Can a machine teach itself to think like a human?

Speaker A: You have to create a sense of safety and a sense of belonging within each individual, within the company, and then give them the accountability and ownership and autonomy over their role. And I've seen that's where the critical thinking really thrives. Innovative ideas emerge in those types of environments.

Speaker B: Uh, Hey everybody, welcome back to Founder Vision. Today I'm speaking with Sravan Puttagunta. He is with Hyperspec AI. And uh, tell us a little bit about what you do, tell us a little bit about what's your role there and what does Hyperspec do?

Speaker A: Hey Brett, thanks for having me here. My name is Shravan and I am the CEO at Hyperspec and I was the founding member for the startup and I'm flanked by two amazing co founders. Problem we're solving is we're enabling machines to have spatial reasoning and spatial context without relying on predefined maps that are built by humans. Typically in robotic systems you need these predefined maps that describe the environment to the robot before the robot is able to interact with the physical world. So we're trying to mimic human behavior and enable robotics to think critically, see and interact with the world the way humans do.

Speaker B: That sounds like there's an interesting mapping onto the founder's journey there of, ah, exploring a space that is unmapped and mapping it as you go. How does that apply in. How do you, in a very broad technical sense, bring your experience as a human into programming this into robotics?

Speaker A: Yeah. So for example, if I were to tell you, hey Brett, I know you asked for directions to the grocery store, it's four blocks down and it's the uh, right turn and the second traffic signal. You take a left turn and the parking lot for the grocery store will be on your right side. That is all the information most humans would need to get from where they are to that grocery store. But for a car, we need to describe all of the geometry that it might encounter and all of the traffic signals and the types of infrastructure and the semantics of how to interact with that situation, uh, is all currently stored in a map. So what we're doing is we're enabling the car to process the sensor data in real time. Using machine learning. We're able to create that context on the fly instead of like loading the data from a database. And we're able to use that context and share it with the planning and actuation part of the self driving stack to enable the car to, to just, to complete the journey with the same level of detail that it gave you. As a human, which is drive four blocks down, take a right turn, take a left on your second traffic signal and the grocery store will be on your right side. So we're reducing the requirements on the level of detail that you need to provide to the car with respect to geometry, with respect to all of the context that you might encounter from where you're starting to where you're going.

Speaker B: Yeah, that's fascinating. So what uh, what brought you to, to found this company? What, what put you into the position where this was what you wanted to do and believed yourself to be in the best position to approach?

Speaker A: Sure. Immediately after college, I've always focused on computer vision and solving really hard problems in computer vision. Uh, my experience at uh, Samba TV was developing video fingerprinting so the TV can identify the shows that are being watched and give analytics back to the broadcasters so they have feedback on what type of content is resonating well with their users and they can use that to adjust the types of shows they double down on Then I was the founder uh, of Civil Maps, which was a startup focused on 3D mapping and building map infrastructure for both heavy industries. So think about like trains, PG&E, power transmission, construction, mining industries, they all use maps. But as we started building that out we also had automotive companies approach us and say hey, can we leverage your technology for self driving vehicle? And that's where I uh, did a deep dive into the automotive space, understood exactly how the self driving stack in a car would use these maps. And so my journey at Civil Maps sort of gave me exposure to how maps actually help the self driving stack make decisions and enable the car to get from A to B. Uh, but the problem is those maps are only deployed on 3% of US roads, uh, mostly highways. So if you for example, or uh, small cities like Phoenix or San uh, Francisco and Mountain View. But if you take like a Waymo car, let's say from San Francisco, and place it in Berkeley, uh, it wouldn't be able to drive itself because there's no HD map that exists for Berkeley and that applies for any self driving program that doesn't use real time maps. So the concept of being able to create maps on the fly was something that I started pursuing in 2019 and I started looking at various methods to do it and I started looking at the computational budget that you have on a car and that's been growing exponentially with, with neural network accelerators and GPUs and various types of research that's happening in companies like Nvidia or Qualcomm et cetera. They're all building these self driving compute stacks that are uh, basically supercomputers that will be in your car. And so there's a world out there now where the computing budget that you have available to run your algorithms is equivalent to a small data center. And so we can start building these very complex machine vision algorithms that enable the car to see and think critically like a human can. So that's when I started really aggressively pursuing real time maps.

Speaker B: M that's interesting. There's almost an oxymoron in that phrase, real time maps. Because the moment you have a map, it's no longer real time. It's actually like the purpose of it is to have a model of the world. And the moment you're using a model and not the direct world, you're starting to separate from it and things can change out from under you. And so there's this, this idea that it's like a, it's like a map that's approaching real time but also still develop, like still being able to be used as a model for the world. And so that's something that's interesting to me.

Speaker A: Yeah, so think about how humans navigate the space. Like when you get up from your desk and walk to the kitchen table, you're probably seeing all the objects in real time and you're classifying those objects as a door and as a hallway, but you're not necessarily memorizing the exact geometry of the door, but roughly where the, where to expect the door. And you have a sense of orientation. Right. So the map that you're building, there's two layers to it. One is tactical, which is, okay, I see what's in front of me, I'm approaching this object, I'm going to interact with it. And so that's like very low latency. And you're able to interact with that object through, through a direct feedback loop. Then you have a strategic map which is like, I know the kitchen is roughly a hundred feet that way and I need to exit the store and take a right. So that is what's called strategic sort of instructions. So if you divide the self driving stack in the same way, the strategic map would be like the navigation map that you use today, like Google Maps or Apple Maps or Here Maps or Waze. Those all give you turn by turn instructions on how to get from A to B. But they're meant for humans to consume. A tactical decision making of how you actually drive on the street is what we're doing. And that's the capability that we're creating through this thing called real time maps. But if you were to like distill it down to its basic definition, it's the ability to do spatial reasoning. So you see something, you're able to contextualize it and you're able to reason with it and then you build context from that and then that's how you interact with the environment is on the basis of this context.

Speaker B: Yeah. I'm curious, to what extent in the spatial reasoning do you, do you kind of collapse the space into concepts? And for example, if I'm walking from my bedroom out to the kitchen, there's a door and there's a thing that it means to be a door. And I understand that there's certain functions of doors, they tend to be on hinges and they can swing open and close. And if there's one that's closed in front of me that might be locked and there's like certain properties and characteristics that I might expect. And that's a very conceptual system where there's like a, a class and its characteristics and its properties, subclasses and all of that. And I'm curious to what extent that's true in what you're doing. Or if it's just pure spatial reasoning where it's like I'm detecting a solid object and that object is of a shape that maybe a round object near a road is something that might be expected to move in front of the road and be chased by a child shaped object. Uh, uh, how far does the compression go into categories in this software?

Speaker A: I think you hit on something really interesting because in uh, biological systems spatial reasoning has two categories. One is your subconscious and one is your conscious. So the first part that you described, which is the ability to classify that this is a door, it has hinges and this is the way it opens is typically in the conscious layer where you're able to recognize and reason with things that are higher level. There's a subconscious sort of spatial reasoning which says there's an object in my way and I can't get past that object. And that's sort of on autopilot. We don't have to spend many cognitive cycles determining what the strategy is for that. So if there's a object rapidly approaching you, you move out of the way, even maybe before you recognize what the object is. And so this is basically human instinct. Right. So if you draw parallels to vehicles, we can detect blobs and in a 3D scene and track how these blobs move through the 3D scene and then predict where they'll be in the future. That's kind of the subconscious aspect of the car. And it can do its motion planning to get around these blobs. But then there's also the semantic aspect, which describes the context, which is, okay, I'm in the leftmost lane and the lane next adjacent to me has a broken white line in between. That means I can merge over. I also know that there's five traffic signals at the intersection, but the one I care about is on the leftmost side. So that's what we call connectors and relationships. So geometry describes like the most basic aspect of spatial reasoning, which is you can describe things with the line, a point or a polygon, Whereas connectors and relationships describe that higher level abstract information. And so that information we actually are able to observe through how other vehicles interact with the traffic signal change. If I draw a parallel back to humans, the first time you learn how to drive is not when you took the DMV test. It's when you observed your parents drive. You observed that they go forward when maybe the traffic signal turns green. And you created these correlations and a mental model on how to interpret that information. In a self driving stack. The same thing can happen. We can observe how other vehicles interact with traffic signals and then create correlations between the leftmost signal actuating all the cars on the left most lane to go forward. That type of passive observation, unsupervised machine learning can help you build that context passively, even if the car is being driven by a human. And then over time you push an over the air update to enable the car to do it by itself.

Speaker B: Yeah, that's interesting about how you describe the way that we learn driving is by seeing it modeled by our parents or by whoever else is driving. And one of the things that I think that we pick up on a lot that's really important, that can't be. There's just too much of this information is too diffused to be taught in a driving course. Is all the context aware stuff where an obvious one is, oh, my parents tend to go slower when there's cops around or they sometimes enter a construction site and the rules break down for a moment and have to drive over the double yellow line to go around some cones. And like there's certain broader context clues that mean that you can break certain basic rules. And this makes me think of a number of years ago, I saw a meme that it was like an autonomous car trap and it had a car inside a basically a circular yellow line. With one solid line on the inside and then dashed lines on the outside, which would make a car obviously think that it could enter this circle but never leave. And so I'm thinking about that in terms of how do you, how do you handle it when you have uh, some kind of construction going on and you're entering a highway and there's a solid line, like two solid lines that are just converging on each other and two lanes that would be converging on each other, but one of them would then be going off of a berm and the other one would be going onto the highway correctly. And how does the software learn the broader context within which it can learn to break various rules? Because it's following a broader principle.

Speaker A: So our initial integration will be with cars that are being driven by humans. And the role of the system is to provide driver assistance and improve the overall safety for the driver. As the car autonomous system is entering these use cases, typically the user will disengage because they're not happy with how it's performing. So what happens is you take uh, uh, a section of sensor data that's been collected in the car before and after that disengagement and that becomes a use case that you have to solve for in the backend infrastructure. So the data is offloaded from the car into the cloud and that's a use case that the models are retrained against. Um, and so that feedback loop is essentially reinforcement learning where the user's disengagement is giving feedback to the self driving system on where it's not doing well. And to be humble here, I don't have all the answers because we're pretty nascent company. There's a lot more we need to do in terms of increasing the test coverage and uh, solving for all the different edge cases and corner cases. But I think the market we're going after is the ADAS market where cars will still be personally owned. And the benefit of this system is that it takes a lot of the workload away from the driver, but they're ultimately still monitoring the car and driving the car.

Speaker B: So tell me what adas, what was that uh, acronym you just described?

Speaker A: ADAS stands for Advanced Driver Assistance System. It's meant for like, think about like dynamic cruise control or lane assist or automated braking.

Speaker B: Got it.

Speaker A: These different safety features that uh, someone's feeling drowsy, it can keep the car within the lane and automatically brake if the car, there's a vehicle stopped in front of the car. Uh, these are safety features that we can introduce. So those features are mainly limited to highway driving. And so what we're trying to do is expand it from the highway to the driveway where you can actually use uh, our system in non highway conditions where you have intersections or four way stop signs and it's able to do the motion planning and navigate more complex environments. But the system is still meant to be monitored by humans and it's meant to assist human, not necessarily replace the human driver. And as we have more comprehensive test coverage, might be three or five years from now, I think we'll be in a better place to deploy a full self driving system.

Speaker B: Yeah, something that's interesting about how you've described that it learns from these interrupt moments and that seems similar to the way we learn too. It's when we have something that doesn't match our expected reality that our monoamines are released in our brain and we turn up the learning rate temporarily for that window and learn the new pattern. And if you're, if you're training cars on the interrupts of human drivers everywhere, then on some level you're initially going to be learning the statistical errors that humans tend to make as well. And then maybe there's a second step there where you're correlating that with the incidence of accidents and then you can use that as a error signal. So that's interesting to me. There's a way. Yeah, go ahead.

Speaker A: Yeah. One other way to sort of think about that is like a safety bubble around the car. And whenever the safety bubble gets compromised, whether it's a human driver or autonomous system, the car's operating envelope should shrink or expand. Going back to your previous example where you mentioned that your parents slow down if they see a cop car, there's all these different variations of how external factors can affect your sense of safety. And that is the main feedback to whether you drive more aggressively or more conservatively or in a neutral setting. Right. So the safety envelope is something that can be learned in an unsupervised and it can be trained based on these disengagements or some, some form of reinforcement learning that we introduce based, based on the driver monitoring system. And essentially that allows the vehicle to analyze if the driver is feeling anxious or if they're feeling relaxed. And the system would then improve itself over time. That, uh, safety bubble is also can be used in the future to expand how much freedom the self driving stack has to explore, uh, undefined conditions. So as long as it feels safe, maybe it's okay to, as you mentioned, slightly break the law or if you need to creep forward to get better visibility or cross a double yellow line in order to get past that infinite loop situation, you need some sort of flexibility to handle those situations that are would otherwise require human intervention.

Speaker B: So there's an interesting thread that comes up in my head now and it comes from this, that phrase of like that I can learn in an unsupervised way. And it does seem that there's, there's a certain value to being able to trust unsupervised learning to handle more and more of the, in any kind of autonomous system and also also individually within ourselves. And this is something where it kind of ties into my work as a, as a coach, which is to find the ways that people are using supervised learning in their consciousness when they actually would benefit more from relaxing those constraints and trusting their, the, their subconscious to unsupervisedly learn more of learn and map their environment without having to constrain it through what could be labeled and what you could use as definable error correction data. And I'm m curious for you, have you seen some kind of a trend towards from the early versions of the system or the early versions of self driving being primarily maybe symbolic approach or supervised learning to more and more subsystems being relaxed into an unsupervised type of learning and finding that to be the source of a lot of the power and benefit?

Speaker A: Yeah, if you, you can directly draw parallels between what you said to HD maps versus real time maps. So HD maps are pre computed and all of the different paths that you could potentially take are pre drawn. And so you could think about it like a virtual train. So if you were to build train tracks on the roadways, digital train tracks throughout the entire US roadway system, then the car can only follow the tracks and it's not able to sort of change tracks at will. So that's a very rigid approach and that's how autonomy was developed back in the day. Mostly because computational resources were not there. And so to compute in a, in a low latency environment where the car can have spatial reasoning, they, they, they, they had to rely on these pre coded maps to just inherit the spatial reasoning. And if something in the environment changes or if a lane gets closed, then the car just gives up and gives control back to the user. Not only that, if the localization um, has an error where the map and the perception data is not perfectly aligned, that also led to disengagements. So now as we're going into real time mapping the same sensor data is providing you the perception output and the map output. And so localization is no longer a concern. The environment changing is no longer concerned because the sensor data is capturing the information in real time. So as long as you can parse that information, make sense of it, you're able to get the context. But there still will be situations where you don't have perfect understanding of the environment and the car still has to decide whether or not it feels safe and whether or not it should disengage. Just because uh, it doesn't recognize a certain object doesn't mean it should just give up. Right. So this is the critical problem that a lot of people have today is there's no tertiary system that's monitoring the sense of safety, which is agnostic to context. And so that's where the safety envelope and the real time maps overlap. Because if you have the ability to get your sense of safety, then you can throttle the freedom that the Dynamic Motion Planner has which is consuming this real time map data. And if you can balance those two things really well and still have ah, a system that has a high degree of safety, then you can deploy an unsupervised machine learning solution and sort of deploy the system on all roadways without having to worry about the rigidity and understanding every square inch of the US roadway system and making a digital twin.

Speaker B: Yeah, once again that's just fascinating how that relates. And I'm sure you've noticed that I have a hard time not relating anything about AI to the self, to consciousness, just because it's just so fascinating to me how it just is the natural course of things that uh, the AI that we build tends to trend towards something that's kind of like the way that we operate and we learn so much about one from the other. And one thing that you said about this like safety bubble is again, in coaching or in life, if we were constantly determining what our level of threat is and what our safety level is. And the safer we are, the bigger our bubble is, the more exploratory we are, the more we allow unsupervised learning to take over unsupervised processes. Let our like, trust our subconscious. And if we find ourselves in a uh, like a long term autonomic freeze state or a fight state or flight state, then we can spend years in a, uh, in a place where we're just cutting off that unsupervised learning. And then we've reinforced the belief that we have to understand literally everything around us in order to be safe. And then that becomes a, you know, a big cycle hog, essentially a computational hog. And it ends up making us not fully present and aware in our life. I'm curious for you how much being that you're swimming in these AI questions and you're seeing the world through the windshield of an autonomous car, how has that interplayed with your awareness of your own, the way that you operate as a human? And how has that fed the understanding of AI for you?

Speaker A: Yeah, definitely. Being CEO, the other hat I have to wear is understanding human psychology and understanding what motivates my team and my own personal motivations. And you're right. You're absolutely right. It's a sense of safety. So if you have to create a sense of safety and a sense of belonging within each individual, within the company, and then give them the accountability and ownership over, an autonomy over, over their role, and I've seen that's where the critical thinking really thrives. Innovative ideas emerge in those types of environments. The more rigidity we introduce. Yes, we might get a short term burst in performance, but I've seen long term team cadence take a hit in those types of situations because, um, the individual is no longer exercising their critical thinking, but rather being told what to do. So I think company culture can draw a lot of parallels to how scalable AI is being built. And I think the reinforcement learning that happens in terms of personal growth has a lot of parallels to how a self driving car learns how to drive on the road as well.

Speaker B: Yeah, fascinating. So I want to talk a little bit more about your journey, about your journey, founding the company and getting to this, getting to what you've learned about being a CEO and about how all these lessons and how all this applies to running an organization. And um, tell me a little bit about how you explored starting a company and founding a company and building a real time map of the space as you entered it.

Speaker A: Yeah. So I think my first company was Civil Maps and I want to briefly describe my personal journey there before jumping into hyperspecial. So at Silver Maps, I was a first time founder. Prior to Civil Maps, I was the early employee at Samba tv. So I was observing other people start their company and I was the employee there. But at Silver Maps I was the founding CEO. So that was my first time actually going through that journey. And it was actually, I didn't have the approval of my parents. They were like, hey Shrevan, why don't you just get a, uh, job, you're pretty qualified and leave all of this hard, hard work to someone else who's crazy enough to do it, but I had to sort of the, I think the first challenge most entrepreneurs face is getting friends and family on board and creating that support network. And part of it strategy that worked for me was to get them invested into the vision, get them invested into my journey and, and get them involved and that really helped a get the approval. The, the second thing I would say that was really important was to separate my ego from, from the business. So understanding market patterns, uh, understanding maybe that my initial idea was not the correct idea for what the market wants. And there could be multiple vectors like timing or I'm going after the wrong market segment or I'm going after the wrong value proposition. So allowing the idea to evolve and the quicker you can do that, the faster you can sync with what the market expects. So I would say most of the civil Maps journey was focused on building those skills and assembling a team and fundraising as well. But fundraising in 2015 is much easier than in 2019 I would say because the market matured, the hype cycle is gone and you have to really show traction and take a merit based approach towards fundraising which I actually agree with in terms of how the funding cycle should operate for companies. At Hyperspec, I would say the, the biggest growth I've had is being a first time founder at somaps. I was very protective of the startup and making sure things worked as according to plan. But I think at Hyperspec I've learned to let go a little bit. Which goes back to your theme of uh, unsupervised machine learning is letting people explore the space that they've been allocated and letting people grow without intervention as much. And that's a skill that I picked up as a second time founder where I'm able to maybe like if you think about a bowling alley, act as guardrails when people are learning to bowl but not necessarily holding their hand and showing them how to do it, which has been really beneficial. I've learned that I'm less of a bottleneck in most cases whether it comes to sales or fundraising or engineering. My team is able to share the burden and, and sort of buy into the vision and have ownership over the future of the company. And I think that's been really helpful in uh, starting Hyperspec and growing hyperspec.

Speaker B: Yeah. Something about the theme of this conversation then what you just described for your journey, going from the first company to Hyperspec reminds me of something that somebody once told me about raising kids where they said your job of Raising kids is to, to selectively unfilter the world for them. And that speaks to, there's. You start with these like, very goal oriented. These are the things that are important and these are the, these are the, the bumpers if you're bowling so you don't end up in the gutter and make an unrecoverable error, fall off the table and die as a one year old. Things like that touch the hot burner and then just over time you, ideally you're trusting both your children and your employees and yourself and your AI, your autonomous cars to take on more and more of its own unsupervised. Just trust it to figure out what's going on with the, with the uh, filters and the guide guardrails being slowly relaxed over time.

Speaker A: Yeah, I think that's a really important theme in entrepreneurship and it might boil up to the surface in different ways for different people. I think to build a scalable team, having people that are able to take on the lessons that you've imparted onto them and they can carry it forward to people that they hire, work with is really important as well. So, so having Buy in into the culture, it's possible through like the hiring process, through the mentorship process that you have within the team. And then it goes back to the original point about culture of safety and culture of belonging. If that's the common thread that everyone builds up from, then it's sort of the role of the, as a founder to cultivate that and cherish that, uh, and sort of encourage people to champion that within the organization. And as the company scales, that will become 30 to 40% of my time will be spent doing that and that'll have greater dividends than any individual performance that I would have as, uh, a person doing specific things within the company.

Speaker B: Something I'm also noticing here is that in describing the decision to start your first company, you used the word approval as something that you were trying to get from your friends and family. And later on, what you describe in Hyperspec is buy in. And I'm curious if there's a gradient there over your journey from seeking approval to wanting but not needing buy in and how that's. What's the difference between those concepts for you and how has that shifted for you?

Speaker A: Sure. I think getting approval was more about me validating to myself that I'm capable of starting a company and fundraising and getting customers and building out a team. I think getting Buy in is uh, making that process sort of repeatable. So having done it Before I kind of know what works, what doesn't work. So I want to cultivate culture. I want to have people contribute to culture because the culture is defined by the set of beliefs that a group of people holds. And so if, uh, if it's a different set of people in your second company, you're not going to reproduce the culture that you have in your first because the beliefs and behaviors that people have will be different because it's a different set of people. So buy in is essentially relates to not my vision of culture, but rather what the group is committing to, like the social contract that we're creating. And I think that's one of the misconceptions that I had about culture. It's not top down, it's, it's actually bottoms up.

Speaker B: And by definition it is the groups. It is the group's culture. Yeah, yeah.

Speaker A: And the second concept that I learned about is, I would say culture is highly dynamic and volatile, uh, up to the first 25 people. And then I think it becomes sort of stable because there's a gravity that forms within the organization. So each new individual that's joining the company will gravitate towards the company culture versus, uh, and the individuals that joined before the first 21st, ah, the first 25 set of people will have a disproportionate effect on culture. So hiring the right first 25 people can set you up for success. And so I think that is one of the big lessons that I've learned as an entrepreneur is in the hiring process. Emphasize culture heavily in the beginning and continue to do so, but know that you have some center of gravity around your company's culture. As more people join, it starts to stabilize.

Speaker B: Yeah, there's an interesting feedback cycle there. Even paradox where culture is defined as the behaviors and agreements and aggregate. Okay, I'll use the word culture again in the definition of the group. And yet it's still in a company. The culture takes the shape of the founder or the leader simply because of selection bias. You end up hiring the people that share the same values as you. And you aren't, aren't able to work with certain people with different kinds of values. There's certain behavior that you allow to fly and certain behavior that you don't let fly. There's certain creativity that you either encourage or stifle, whether or not you're aware of it. You have blind spots and those, those blind spots become projected into the company and then it's then it's even then it's kind of self reproducing. Like. Like you said, once you've hired the first 25 people and you've grown to like, 75, then whatever your consciousness was at that time is something that you're then sort of constantly having to run into and be like, oh, man, I see this thing bubbling up that reminds me of exactly the way I was. I mean, this has happened for me at least, like, yep, this is the way that I was a year ago. And now that is coming to burn us all in the ass. And we all get to experience my lesson from last year together, even if it didn't come from anywhere. That was in my direct performance in the company.

Speaker A: Yeah, definitely. And one strategy we're trying out, I don't know if it's going to play out the way I want it to, is actually making it acceptable to have disagreements or discourse. So it's okay to disagree, but you should still commit to what the team agrees to. And what we actually do is we document disagreements. So we have, like, role definitions and people who own various aspects and responsibilities. But. And then we have criteria where, if certain responsibilities that people have, if the situation demands the need for sort of intervention from other people, that that criteria has to be crossed before they can interject. But until that point, the person has autonomy to make decisions. And you have the right to have discourse with that person, provide them feedback, but they have the ability to drive that process without too much oversight. And they have the flexibility to take feedback and incorporate it as they wish. And they essentially also have the right to disagree or the right to overrule other people's disagreements. But they have to create a space where people can voice their concerns and share their opinions and let them know how they feel. And so I think I'm trying to find a balance between what you described, which is like, everyone's a clone of myself, versus having some diversity and friction, healthy friction, but still have mechanics in place where it just doesn't blow up into a big conflict every time.

Speaker B: Yeah, yeah, there's something that. This reminds me. What you were just describing reminds me of a concept called Lead Partner follow and the advice process. I don't know if you've heard of these.

Speaker A: No, I would. I'd love to hear a little bit about that.

Speaker B: Yeah. So Lead Partner follow is a concept where if there's an initiative that comes up, somebody takes leadership of it and that person makes all the decisions. And then people can elect to partner with them, and the partners are consulted, but they don't make the decisions. The leader makes the decisions. So that there's a space for, There's a space for information to aggregate, but not to get locked in a consensus situation and then follow. People can follow and be updated. They're like, I want to know how the decision made here impacts me, but I don't need to be a part of making it. And so that's one thing that seems kind of relevant to what you're saying. And then another one is just the advice process, which kind of happens within that framework where if you're about to make a decision and you see that it is likely to affect somebody, you ask their advice, but not for them to overrule what you're doing. But like, and if you're, of course it really comes, it really matters where you're coming from in this, because if you're just trying to tick the box of I'm, um, getting their advice, but you don't actually give a fuck whatever they're going to say, then it doesn't work. But it really does work if you actually do recognize that you care about how, what your decision you're about to make impacts people. And from your. If you're the one that's closest to that decision and most capable of making that decision and you've seen the way it's going to impact others and they're aware of how it's going to impact them, um, and they've had a chance to give you some information and some feedback, even if you don't make the decision that makes it feel easy for them, it might still still more likely be better for the company overall. And they're also going to feel heard and they're going to be ready to pivot on whatever they need to do because of the decision. I like, uh, both of those philosophies because they really just allow decisions to be made really, really quickly and for error correction to propagate through the system.

Speaker A: Yeah. And I think having some flexibility for culture to evolve in a healthy way is important. The rigidity of the founder's Persona, uh, sort of stamping over everything else is probably not, uh, a healthy way.

Speaker B: Right.

Speaker A: So that's what I'm trying to sort of nurture at Hyperspec.

Speaker B: Yeah. Beautiful. Well, I feel like that wraps us up to a pretty good close. And I really appreciate you joining us, Shravan. And is there anything else you'd like to say to listeners?

Speaker A: Yeah, I think I would like to say that entrepreneurship is a pretty hard journey and it's, and it's really rewarding and there's often, uh, times where you feel like it's a rollercoaster ride. But I do think the world needs more innovators and more entrepreneurs and more mentors. And so I think just being part of the community and contributing back is something I really cherish. And, Brett, thank you for giving me the opportunity to be here and have this conversation and was very insightful. Yeah.

Speaker B: Um, thank you so much.

Speaker A: Thank you.