Subscribe
Share
Share
Embed
Deep conversations with the founders, investors, and operators building real-world AI - robotics, automation, industrial systems & AI infrastructure. Past the headlines, into how these technologies are really built, deployed, and scaled. Hosted by Bogdan Cristei, venture partner and former systems engineer.
THE OPTIM UPDATE - Episode 1
How AI Models for Robotics Are Really Built, Deployed, and Sold
Guest: Dr. Asad Tirmizi, Co-Founder & CEO, Trener Robotics
Host: Bogdan Cristei
---
Bogdan Cristei (00:00)
This is The OPTIM Update. I'm Bogdan Cristei, and today's guest is Dr. Asad Tirmizi, Co-Founder and CEO of Trener Robotics. They work across ABB, Universal Robots, FANUC, and they've built up over 15 integration partners across Europe and the US. Asad came out of Vicarious, the robotics company that was acquired by Google. And before Trener, he helped build out the robotics and haptics program at ByteDance. This is someone who has actually built the models, not just talked about them.
I first met Asad on day one at UC Berkeley SkyDeck. Even back then, it was super clear to me he has something very special, and it's been so incredible to watch the journey from that first meeting to now a $32 million Series A with backing from Engine Ventures, IAG Capital, Cadence, and Nikon. So huge congrats to Asad and the whole Trener team.
We get into how AI models for robotics are actually built, what happens when they hit a real factory floor, and the part nobody wants to discuss - how do you actually sell this stuff? Without further ado, here's my conversation with Dr. Asad Tirmizi.
Bogdan Cristei (01:22)
Asad, welcome in. I'm really looking forward to this. How do you feel today?
Asad Tirmizi
I'm feeling great. And it's always great to be in your company, Bogdan.
Bogdan Cristei
Thank you, sir. So here's what I want to cover today. I want to walk through with you the full lifecycle of an AI model in robotics. First, how models are actually built - the real version, not the pitch deck version. Then, how do they get deployed on the factory floor? What breaks, what works, what "robot agnostic" actually means. And finally, how they get sold - the go-to-market, the channel strategy, what manufacturers need to see before they say yes. And then we'll close with some hot takes. Does that sound good?
Asad Tirmizi
Sounds great, man. Let's get into it.
Bogdan Cristei
Awesome. So the first topic, let's get into it - how models are built. I prepared a few questions for you, but we can go off script anytime. So, first question: if someone is building a VLA for robotics today, what does the training pipeline generally look like?
Asad Tirmizi
So, first thing - when you're building a robot model, it's a long process. You can look at it as a recipe, like a thick spaghetti bolognese, you know.
Bogdan Cristei
I see, I see.
Asad Tirmizi
Everybody knows that there are some main components that go into this recipe. But you can always have different unique touches based on who your audience is. So when you're talking specifically about the training pipeline, it depends on what you mean. If you mean the end-to-end training pipeline - data comes in and out go the actions that are hopefully deployable - then you're looking at a pretty big set of actions.
But let's just start from where it all starts. At some point around 2015, 2016, we found out that when you get data, you have this ability to create emergent behaviors out of it. So let me put it in perspective. You have this classic sense, plan, act, move. This is how it started for me - you have what you call classic sensor-based control. And at some point you start thinking, all right, that sensor data that comes in, if we start storing or concatenating this sensor data, there is learning that starts happening. And if learning starts happening, you can control this learning to create emergent behaviors.
Now, this is a very simplistic way where you see that there is a scaling factor applied - data comes in, we can collect a lot of data, and then we can make the robots do something smart. That is a sea change from how robots have been for the last 70 years, where the human has to think, we have to structure the environment, and that's how robotics is done. So I think this is where the interest in robotics and the opportunity in robotics is coming from.
But of course, I can go into much more detail about what this data is that's coming in and then what happens to it. So I'll wait for your questions before I dive into that part.
Bogdan Cristei
I would love to dive into that. Maybe a good question to start us off is: what's the tell that someone has actually built models versus just read about them?
Asad Tirmizi
They just keep talking about data, but not about how difficult processing that data into something useful is. Because sometimes - and please, I'm not targeting anyone - but people who have not built a full-fledged deployable or at least demonstrable VLA, they often miss how much of that data is crap, or how much effort goes into making that data into a model that is usable.
Because the fact is this: if your timestamps are off or if your data episodes are wrong, you are putting it into essentially a black box. You are putting in millions of episodes, and once you get garbage in, you're certainly getting garbage out.
So I think that's the telltale sign. It's not like data comes in and you have a VLA or a model that goes out and deploys. That's how I check for whether somebody is actually building it or is just interested in the field. But either way, no judgment on my end.
Bogdan Cristei
But maybe we can educate the listeners a little bit. How does this data look? Is it in a table? How do you normalize across everything? How do you make it searchable? How do you know which data points are crap and you need to take them out? How do you approach these things?
Asad Tirmizi
How it smells, that I can't tell. But how it looks, that I can tell. So, there's a lot of things to unpack here. Let's try to make it simple yet deep enough that it covers a broad range of audience.
The first thing is raw robot data collection. You're looking at basically timesteps that include observations - could be camera images, robot joint positions, robot joint velocities, torques. Typically what you have is an observation part of the data and then the action part of the data. So you'll have joint torques, joint position deltas that go into the action head. And then you have the observation part, where typically it's encoder positions and camera images. So that's the raw data stream you're looking at.
And then the next part is the dataset construction. And this is the hard one. Frankly speaking, in my opinion, if this thing takes off, then this whole dataset construction - you could create a company out of it. Because when we started doing it, aligning timestamps, interpolating missing frames, normalizing joint coordinates, removing corrupted runs - you're looking at a huge amount of data. You have teleoperated data, you have open-source data, you have simulated data. And all of that has to converge into what you can call a training sample - an observation and then an associated action.
Does that make sense? A lot of data comes in. You have to make that into a usable dataset. And then this dataset becomes a set of observations and actions. And that is where the GPU part comes in, so you can start making embeddings out of it.
Bogdan Cristei
Yeah, that's super interesting. That kind of leads me toward my next topic: how are models deployed? We've been talking a lot in the robotics space about skills. So I'm curious, how would you define a skill in practice? For example, if I'm a shop floor operator, what changes for me personally?
Asad Tirmizi
So if you were a shop floor operator, previously you would buy a robot and then you would go to somebody who knows how to operate robots before that robot is of any use to you. And the whole idea is this - with all this dataset, and I mean, I didn't go very deep into it because there is a part of the training pipeline that is essentially very slow. There is still a lot of work that needs to be put in to make it fast enough.
But once that is done and done correctly, and now you go to a machine shop, what comes out as an advantage is this: now suddenly a skill has intrinsic knowledge on how, for example, CNC machining is done. So if you as an operator do not know what is going on, the model knows what is going on because we trained it on all the visuals of machining, all the actions of machining, haptics of machining, the language of machining. Because for an average person, if I say "let's clamp a vise device," what does it mean? But you need to know what that means if you have to deploy a robot to do that.
So in my opinion, the robot skill is that particular unit of software that goes in and brings the intrinsic knowledge of the task, the domain expertise of the task, and then the ability to generalize no matter what happens in this task. And that's a very successful business model.
Bogdan Cristei
Got it. So maybe a good follow-up here is: where does abstraction win, and where does it leak?
Asad Tirmizi
I think abstraction, if done right, always wins. Except when it's not done right. Because there are things that if you haven't abstracted them well, you are bound to create customer dissatisfaction. Now, it is also true that there are things that do not abstract very well.
So at least in our company, the business model that we have is that we partner with system integrators. They are a very essential component of our go-to-market. Because the idea is that things that can be abstracted away, we abstract them away, and we do them well. But then we're very open about things that cannot be abstracted - out-of-domain knowledge. It creates this perfect mix where system integrators and OEMs bring what they are best at when it comes to hardware or fine manipulations, and then the overall setup is abstracted away to the AI model. So at least for now, we have seen good success with that approach.
Bogdan Cristei
Got it. And what's your experience been like with high-mix manufacturing? Are there any breaks in traditional automation for that?
Asad Tirmizi
I think high-mix is a precursor to going toward unstructured and semi-structured environments. My opinion on this is that if you are going into robotics today to make robots better at just repeatability, then you're solving a solved problem. The real prize here is not that robots can do CNC machining - that's a good entry point - but that robots can do much more than what they have traditionally been doing. You can start from a traditional space, but you can expand very much.
High-mix is typically something that robots were not good at. So that's a natural point for AI robotic companies to start their journey. But then this journey is limitless in my opinion, because there is so much automation labor need that if you get this flywheel going, it's almost never-ending.
Bogdan Cristei
Got it. Well, that kind of takes us into the factory reality. In your opinion, where do systems actually fail? What's been surprising you?
Asad Tirmizi
So a factory is not just a robot. That's number one. A robot is a very small part of the big factory. And frankly speaking, those who are working in factories, they do not look at a single unit doing a part of the job. They look at the overall production. So a lot of things that look really great in demos are either useless or nobody cares about them.
What they care about is predictability, determinism, and that this robot will be able to do the task in this amount of time. I think when it comes to factory, the simple answer is this: if you do not have reliability, you do not have anything. So it is much better to keep yourself feature-light so that you can do the job at hand rather than have a super feature-rich product, half of whose features are never used, and then dig yourself a hole by basically stopping the entire production floor.
Bogdan Cristei
I'm curious, how do you catch drift and silent failures in production?
Asad Tirmizi
There is no doubt that monitoring your system is extremely important. And this is actually a very difficult part of the job because factories have their own IT/OT systems. You have to be very mindful about that because you can't walk into Mitsubishi or Siemens and tell them, "I come with something fancy, but you have to take your whole IT system down." It's not going to happen.
Typically what happens is that you need to have a baseline system for monitoring, for QA testing...
Bogdan Cristei
Make some friends first, right?
Asad Tirmizi
Right. If you are getting into an enterprise contract where the firewalls around IT/OT are really severe, then you need to have a local server system where these things can be handled, and then some special arrangement to get this information to you. Because if you don't have visibility, you will for sure lose that line. These things will happen. Sensor drifts will happen. Grippers will break. A joint failure is either your responsibility or your partner's responsibility.
How we deal with it is that integrators have been doing this job for a very, very long time. So there is a lot of day-to-day maintenance that no matter how awesome your AI is, it will still be required. And you need to look at the ecosystem and have friends in the ecosystem - all the way from the buyers to your partners - to basically give a good service level.
Bogdan Cristei
Got it. On that line of thinking, in your view, what's the definition of "reliable enough"?
Asad Tirmizi
That's something you have to very gainfully define for the domain of applications that you are going after. I can give you an example from one of the domains that we have, which is the machining models that we are deploying. The mean time between failure, part changeover time - these are extremely important things. And calculating that realistically, because it doesn't work like, "Oh, my model didn't fail - it's actually the PLC that had a failure." It's also system verification around what kind of system have I put together and how much of these systems are natively supported by my model.
So those are some really key metrics, but then there are some side metrics as well - what is the time to deploy? What is the time before we pass the factory acceptance test? Those are more like operational checks. But after that, we get a comprehensive picture. And most important for the business side of things is: what is the ROI for the end user? Because that is something that you will use over and over again. And with these base metrics, you can very easily calculate the ROI. And once you have that, you have very good marketing material, and also a very good understanding of whether you're pricing yourself correctly or not.
Bogdan Cristei
On the topic of marketing material and pitch decks, I always find it interesting to compare the reality between the shop floor and the pitch deck metrics. So what matters in evaluation the most, do you think?
Asad Tirmizi
You mean the evaluation of the pitch deck? I didn't fully understand the question. Can you repeat it?
Bogdan Cristei
Evaluation of the technology. When you're talking about deployments, when you're talking about hitting your metrics, you have the pitch deck numbers and then you have the shop floor numbers. So I'm curious how you think about evaluation in general - your day-to-day business, what's actually happening on the floor, and then how do you communicate that in marketing materials?
Asad Tirmizi
I think this is where - at least I'm very clear about it - inflating numbers or showing a better version of your product basically hurts you in the long run. Typically, the companies that adopt your solution are those companies that have a forward-looking outlook. These are the customers that we have seen - there is a lot of passion about it. And I'm not talking about just big customers. I'm talking about the small customers. It's really fascinating for them to have a robot that they are talking to, that is seeing around, that is changing parts, whose movements remind them of the movements of a fellow human.
So personally - or maybe I'm just lucky at this point at this scale - we have investors and we have customers who really expect that the numbers we are seeing are what we are marketing. And that's exactly what we are doing. At least in our case, we are not trying to paint a rosier picture than reality.
Bogdan Cristei
I wonder, what's the most misunderstood thing today about deploying AI in factories?
Asad Tirmizi
AI models, as we typically know them, are probabilistic systems. It's a probabilistic output that comes out. And robots are systems that are, or should be, deterministic systems, especially in factory settings. So this is a very important distinction that model builders need to be clear about - once you put in something that is fundamentally probabilistic in nature, you are creating issues for yourself.
This is an architectural question for most CTOs or product executives - how much do they want to invest in the layer that at least ensures that safety is deterministic? Because otherwise, you're playing with somebody's life.
But then there are also deterministic actions that you can take for higher-level parts of your stack that will ensure reliability. Frankly speaking, in our company, this has been the main part of our research for the last 14 years, because we always knew that learning and emergent behaviors were coming, but how do you make them factory-ready or workshop-ready? And how do you change something fundamentally probabilistic into something that at least has some critical functionalities that are deterministic?
Bogdan Cristei
So if you had to name one bottleneck to scaling real-world robotics, what would that be?
Asad Tirmizi
I'm thinking, because there are many bottlenecks. I think the number one bottleneck is the cross-embodiment problem, because models need to be - if anybody says right now that cross-embodiment does not degrade your model performance, that's a lie. I stand on that very strongly, because I've been doing this for a while. This is a really important problem.
There are good signs that this will be solved, but it is fundamentally unsolved right now. There are ways around it - for example, we have to fine-tune our models to the robot brands that we support. We have to do this additional training. You can look at it as baseline training, and then there is training for the FANUC robot and training for the UR robot, until they come to a point where they are behaving at acceptable levels.
The other one is often repeated, which is the data problem. Data is indeed critical for these models that feed on it. But it is extremely difficult and expensive to collect that data. And when I say collecting that data, it's not that collecting data is difficult. Collecting the right data - that is the challenge.
Companies come and say, "I have this data foundry somewhere in some country, and they can give you 5,000 episodes." I don't know how you do quality control over what went into your model, because once they are in the black box, finding out what went wrong - just five episodes could make a million good episodes redundant. That's just the truth of how this works.
There is a third one as well, which is the compute problem. We as roboticists have never had a chip that is made for robotics. We have always worked with something that was fundamentally built for some other type of computation that then got accepted as the de facto compute for robotics.
For the last 30 to 40 years, it has been CPUs. Those who have been in this business for a while know that until five or six years ago, there was still Windows XP on a lot of major controllers. And whenever you have intelligent behavior that has to run on a robot, this computer is a bottleneck.
Even today, this is very true because fundamentally we are working on NVIDIA GPUs, and they are general-purpose GPUs which are not built for robotics. The industrial I/Os, the safety certifications, the real-time control aspect - there is just a lot lacking. Getting the right compute chip that can power this AI revolution is very important.
And I don't see it coming from NVIDIA, because I fundamentally think that their business model is to have general-purpose GPUs - the same for crypto, the same for data centers, and the same for robots. I think a lot of robotic companies will agree that they face a lot of problems with the standard chips like Jetson Orin or Jetson right now.
Bogdan Cristei
Is the issue around cost, or are they just not as efficient?
Asad Tirmizi
It's both. Cost-wise, you're talking about anywhere between $4,000 to $7,000 based on which flavor of these GPUs you're looking at. That's quite a big cost. When you think about every deployment having to get this compute box - I think it's too early in the AI game to have such huge margins on these chips. We should really try to create a market first. So I would be very happy if there was a $500 chip that we can just add and run the model on. That's number one.
But number two is the performance point of view - the safety certifications, the industrial I/Os, the real-time control aspect, especially how shared memory works. There is just a lot left to be desired. I wish somebody would solve that.
Bogdan Cristei
All right. Well, as an investor, I will bookmark that in my head and go out there and look for some founders doing this.
Asad Tirmizi
Yeah, please do. And if you do find somebody, ask them to order a hundred chips for me already in advance if they're doing good.
Bogdan Cristei
All right. They'll have their first PO.
Asad Tirmizi
That's the first PO. Ready to go.
Bogdan Cristei
Okay. So talking about POs, let's move into what I would consider the most fun conversation topic here - how models are sold. I think the go-to-market question doesn't get enough airtime in robotics conversations. Building a great model is one thing. Figuring out who buys it, through what channel, at what price - it's a completely different challenge. So maybe one question to start us off: you've chosen to go through integrators and OEMs rather than direct. Why? What does that unlock for you guys?
Asad Tirmizi
Some integrators are just better than us at integrating systems. That is the truth. And this is not just true about us - this is true about, dare I say, 99% of Silicon Valley companies. What is our strength? Our strength is software. Our strength is AI. So we build a model. This model is working. It depends on what you want to make it work for. Some want to make it work for everything in the world and then just release it. I think that's the pitch of some of the hyper-funded companies.
For us, we create these small models that are edge-deployable and are really good at a certain domain of tasks. So let's say we created a machining model. Last year, that model won the Machine Tool Innovation Award. So we know it is doing great machining and machine tending. But now what do we do? Should we go and start deploying it ourselves?
That's where you have to look at the margins. The software part is an 83% margin on deployments. For us, integration is somewhere around 17% to 18% margin. The margin is not the only thing - we're just not as good as a system integrator who has been deploying machine tending for a very long time.
This also goes for hardware. A lot of Silicon Valley companies get inspired by Apple and say, "We need to have the full stack." And all power to them. But I personally feel - should I compete with FANUC on hardware? ABB on hardware? No way. They are great robots. They are amazing. I love those robots. I've worked with them for a long time. I don't want to recreate things or try to get good at things that there are already players in the market who are good at.
Our models empower them. That's what I'm after. Empower them so that combined, we can create one of the greatest solutions ever in robotics that solves the real problem.
And what is the real problem? The real problem is that there are so many dull, dirty, and dangerous jobs in the world that humans should not be doing. Humans can still have these jobs, but the labor part should be done by a robot. We as a species are getting on average older. Some countries' average population is going to hit 60 years. So to be able to have the quality of life that we have, robots are necessary.
That's the problem we are solving. And this is going to be the biggest market that there is. Labor is the biggest market. So there are no shortcuts. And there is room for a lot of players. But if you try to take over everything too early, you're either going to kill yourself or you're going to stop the market from developing.
Bogdan Cristei
You touched on this a little bit, but we talked about CNC tending. I'm curious, what's the expansion playbook after that?
Asad Tirmizi
When we started, we had this great software - took us 14 years to develop, it's working very nicely. Where do we take it? The cool part would have been to take it to a humanoid that's jumping around and falling off roofs. That is cool to see. But I personally believe there is such a huge market here that there is a part of me that does not want to rely on venture dollars to make it into a business. I want to bring it to a point where I can see it as a business that grows on its own revenue. There is a data flywheel and revenue flywheel that I can see clearly as day, and I just want to reach it as fast as possible.
So if you have this mindset, you can't go into a market where robots are not deployed because there are regulatory issues. You also cannot go into robots where there is not a big market - humanoid or mobile robots have a smaller market. So CNC was this perfect thing where you had a huge market, a lot of robots already do this, they don't do it well, and yet they still get sold. So if you do it well, you have a pretty good market share.
And then when you deploy in CNC, what you are training is not a CNC model. What you are training are atomic skills that combine to make CNC. How do you pick a part? How do you place a part? How do you insert something? This is where the data is going - the atomic skills. And then this data can be combined to make more and more skills.
We have a very ambitious roadmap of around 20 skills that we want to put out, which are typically skills needed in unstructured environments. And after that, we want to open up - there are only so many skills we can train ourselves. So we have big ambitions on how we open up what we're doing to enable skills that we don't even have an idea about.
For example, there is a founder that came to me and said, "Hey, I got some venture dollars a while ago to pick apples, but picking apples is very hard. Can I train an apple-picking skill and put it on my robot?" And I was like, okay, there's no world in which I would know anything about apple picking, but if I can enable this person with our tools and our infrastructure to pick apples? That's great. That's a great business model. That's great for the world. Maybe apples will be cheaper - you'll get cheaper apple juice.
So we have a very ambitious, aggressive plan to go to more semi-structured manufacturing tasks, and then keep increasing to wherever there is a market for a skill.
Bogdan Cristei
You said 20 skills on the roadmap. What are three or four of your favorites?
Asad Tirmizi
Inspection is a really huge one for us. One of our investors is Nikon. They have incredible vision systems for inspection and they have a ton of data. So I saw that and could clearly make the connection - we can make inspection skills that have never existed before.
Then there are things that we have been doing for a long time but can do much better now. Bin picking in harsh lighting environments, especially when you're dealing with metallic pieces. I'm very interested in that.
Also, one of our investors is Cadence, because of whom we have access to really amazing physics models. We could see that we can put really amazing physics into a lot of things that we are simulating. You can see things like not just pick and pack, but being able to pack things that were not possible with robots before - like squishy toys or flexible material.
If you get into this frame of mind and you get the basic recipe right, I'm walking around and I'm seeing things that robots can do and a human should not be doing. So there are some really interesting things on the roadmap. And there are some really hard things that cannot be solved today, but we are very passionate about solving them.
For example, assembly - that's a very hard one, because you simply do not have the dexterity right now to do the kind of assembly that you need to do. You can do bits and pieces, but to make a model that passes our criteria for proper assembly? It doesn't exist. There is a very intricate connection between what you can train and what the hardware capabilities are for actuation. That's something to keep in mind.
Bogdan Cristei
That's interesting. Something I always think about is: what does a manufacturer actually need to see before they say yes to a project?
Asad Tirmizi
Our experience has been mixed on this. In the early days, when we were trying to get that first customer - that was the hardest. Especially when there is an upfront CapEx that they have to deploy, and it's something that affects their production. It's a very hard thing.
They would ask, "How many have you deployed? Do you have any data, anybody we can call that can say this thing works?" That was really hard. But once you have those case studies, once you have those metrics, once you see some success, you also invest into confidence-building measures. How can I invite my customer to our premises and show them this thing in action? How can I put that machinist on the system and make them ask their boss to buy this?
I think success begets success. I was just looking at it - it's the first quarter where we have inbounds that are converting into sales for our system integrators. We don't do direct deployments, we go through system integrators, but it's quite something to see that in France, our system integrators are getting inbound interest. In Portugal too. I think as a founder, it was always fascinating to see that this thing is wanted. That means much more than anything else.
Bogdan Cristei
Well, after 14 years of building this software, you're an overnight success, man. Congrats.
Asad Tirmizi
Overnight. That's a tough one. But I also don't want to say that everything is solved. There is literally a mountain in front of us that we have to climb. Because if we and other companies in this space play our cards right and develop this technology, I can tell you that this will have a very strong effect on the world economy.
And that's fascinating. I think that's why you become a founder, right? You see something that you think you can do and you go start. I mean, to be honest, I was awake all night last night. I haven't slept, so I have brain fog right now.
Bogdan Cristei
Go build it!
Asad Tirmizi
Because I was so engrossed in a problem that I worked until 2 AM, and then I couldn't sleep because I was still thinking about it. So at 4 AM I just gave up on sleeping. And I was thinking about you, because I was like, okay, how am I going to give a podcast with such brain fog? So I hope I did a good job.
Bogdan Cristei
You're doing great. Actually, we only have a few things left. We're right on time. So let's keep going with the fun part of our conversation. I'm curious - I want to get a couple of hot takes from you. Maybe I'll ask a few questions. Don't think about it too long, just be honest and jump right in.
So the first question: what do you think will look obvious in five years that sounds non-obvious today?
Asad Tirmizi
Today we do functional programming. I think in five years we'll be doing declarative programming across robots. In easy terms - today we code or program robots. In five years, I think we will declare our intentions to the robot. And even these intentions will be partially declared, and the robot will be able to fill in the rest of the declarations.
Bogdan Cristei
Got it. All right. One piece of advice for a robotics founder starting today?
Asad Tirmizi
Don't underestimate the last 10%.
It's one of those things - you're going to feel as if this whole field is made to suck you in. You'll think, "Oh my God, I got it right!" Because the amount of data that you have and the performance that you have - it gives you this feeling that you might be doing something nobody else is doing. Only to find out the diminishing returns as you go into the 70%, 80%.
And the last 10%, man, that just makes you cry. Because now you have this huge dataset and the black box and you just... yeah. The real challenges appear in the last 10%, and the later you go in percentage points, the bigger the challenges become.
Bogdan Cristei
So this may already be the answer to the next question, but I'll ask it anyway. What do most people get wrong about robotics right now?
Asad Tirmizi
I think this is related. My CTO and co-founder Lars puts it in a very nice way: a 90% result in academia is a groundbreaking result. A 99% result is a great demonstrator. But it is still not something that you can deploy. It's only when you go to 99.99% that you have something you can effectively scale.
And one of my fears for founders and for the robotics community in general is this: at 90%, the story becomes very convincing. It just looks like, "Oh my God, this is going to work." But there are so many moving pieces that to get it to 99.9% consistently and scaling - this is art.
So my advice, if I'm qualified to give it, would be this: in the grand scheme of things, getting things right even at small scale serves you much better than going out with something too big that you just can't control. You can have a valuation of $39 billion, but you wouldn't be solving the same problems - you're just solving a much bigger set of problems, and the impact is still very limited because you're stuck on something small.
Bogdan Cristei
Well, having seen you from your early days at Berkeley SkyDeck, from before anybody ever wrote you a check to now a $32 million successful Series A with amazing investors and people, it's been fascinating to watch your journey. So thank you so much for still answering my texts after you're becoming quite famous in the robotics community.
Asad Tirmizi
Really? Appreciate it. Bogdan, even from the early days, you've been a big supporter for not just me and my company, but I've also seen that you have done this across the board for many in the robotics community. So deeply from my heart, I'm very thankful to you for the support to me and all the others.
And I hope this talk might give some more ideas, some new takes on what needs to be done and who can do it better than even you and I. So really very thankful to be on this, and hopefully we can continue this journey of innovation.
Bogdan Cristei
Absolutely. Well, thank you so much for your time today. Really appreciate it. Thanks for the conversation.