You Can’t Vibe Code a 100-Ton Truck: Inside Applied Intuition’s Approach to Safety-Critical AI

27:02Yeah, you said two things that kind of mix together, but they're actually quite different, is the interaction with the machine and self-driving. And so if you are in a situation where it's like, you know, what they call an L2++ system. So you're still expected to be in the loop of the drive, which means if the car disengages, you have to take over, i.e. you have to be in the driver's wheel. And that interface is going to be different than you don't have a steering wheel and you don't have pedals. And I do think we're seeing the early kind of rays of light of that future emerging, which is a future where there will be vehicles that won't have steering wheels and pedals.

27:41I That must be more like, you know, practical in like, is that actually the best UI? Like I would imagine you're going to a Dumber. You might want to press a button once in a while just because it's going to be weird to say like, hey, Dumber, drive right. You know, like, so as I'm sure you do like a lot of kind of like evaluation or like what is the most effective kind of like interface? Where does that kind of take you? It's a mix of voice and actual interaction that way you'd consider, you know, traditional like human computer interaction. HCI, I think over time, over a longer period of time, it'll be, I think it'll be almost all voice.

28:18The main reason is the example you just use like turn left or turn right. The vehicle is going to know that already. So it's, it's much more of a context kind of question. Like there, there, there's a faster way or there's a smoother way. I'm fascinated by that. So you think even like in a fighter plane, 30 years from now, they're going to be like, shoot the bad guy. That's very different yeah that's very different we just made a huge context change uh yeah yeah so so in those types of situations i think the human uh because now you're going from uh using the example you just use is like going from using chat gpt to somebody who's writing software those are two different so the interfaces are going to be different um i i don't think they'll merge anytime soon and also in 30 years from now frank is quite unlikely that there will be any people in fighter jets, right?

29:08Yeah. I think that'll be sooner than 30 years. I think we're looking at the last true fighter jet being developed right now. I think afterwards it will... Yeah. It'll be all drones, autonomous aircraft. Total drone. Yeah. My daughter wants to join the Air Force. You're telling me that I might need to redirect her career-wise? Is that what you're saying? I mean, there will be people who are in the Air Force. They just might not be pilots. So I think that as a military branch, it might even become more important than it ever has. And I think Space Force, I think, will become more important than it ever has just because of the nature of conflict.

29:42Yeah. And no matter what, I mean, the Air Force is going to be the primary procurer of drones and the drone warfare systems. Yeah, exactly. So we've been talking about how AI can kind of revolutionize your deployment of AI. Can we talk about your use of AI at Applied? What are the coolest, most interesting ways you've seen your team or yourselves or your team deploy AI to kind of get that 10x human kind of outcome? Yeah, I mean, the most obvious and apparent, of course, is the use of AI coding tools, which are super interesting and valuable. I think it's actually probably the killer enterprise use case for LLMs.

30:23And I see Anthropic is seeing, I think, good success from that and them choosing to focus on that area with Claude, it was a good strategic move because like every software engineer in the world can get a productivity boost from using these models. And it just takes care of a lot of the boilerplate. I think you do end up seeing there is a certain complexity ceiling that you hit with those tools and that ceiling keeps rising, which is great. But when you're working on really hard problems, the LLMs become less useful still. But again, And there's a lot of boilerplate, which the tools are, it's still a productivity boost no matter what.

30:58And just to keep it entertaining, rather than saying what everybody says, the bare case on CodeComplete tools is, you know, you don't have to write a lot of software to know more software doesn't necessarily mean better software, especially in like systems we built, heavily optimized performance systems, because you don't have endless compute available when you're on the vehicle, when it's actually in inference in the real world. Now you've got to pay for those chips. Every line of code has a lifetime maintenance cost. And so more code does mean more maintenance. It's technical debt, right? Technical debt's a real thing.

31:33Do you have to make your own LMMs? I would imagine a lot of the systems that you're using, you have to write your own models, right? Because they have to run on hardware. Oh, yeah. We train a lot of models. The models that we train, though, they are more specific to the vehicle technology aspect. For generic things like coding, the off-the-shelf models are pretty great. And you can do a lot of prompting, prompt engineering to those models to sort of get what you want at the end of the day. Is CodeGen the primary use or are there other cool uses? I mean, I would, here, I'll project for a moment and you guys correct me if I'm wrong.

32:09Radical AI autonomy company in Mountain View, California, where I've lived for 12 plus years myself, by the way. But radical, futuristic company. you're telling me the coolest use cases of AI among your workforce is CodeGen? Well, the reason I say that is because that's sort of like the infinity use case, right? If you can do CodeGen, you can do anything. And so this allows us to build - That's a Sam Altman answer. That's a Sam Altman answer. If we just get it to write better AI, then that solves everything. There's not like performance review lists. There's not like, I mean, you're just Peter, right before Kasser joined, Peter was selling us out.

32:49All of a sudden you're at professionalizing your recruiting process and the brand is out there in a way that wasn't before. There aren't fascinating ways that those teams are getting a huge augmentation. You've got to spill the beans. Yeah. I guess we're like two Detroit guys who are like, I think we lean into being boring, but let's give you some exciting answers. No more bears. No more bears. I want both. Yeah, exactly. By the way, all of our values, we boil them down into two words, which is radical pragmatism. And so we - I was going to ask you about that. I have that in my notes. You see my notes right there?

33:25Oh, nice, nice, nice. I was going to ask you about those words. So the way that we've, any company, we're getting to scale as a company. We're over a thousand engineers. We do hundreds of millions in revenue. And so we're already, let's just say, becoming a stodgy company, for the lack of a better word, right? You're already starting to ossify. And so So as this revolution hits us, the way that we've done it on the non-technical side, now remember 82 % of our company is software engineering. So it is a very technical company. Google is 50-50 to give you some context here. We've told all the commercial departments, legal, design, people operations, et cetera, that they have to first, the first request, gentle request, even Peter and I are gentle leaders.

34:10Our first gentle request was, you know, use AI in your workflows. And you know what happened is, especially experiencing folks, they're like, actually, this thing that we do works pretty damn well. And we've already, you know, made it as efficient as possible. And so our kind of second wave and third wave is doing things like we have, you know, weekly live all hands. And so having departments go up and actually showcase, hey, this is how we're using AI in our function, in the legal function, within the constraints of what a legal team has to do or within, you know, design function, et cetera. that creates an environment where everybody thinks every problem I interact with my knee-jerk reaction should be is there a new tool that solves this rather than an old tool and you do that every week it's your all hands you have one of the teams get up and share how we use AI exactly and so all it is is keeping top of mind and at some point that lift just started happening where the knee-jerk reaction and then the thing that's the most I would say exciting or things that we typically don't talk about, but it's the most actual practical implementation is, well, a long time ago, we started an in-house software team that I personally lead, which is, the shorthand is the software that runs the company.

35:21And the software that runs the company obviously shouldn't be just a web app with org charts. And so making that more and more intelligent, it really is becoming the brain. So we dump, you know, our own, it's becoming like the knowledge center. Like literally last night, the person who runs that product, you know, sent me a voice message and said, hey, I think We're ready to like roll out essentially like a brain version of this based on these inputs. So I think it's still early days. I think if you saw it, it's interesting. I think we keep compounding, keep working on that software. Our enterprise will literally be a more advanced enterprise.

35:56And we can do that because we just use AI a lot in the company. We develop it, we make it. So then we point some of it towards enterprise. Talk about gentle because I want to come back to this. Or push on it, pull on it. I don't know. how do you wrestle with the question of gentle because that's a word you emphasized and you know you got ceos like toby luke you know saying hey you cannot hire unless you can demonstrate that ai can't do the job right you've got fiverr ceo you've got duolingo ceo making very kind of old pronouncements how do you think about leading the they call it organizational transformation like if you look five years in the future you can't imagine everyone not being ai augmented Yeah.

36:37You think it just happens naturally or how do you, how do you gently? I said that, uh, I said that gentle word sarcastically. Yeah. Anybody who works with applied or at applied would ever, you know, I think the most common word they talk about is intense. Uh, but I think it's actually good because people like listeners don't have that context. You're saying you're so clearly not gentle that to say the word gentle is actually a joke. So when you say you gently recommend, I'm just looking back at my notes, right? You gently recommend that all functional leaders use AI in their workflows. Here's where I want to get very specific.

37:12What's the consequence of not doing so? I mean, so like you do a lingo set of examples that you use. I've caught a little bit. Another kind of controversial thing is me and Peter are not on social media. We tend not to like just need to hear what like thing has to really hit a level of certain frequency for, you know, our ears to hear it. My assessment, you know, I worked at YC for many years and I was very deep in the startup ecosystem. Half of that stuff is entertainment. So let's just leave the entertainment and bombasticness to the side. The practical reality of leading an organization is if somebody's a great head of design or somebody's a phenomenal general counsel, you're not going to fire them because they didn't try some startup that says they can read contracts faster.

37:56So I want to have that like direct conversation of what are the actual limitations of these products that are out in the market. Now, the leader should be able to articulate that. If they shrug and say, don't know boss, then we have bigger problems. So I think it's not these simple use this or get fired. Also, our company, I mean, it's no disrespect to a Duolingo. Our company is an order of magnitude, probably more complex, if not multiple order of magnitude. We work on AI safety systems that can deploy to machines that humans use. It is the most complex. We are as technical as a technical company can be.

38:33Like there's the, we are at that level of an anthropic or an open AI in terms of raw technical. Double digital. Percentage of our company have technical PhDs. Majority of the company has graduate degrees. So the caliber is very high. And then because of that, in each of these fields, there's just a lot of adoption. An example is like in IT, you can automate 80 % of the IT tickets that come in for basic stuff. in project management, you can actually get a better organizational structure of deliverables using that compensation analysis, right? We can use that to figure out if there are outliers in the leveling compensation scheme.

39:08So there's all of these things that sort of just fall into place since we have the right leaders in place. Yeah. As other leaders who are listening to this, I think the question isn't, you can use AI adoption as a proxy for how good your leads are, but it should not be the sole like decision maker, right? Because then you're also going to get, you know, your organization gets to any level of size, 50 people, 100 people, your employee, the human nature, they're going to start filtering and saying things to you in certain ways. And so if you do something simplistic of use AI or get fired, guess what?

39:44You're going to have a crappy lawyer who's going to say the right things to you. And suddenly they're the GC because they're quote unquote adopting AI. and the actual lawyer who's really good at their job, who hasn't, who gets fired, and then you put your company actually at greater risk. Can I go back on the human side? You know, obviously, a lot of the stuff that everybody's building, including you, is stuff that at the end of it is in service of a human, right? Like we built autonomous drive so that they can drive us around. And then we build something that probably works and then lawmakers change the law.

40:18And so like the software update comes in and suddenly it doesn't drive automatically because, you know, like there was a bug somewhere that killed the person, which is understandable. But it would suggest to me that a lot of the stuff that you do is actually solving the technical problem. But I would imagine that you guys have as much kind of challenges understanding the human problem. Why is it that the person sitting in the big number doesn't do X and all those things? How do you actually like, from an organization point of view, start to internalize that into an organization that is so technical?

40:52Like how does the human understanding and the technical understanding overlap? I mean, what's also different about our company being an enterprise company is for all of these technologies, we're partnering with the manufacturer who makes these machines and knows these marks well. We don't know anything about mining. We don't know anything about commercial trucking. But when you work with the literally top commercial trucking companies on the planet, As an example, commercial trucking, the real problem as much as in the cabin is actually they're small businesses. A lot of commercial trucks are owned by a LLC that owns maybe one to three trucks.

41:23And so the software and AI problem, as much as in cabin and it's the fleet management stuff and it's the maintenance. And for a lot of people, that's their biggest, most expensive asset. As a family, that truck or those two trucks. You get them to work for you. Yeah, exactly. And so our partners are the ones who say, point your AI technology towards this set of problems. And then we put those set of problems. And it's all under the vehicle intelligence umbrella. But that's how we have been able to work in lots and lots of different fields. It's not because we know all these fields really well.

41:54We actually are just fundamentally an AI company. But instead of our AI focus being LLMs, it's within vehicles. And then we partner with manufacturers. Can I ask an innovation kind of level question? Henrik and I both have kind of a shared passion and curiosity and morbid interest, you could say, in the challenges of organizations in innovating. I'd be curious here, you mentioned earlier that you studied at the GM Institute. Why is it applied integration coming out of GM? Can you talk for a second about the challenges of innovation and your observations? Yeah, that's interesting when you say that.

42:30So, I mean, in some ways it did. I'm an alumni of General Motors as well, not only the GMI, but also the company. So in some ways it is the case. Even earlier, we were talking about all these self-driving companies that ultimately didn't make it. You know, where they do make it is a lot of those alums who learned all that are in other companies and they're pushing Tesla and Waymo and Applied Intuition to production. So I think it's incorrect to say like that suddenly that, you know, if somebody works at General Motors and General Motors doesn't create an AI that somehow General Motors didn't contribute to it.

43:02It did, but not in a direct way. That's kind of an accomplished model. It's not affecting, say, their market cap. Exactly. So that question is actually so a separate question, which is you could actually have the talent inside the company, but why can't the company almost extract that talent? And that's actually not a General Motors or a manufacturer problem. That is a large organization problem because a better example than General Motors is Google. Google invented this technology, OpenAI is a little group of people who've actually monetized it. Why did that happen? Because that's actually a way more complex thing because it's literally Google's job to make that technology and make it great.

43:39And they didn't. And so I think it's really, if there's something about the reason Silicon Valley is successful is there's something about small teams that are incentivized through equity that just do really well. And it's been 75 years of proving that case again and again. But when I was at YC, the question always happened, would there ever be an era where there won't be startups? No. I actually think that they're more likely to be an era that there won't be really large tech companies before than there would be an era that would be just because the incentive structure. And the reality of is, if when you work as a group of five people, why do five people don't, they don't need to do sync meetings and they don't need to do like, you know, Monday morning, what happened last week?

44:19Because you're just literally, when we were in the living room with five of us, our all hands would be, we would twirl the chairs and we'd say, okay, let's talk about what's going on. And often those all hands were so in line with each other, there isn't much to talk about. Yeah. And so then we'd be like, is there anything we haven't discussed? So now the question is, why doesn't it happen in a large company? It's because it's the structure of communication. Once you start adding layers of managers and you have to have a central control. So the hypothesis would be if you had tons of small groups that were loosely affiliated, that could be in a corporate a better enterprise design.

44:58Yeah. Enterprise. Now there's companies have tried this. I think Steam had this one version of like a managerless office and pods and stuff. We tried actually without an org chart when the company was, you know, sub 75 people, extremely difficult. There's a reason these org charts. It's kind of like you look at human societies around the globe. The concept of a nuclear family with parents educating their children, school systems and hospitals, a huge amount of cultures. And they all look shockingly similar, right? They all have a job. rarely in societies today, you don't have some version of an occupation that you specialize in.

45:32You don't just stay at home. So why does it happen? It's because somehow the environment gets created that. So I think the trillion dollar question is, how can you make large organizations operate like small organizations? And that's like, how do you make a tall person operate like a small person? Or how do you make a, you know, it's almost paradoxical because the incentives are not correct and the communication structures are not ideal for that reality. As a fellow kind of student of this topic, if you're curious, one book I'd recommend is Safi Bakal's Loonshots. I don't know if you've read that, but he kind of talks about this.

46:03I thought you were about to pitch my old book, Jeremy, The Acorn Method, which makes this exact point. The podcast would not be complete without us sharing at least one book that we've written. Safi Bakal's book, Loonshots, is quite interesting. And one thing that he gets to, Kasser, which is just to your point is, I think it's Dunbar's equation. Yeah. But it's this idea of about 150 people is kind of the max before things start to break down. Anyway, you're reminding me of that as you were talking. Yeah. One way, by the way, Google, we're both ex-Googlers and a huge part of the company is from Google.

46:40At some point, when the company was a few hundred people, literally still, I think up to 400 people, you know, we're still a majority Googler. So we have a lot of Google people in the company. and so you know we know the company quite well if google had really in the era of 20 let's say 10 as the company was now at scale making like a billion in cash flow a month what the company did at that time was they started the moonshots you know ultimately were waybo and x right yeah that came out of there and then alphabet kind of emerged another five years later if you could go back in time and let's say google instead did something different they believed a large organizations are not ever going to be effective.

47:17And we're basically going to be like a holding company venture capital fund. If you look at YouTube, YouTube is a great case study. It was left alone and it grew into this thing. If YouTube had become Google video, YouTube would not be what YouTube is today. And I think everybody, you don't have to, you know, you don't have to have a PhD in tech strategy in Silicon Valley to get that. And so if Google could have replicated that across hundreds of companies, there's a version that Google is actually just a, almost like a guild of companies. And so one of the reasons that large organizations are more efficient is they're sharing learnings.

47:50You share recruiting learnings across, we have 30 some products, all the products learn when one product learns something about marketing, all the products benefit from that. So if you have this guild of an organization where you share this information, incentive to share, maybe that could work, but that's not our business. Our business is vehicle intelligence. So I think that might be, I know everybody has a hot stuff, so I will be the downer that kind of like throws that in there. Cause a funny way of saying bored. But yeah, let's, I mean, I just couldn't keep my eyes open anymore. So it was that, or we often in these conversations, like the most interesting point when everybody's like, finally, you know, we're like, Oh, join us again for another interesting, it's a, it's, it's Seinfeld, right?

48:33You got to leave on a high note. That's awesome. you guys are amazing thank you Mr. Oddly is Nerdistalk about something that wasn't kind of like another thing that came up like comes up in a text prompt right this is something that is different you know I mean I thought it was hysterical the number of times we talk about human tank interaction it's hysterical you know one thing I find myself kind of wondering about you know as Kasser painted the bear case so to speak is just really You know, I think the question of where will the average shake out? It's hard for me. And maybe I'm optimistic or Pollyanna about this because becoming a good collaborator to AI is so accessible.

49:18It requires no advanced degree. It requires minimal advanced training. It just requires some intention and maybe a little bit of practice. But everyone can be a top 1 % collaborator or get top 1 % outputs out of AI if they give it a little bit of attention. And so the bear case is saying that it's effectively for most people just going to remain information retrieval. That makes me sad to think about because it just means the vast majority of people will hardly scratch the surface of possibility, hardly scratch the surface of capability. That's something I'm thinking about. I don't know if you have any thoughts on that from where you sit.

50:01No, I mean, I do think that he just posted this as a fair case, right? You know, like, and where my head kind of went a lot was, I think when we talk about specifically generative AI, a lot of us are just very anchored in a congenital image and text and code. And I think as I was kind of reading up for this conversation, I was trying to understand a little bit more like how is the ai being used in all the places that we don't normally talk about and i stumble into this case which i'm not sure is like how real was but basically one of the issues of creating fusion energy like unlimited energy is to keep like the plasma stable and people are now using ai to really help you know i don't know how that actually works but like you know keep the plasma stable so that when they have these like two particles kind of colliding they can have like almost like it's not a force field but like the force field stable and so one of the kind of like unsung benefits of ai is that it now kind of like work as a participating actor in some of these things that could obviously revolutionize the whole world because if we have unlimited pollution free energy that would be wild and so what i think this conversation really kind of made me think of is like ah there's probably this whole universe outside the stuff that we normally talk about, like having kind of like autonomous costs, being like just that first step removed.

51:31But as we keep that step, like removing ourself from the coal of just generating text code and images and sound, then yeah, like my head just normally don't go there. And so it was nice to kind of like explore that. Yeah, I think it was wild to me. I mean, it just shows kind of what maybe sheltered lives we live. I mean, I work with Komatsu as an example, so I'm familiar with them as a business, certainly. And yet the thought that there are machines running 24 seven in cities in the middle of nowhere that for decades and people who are spending months of their lives at a time there. That is, it's such a different, and where visibility is limited because of dust, where humans are at risk because tires alone are 12 feet in diameter.

52:28That is such a different world from our day to day that, that it's in some way, it's really exciting and invigorating to realize there are people like Kasser and Peter who have dedicated themselves to improving the safety. And you could say the humaneness of those conditions. I mean, it's, it's really cool. It's like, it was in a conversation that made me feel more optimistic actually about the impact of technology. The other thing that I was thinking about is, I mean, like I like to hate in San Francisco because I don't live there. Um, but it is, it is incredible. You always hate on it publicly, but folks privately, Henrik's always texting me going, you're so lucky that you live so close to But it is incredible, right?

53:12Like you have people like these two folks that, you know, clearly are very smart and then very ambitious, right? They, you know, I had in full transparency, never heard about their company before. And one of my friends joined and then we heard about them. And it's like a, I think,$15 billion company. They do incredible work. They have very big ambitions. They're doing something that's very complicated. And, you know, you have to just admire kind of like the ambition level and the ability to think kind of just really big of a large group of people, specifically kind of like on the West Coast of the U.S.

53:50And so, yeah, like I was there's another one where I want to hate on it, but it's actually pretty impressive. You know, speaking of large, you know, the comments, as you know, and probably is the case for you, the end of the conversation is always my favorite part. We were talking about what does it look like to enable innovation in an organization? Kester's comments on the size of the organization matter. The fact that, I mean, they're a startup, right? I mean, they've raised, whatever, 500 million bucks. They're not nothing. But they've got 1 ,000 people. And he said, quote, we are already stodgy.

54:20Yeah, I think anybody who's got a 10 ,000 person or 100 ,000 person organization goes, wow, what I wouldn't give to only have a thousand. And yet he's noticing how many of their functions are starting to be, what did he say, sclerotic, right? And just calcified. And to me, it's a really interesting, I mean, they are still able to do something like, which I think is a great technique at the weekly all hands, do a showcase of how folks are working with AI to keep it top of mind. But it is an interesting challenge. what is the ideal organizational configuration to continue to invent the future and to future-proof yourself.

54:55I liked his kind of alternative realities, kind of black mirror version of Google, which is a guild of small teams. How many more YouTubes might there be? I don't know. What would the cost to Google search be? I don't know, but it's an interesting kind of alternate reality to kind of entertain. I think that's actually a perfect note to end on. Can we have the secret code word be Alternate reality or black mirror, I think is a perfect code word to end on. It's actually two words. Just pointing it out. Code words to end on. It's two sets of two words. It's deeply confusing. If you're confused and you only want one code word, just say confused.

55:36But whatever you do, please like and subscribe and share with the friends and put it on LinkedIn. And email us if you have any good ideas or questions. Tell us who we should interview next. Yes. And with that, bye.