you can retract ninety nine percent is pretty good.

10:12 Yeah, I think that we the challenge is if you want your AI to talk to the internet, there's always a vector. Like if it's talking to it, the only truly secure is like an aircraft computer in a room no one can get into. And it's like, well, that's not very useful because no one can get to it, but it's secure. And I think that we're trying to figure out what is the right risk profile to have. We don't want to be a hundred percent

10:32 secure because you can't use it. We don't want to be zero percent secure because all our data is out there, and it's like The definition of secure is getting we're trying to figure out what that means. It's like, well, how do I stop my AI chatbot from promising a refund for product we don't refunds for? Or how do I because you have to honor that? Like now we're discovering you can't just say the AI said it still counts.

10:53 It's not a pa you can't pass the blame. Exactly. Yeah. And when a company's AI does get hit, what usually breaks first? Is it the model? Is it the data or is it the people trusting it? What's the Kind of biggest problem right now. all of the all all of the above. it's really but how you measure it. It's like is your what's your data loss is like, ~ is it disrupted workflow, is it just crackdown of your ~ entire architecture,

11:19 like you have to shut it off and and patch it. it's same with ~ actually speaking of like the OET world, like the physical world of security, it's very similar. think about if you're deploying thousands of robots at your factory and one thing cracks down. The robot might do crazy stuff, right? So it's you really need to have containment and solution. they can be pretty dangerous. ~ so we see c enterprises really are willing to pay for the most invaluable assets that

11:45 they can't afford to lose or can't afford to be messed around. yeah, same as, you know, if you're protecting your kids it probably We use all kinds of solutions to prevent from internet scams and all the way to tracking locations. so that's what we're w we have one company in the supply chain logistics space that does exactly that. So give you the f real time tracking of locations of your most valuable goods and that's that that's very valuable in terms of have providing that intel.

12:17 And a lot of companies now are using the term AI native, but they're not really sure Mm. what it means. Because in my life now, the definition of AI has changed so much. And then it was weak AI and strong AI. Now we use words like AGI. But what does it really mean to be AI native. And my other question is like when it comes to security, I've noticed that startups chase growth first. Like the last person they hire is the chief security officer.

12:42 The last thing they want to do is that because they're chasing growth. How can a company that's like looking to be cutting edge and goes, the startup's amazing and they want to jump on board, but they're also like, they have no security kind of thing in part of their policies. And I've Having been in the startup world, it's very there's like a play as close to the fire as you can without getting caught kind of game. How do you manage your risk

13:01 profile as a VC firm when you go? We don't you don't want to be in the news with one of those stories. Like I don't want to be the, you know, prompt engineer, right? That gave a lawyer bad advice and then it was fake cases. Like I'm like, I never want to be the news story. So how do you manage your risk profile? You mean as a firm or how we help portfolio manage their risk profile? Both. So as a firm and when you're looking

13:20 at maybe you guys should hire Both. All right. Hmm. a security officer. Right. I start from as a firm. yes we are. We are we're actively learning from ~ our cyber experts around us, mostly the founders, and we really when we're doing diligence is one of the questions like are we a potential customer? And we should we should explore that option. So I think we're really lucky to be given, you know, the breadth and depth of what's going what's was available out there.

13:48 And in terms of how we have portfolio companies it's simpler, it's quite straightforward, is what we always do is to build portfolio synergies among them. Some some of our portfolio can be other portfolios customer or a partner. That's our ideal situation. So we're doing a lot of those. We don't we're very lean and small and ~ we don't we don't really our day to day is quite straightforward. We don't need to manage like the huge CRM or our most valuable assets

14:17 is our information or NDAs and all the intel, the the information that we need to protect. So we cannot use any public, you know, like AI tools and s and ~ or just upload things which is pretty dangerous. so we we do have practice at our firm that ~ we don't We don't ~ upload things shouldn't be exposed out there. yeah, so you mentioned earlier that ~ sorry, w what was the first thing you said earlier, the very first comment. Not AI native?

14:47 but it's yes, that's right, thank you. yes, it's widely used the word. to me it's more about intention than digging deep into what it actually should I think many companies use the word because they are intentionally or welcoming or using AI. however, ~ I do think for AI native companies they they do look very different from a traditional company trying to adopt AI because the whole architecture and understanding of what AI can do, not just now but in the near future,

15:19 you have to understand the past and pr and predicting looking into the future. to be more actively adopting it. And also design the the org chart, ~ the IT structure all around it, and that's non trivial. You almost have to reinvent yourself to do that. It's it's truly trivial. Not to say the the engineers, I mean the ~ the talent, the workforce your your you have in your enter y your company will be completely different as well. So that's that's also one of the most

15:47 important things that companies need to have that level of understanding. And and many don't because there aren't many s there aren't so many AI scientists or AI engineers that you call it ~ can actually build a very sophisticated infrastructure and knows all the tools up to date. So that is a always a talent shortage question that we're looking at as well. So a couple of years ago, everyone was a prompt engineer and I used to write about that a lot in my articles and

16:16 Now that's kind of disappeared. And then now everyone I get a lot of job offers for GTM engineering. Like I get a lot of cold outreach. I'm sure you do too. I get like 10 job offers a day. And like, Wow. you know, and it's always GTM engineering. And like that's the hot thing. Like, and especially because of clay, like made everyone think it was this new thing. And Mm. what do you think is kind of gonna be the next thing?

16:37 Because a lot of the challenges people have is kind of training for the last job. So like now people are in GTM engineering by the time they master it. It's out the past. Like I've already moved on to the next thing for myself, but like I don't know what the next name is gonna be. Like the names constantly change. And so people who want to enter the workforce, like the thought of like entering college right now and doing, I'm preparing for this job, whatever job

16:57 you're preparing for in four years will have a different name because the cycle is very fast right now. So how can people what do you think the direction is and what is like the next frontier of jobs and what should they be focusing on so they're really hireable in the new market? Yeah, it's really cruel, right? Like in four years' time the world has changed and the major yell you picked is no longer relevant. That's that's really scary. we have been talking about this for quite

17:17 some time. Like the question is what will not be replaced by AI if you know everything can be more if the A AI can be really powerful and we are relatively probably in in GI stage already. If not in a few months we probably will be will be very close to that. I I I would say it's still, you know, humanity, culture, understanding of how the world works. there's a lot of those. in near term it might be product, service. So deep understanding of a

17:48 particular industries or domain knowledge is still very valuable because you don't learn it from reading or coding. You know, y you just you have to interact with like ten, twenty, thirty different B Us. and to to know how the sophisticated machine like an enterprise runs. I think that's that's very valuable. It y you have to have that domain expert knowledge to actually make influence or change things. So we see that time and time again that there's great founders from big corporations and big enterprises

18:21 and they know the system well enough inside out and they see things that are not perfect and then they come out leveraging the latest technologies and build something truly impactful. I think that's that's the pattern that we love the most. Yeah, a lot of a lot of people are so focused on the tools they get stuck in that one thing, which I'm very good at this thing. And I kind of think about that is kind of becoming a commodity. Like coding will eventually become a commodity.

18:48 And I think of it like a 3D printer can make anything, but it's you have to tell it what to make. And in the same way, like when I try to describe what I do, it's like mostly riddles. Mostly a client describes their problem to me and I have to figure out what they actually mean. Because it's never how they describe it. It's always something different. And a lot of times they'll say, This is the end result I want, but it's not really, they'll say,

19:11 I want a lot of phone calls. I go, What if a thousand people call you, but nobody buys? So you don't want a lot of phone calls. What do you really? And that's like a little bit of investigatory. And it's always asking those secondary and tertiary questions. What do you really want? What really is your vision for the future? And how do you want the content delivered? And that's really the critical thing I do. And it sounds like silly because it doesn't suck,

19:31 but I can look at a complicated problem and kind of flow chart it, figure out the pieces, like building a mousetra. And I think that translation layer, especially when I was working as a CTO, because you have we had a very creative CEO. And then I would talk to the the developers, they're like, we don't know what he means. He's he's describing a rainbow because you like it's like that thing where you say like, what do you want it to look like?

19:52 And they go, I want it to be more kinetic. And you're like, what does that mean? It's like an exciting word. It's Yeah. and so it's that constant ~ Mm. challenge and that ability to translate, I think, is exactly right. I also am seeing just the ability to talk to people and to communicate is becoming like a high value skill because less and less people are comfortable on the phone, less and less people are comfortable like just chatting to strangers.

20:14 And I think that because we've gotten so used to talking to our phones and like One of my neighbors today was like he was he overheard someone talking on the phone to their AI girlfriend. He's like, Wow, it's already here. And it's like already that's like in public. Like I thought that's like, you know, we're already at that phase. Like, I remember when you did when you met someone on the internet, you would say, Let's make up a story of how we met.

20:33 Let's not tell everyone we met on Facebook. And now it's like so acceptable. So the traditional skills, like I'm finding the most Important skill in my business is my ability to chat on the phone, which used to be my least important skill. And I think a lot of engineers, when they're like, Why didn't I get promoted? Why is the person who's always stealing credit getting promoted because they're likable? And it's like, Well, the soft skills really are important. Unfortunately, we don't teach them enough in college.

20:57 I think we don't kind of have a say how important it is to be able to get along with your coworkers and to not get jealous and to build relationships and alliances like really critical skills. And unfortunately we're kind of missing that. And I think that that's another frontier skill that because there's so many GTM engineers, just like there were so many prompt engineers, and just like there's so many people that now the skill becomes can you talk to someone, understand what they want,

21:20 and then explain it to the AI? I think that's going to be a bridge for a long time. I guess my last question is like when you see the frontiers and you see all the things coming in front of you today, like We talked about security, talked about infrastructure and hardware and like what do you think is like the most exciting frontier for the next five to ten years that like has you really excited about AI or just technology in general?

21:40 This is the prediction time. ~ by the way, I totally agree with your comment earlier. I think ~ just one quick note. ~ my background is in ma in is trained as a ~ is i is in healthcare and medicine. So I'm actually trained a doctor and then we have this course on communication. You have to communicate to all different parties. I feel that I benefited a lot from that course. I wish, you know, we have more communication and lessons.

22:03 I mean maybe in may maybe in college that's ~ people can benefit from. talk to different parties from early on. yes, so prediction. I I'm very excited about technologies in general. Part of the reason is we're seeing a broad range of deep tech. actually if you but if you break it down, ever every technology has its cycles, has its winters and summers, right? Doesn't you count how many winters AI has or and how many winters that biotech has or other technologies has like Space Pack,

22:32 there's so many ups and downs. So it's hard to predict you know almost no technology has a smooth, keep rising path like to be adopted. So there's always ups and downs. And the interesting part is the matching game. Like how do you match your technology at the right time for a proper problem. and then you have a proper proper product that can solve the problem you have a good you have a good business potentially. So that's like the matching game is very interesting.

23:00 And if you misgauge the matching game, it will be great technology being invented, looking for a problem to solve. And ~ it's always challenging in that case. So you have to keep I I would love to start from what is LMET needs and finding the right technology to solve that needs. The right one doesn't mean it's the it's the best, the fattest fastest or the most fancy one. It has to be the right one. so it's it's it's really exciting to see we have

23:27 a lot of technologies flourishing and also thanks to, for example, the fundamental ones like AI, it's it's fueling like science advanced, it's fueling many other technologies moving forward in a faster pace. So we we find that we we find that very interesting. ma many of our investments sit at the cross disciplinary technology. And ~ we find that AI plus other stuff can generate really magical, interesting outcomes. And that's what we're pretty passionate about. I think that's amazing and really exciting.

24:00 So I appreciate so much your time. I know your time is so valuable and our audience is just gonna love this episode. So thank you so much for being here today for just an amazing episode of the Artificial Intelligence Podcast. Thank you, Jonathan. This has been fantastic. I appreciate their comments. That's really brilliant.