Retool CEO David Hsu on Why AI Models Are No Longer the Bottleneck

0:00And the reality is actually what is blocking a company, an enterprise, from actually getting value from AI at this point is no longer how smart the model is. The model is smarter than any one of us put together.

0:14Peter High: Welcome to Technovation. I'm Peter High, president of Metastrategy. My guest today is David Hsu. David is the co-founder and chief executive officer of Retool. Founded nine years ago, Retool has grown into one of the leading enterprise application platforms, helping organizations build internal software and AI-powered applications faster and more securely. A graduate of Y Combinator, the company attracted early backing from the co-founders of Stripe and following its most recent funding round, led by Sequoia Capital, is valued at more than$3 billion. As Chief Executive Officer, David is leading Retool through its next phase of growth as enterprises embrace AI-native software development.

0:50Peter High: His team is focused on helping organizations harness the speed and creativity of large language models while providing the governance, security, and centralized controls needed to deploy AI responsibly at scale. As AI makes it possible for nearly anyone to build applications, Retool is helping our enterprises ensure those applications are created in a way that's secure, compliant, and manageable. David co-founded Retool in 2017 after beginning his career as a software engineer at Palantir, where he saw firsthand how difficult and time-consuming it was for organizations to build the internal tools their teams depended on every day.

1:24Peter High: That experience became the inspiration for Retool's mission to dramatically simplify enterprise software development. David Hsu, welcome to Technovation. It's great to speak with you today. Great to be here. Today's episode is brought to you by Eleven Labs. Most enterprises trying to deploy AI agents hit the same wall. The agent itself is fine. The problem is that it sounds wrong. Robotic, flat, off-brand, and the moment a customer notices, the trust is gone. Eleven Labs develops the AI voice and audio models that give agents a human-level voice, the tone, the emotion, the rhythm of real speech, and wraps them in the full stack you need to actually deploy.

2:03Peter High: Knowledge integrations, workflow tooling, multilingual support, and enterprise-grade reliability. Build human-level conversational experiences into your products, services, and operations. Head to 11labs.io backslash technovation to get started for free. And now on to the interview. Well, David, I wanted to ask you, I thought we'd begin with Retool, the company that nine years ago you co-founded and that you currently lead. After your most recent funding round, the company's valued at over$3 billion. Co-founders of Stripe were angel investors. You were a Y Combinator company. Sequoia led a recent funding round as well.

2:45Peter High: So grade A investors behind you. Take a bit of time, if you would, for those who may be less familiar with it, to describe Retool's business, if you would. Yeah. So today, what Retool is really focused on is how do you secure AI in your company? Specifically, how do you secure vibe coding in your company? And the core idea here is that with LLMs and with AI today, everyone in your company can go and build applications, which is great. But the challenge is that when you don't have the guardrails, when you can't control who is building what, what data they're connected to, or they're hosting the different applications, it becomes actually more of a headache than a boon to your business.

3:30And so is there a way to go leverage the power and creativity of AIs for all your employees, for them to be able to create all the ideas they want, but do so in a centralized, governed way that IT or some centralized team can actually go and manage? That's basically what it requires.

3:49Peter High: And I mentioned you co-founded the company nine years ago. How has it evolved across that time? I'd imagine there are probably some pivots to what you're just now describing. Talk a little bit about the journey across most of the past decade. Yeah. So there's been quite a bit. So today, Retool is maybe around 10 ,000 or so customers. And there are companies ranging from Amazon to Taco Bell to ConAgra to Boeing to DoorDash to Stripe to OpenAI, et cetera. So many, many companies today use Retool. And when we first started the company, it was born. So my co-founder and I were actually both engineers.

4:30It was born out of an idea that we were building a lot of business software, sort of business apps, if you will, internal apps, as they say in Selected Valley. And after building so many of these apps, eventually we thought there's got to be a faster way of building all these different applications. Because if you look at enterprise applications, they all kind of have the same building blocks. They have authentication, authorization, audit logs, but on the front end, really it's just buttons, tables, drop downs, forms, stuff like that. And so there's a lot of repetition, if you will, across this large class of software, enterprise apps.

5:10And so we thought, could there be a more effective or efficient way you've built all these different applications instead of writing code from scratch all the time? So that's kind of where the idea of a retool first came from. And we discovered pretty quickly is that we started actually by trying to sell to other startups. But quickly, we discovered that startups don't have that many enterprise apps, which is not too surprising. And instead, one of our first customers was a company called Tyco Electronics. I think it's around 140 ,000 person company. So, you know, sort of a mid-sized Fortune 500 company.

5:46And they told us they have so many apps to go and build. and instead of building all these apps from scratch, if they could use a platform like Retool, they would really love that. And so that's where sort of Retool first came from. And what's changed over the last few years with the advent of LLMs, today, no everyone can be a builder. It used to be the case actually with Retool that actually we mostly targeted engineers because we believed engineers are the ones building all the internal applications or all the enterprise apps. And this was true up until probably this year, actually. This year, we've seen in our customer base, especially amongst the enterprise customers, there's been a lot of interest in, given how powerful LLMs are, can we allow these non-technical, these non-software engineers to actually do and build apps?

6:34And of course, everyone wants the answer to be yes. The answer should be yes. But I think in the enterprise, there's really specific needs. And it's not just, let's go right into coding chat GPT, throw it up on the public internet and the Apple work. You have to worry about what data to connect to. How do you connect to the data? Can you connect to the data if it's on-premise, for example? Who can use the application? How do I hook in a single sign-on? So which groups in my company can actually go use this application? How do I make sure everything is audited and secure? How do I make sure it's hosted on my own VPC?

7:07So, you know, you have to be on a VPN to even go and use this application. There's so many things you have to worry about besides just the code itself. But the really incredible things with LLM, now the code itself is actually so easy to write. So now anyone can be a builder.

7:20Peter High: It's so fascinating and an exciting space that we're in that we've reached. And I wonder, you know, now having worked with so many massive companies across B2C, B2B, business to government, What do you find to be the impediments to more broadly deriving value out of AI, generally speaking, Vibe coding more specifically, and frankly, developing a more broadly democratic software development methodology, if you will, akin to what you're describing now? Yeah, so the really surprising thing is, if you look at most industry discourse, whether in the LVM industry or broadly in the technology industry, I think it's wrong.

8:02What most people are interested in talking about is how smart are the most recent models. And we're recording this in late July. Fable has come out, GPT 5.6. Sol has recently come out as well. And there's a lot of discussion. And CIOs that I meet with would love to learn about, you know, how smart are these models? How powerful are they? And the reality is, actually, what is blocking a company, an enterprise, from actually getting value from AI at this point is no longer how smart the model is. The model is smarter than any one of us put together. It's way smarter, actually. If you think about the diversity of knowledge it has, if you think about the fact that it's able to give a pretty intelligent reply to practically any question, I would say it's pretty clearly smarter than any individual QMIT at this point at probably 99, if not 100 % of tasks.

8:57And so it's great that we're pushing the front here with Fable and with 5.6, but to be honest, if you go back to 5.5, is it that much less capable? It's not. And what's really bottlenecking or what's really preventing enterprises from getting the full value from AI or LLMs today is not, in fact, LLM itself anymore. It's what we would call the harness around the LLM, which is, how does the LLM actually affect change or do work in your organization? Just the raw LLM by itself. You can imagine it just being sort of a chat window. And as we know, chat is great. Again, chat can be very smart. And in fact, I think at this point, chat can get smarter.

9:39and I don't think it would make any impact on our business or your business at this point. But what's really powerful is the idea of the harness, which is it allows the AI to autonomously go do things in your organization. One analogy I would give, which I think is quite interesting, is if you look at the difference between something like a chat GPT versus a cloud code, for example, or even, you know, cloud chat versus cloud code, they're powered by the same model. You can go use Fable in cloud chat. and could use Fable Cloud Code. And the innovation behind something like a Cloud Code or a cursor or a codex is not a better model.

10:18It's actually a better way for the model to go and do work in your code base. Because you can imagine, you know, before Cloud Code came out, what you could do is you could just chat to chat to you. And you could say, hey, you're running this program. It writes to this program. You copy paste that program into your terminal window. You run the program, you say there's an error. Let me copy and paste the error back into chat. And the chat says, okay, well, you're right. This is error. Let me fix it. Okay, here's the fix. Here's this one-line fix. You have to copy the one-line over again into the text editor.

10:51So you're kind of going back and forth, right? You have effectively, the engineer has effectively become the API for the LLM. Because the LLM is using the engineer to effect change in the code base, basically. And today, obviously no one does that anymore because we have a harness that allows the LLM to directly go make change in the code base. And in fact, the LLM is even able to not only change the code, but even run the code. And so now you can actually see engineers are substantially more productive than having to copy, paste back and flip them, right? What's really remarkable about this is that this change has only happened for engineers.

11:30If you look at how non-engineers in your company are using AI today, largely speaking, it's a lot of copy and paste. And so if you think about it that way, what your non-engineers, what the majority of your company is doing when they use AI, is they are effectively acting as the API for AI to go affect change in your company. Which, if you think about it, it seems totally ridiculous. It's almost like the human has become the API. The human's not a very good API, right? It takes a lot of work. You pay them a lot. It's not very efficient. They make mistakes even sometimes. It would be far more effective as if you could just somehow glue the LLM directly into doing actual business tasks.

12:14Now, the problem is that LLMs do make mistakes. And so you then need a governance layer. How do you supervise the LLM? But that, I think, is really the bottleneck today is if we can figure out that harness for non-knowledge work as an industry, I think what you saw happen in coding, LMS, I would say that LMS have really totally reinvented what coding looks like in a span of 12 months. I think that kind of change will come for every single other job there is. But what's really missing is that harness layer. It's no longer the intelligence layer.

12:49Peter High: What do you see as the limitations of artificial intelligence? As you contemplate and have had reason within your own company and as well as seeing now in its implementation across so many major enterprises, David, you've articulated some of the most important and accretive value propositions in your description. What are the limitations for the foreseeable future that you see?

13:16hmm i would say that again from an intelligence perspective i'm actually not sure there are too many limitations at this point uh i recently saw uh a thread uh a chachi pt thread uh from this mathematician and terence tau and he was proving a new theorem uh or he was uh validating the proof of a of a new theorem with chachi pt and if you look at the conversation i'll send it to you later at the show notes, you can see that the LM is really, really, really smart. Terence Tao, who's one of the preeminent mathematicians of our time, is almost an equal, is treating the LM as equal. And so it's, you know, if you look at that, you're like, wow, there is not much limitation there.

14:00Instead, I would say the limitation really is, are you able to effectively monitor what the LM is doing? Are you effectively able to give it the context that it needs? Are you sometimes able to roll things back if necessary? Are you able to work with humans effectively? So some things will, I think, always require human approval because LMs cannot be held responsible for certain things. And that's, I think, those are the features, I would say, of a harness. Where you think about a coding harness, let's say like a ClotCode, for example. What ClotCode does is not only does it allow, does it give it the tools to go and do things, but it also gives it the context because it gives it the whole code base that it can read and create index of, for example.

14:40the fact that it's code also makes it nice because you can actually write it and put it in staging and so it doesn't directly affect anything yet until you promote it to production and I suspect the promotion to production will be human-gated for some time to come I think and the reason is again because we just say who's responsible for doing some dangerous action and so I think those if you sort of think about the analogy between coding and the rest of knowledge work and you think what has made the coding harness so effective, the context engineering, the gating, the human approvals, if you're able to bring that over to knowledge work, I think those are the bottlenecks at this point rather than the model itself.

15:19Peter High: As I contemplate topics of curiosity and worry among CIOs, David, among those that rises high these days is tokenomics. And as demand for the use of the best models increases, especially among technical staff, but increasingly, as you point out to non-technical staff as well, the possibility that the best model available is used for, you know, the frontier model used for simple tasks, for example, can really drive up the costs, among other ways as well. And I wonder, what advice do you have for organizations and for executives who are worried about the total cost of ownership of artificial intelligence?

16:09Yeah, two points. One is, first, there's really great news, which is that there's a lot of innovation right now in the LM space. It used to be more of an ugly LP with just open AI and thought they maybe arguably even just open AI for the first year or so. But today, there's so many options, whether it's the open source models from the US or from China. There's a lot of competition now in the model industry. And to what you said, what we discovered internally at Retool, but also with many of our customers, is that you oftentimes do not need the best model, the most intelligent model to do a lot of these tasks.

16:49What we've noticed is in our harness when we do some evals, when we eval, for example, the latest CRN models, the latest open source models, and compare them actually to models such as in a Fable, for example, we see that actually in probably 90 % of tasks, the success rate, we write the evals, the success rate of the task with an open source model actually is around about 90 % to 95 % as good as something like a Fable at something like 10 % of the cost. And so it's really shocking that you're able, and I think what's happening is that, again, intelligence is not the model anymore. Models are getting smarter and smarter every day, but do you really need to pay 10x more for that 10 % more intelligence?

17:43Probably not, actually. Let's give actually one story, one anecdote here. We were working with one company that had access to Fable, sorry, access to Mythos, the more dangerous version, cybersecurity version of Fable. and they were one of 20 companies or something worldwide have access to this model, and they were using it, and they discovered many use cases for coding purposes from a security perspective for coding purposes, but they actually weren't really able to use the model in any other way, actually. They just didn't know what to do with it. We have a really smart model, and we can chat with it, and we can use it in a coding agent, but there's not much else we can do with it.

18:29And you can compare that with another company who did not have access to Mythos. This was around a month or two ago. And they were substantially more advanced from an AI perspective. They had automated, I think, on the order of maybe 100 or so different workflows and tasks and jobs with AI. and they had no access to a model like Mythos. So that I think is a good example where one company is substantially more AI advanced and is able to adopt AI much more across the organization to achieve much more ROI from AI and actually at a much cheaper cost. I believe they were using GLF actually, the open source model, despite not having access to the smartest model.

19:13And so again, I think especially for enterprise use cases, a lot of the time you actually prefer these open source models, that you can run in your own infrastructure, for example. You can actually air gap to make sure that no data goes in or out. And the bottleneck is really not the intelligence. It's the fact that you can get 90 % of the intelligence or 10 % of the cost is, I think, kind of a no-brainer. So that's number one. So I think the fact that there's so many models is actually great for us as consumers. There's so much competition. That's a great thing. The second point, though, is, And this is where I think Palantir CEO Alex Karp has talked a little bit about, which is, I think, given how the model industry is developing, how much competition there is, and how quickly things are moving there, I think hitching your wagon to any one particular model provider at this point is basically suicide.

20:11you see a lot of stories now there was one in the journal I think just last week about how was this the case with startups I don't know if this is bubble to the enterprise level yet but you see companies like Anthropocon OpenAI offering crazy discounts to startups on the tokens and the purpose of that is to lock them in they want to lock you into a multi-year contract such that when new models like a GLM or when you come out, you aren't able to switch to them. And most startups that I know are not agreeing to these terms because they realize that even if you get a, let's say, 30 % discount on Anthropic or OpenAI tokens, if GeoLab comes out with a model that is 90 % cheaper and is 90 % as good, you're still going to want to use that, actually.

21:02And so what I see as a danger, I think, especially for enterprises, is that if you are so attached to one particular company, whether Anthropoc or OpenAI, you become totally beholden, not only to the future pricing, but also to the future research. But also, you almost outsource your company's most precious asset. It's your memories, your contacts, all to one. And once they have you there, that's exactly what they want you. They want you having all your Microsoft Teams messages synced into a particular place. They want access to all your databases and APIs. And once they've built up the contextual graph, it becomes almost impossible to escape them, no matter how expensive or how crappy their models become in the future.

21:51And so that's something I would really caution enterprises against. And so the cost of the model actually is not even just the token cost. It is sort of how locked in are you? and the future sort of opportunity cost of being so locked in to one particular provider. And that I think is a really hidden and really dangerous cost. That I suspect is something we'll be talking more about, especially as if they're able to come as enterprises, sorry, multi-year, five, 10-year contracts. That would be really scary, I think.

22:23Peter High: Well, I want to, this raises another point, David, that I wanted to raise with you. And that is the traditional SaaS players, They certainly historically have wanted to lock you in. They do today as well. Lock you in for multi-year contracts where your relationship with them gets deeper and more complex as you leverage and provide more and more data such that it makes it difficult to unwind. I don't mean to say this cynically. It's a good business model, obviously. And these have been very successful companies that have occupied these spaces. There are those who say certainly, and Wall Street is certainly reacting accordingly, that those advantages are dissipating, especially as a solid number of startups are coming to market at enterprise levels with a different sort of model that they are employing that directly takes on the traditional SaaS players.

23:21Peter High: and thus the SaaSpocalypse, as it's sometimes referred to, is rising, let's say. I'd love your own thought as somebody who's with an organization that is at times taking on the behemoths, as well as observing this on the heels of the points you just made about the importance for CIOs not to lock in to multi-year deals that provide data, where the data becomes so woven into solutions that it's very difficult to unwind those contracts. Talk a bit about your own reaction to that. Yeah, this is fascinating because I'm to some extent on both sides. Retool is a SaaS company or a software company, just like any other company in Silicon Valley.

24:09So to some extent, you could even say we're affected by SaaS companies, you could argue. On the other hand, we see is that many of our customers today are actually replacing SaaS with Retool because they're able to build these applications internally much faster in a much more agile manner. And it's specifically suited to them. It's not something they buy off the shelf. It's way cheaper as well. So we see both sides. Where I'd net out is, it's a little nuanced. I think that for the very large system of records, especially in areas where there's high compliance and governance requirements, an example I would say is Workday.

24:52I think those platforms, largely speaking from a centralized perspective, will actually thrive, I think. And the reason is there is kind of just too much lock in there. I think for someone to imagine vibe coding a Workday, it's ridiculous. because, again, the hard part about Workday is not the actual interfaces. It's not the interfaces of, you know, how do you change something for a particular employee. That's actually easy. The much harder part is how do you deal with payroll compliance in 50 different states and all these different countries? That's the stuff that is really hard. And I cannot imagine that anyone would want to take on the responsibility of vibe coding that, maybe because you're probably going to be wrong about it.

25:35But even if you are right about it, it's going to require so much effort then you might as well go pay Workday, honestly, to go take on the responsibility. So I think SaaS platforms like that, I think, honestly, at least the core of that platform is going to be pretty safe. Where I think it gets a little trickier is when you start looking at applications on top of those platforms, because what we've seen amongst many of our customers is that many CIOs have purchased these platforms. And in fact, if you look at the ServiceNow, the sales forces, the SAPs, the Workday, pretty much every large company has purchased all of these different platforms at this point.

26:15And again, I think that the sort of central idea of, I want to hire an employee and that lives on Workday, that will not change, I don't think. But what we see is if you're looking to build a simple application on top of your Workday data, for example, we have one customer, Stripe actually, a Silicon Valley company, I think around 10 ,000 or so employees, so of a moderate size. And they love Workday, but they built a bunch of applications at Retool on top of Workday data. Because the fact is that a lot of these platforms, they have really crappy developer experiences. And that's true if you're a developer.

26:54And if you're a non-developer, you might as well give up. There's no way for you to go learn Workday Extend, to go learn the SFDC Lightning app builder. It's just so much work to go learn an app builder like that. And, you know, as an HR person, you may just have a very simple app, a very simple analytics dashboard you want to go and build. In Stripe's case, they want to go build performance management dashboards on top of their Workday data. So they could actually allow individuals to go enter the performance reviews in a nicer looking interface, basically, than Workday. And doing so in something like a RIT tool where you can very easily prompt, hey, you know, I want to build this kind of application on this kind of data, so you sort of have the power of AI.

27:36But on top of your Workday, Salesforce, ServiceNow data, that is really powerful. And so what I imagine is I think the platform will stay there, but when people start building more and more applications, they will probably shy away from Workday Extend, from Salesforce Lightning, from the ServiceNow development platform. And instead, they'll probably build in something more like Retool and it'll just connect to that data. So that's how I see the big player is playing out. I think there's a lot of sort of micro SaaS or SaaS that's maybe less than a billion dollars in revenue, I would say. So maybe SaaS that is specific, that someone in your business purchased for a specific problem.

28:21A lot of that I do see probably going away. And the reason for that is it may be at this point actually easier for, at least, you know, if not today, I think in a year's time, once you have more of the governance card relative to the enterprise harness, for the individual, let's say, person in a analytics function or in a people-oriented function, for them to go, honestly, go build an app with an LLM, I think it's going to be quite possible, if not now, I think certainly within a year. And a lot of those sort of small applications they might have bought for, you know, one particular thing, I think that'll go away.

28:58So I think the core of the big platforms are safe. I think the building on top of the big platforms with their development platforms, I think that will probably go away. And I think this micro SaaS, if you will, will also go away.

29:12Peter High: Great overview. Thank you for sharing your insights there. You mentioned mythos earlier, which certainly leads to topics centered around cybersecurity, which, of course, becomes as always more complex as time goes on, but all the more so now with the emergence of both generative as well as agentic artificial intelligence. intelligence. And you pointed out, of course, that defenders have to protect everything while attackers only need to find one gap. And I wonder, you know, as you contemplate this asymmetry, what are your observations or recommendations for heads of tech and digital at major enterprises to ensure they are both getting the advantages naturally out of artificial intelligence while safeguarding their most important assets at the same time.

30:00Yeah. So this, unfortunately, I think is mostly bad news because the fact of the matter is that compared to where software engineering or software security or cyber security was a few years ago, for software engineering, at least where we are now, is not only do we have far more code being written. So if you, I think GitHub released a stat that said that this year, I think their code volume was up around 10 times since last year. That's pretty, it's almost, it's a horror magazine, but it's kind of hard to imagine really. There's so much more code being written. So that's one fact. And the code quality is actually substantially worse than we were two years ago.

30:49And the reason is it used to be the case that any line of code that was written was certainly written by a human. And so some human really thought about it, but probably it was even reviewed by a human too before it made it to prod. Today, certainly at this point, humans are not writing all the code. Certainly that's the case already. And sooner or later, it's going to become the case that humans are not going to be reviewing the code anymore. And so on the positive side, you can say, well, code is just so cheap. You can write so much of it and we can just ship it. The downside, it's true, code is cheap, but the problem of more code and worse code is actually, that should be very frightening to any CI or any CSIP because you just don't know what's going on.

31:39And this is exacerbated by the fact that a lot of people, and I think for good reasons, want non-engineers to go build software. You could argue that at least when the engineer is writing a lot more code, despite the code being probably lower quality, the engineer is responsible for it. If an engineer introduces a vulnerability, which I think they are, increasingly introducing more vulnerabilities accidentally, at least they're somewhat responsible for it. You can give them feedback, you can hold them accountable, and they will learn. With a non-engineer, it's practically impossible to hold them accountable to the code.

32:16They don't even understand the code. And so you basically have a black and white decision. Either you allow them to go and go create their own apps, or you don't. And what I see a lot of CIOs and CISOs especially struggle with now is how do you achieve both? Is there any path where you can actually do both? And I think there is. And I think this comes down again to the enterprise hearts, which is if you are able to go find a way where non-engineers, you can set up guardrails for them, guardrails that probably an engineer may not actually need, but the non-engineer really does need, if you're able to set up those guardrails, I believe you are actually able to democratize building.

33:01And to give you some examples of these guardrails that actually Retool provides, for example, we allow you to host Retool fully on-premise. and so you can deploy it in your own VPC, in your own AWS or GCP as you're wherever you want. You can have a fully air gap, actually. And so now you have sort of a solution where anyone can go prompt and go build applications and it's fully air gapped within your own environment, which is very powerful. Not only does that prevent other people from getting in, because now I talked to a CIO recently who told me that he recently discovered that a bunch of his company was actually vibe coding.

33:40And not only were they vibe coding, they had actually deployed those applications onto the public internet. And so if you just go visit this URL, you'll see confidential public information from that company. And what happened was the, I think it was a business person. He had tried to say, hey, I want to build an app to analyze this particular data. Well, the data is not available. I can't connect to it. So you know what? I'm just going to export that data from Salesforce. So he exported it as a CSV, he uploaded it into his application, and now the app is on the internet. And it's visible for anyone to see.

34:12In fact, it was being indexed by Google, actually. And so that's the example where it's really bad. And so I think actually a lot of this stuff is very hard and actually not impossible to shut down because in that particular case, that business person actually could use their personal account to go sign up for that Vibe coding platform. And so what you really want, I think trying to shut this stuff down is actually quite hard. What you instead want to do is actually give them a sanctioned place where they can do all of this. And so if you had, and actually they're a customer now, this particular company, and now that company is able to say, hey, everyone that wants to go build applications, go build applications here.

34:53You can hook it up to a Cloud or a ChatGPT if you want to, or we can build inside, doesn't matter, You can build anywhere. They just say, hey, Claude, go build me this application. You'll go build it on the Retail platform. And that guarantees that we're connected to data in the right way, in a centralized way, where you can actually monitor what is going on. It's deployed in your VPC, so it's never accessible on the internet. It's only accessible via VPN, et cetera. And so with a platform like this, it becomes actually possible to go democratize software engineering and software creation. And so that's, I think, how you sort of get past how you sort of achieve the power of AI while maintaining security.

35:35Now, you also asked a different question, which I also find fascinating, which is how does this change the equilibrium between attackers and defenders? And I would say over the last few years, we're roughly at an equilibrium where if you look at large companies today, large companies get hacked every now and then. The average company probably will not suffer a devastating hack over the course of a decade. Probably not. But every now and then, you do hear about some really bad hacks every now and then. But it's the minority of cases. And so, largely speaking, it's an okay world for us to live in.

36:13And we've gotten here by investing a lot in security. CIOs have invested a lot in security. We've bought a lot of solutions. We've figured out how to lock things down across our company. So after a lot of investment, things are roughly at the equilibrium.

36:29Peter High: Now we have AI. Now you have LLMs. And LLMs, unfortunately, are, as we discussed earlier before the show, all they have to do is they have to find one vulnerability. They find one vulnerability, they're in. And you, as a defender, actually have to defend against all vulnerabilities. And so it's a pretty asymmetric position to be in, if you will. And the ability to go find one hole in your infrastructure is now much easier because you can actually automate that with an LLM now. So it's, I think, a very scary place to be. And so that's one fact. The second is that bad stuff is already happening.

37:07You've seen, I think it was last week, when Hugging Face got hacked accidentally by OpenAI, actually. What OpenAI did was they were testing a new model and they had removed some of the guardrails And they said, you know, please go try to find answers or, you know, do this test, basically, is what they said. And what the model decided is that, well, to do this test, I believe that actually I can find the answers to this test actually on Hugging Face. And so let me go hack into Hugging Face and try to go find the answers that way. And what was remarkable is that actually that the container in which that test was running was supposed to be air-gapped.

37:52It was air-gapped via software. So there was no sort of hard air-gapped from a hardware perspective, but it was, you know, air-gapped from software. And it turns out that that model, I think it was 5.6 at least, if not a greater model, actually found a zero-day basically in that software. and it managed to escape the sandbox, the software sandbox, and actually found its way into Hockingface's infrastructure. And that's the example where I think in the past, finding zero days in software was prohibitively expensive. I think if you want to go buy a zero day for iOS, for example, it's in the order of a few million to tens of billions of dollars to go find a vulnerability like that.

38:32And there's a shelf life to it. By the time you use it, it's probably going to be patched by Apple. So it's kind of a bad deal to go buy vulnerabilities in that way. And today, you're going to go find a vulnerability like this. I was reading another article where I think actually GLM, actually the OpenSource model, found a vulnerability in WordPress, in the latest version of WordPress. And WordPress is very commonly used for work that almost every enterprise has some WordPress deployment somewhere. There was a vulnerability. Zipridae found that again. And so I think what's going to happen is that the, I think right now, we're actually a little bit early, but I think in the next six to 12 months, there probably will be so many zero days discovered that unless the enterprise is able to effectively invest in cyber defense via sort of this LM, you know, it's almost like if you don't hack yourself, you will get hacked by someone else, basically.

39:28I think that, again, we're in a lull right now, but I think it's almost the calm before the storm. I think very soon the storm will hit and we should all be prepared for that. My best way of preparing for this right now is I think actually old school security still works pretty well. And the way to think about it is the most secure thing you can do is honestly just not be on the public internet. because if you're not on the public internet, it's actually practically impossible to hack you. If the only way to access your systems, for example, is by getting into the VPC, that is, you know, the sort of service area you have is no longer, oh, well, I host a WordPress over here, and I host this thing over here, and I host this thing over here.

40:12Again, you know, the average enterprise at this point probably has 200, 300 different SaaS applications, not more. And all of these SaaS applications, I'm sure, have zero days. And it's actually not their fault. It's just, you know, the fact is that finding zero days is going to become so cheap over the next 12 months. And so instead of saying, well, you know, let's go try to patch every single SaaS application or, you know, let's not use any SaaS whatsoever, it's not going to help you. Because, you know, if you build off yourself, it's awesome to have zero days too, right? You could have even more vulnerabilities problem.

40:39I think the only way to stay safe, really, is to say, well, all sorts of data really has to stay air-gapped if we can. And this is also where Retool's investing quite a bit, talking about the book a little bit, which is we don't believe that when you have people coding these applications, they should be hosted on the public internet. They should not, we believe, even be hosted on something like Eversell or anything like that. We believe it should fully be within your VPC. And if you're able to say that your app is fully within your VPC, even if it has vulnerabilities, the only way they're going to be exploited is by an employee, which is a lot better than being on the public internet where anyone can go exploit you.

41:23Peter High: A great overview. Thank you so much, David, for walking through that. One of the fascinating things about your background is you studied both philosophy and computer science at the University of Oxford. Computer science certainly seems logical given what you do, but I wonder to what extent you're leveraging the philosophy part of your background in the work that you do. Yeah, I feel that actually much of our discussion so far has been somewhat philosophical in nature in the sense that I think philosophy is all about when you change some sort of inputs, how does the output change, if you will? And I think in the case of AI, the great thing is the inputs are changing all the time because AI is moving so quickly.

42:02So many predictions have come and gone and, you know, been true or been false. And it's really exciting, actually, to be able to reason through. Well, you know, if, again, cybersecurity is a great example where I think I really believe we've kind of been at an equilibrium. You know, it's been OK. Attackers have gotten a bit better. Defense has gotten quite a bit better as well. And so overall, you know, we're not too scared, if you will. We are somewhat scared, but largely speaking, most people are not getting hacked most of the time. And that will change dramatically because, again, we've talked about this, but the sort of idea that attacking becomes so cheap now, it's just tokens, basically.

42:44And tokens are getting cheaper and cheaper every day. And you have open source models that are totally willing to go and go attack. and you have the internet, everything's fully connected. And so it is, I think, almost a perfect store, if you will. And thinking about how security will change, what will be the new equilibrium, what will happen in the short term, what will happen in the long term. I think that's so interesting. And I think it's a really great application of philosophy.

43:14Peter High: Is there a philosopher that you particularly admire most or who has been most influential on you or perhaps put a different way that you find yourself rereading these days and a lot of the things you're thinking about in your day job? Yeah. So my favorite book is Gödel, Escher, and Bach. And it's written by this guy named Douglas Hofstadter. And it's a pretty large, pretty long book, I would say. But it's kind of, it actually, that's the reason, actually. I first read it, I think, when I was in high school. And that's what actually got me interested in studying computer science of philosophy together.

43:49It asks a lot of interesting philosophical questions that are perhaps somewhat related to what we've talked about. So one of the questions that it asks is, what is intelligence? And what does it mean to have intelligence? And is it possible to create intelligence out of things that don't have intelligence? And so, for example, in the case of a human, which we're a little bit more familiar with, individually at a neuron level, no neuron seems intelligent. And, you know, if you look at the neuron, what it's doing is very straightforward, actually deterministic even. It fires when sort of the conditions are true.

44:25And so, great, you know, that's a neuron. It doesn't seem very smart. And through, when you wire all these neurons together, somehow they turn into a brain. And the brain, actually, the human brain, seems pretty smart, actually. And so, you know, that's, you know, kind of almost as paradoxical. How is it possible that, you know, from sort of such dumb things can intelligence come? So that's the human version of it. But actually there's an analogy in nature, actually, which is ant colonies. And so ant colonies, actually as a whole, actually display intentionality. So ant colonies will go and explore a forest.

45:03They will actually go and have certain desires. They'll go do things. And the ant colony, if you sort of treat it as one sort of animal, actually is quite intelligent. You know, if you sort of poke it over here, something else will happen over there. You can watch it, be intelligent. When it's hungry, it will go do this. So it's quite intelligent. But the individual ant itself is actually not very intelligent. If you look at the individual ant, you know, remove it from the colony, you know, it crawls around, doesn't really do very much. It's not very smart. It will probably try to find the colony, actually.

45:34But besides that, you know, it doesn't do much, right? And so it's a very similar, it's an analogous thing of where does that intelligence arise from? where does the intentionality arise from? At an even higher level, you could say, you know, if you put groups of humans together, is the organization itself intelligent? Is that more intelligent than an individual human? I'm not sure. And so there's a lot of thinking out there. But yeah, I think especially with LLMs and thinking about, are LLMs intelligent? I think for sure they are. Personally, back in the day when I was studying computer science and philosophy, the commonly accepted test was the turning test was if um a you talked you chat to a chat bot or a computer and the computer and you are not able to tell if the chat bot is human or a computer then you would say it's intelligent and certainly and you know this i suppose you know informs some of my beliefs around intelligence development bottleneck you know lms are far intelligent enough.

46:34But if you talk to an LLM today, I mean, it is extremely intelligent. There is no way, I think, in fact, if you had to ask me, who's more intelligent? You know, the average colleague, and I work with pretty smart colleagues, or the LLM, I would say the LLM, honestly. And if you think about the power of language, language is how we communicate to each other. Once you've mastered language, really, there's not much else that is required for you to say that someone is intelligent, and you can imagine if I were fully disabled, for example, I had no, you know, I could not use any of my senses. All I could do is I could just answer questions.

47:10I could use my mouth and use my ears, let's say, and respond to you. You would still say that I'm intelligent. You know, I don't think you would say, well, you know, David doesn't have arms. He doesn't have legs. Therefore, he's an idiot now. You would ask me a question. I would respond, you know, hopefully intelligently. You say, oh, you know, David is intelligent. And that's kind of where LMs are at. And I think it's, great that LLV's getting even more intelligent, but honestly, who cares? At this point, they are far intelligent enough that they should be having an impact on your business, and they're not.

47:40And so I think all that is really informed by philosophy, which is really cool. So I'm really passionate about philosophy still.

47:47Peter High: Well, thank you for those suggestions. And actually, coincidentally, it's the second time in a month that Gödel, Echel, and Bach has been recommended a 1979 Pulitzer Prize winning book. So thank you for mentioning that. You mentioned earlier, the work of Terence Tao as being particularly important to you as well, or at least that you read his work. So thank you for those suggestions, certainly, David. And David, thank you so much, more generally speaking, for a great conversation covering a wide array of topics of great relevance to the technology community, both those that you're influencing and others that you're thinking about.

48:16Peter High: I appreciate a really thoughtful dialogue today. Yeah, Peter, thanks for having me. Thank you. Thanks for listening to Technovation. If you enjoyed this conversation, subscribe and leave a rating that helps more leaders find our podcast thank you and i look forward to seeing you next time