DeAndre + Grant (00:00.174)
So you're up late and then you all of a sudden wake up and it's morning again. That's tough. Yeah. I think it's the e I think it is the East Coast time zone, is the cheat code going to Europe. But then coming home is horrible because you fly all day and you get home in like the afternoon. So you're really tired, but you shouldn't go to bed yet. And if you go to bed before like 10 o'clock at night when you get home, your clock is just broken for two or three days. It's just cooked.

Peter Tuszynski (00:04.252)
Exactly. So

Peter Tuszynski (00:24.208)
Yeah. But I can't complain. It's all good.

DeAndre + Grant (00:27.702)
No. No. And then and then there's the let's go to the other side of the planet, you know, the thirty two hour door to door trips to Asia. And then in which point you don't know what year it is. You're like, I don't know what day of the week it is. Like we are just here and the sun is up and therefore I should be awake.

Peter Tuszynski (00:39.334)
The worst is those long flights between like Hong Kong and LA, which are like sixteen hours, and you f I suppose that like you you f you s you you sleep and you're excited 'cause you just wake up and you look at the the watch. It's it's like I slept for six hours and I have ten to go.

DeAndre + Grant (00:44.638)
Those those are great. Yeah, those are hard.

DeAndre + Grant (00:53.624)
Clock.

DeAndre + Grant (00:58.37)
Let's see. The thing is when I get on those inner like the anything over a seven hour flight, it is time travel. I get in and I put in the seat, no matter if I'm in basic economy or like we win the business upgrade lottery with our United status, like full lay down. Doesn't matter. The moment the first meal is served and like the tray goes away, I am unconscious until they're serving breakfast whenever we land. It's audiobooks, man. Audio audio audiobooks and sleep headphones can just put you right back to sleep.

Filip (01:20.049)
No way. That's a superpower.

Peter Tuszynski (01:25.924)
I should try that. I I use I I typically do podcasts, but maybe audiobooks are the trick.

DeAndre + Grant (01:31.748)
yeah, no, you can't do something too interesting. You need something that like you can absolutely fall asleep to and you have to have sleep headphones. Like I have tried it with AirPods, I can't do it. They're too uncomfortable. So let me let me know if you want recommendations. I can drop you some in in an email. There there's a couple of brands now out that are pushing them and some of them are really, really good.

Peter Tuszynski (01:40.496)
Yeah. Yeah.

Peter Tuszynski (01:49.18)
That's the deal. I'll reach out.

DeAndre + Grant (01:50.85)
Yep. Well sweet guys. excited to get this rolling. So Philip, on your side, I can hide you. I don't know what that does on your side, but if that's what you wanted, then I can But I think we g we can we get each individual video stream, right? Yeah. Yeah. So yeah, what you can just cut let's let's keep them on. We just cut later. that way we don't miss any of his awesome reactions if we do need to cut it, cut cut to cut to Philip.

Filip (02:10.663)
No.

DeAndre + Grant (02:10.933)
Yeah, Peter I was telling Philip, I said I he I got the email saying that you you were he's gonna be a fly on the wall, but no promises you don't catch a couple of strays and and some questions in the in the meantime, so

thrown right under the bus.

Filip (02:22.37)
I will make good sound reactions again, so don't worry.

DeAndre + Grant (02:25.677)
Love it. Perfect. Awesome guys. I can get us rolling. and yeah, just kind of kick us we'll kind of kick off on some intros of the company. I think this is a really cool one just 'cause it really is like the peanut butter and jelly of how you bring an IoT device to life. And I think just that general narrative and obviously how AI is impacting that, and we can just kind go from there. I can't wait. We're gonna go down so many rabbit holes.

Peter Tuszynski (02:45.434)
Yeah, hopefully, right? Yeah, I'm excited.

DeAndre + Grant (02:48.461)
All right, let's do it. And everybody, welcome back to the Hard Tech Podcast. I'm your host, DeAndre Herakus, with my usual suspect, Grant Chapman. Welcome back, everybody. Super excited for today's one. This is, you know, my AI pilled CEO role. Actually, it's been Jones aimed to actually dig into this one on a show. Yes, me too. just if if you guys are just now tuning in to the Hard Tech Podcast, we're currently in season three of the podcast. Season two, we did forty two different episodes with founders and product leaders all across the country. I think we're up to like fifty-seven now, with season three included. So it's been a ton of fun.

stay tuned for for the episodes dropping. But super excited for today. We've got Peter and Philip from the intent team over across the pond. We actually just talked about that before before the podcast got rolling. Welcome to the show.

Peter Tuszynski (03:29.423)
Hey, thank you for having us.

DeAndre + Grant (03:31.53)
Absolutely. Well, yeah, guys, I would love to get a quick intro from you guys and just to kind of tee up the conversation for those listening. This conversation really is the peanut butter and jelly of what it means to bring an IoT product to life. You have the the hardware development firm, you've got the software development firm, all in the world of connected IoT devices. And so this could be really interesting. if you're if you're wanting to bring a product to market, this you get these guys in the room, you're probably going to be able to do it with some with some amount of haste. And so that's exciting. but yeah, we'd love to get a quick intro from you, Peter, on intent and yourself.

Peter Tuszynski (03:59.631)
Absolutely. so hey everybody, my name is Peter Tushinski. I work as a CTO at Intent, and Intent is this consultancy focused on software, specifically targeting hardware startups and scale-ups and products in general. and are the the thesis behind a company is essentially we help busy founders who are super passionate and super excited about hardware take the burden of building software that goes along with it.

DeAndre + Grant (04:30.059)
Yeah, but how hard is software, right? You guys just hit like, you know, if crash then don't compile and like it works, right? That's that's how software coding happens today.

Peter Tuszynski (04:38.084)
I wish it did. I wish it did. I think we're we're we're kind of getting close to that. or or maybe, you know, just just right into vicinity. But software is hard. I I I would say it's it's you know, it's it's it's equally hard as hard as hard.

DeAndre + Grant (04:51.925)
Well and it it's fun because we have embedded software here at Glassboard. So we're like in the middle of that level, you know, between like the hardware side, like the electronics and the like plastics and metals. And then there's the firmware and there's the software. And what we discovered, because we always used to make fun of the software guys for their jobs being easy, because they could just hit compile and test and find out if they're right right away. And that is a superpower that you guys have that we're catching up with, right? 3D printers are getting faster. We can compile faster to test. But

The thing that is a bigger challenge for you guys is actually how abstracted you are from the core and how much of a knowledge repository you have to work with in software. So like in hardware, mechanical engineers literally have the laws of physics and like mechanical engineering textbooks, which there aren't that many of. Right. and the electronics engineers, yes, they are saddled with 1000 and one data sheets, right, for every circuit board they make of every microcontroller and you know, LDO and power supply. They have to understand all these, you know, temperature ranges and current limits and all these things.

But then you go to the software side, both embedded and, you know, what I'll call cloud software, other software. Man, you guys are standing on the shoulders of a thousand giants, not just a couple. You have to understand or not understand and use anyways these tools written by everyone else and somehow make this massive amount of information work cohesively. That's like software versus hardware for me.

Peter Tuszynski (06:07.682)
One hundred percent. And there are different different flavors of software, right? Because there are, you know, the the apps that sit on a desktop computer or apps that live in your browser, which are I would make a, you know, it I I would say it they're slightly more predictable, right? Because the most software that we build sits on top of a phone. And the phone is such a such a platform that exists or coexists in so many contexts, right?

And we and our partners, you know, there there's been so many super interesting edge cases that would otherwise worked in the lab or on the test bench or you know, whenever exactly turned out not to be working all that well, right? Because someone had, you know, like a do not disturb turn on, or someone's you know, used a legacy operating system, or they had

DeAndre + Grant (06:49.483)
Or on the phone you compiled it for.

Peter Tuszynski (07:03.884)
some sort of low power mode setting obscured, you know, way back, you know, down in the settings that you only select few know of. and every every one of those, you know, elements that make up this whole context impacts the way the app communicates with a peripheral.

DeAndre + Grant (07:22.805)
or the person who's using the app. I mean this is the other thing that we both deal with, is that user experience is actually like ninety-eight percent of the experience of a product. The physics can work, the code can work, the database can have all the right data in it. If the user doesn't know how to get to it well, product failed. It's the worst.

Peter Tuszynski (07:37.391)
One hundred percent. Yeah, yeah. And and you know, smartphones are they used to be a, you know, this this very like only select few could have a smartphone, right? back in in, you know, even even before the iPhone, back in like Palmo S days, right? You could all there was a select group of people that used these tools, those devices in the select in a specific context. And there was far more predictability, even though those devices were

DeAndre + Grant (07:49.494)
Mm-hmm.

Peter Tuszynski (08:05.818)
you know, super underpowered. you had, you know, much less tools to work with. Like, you know, you didn't have garbage collection. You didn't have automatic memory management in these devices. And they're super underpowered. Now like you have so many different smartphones targeted at so many different markets that it's really hard. It gets really hard to come up with just one binary that works for everybody.

DeAndre + Grant (08:12.909)
Mm-hmm.

DeAndre + Grant (08:30.935)
When it's you have to build the lowest common denominator, the lowest link in the chain, like the weakest link is is what your your ceiling now is for performance and for experience and things like that. And I think this is the the beauty of design and art isn't actually what art can do or design can do when given the most, you know, knobs to turn or the most tools in the toolbox. It's actually giving them, you know, given a small amount of chisels, how do you turn this piece of marble into something beautiful? And that's what makes, you know, the separates good from bad, like app and software design.

Is right, making things elegantly simple, which is the hardest thing to do. It's way easier to throw fable fable and horsepower out and be like, I don't care how inefficient it is, just make it work. Speaking of fable, Peter, from your guys' perspective, how have you seen AI change connected device development like in the end? What has that really looked like? I guess, Peter, on your side on the software front, and then after Grant on the hardware front, what are you guys really seeing and how is that changing?

Peter Tuszynski (09:20.728)
It's I would say across the board, it changed, you know, almost I I don't think there is a single element of the process that is untouched by AI. I, you know, obviously something that started it all was the coding agents, right? and before that you had a lot of you know the a lot of models trying to one shot or you know just return an answer to a specific question.

and you had a lot of hallucinations back then, right? But I think the turning point was that started it all was the moment when you started having these models become equipped with tools and be able to verify owns work, provided you give it enough constraints to work around those harnesses that then were introduced, right? so it all started with software engineering and I feel as though it it all started

right around fall last year, right? You've had

DeAndre + Grant (10:20.317)
It was insane. It was like a light switch. It was like, AI is cool for autocomplete. And it's really cool like searching for like what kind of software should I write? I'm having this bug. Where do I go look on Stack Overflow to get the answer? That was pre fall of twenty five. And then post fall of twenty five, like, hey, I have this problem. And then it would go think for a bit and come back like, I tried twelve things. I think it's this last thirteen thing that may work. Go try it.

Peter Tuszynski (10:42.14)
Yes, ex exact exactly that. And then I feel as though the you know the snowball just got so much momentum. and you know, people started building these connectors and extensions for Claude or Auto models to talk to Figma, to talk to, you know, other services that you, you know, you build your product with. Right. So starting with engineering, across design, even some research, right? I think it

those LLMs are such a powerful research tool, right? And there there are so many shades of gray with just within like every domain that you know it touches. So some people I see they use LLMs to research some concepts or you know demographics. But I've seen people, you know, create user cohorts with LLMs and try to do like user research. Actual user of course it's an LLM, right? But you can

can prompt it so that it kind of tries to impersonate your target user. And

DeAndre + Grant (11:44.77)
And I I think you nailed something. It kind of impersonates your target user today. Right. Like we're in this like weird valley of death between like it's perfect and it's trash. So we're trusting it a lot, but it's not right yet. And I think the the what separates people that are using AI like really poorly and making a bunch of AI slop is they don't check anything. They don't put a human in the loop. And the the thing that is the like I'm seeing like rocket fuel on teams is like, no, I know this AI can make some good and some garbage, but it can do it really fast.

Let's have it make a bunch of stuff and make a human in the middle. And Dana I was telling you about like the Tinder app, make a yesno interface for yourself for the outputs, whether this is code that you're reviewing or things like this, find a way to get the the agent to present the data to you, whether it's research or you know, user user summaries. How do you make a user review it quickly and not burn out? Cause I think, Peter, if you suffered this, like reviewing AI's work is exhausting.

Peter Tuszynski (12:36.634)
yeah. yeah. Especially because, you know, it's it it it scales almost infinitely, right? We're we're just I think touching the caps of, you know, how much is it gonna cost and the token maxing, right? I it it swings back and forth between, you know, talk and maxing and optimizing. But I don't think we've ever seen value or, you know, something that perceived that you know you can perceive as value be created so quickly before.

And and your your question is is 100% right. And I don't think anybody has yet found a golden standard of how do we validate the output of the AI? Because as you know, you can of course build out, you know, build build tests around code. And that's very deterministic. But how do you validate there's so much talk around taste nowadays, right? People, you know, people start hiring.

taste officers in their companies, right? To to to tell, you know, between what's in the AI slab and what isn't.

DeAndre + Grant (13:41.836)
Well, and I I I mean, digging into that, I think this is why I'm not afraid of all the jobs disappearing. Right? Everyone is like, you know, doomerism, the world's gonna go away and all the jobs are gonna go away. I'm like, no, humanity still needs tastemakers and decision makers and fil the great filter of I mean, humans consume all the things that we're making, right? AI doesn't consume the products we're all engineering, whether it's software product or hardware product, unless you work NVIDIA and then AI is consuming all of it. but

It's all about how do how do we make the tools make the best stuff for us and the people that are like us. And even if it automates some of your job, you need to pivot to a way to find joy in creating more, better, faster. And I think we're we're starting to see like the turn, right? I have some software engineers on my team that I think, you know, maybe about May was probably the depth of like, my career is over, everything is changing, I don't get to do what I love anymore. And we're seeing this, you know.

Up the hill now of excitement of like, man, look at this cool thing that I did so fast. It would have taken me eight weeks to do this horrible, you know, driver integration. We did it in two days. And now I can do the fun part of architecting how does this product actually work. I do think it really comes back to velocity, right? Because to your point, Peter, you have these experiences. I experienced it a lot with like talking with founders and their pitch decks, for example. And you can tell within about a fraction of a second that as soon as you see this deck and you see some long like

Peter Tuszynski (14:50.788)
Is it?

DeAndre + Grant (15:05.651)
M-dash, it's like, okay, this entire thing is has been AI generated. But if you just took a little bit of time and took that as like your template, your customized template that you just had, you know, Claude create, for example, and then you go and add in your own personal flair with a little bit like understanding design, understanding how you want it to look and feel, all of a sudden it goes from this is AI to this is just a good looking presentation that would have otherwise you would have needed a designer to even just get to that point to the point where you could start to hone it in.

And so I do think that what we're seeing on the velocity front is truly what was the outcome versus people losing their entire roles. Now, granted, we're still early days with AI that could obviously change and it could all of a sudden develop taste and then, you know, we can all just start farming and enjoy our lives, right? But until then, universal basic income. But I guess the other question I got for you guys is just around embedded. I feel like that's kind of like the the front final frontier of like AI, you know, code development because you're not just touching, you know, a software stack.

Right, you're also touching like a microcontroller or embedded Linux on a physical device. And so how have you found AI's, you know, playing a role in in that process, you know, when it goes from the the actual code being written on a device to, you know, the device doing it. The the device itself. And how do you test that?

Peter Tuszynski (16:14.95)
So I'll I'll start with there's there was this misconception that AI is not really good for embedded and that the models are not you know qualified enough and and there there is some truth to it because that there is so much that the corpse of data for web development is ginormous compared to the code the embedded code that you have available on GitHub and elsewhere. there's been so many, I think, materials

that were just not available to AI because they've been MD8 or they've been shared. Exactly. And and they would rarely make it out of you know email threads or whatnot, like private repositories. But fact of the matter is I think the the the unlock for embedded development happened when the model was started to be able to iterate on its own.

DeAndre + Grant (16:47.521)
They're closed sourced of some kind.

Peter Tuszynski (17:09.596)
Right. So as soon as you connected a and you know, let's just take an ESP thirty-two board to your computer and gave the model enough tooling, like specifically CLI tooling, to flash the the device and then also give it enough tools to validate if you know whatever you're flashed on this board actually works. So you know, give it a a an an Nordic dongle to

you know, to validate if certain Bluetooth characteristics actually stood up on that board. that's when I think the the unlock happened for for the embedded. And I think all of all of all of Intense engineers are currently working with you know pair with with Claude or or other tools that have an AI component in there.

DeAndre + Grant (17:59.438)
And I I think the the magic that I love to watch in this entire like evolving chain is we were pretty bad at writing software with AI in twenty-four. And then through twenty five it was great for autocomplete, but didn't write it on its own. It helped us write it faster. End of twenty five, it started writing software on its own and kind of wrote bad embedded software. Early twenty five, it started ear early twenty six, it started writing embedded software and you had to be there to like babysit it and like unplug the thing and plug it in for your AI. Like you were the you were the technician and it was now the coder. And

Now we've used the LMs to write drivers for, you know, custom integrations to our power supplies, to our oscilloscopes, to our, you know, giving it control of the physical world. And it's our job to architect the loop. Right. How do you how do you take all the outputs that you're trying to measure, whether that is a voltage, right? Truly in the power domain or battery percentage or these things that you need to validate with a piece of lab equipment, all the way to the user experience side. Hey, what does the UX display on the display? Do we need to put a camera up with OC with like an image capture and do? Like, does it do it? Did it

have any jagged edges, does it load slow? And the more you can close this loop between the physical world and what that AI can touch, it's now able to verify its outc outputs. And who cares how long it takes? Let it iterate ten thousand times overnight. I'm taking a nap or having a coffee. It is doing the thing. And Peter, I think what you and I are so excited about is closing that side of the loop, right, for our world. Because the software guys have had that for about six or nine months and we were we were hungry to get caught up.

Peter Tuszynski (19:23.098)
Yes, yes, well one hundred percent and and and exact exactly right. And you you actually touched on on you know on this loop buzzword. There everybody's now, you know, we've been prompt engineering, then we've been context engineering, harness engineering, and then finally we're right with loop engineering, right? And as as long as you give those models, that's our experience, because th those models are I would say, you know, eighty percent roughly on par from the frontier labs, right? Claude

DeAndre + Grant (19:51.917)
Mm-hmm.

Peter Tuszynski (19:53.275)
really good there's been some buzz about the latest model release from from from entropic not being you know as big of a jump forward as people expected some skills didn't work right but it but it's these are growing pains right and especially if you iterate as fast as them some of that stuff is is expected but nonetheless if you if you give this the these these

models enough tooling and just leave them there and provide them with a clear goal for them to to to arrive at it doesn't really matter how long it takes right it's it usually our at least our experiences it takes less than it would have taken us right

DeAndre + Grant (20:41.387)
Mm-hmm. Mm-hmm. Well, and and I think the the the double dip in that is what you said was really interesting. the next model didn't have the jump we all expected. Man, we all need to look in a mirror. Like the thing we expected six months ago is one percent of what we expect today. Like it's not 2x or 10. Like it is our the pace of change in this world is staggering. And Peter, you and I have been at the forefront of innovation and development for our entire careers, is what we do. I don't know how to wrap my hands around how fast it's moving.

But the thing I do know how to wrap my head around and my hands around, like how do I like vision this is this is a tool that makes tools. Sometimes there's just, you know, a one-step change, right? Insert new technology comes out and now everything's built on it. I'm calling this the iPhone moment. All of a sudden there was a color touchscreen in your pocket that could access the internet and have compute on it. Therefore, the last 15 years of all app connected devices happened because there was an iPhone platform. But that happened at once, right? There's one big step change and very little incremental gains after that. It was just all this searching that ocean.

For the application we could, you know, pr provide our service in and you know build products for. But I think what's neat now is when you say give it tooling, the fun fact is it's not waiting for Anthropic to give it tooling. It's not waiting for open AI. It's also not waiting for bootloop AI, who's great. I they're they're they're building these harnesses and these connectors that actually enable the hardware and and AI connection that if your team didn't build it from scratch or wants to maintain it, which we're finding is a challenge.

Right. You build this cool custom tool in and you're like, man, I don't want to maintain this. I want someone else. I want to pay someone twenty dollars a month. So SAS is not dead, guys. there's still people that I would love to pay SaaS tools to to automate this. But the end game is I can help make the tools better right now, not just over time. Like spending 30% of your time recursively making the tool better is worth exponential gains in the time it takes to solve the problem. Right. So I think the what is it? A Wilincolin has, if I was gonna cut down a tree and I had six hours, I'd use four of them to sharpen the axe.

This is a a moment in time where I think it's even more important than ever.

Peter Tuszynski (22:40.216)
Exactly. And I've seen some really amazing moments when non-technical people started building tools to help themselves out. So not only do we have now tech, you know, engineers and and tech, you know, tech savvy people building their own tools and speeding up the models output and making those models more capable. But also we've seen, you know, a across our entire company, you know, we've seen

Folks that are dramatically not technical build Slack bots to automate some stuff. It's just incredible.

DeAndre + Grant (23:12.937)
Mm-hmm. Yeah, whenever Philip and I do a whole you know, podcast over GoTarket, we're gonna dive deep on that. but I the the question I've got for you guys is I'm gonna dive into the QA in a second, because you guys have built a really cool tool I want you guys to come talk about. but before that, I kinda wanna zoom out and think about it from call it the client or the product lens, right? So with AI enabling faster QA, enabling faster development, enabling insert this. So we we talked about velocity earlier.

Is the velocity of bringing a connected cause a connected hardware product to market actually getting faster? Or like what is that true delta? How has that really changed in terms of it used to take you eighteen months? Now does it take you seventeen, sixteen? You I'm just kind of curious, like if you're thinking about going out and developing this, you know, MPI, you know, bringing out a new product and with these new tools, what is the the real change of velocity on the macro scale versus like these micro, like in the process,

way to look at it. Kind of curious from your guys' perspective. I know time li as a consultant timeline and you get slapped on the wrist, but I'm gonna do it anyways. Yeah, yeah. Go go go put your you know your pin in the timeline and say this is what it's gonna do now. Peter, you're up first. I'll follow

Peter Tuszynski (24:22.31)
Sure. I I don't think we're we're these things were around long enough for us to understand the long, you know, the macro impact, right? There are there are so many signals that we have that we've captured over you know the past year that certain things get done dramatically faster. But because we're introducing

A dramatic increase in velocity at certain parts of the process, that means that the bottlenecks sometimes appear, you know, further down. And that's why that's why we we we introduced we started building QAGEN, which we're probably gonna talk in about in a little bit. But I think what what I was trying to say is the fact that we're speeding up development and so it it really differs between individuals. We've seen, you know, anywhere between 1.5x all the way to you know

DeAndre + Grant (24:54.637)
Fascinating.

Peter Tuszynski (25:16.196)
Something's like 7x or 8x output of an engineer. all of those things are very valuable, but it's like unreal unrealized gains, capital gains, right? Until you ship a product, like a lot of that stuff is on paper, right? So so if you ship more code, that just means you ship more code that is not verified, that is not Q8. That doesn't necessarily mean that you have.

you know, absolute velocity increase of whatever your engineers are doing faster, right? You just make the you just have to make sure that the whole process, you know, gets the benefit of that velocity.

DeAndre + Grant (25:58.494)
No, a hundred percent. I I think the hearkening back to our first conversation, this is about making a loop where you can check the outputs fast. And unfortunately in hardware, like product, that's until you ship a product and it sells well, right? None of us have been through half of a cycle with these tools. And again, even within the cycle we're in, we started the cycle with really shitty tools. We have what we think are awesome tools now, but are probably mediocre. And in six months by the time that product ships, the tools will be, you know, 10x as good as they are today.

So I think we're in a in a world where none of us can make a call of how much is this helping? Because we're finding the next bottleneck really fast. Right. You solve problem A and then swack them all. The next one comes up that was usually running in parallel, but now one got done early. Now this is your, you know, long pole in the tent. And you keep cutting down poles and you keep finding longer poles in the tent as you go. And you realize that, you know, the spread from the longest task that you had, like in a calendar time, to the one that you're at now isn't as big as you thought it was.

Right. We aren't seeing like this insane velocity increase in like shipping products, but what we are seeing is reduction in risk. This is this is what I see AI doing today. Like again, I think the next generation will be faster, but in today's generation, people using AI well are reducing risk because they're looking under rocks they wouldn't have turned over earlier. Right. Hey, I mean one of our favorite ways to use AI actually isn't engineering at all. It is requirements gathering and requirements tracking. Hey, we record every meeting, every transcript goes into the LM and it says, hey.

Did anything we talked about today change any requirement? Or even, you know, yellow light, not even red light, yellow light. Did we talk about something that wasn't exactly aligned? And then some human has to play the Tinder game. Yes, we changed something. Let's go write it down how we did it when. Or no, we didn't change anything. L L U or you know, conflating things that are unrelated. And doing that really fast is just a bunch of double checking that a human doesn't have to do. Right? The AI is asking every ISO standard, did we break one section of the standard that none of us have actually remembered reading? Because that was a you know book this thick. And

That is the reduction risk that I see being nonlinear today. Just, you know, are you meeting all the things you set out to meet really early? And the humans are still responsible for making the product great. It's super fascinating and it it

Peter Tuszynski (28:01.818)
Yeah. A little bit about this this de-risking because it's it's something we're super passionate about. But it from the other from the other side, from the engineering side, because now AI, even though a lot of the I probably have to count the tokens that we burn through, but I would I would make an assumption that a pretty comfortable assumption that most of the tokens that we burn are not the production code. There are POCs that we do just to de-risk certain aspects of the product. And

DeAndre + Grant (28:07.533)
Mm-hmm.

Peter Tuszynski (28:31.388)
Because of AI, because of you can paralyze so like so many POCs, you know, in building

DeAndre + Grant (28:37.655)
Proof of concept for everyone that doesn't live in our world.

Peter Tuszynski (28:40.26)
Exactly. Like throwaway code that just basically the only the only aim of that POC of proof of concept is to validate or disvalidate a certain assumption that we had before. So so now we don't have to build things, we can just test them over a course of, you know, sometimes even a couple of hours.

DeAndre + Grant (29:01.397)
Sometimes seconds. I I one shot at a CFD viewer HTML page last night, not because I needed one to work. I needed to show a software firm, I want you to build a tool like this to view the data we're generating out of this IoT product, but it's not CFD data, but this is the, you know, here are the knobs I need. And I pointed it at a website of a demo. I'm like, go make me an HTML page. It gives me these controls over a data field. And in 30 seconds, I had the demo I needed to translate an idea. Right. It's not production code. It is to show a concept to a tastemaker. Right.

In some ways it almost here it sounds like, you know, the lower levels of like AI development almost has become like a commodity, right? And then once you get more advanced and get to production, obviously that's where a lot of the value is. But as you guys described, it's super interesting to kind of see this happening in real time, right? Like I'm imagining like a hose that has like a ton of water in one part and it's like, you know, ballooning up and then like a little spigot. The only spigot is like the normal little spigot and like

Every month that that goes by, we're kind of fully understanding and the hose is the full development cycle, right? And like once that full hose is, you know, both big on that end and also no longer a balloon on the other, we'll really kind of understand what that velocity changes from the full product cycle, which is super interesting. but it's it's a bottleneck, you know, fundamentally. And so I'm curious, you you mentioned, Peter, that the next bottleneck for you guys, what you've seen, okay, our code velocity is, you know, either 1.5 to 7x, which is incredible.

Are we now building tools to do s the POCs and so on? But the QA still has to be done. You guys actually have built something to solve this this yet ni net new bottleneck. I'm curious what that is and kinda how you guys stumbled on it and what ultimately does it get done.

Peter Tuszynski (30:33.894)
Sure. So so there there are a couple of stages in the in the product engineering process that we follow and and some got you know a dramatic increase of speed like software engineering or code reviews, right? Because you know you now can engage a a a

A model to do a code review, hopefully not the same model that build your code, right? Because it kind of sucks to be a judge in what's case. But where we found the bottleneck to be the most limiting factor to our velocity was the manual QA, right? Because in those simpler software engineering process pro projects where you build a website or you build a desktop app, you can relatively

you you can provide a relatively reassuring code test coverage that will probably get a lot of the QA issues caught quickly over you know, even even in an automated way. The problem with the projects that we build is they're usually have they usually have a peripheral, right? So you have a phone in your pocket or your hand and a peripheral, and a lot of the

DeAndre + Grant (31:46.077)
And a database and a web c web client where they connect to. Like it is actually a like three headed Hydra, not, hey, this is a web page. Here are the inputs, here are the outputs the user can touch. Just cycle through all of them and see if you find any bugs.

Peter Tuszynski (31:57.765)
Exactly. So so and and the phone and the and the cloud are, you know, we we we've done it so many times and there are there are so many, I think, really good solutions to make sure that there's data integrity and whatnot, right? but we found that the most problems happen in between the phone and the peripheral, right? Because the Bluetooth is fairly low in bandwidth. There are some, you know, aspects of Bluetooth that, you know, haven't been

improving all that well, right? Over years. There are so many headsets that headsets that are stuck on legacy Bluetooth version.

DeAndre + Grant (32:36.309)
Or your Apple and they only give you access to some of the Bluetooth knobs and not the whole radio, so you can't do half the things you would like to do for most of your you know, fifty one percent of your users or whatever the Apple ecosystem is in the US. Right.

Peter Tuszynski (32:47.708)
even more probably right now, yeah. And in though those problems usually have to be tested by, you know, by by actually taking a phone and a peripheral and actually working through a set of test cases and making sure that, you know, XYZ condition is satisfied. and because of that dramatic increase of velo in velocity in our engineering teams, we sat down and tried to

come up with a solution that would also empower our QA engine engineers to do maybe less of that hand validation or maybe sp you know allow them to dedicate more time towards more meaningful, more critical paths of the user journey and then maybe offset those less critical, rest re less relevant user journeys to something else. And iterating on that, we built and and and

That is a reference to what we just talked before about w moment you give model proper tooling, it's you you you almost take it to the next level. So we built a QA agent, which is essentially a an agent that is agnostic of a model. You can substitute models underneath, and it has a lot of tooling around it to allow it to interact with the phone, no matter if it's an iPhone or an Android phone, there are protocols too.

sort of simulate user behavior on either. And you also can give it a CLI and a peripheral that is tethered to a computer that it's running on. And then the third piece of the puzzle and sort of the ICM in the cake is you also give it a test case repository like case.io or similar, and then you ask it to do a test run. And lo and behold, those newer models that have you know far more intelligence

They are actually able to parse a human readable test case and act on it as if they were a QA engineer.

DeAndre + Grant (34:50.06)
Mm-hmm.

DeAndre + Grant (34:56.021)
Well, I I think where this is going is where I get actually like chills and excited is that's great, except if your hardware wasn't built to do this, to interface with this. But you know, as you have teams that are on this cutting edge and we are now like harassing our hardware engineers like, no, you need a single data port that we're not gonna populate in production that has all these test points on some dense connector I can run back to the computer. So the computer can then, you know, flip bits on this test port that is a

Button press, right? Is a you know, all the I.O., hey, I can unplug the thermistor that's connected to my device and fake thermistor values so I can fake my temperature readings to test my software. Like, is it reporting, is doing the right thing? And learning how to preempt your test plans before you've ever made the board, which is ironically what you're supposed to in like FDA development, like in quality and like, you know, user need, design input, and then your verification, validation test points, like you should have an idea what those look like before you've ever put pen to paper.

Because you have call your shot. I'm going to do this. If we eat our veggies early, we don't have to do the testing that none of us like. It is this true, like the final implementation of like, you know, work out so you feel better and then you'll be more have more energy, even though you spend energy to work out. We're seeing that in our jobs because these tools are starting to take off of our plate, the tedious stuff, the the the repetitive stuff that isn't creative, it isn't problem solving. And if you can I think the right word is gaslight your engineers to solving the problem of solving the problem.

It gets really nonlinear. I mean, Peter, I'm sure you've done this like tuning a PID loop, proportional integr integration differential like loop, like a PID loop, is nonlinear. You you touch one of those knobs just a little bit the wrong way in your system, and the answer goes way far away from the target or from what you were at. But we're seeing with these like you do basically described building a great QA harness, right? You've built an amazing harness for this agent that doesn't matter what models inside. You've just given it the scaffolding that tells it what rails to live on, and you're giving it great tooling on the other side.

Each one of those is one of the nonlinear knobs and it's not two plus two is four, right? Sometimes it's the right two plus the right two is forty or four hundred. Is that what you you're kind of seeing in the outputs?

Peter Tuszynski (37:04.568)
One hundred percent. And it's sometimes even it's sometimes even easier than you know running out to your engineering team and asking and begging for, you know, to to to include a certain port in, you know, the in the in the device model because we over over the course of you know, however we were in business for, we learned that there are a lot of things that you can simulate. Of course it's not one to one

alignment with reality because you have differences that you know that are just there because of the chips chip difference, right? And and and you know the the way the BLE stack is implemented on Nordic versus other vendor. But fact of the matter is you can really thoroughly test an iOS app against a simulator which is based on an ESP32 board that talks through the same kind of APIs that the real device would talk through. So as long as kind of like you said

you you you do your homework before that, right? Eat the veget before and write down those API contracts, right? That you know app and peripheral communicate with or the app and the cloud communicate with. You can then simulate a lot of that. And rest assured that you will catch a lot of really interesting things j even before you, you know, you lay your hands on a hard

DeAndre + Grant (38:23.329)
Yeah. Well, and and this is the fun part. What you just described is the assumption that the hardware's going to work. And I have to do the other thing around. We have to assume the software's going to work, right? And the app. So we have to like find ways to test our hardware. And it's like two sides of the exact same coin that is converging at the same output. And this is this is what I think is like the most magic side of like what A is actually doing is I see it doing less and less of the stuff we all actually like to do and enabling more and more very creative ways to solve the problem in a way that we all enjoy.

Right. You guys are truly taking this insane technology of LLMs and models and harnesses like clay and molding it to what you all need it to do for you. Right. We're doing some of the same thing here with some of our requirements, tooling, and things like that. And this is why I everyone's like, no, this is like, isn't this just Google? I'm like, yeah. Do you want to know what Google did to all of our lives over the last 30 years? It's just happening at a rate that is so different because it allows us to make

tools faster to help build things like it's an a self compounding curve.

Peter Tuszynski (39:24.44)
Exactly. Yeah. And and maybe just talking a little about the the mindset, right? Because I think a lot of people are uncomfortable with how, you know, with the pace of everything just accelerating, right? You feel you know, I even I get a lot of FOMO every time, you know, I wake up in the morning and I just try to scroll news, specifically AI news. Thanks. Thank you. and it it and I was just talking to someone the other day.

DeAndre + Grant (39:45.237)
At least I'm not alone.

Yeah.

Peter Tuszynski (39:54.011)
that and they expressed a lot of frustration because their passion is gone, right? Like they they've been passionate about writing code. And it's funny that there's this this this quote from Sam Altman that's stuck with me all the time about, you him having a lot of empathy towards people who are who used to be writing code, you know, character by character, which is kind of, you know, the way I did it for two decades. but I think it's it's really important to be successful nowadays. It's is to l let go of certain, you know

DeAndre + Grant (40:15.661)
Mm-hmm.

Peter Tuszynski (40:23.94)
certain ways of you you you doing stuff and be very, very open to you know to the to things you're gonna be unable to do because you're letting go of the things that are tedious and boring and sometimes, you know, frustrating.

DeAndre + Grant (40:38.955)
I I think my favorite an analogy is the carpenter one, right? It's a carpenter used to have a hand set of tools and carve from wood a beautiful piece of furniture. Right. And very few people could have that piece of furniture because they had to buy, you know, weeks of that guy's time to have this chair. And then that guy got better tools and a, you know, a bandsaw instead of a handsaw. And he can go faster. And he still felt like it was his craft, but he was moving faster. Fast forward to today, a CNC machine can cut that exact same chair out of a block of wood, and it's no less beautiful.

And it took no less design to design the chair. It's just he had to upgrade his tooling and let go of carving the wood grain by hand and go towards I carve the first one by hand to know what I want it to look like. And then we automate making it. I think that's the analogy I like to put in front of all engineers that are afraid of their craft. I'm like, it isn't your craft going away. It's just you using different tools to carve the wood. I would agree. And I think the other thing that would add that is just I think unlike the the gentleman like shaving the wood, right? Like

You could still go and buy the f fully CNC, you know, table, for example. But if you if you got one of those things that got shaved up by that guy, that's gonna be about 10x as expensive. And you can tell everyone this is a handmade, you know, table by a guy in Maine and, you know, he cut the wood himself and, you know, did all the things, stained it. I don't think people are as excited if they hear like, no, I I hand rent the code. Yeah, I I laid out the circuit board by hand. Yeah, I I did it manually. It's that's why it's 10 extra dollars, you know, every time you buy one. And I think that's probably the biggest difference, right? Which is, you know, you don't actually see

It lives underneath versus versus what the actual experience is for the user. I guess which kind of leads me to my next question, guys, around like we talked about in our our call ahead of the podcast that what like a SOP or like prompting AI, you think you need to be as deep and like, you know, give it all the parameters. This is your context. You're an expert at X, Y, and Z. But you guys explained it very differently. Like the way that you actually prompt AI is almost like talking like a caveman.

And so I'm curious like what has been your discovery there as you've began like learning the talk machine?

Peter Tuszynski (42:37.69)
I I I think it changes all the time. I it the the way I observe even myself and and I sometimes scroll through my you know chatbot history or codex history. These things evolve and I think that's what that that's one of the aspects that make this whole revolution so uncomfortable that you really have to revisit the way you do things almost on the weekly almost yeah, exact exactly, every day, right? Because those models they change and the way

DeAndre + Grant (42:58.902)
Every day.

Peter Tuszynski (43:06.426)
they do some work for you also changes. So at the beginning, you know, I was talking to someone from Google DeepMind months ago and the I I was shocked when the guy told me that, you know, they their prompts are many, many pages long. And then I look at that prompts and they're you know, I I thought they were comprehensive, but you know, they're nothing compared to, you know, multiple pages long. now I think that the the the unlock that I'm getting very recently is just me sitting there and talking, just

through my voice into an LLM because I think we as humans we have this this sort of I don't know what it is, but it's it's it's like this this desire to make the written word almost perfect, right? And you just

DeAndre + Grant (43:50.253)
We w we write different than we speak. Have we not been taught to do that? Correct, we have.

Peter Tuszynski (43:55.037)
Exactly. And I I put a lot of you know thinking about like every sentence and when I just yap into that LLM, I sometimes do it in a very verbose way. Sometimes, you know, if you if if I heard it back, it it probably would be confusing to me. But to LLM, you know, it's able to structure, you know, whatever chaotic input that I gave it into something that, you know, really works and it it provides it a lot of detail. So that's that's on the prompting side. and also the the the the the other, you know

that prompting is just one, you know, element of the equation. The other one is the harness, right? So so giving it a goal, giving it a quantifiable metric that it works against.

DeAndre + Grant (44:36.255)
I I I couldn't agree more. I I started yelling at Claude versus typing to Claude a couple of months ago and like it's crazy nonlinear how much better it is. And there like there's three facets. There is you give it more context. It is just getting more context, which in today's AI is like context is cheap. Give it more context. It will throw away what it doesn't need eventually. two, you write differently than you speak. Three, AIs are good at facts. They lack creativity. When you write, you write on facts. When you yell at something, you

Are creative, like, what if we do this? What if this crazy harebrained idea happens and the LM go, that's never gonna work, and just cross out your bad idea. But man, that one out of a hundred bad ideas you wrote down, it's like, I can't find a reason this one doesn't work. Let's go spend some some tokens figuring that out. This is the the unlock that I think that Peter, you you've probably stumbled across and I have too. Do do not put this AI on Rails. Six months ago they were dumb enough, you needed to put them on Rails. Today, they can check their outputs if you give them a good goal.

And the other thing that I think I really want to tell everyone to do today is templates. The thing that everyone hates is that the result is slightly different every time, right? Give me a requirements doc is slightly different. If you just give it an awesome set of templates, like you can't deviate from this template unless you ask me permission and you better have a damn good reason. It will make it in your template, fitting your rules, and you will be able to read it faster and better because you know what to expect. I think there's two bottlenecks in using AI as well today, Peter, and I want your opinion on this.

The first bottleneck is the human giving it data. So this is why I moved from typing to speaking. The second bottleneck is me reviewing its work, the QA, the what do we choose, you know, the I joke the tinder, tenderization of the AI up, but yes, no, yes, no, just move fast. Those the humans are now the bottleneck of like the I.O. of the human, how fast can I talk and how fast can I read is the new bottleneck. And all of my work today is in how do I reduce the friction between those two bottlenecks? Because the LMs are so fast on the other side, it can iterate and work good or bad.

Towards a goal and it is still faster than me.

Peter Tuszynski (46:33.54)
Yeah, yeah. And and to to to speaking part, I think I it's I I don't think there's any other protocol faster than just speaking at an LLM. But you can provide it with so much, you know, ambient captured data. You know, like like you said, you guys are recording every meeting that you do. I was super passionate about this wearable limitless that was acquired by Facebook and shut down that you just do a pendant that you wear and it captures all the conversations that are around you. And

DeAndre + Grant (46:59.831)
Yep. I just bought a plod last night that has 20 hours of constant recording. So you just charge every night and upload it. Yep. Okay. We're in the same boat. I Okay, great. So we were all AI pilled collectively, which is really good. but what the question, so let's let's as we wrap this up, the the question is, why are we? Because I've gone through this valley of like imposter syndrome, right? Where I didn't pay attention enough I felt like and I felt behind the curve. And then I felt too AI pilled that like maybe I am hallucinating and going through AI psychosis, like this isn't as good as it is.

And then you go, you through the roller coaster again, like, no, it's heading there and like I need to figure out how to make it efficient because five percent efficiency gains isn't linear, right? It's not two plus two is four, it's exponential. So DeAndre, where are you living in this space? so I a couple points here that I think are really interesting, around the voice, like text to speech, right? And I was having a conversation with my fiance the other night and I said, I think that the value of written word is just completely being diminished.

Right, because it's really been commoditized. Yeah, having great grammar and articulating yourself and storytelling via written text. Think about LinkedIn, right? LinkedIn's a prime example. We now have a button on LinkedIn that says mark as AI slot. I love that. Because it's been completely overrun. Because the pred like the whole idea of the platform was just professionalism. And professionalism obviously was captured in writing something very professionally. Professional writing. Professional writing. And I think it's really gone the other way. And I think to that point, like

The thing that I will engage with the most now from a from a content perspective, if it is written, is something that sounds a bit like I'm talking. Like it it almost doesn't need I I can't have it be perfect. Philip and I live on the go-to-market front, right? How much how often do you put any amount of like general slang in an email, right? Because that just means that it's a human behind the other behind the wheel versus versus an AI. And so the the more and more that I see, I think is a total side note. My mind went to if you're, you know, if you're listening and you're creating like LinkedIn thought leadership content.

you know, online, you should just te talk to text every morning, create an idea and then just post it. Like that should be it. It doesn't need to be take time to like write anything out. That's why we do this podcast for everyone actually listening. Like the reason we do this podcast is because all of my LinkedIn thought leadership is actually my words. Right. I've said them into this microphone right here on this podcast. And even if it's text on LinkedIn, the AI is not generating it. It is just pulling out of the transcript. And that is the I think the the human like the table, right? The hand carved circuit board.

DeAndre + Grant (49:23.519)
This is the go to market version of that. It came from the human. So I think that's kind of where I sit. I think that just the opportunity is more and more invoice. It's just capturing everything in real time. the bottleneck, I think, fundamentally is context, right? And so how do you have all the context possible? I can imagine there might be a a world in the in the distant future where everyone walks in, they clip on their plod pin and then everything is recorded, right? Down to the whatever you're whoever you're meeting with, the people talking. Maybe they take it off when they, you know, have a phone call with someone personal kind of thing. But

I just think that the we're gonna start feeding these things more and more context so it can just be that much more effective. curious what it looks like. But a couple questions to wrap us up here, because I think that we could go on and on about how my how how we're using it. But it's so funny, Peter and Grant, that you guys are both voice. You guys both have a plot pin. You guys are both doing voice to text. So clearly I see Grant as being kind of bleeding edge, really leaning forward. Obviously, you are as well. So I guess if you're not using a plot pin, then you are officially behind for all those who are worried they are.

You are, unless you have a plug in. we're not sponsored or anything. Yeah, we're not sponsored by Plot Plot, if you'd love to sponsor. Please reach out. if not, we're still gonna buy your product. But lastly, advice for people, founders, scale ups, building in hardware, connected devices in the world of AI. I guess Philip, feel free to to chime in on here as well. Just what what are the core pieces of advice that you give to give to those folks?

before they you know set out on their journey or if they're in the midst of, you know, a product development cycle, just your general advice you always like to share.

Peter Tuszynski (50:56.048)
I would just say, you know, something super simple, but revisit your opinions frequently. The fact that, you know, something h hallucinated for you yesterday does not mean it's gonna hallucinate.

Filip (50:58.652)
you

Peter Tuszynski (51:10.104)
Yeah, town line, right? So I think you r you really have to revisit your opinions, try and come on and try it as soon as possible. And

Peter Tuszynski (51:20.718)
Yeah, so so so so just just be out there, be on the bleeding edge and and as long as you capture, you know, those things with plod, you know, give it to one model, then give it to another model. like you guys said, have it structured in a similar way so we can compare it, right? And and be able to tell which one which one is better, you know, which one is better for you, right? It doesn't matter that it it it can be, you know, my model can be good for me, but not for you. So I think

with the abundance of these models out there right now, it's really important to not lock yourself into one specific, you know, vendor or lab, but just try a lot of them.

DeAndre + Grant (52:01.323)
Yep. No, and I can echo that. I mean, I'm starting to play with local models a lot more than I ever thought I would, a lot earlier than I ever thought I would. And what I'm doing today is just offloading all the menial stuff, right? That I know it doesn't need the intelligence behind because I'm trying to save my claw tokens for smart things. And also I'm just trying to learn. And this is the message on tech home is hey, yes, local models are cool. Please get into them if you want to. They're fun. Be curious. my f my most what's the right word? I don't know, disharmoness.

Thought that ever happened is I was at a an event and someone asked a question of a panel of executives like, hey, why are these these token leaderboards out there? This feels stupid. Why are you like paying people bonuses for burning the most tokens and emitting the most carbon and doing stupid stuff with AI? And ironically, this panel of executives had a terrible answer for it. They like, it was like a big non answer, like, we're just, you know, following the curve. And it hit me in the crowd. So I like raised my hand and spoke up. like,

And I ironically, it wasn't a preconceived thought. It just hit me. Cause I had been on vacation the two weeks prior and I brought my laptop to the beach and I had so much fun just playing with AI to see what it could do. Because I there are there's no rule book. No one knows what this does. And I came out of that two week vacation, like 180 degree change. Let's implement this. There's reasons for it. In a way that all of the consumption of media online, of podcasts and reading news articles of what other companies were doing didn't teach me how to use it. And then I realized that the goal is play.

You actually have to be curious and play with this new technology, not work with it. When we are all using it for work in the moment, I have it deliverable and this is the deadline and I do it. You don't experiment. You don't try weird stuff. You do what you think is the most efficient. And what is the most efficient today is nowhere near what the most efficient thing is tomorrow, next week, next year. And those things are discovered by play. So these token leaderboards were brilliant executives, literally tricking their employees into playing with a tool that they didn't need in their day job yet.

Before it was cool. And they probably are the people that discovered the best use cases. Now, they made 999 use cases that absolutely suck and burn tokens to do nothing other than like put cat emojis in Slack. But there are those one or two percent of use cases that are now deployed across those companies that are saving them risk or money or accelerating development. And what I want everyone to hear is AI is hype. Yes, all of us are AI pilled. And I'm sorry that we all talk about it because we think it's neat. I would encourage everyone to go play with it, not for a deliverable.

DeAndre + Grant (54:27.063)
For at least some real amount of time. That's the takeaway I would like to give everyone. Philip, any closing words on your side?

Filip (54:36.753)
Yeah, one thing actually to build up on top of what you already said at the beginning, you mentioned that 98 % of the experience is no user experience. So for every, you know, young founder and people who are starting their companies, like try these tools, build prototypes and talk with your users as early as possible, because maybe a year ago you would need to spend a couple thousand dollars on, you know, building it with agency of somebody else.

but go ahead of the curve. Talk with your users with something that is clickable. Get their feedback and build something for them that is tailored, that is super niche. And then you can experiment, but keep users in mind and prototype as early as possible.

DeAndre + Grant (55:16.095)
Yeah, the painted door test is so much deeper than it used to be and so much cheaper to put out there. Just paint the door, see if people want to knock on

Amazing. Everybody, this is the Hard Tech Podcast. Peter, Philip, thank you guys so much for jumping on. tune in next week for the next episode. Thank you guys so much.

Peter Tuszynski (55:24.977)
Yeah.