Automate mini series: Whitney Sales

Welcome And The Automate Series

SPEAKER_00 0:00

Welcome to Automation Ladies, the only podcast we know of where girls talk about industrial automation. Another automate series, short episode. Today we have Whitney Dills with us from Thought Forge AI. I had originally reached out, I think, to one of the co-founders of the company, whoever was listed as a speaker at Automate. Was that, did you have a talk, Whitney, or was that somebody else? No, that was me, CEO, Matt. Okay. Yes. But then, you know, sometimes people see the automation ladies and they go, hey, don't you want to talk to a lady? Which honestly, yes, I do. But it depends. Not only for the sole purpose of you being a lady, right? And I do, you know, that's why we try to tell people, like, hey, it's called Automation Ladies because it's hosted by ladies. It's not about ladies. It's not just about women. We just like to talk about nerdy stuff and we found each other and we're friends. Um, and I generally don't like to be have like the one lady in the company shoved at me. But in this case, when Whitney got shoved at me, I was like, heck yes, I will take it. So, Whitney, thank you for coming on the show. Thanks. Thanks for inviting me. Absolutely. So these episodes are meant to be short, like 30 to 40 minutes-ish. And if I'm not careful, I will talk for an entire hour. So I'm just gonna dive straight into it.

Whitney Dills Joins The Show

SPEAKER_00 1:11

Um, can you tell us a little bit about yourself? And we talked earlier before we went on. I would love to hear kind of the founding story of Thought Forge. Um, are you guys solving this problem? What is it? Where do you come from? And and what does it do?

SPEAKER_01 1:26

Yeah, so uh my background, I actually was in venture before I joined Thought Forge. I had co-founded a venture fund. So I've been I've been a male dominant spaces and the lady shoved it at other ladies for a very long time.

SPEAKER_00 1:37

Yeah, yeah, you know what I mean.

SPEAKER_01 1:39

Yeah, I was uh I was a female inventor before the Me Too movement. So I'm co-founder in a fund. And so uh, you know, that's actually how I got connected with Thought Forge. One of our first investments was a company called Bonsai, which was one of the first commercially available deep reinforcement learning platforms. I do remember you. Yeah, my co-founders, Matt and Hernie, were technical and product leads there. Um and we had invested, we were the first uh first funds in uh that company before it got acquired by Microsoft. And so when they got started on their next thing, I was like, oh, this is interesting. Um, and wanted to get into robotics and was kind of done with venture. Um, I'd invested in about 200 different Sassin AI companies. I taught entrepreneurial sales at Harvard, um, like deep background and early stage go-to-market. Um, and uh, as I like to say, it's like beating your head up against a wall until it breaks and really missed operating and wanted to get back into it and uh saw robotics as kind of the next big trend. Uh, and what Matt is working on is uh had been working on for the last 18 years that basically was the foundation of Thought Forge, was really fascinating.

From Venture To Thought Forge

SPEAKER_01 2:39

So, you guys are really familiar with the world of AI uh in robotics and the challenges that come uh come along with it and the general perception that more data equals better model. Um, with Thoughtforge, that's actually not the case. So um, Matt has developed a new type of AI uh that's based on something called active inference, which originates out of MRI brain imaging. Um it's a guy named Carl Friston who actually founded the field. He's the most cited neuroscientist that's still alive today. Um but basically it posits that the brain is always trying to predict its sensory experience and its experience in general, uh, and minimizing the amount of energy that's used in order to do so. Um and so uh what Matt has created is a model that basically does that. So it's constantly predicting its state and updating in real time to uh the sensory feedback from the robot uh to update and adapt in real time to the task it's performing. So instead of uh needing this massive data set in order to train a model, we can train on about a couple thousand data samples, really, like a couple seconds of data. Um, models usually take a couple minutes to train. And then they adapt in real time post-deployment for the task that's being uh that it's actually performing. And so um, kind of fun story about Matt, he basically didn't believe backpropagation, which is how most models train today. Um, like whether you're talking about diffusion models, deeper enforcement learning, um, these teleoperations, for example, all of its backpropagation, which means you have to tell the model everything it needs to know ahead of time in order for it to continue performing a task accurately. Um he didn't believe that was actually how the brain learned, which it isn't. Um, and basically has set out on this journey post-MIT to figure out how the brain learned in like biologically plausible learning, um, and stumbled upon this field in 2019. And so uh, and went and reached out to the founder of the field and was like, Am I even in your space? He flew out to the UK, which is where he's based, and like presented to his PhD students, they're like, Yes, you're in your space. And Carl told him like you need to patent this and get it to market as fast as possible. And he realized the killer application for what he'd been working on is uh was robotic and robotic control specifically. Um, and so the tasks that we're really focused on in robotics, so they're again they're models that update in real time. Uh post-deployment use lots really very little uh low fidelity data that's simulated, um, are insertion and manipulation. So the you know, high contact, high context uh type applications. Um and uh they think like you can wiggle apart in and taking in real-time uh torque feedback and uh joint state feedback in order to kind of continue performing these tasks accurately is countering changes or feedback from the torque so it can kind of feel its way through a task. Um, and then we're partnering on the perception side. So um that's it, that's a little bit about us and kind of what we're doing and uh and and kind of our journey uh into robotics.

SPEAKER_00 5:30

That is well, a like obviously a really unique way to come at it instead of the models and and uh paths that everybody is taking, right? And and I would say then you're probably leapfrogging some of the competition out there quite a bit coming with this to market.

Active Inference In Plain English

SPEAKER_00 5:50

Yep. Um I mean, the lat latest, I just wrote an article on um somebody's claim of like the post-plc era about now using an AI controller instead for like motion planning and vision and all the things that you would typically see with a separate vision system, PLC, HMI, like the traditional stack of like controlling a robot deterministically, um having an all-in-one AI controller with an onboard GPU. So I guess some of the inference is running on the GPU and the training, and then the CPU, there are still like separate CPU cores that actually handle once the the motion and things have been learned by the AI. And I'm I'm not like I haven't researched this well enough to describe this.

SPEAKER_01 6:36

How many people have? Like there's there's three other companies who are really using active inference um in their in their modeling.

SPEAKER_00 6:43

Yeah, but that this um that still to me is like okay, you're kind of just abstracting the PLC onto some CPU cores on an IPC, uh, but it's still, it's I mean, you can call it a post-plc era, it's still a soft PLC in a in a in a way. Um but this is this is this is uh I guess taking it to a degree that I haven't done the research on yet on any companies like yours. And it makes sense then if there's only you say that there's about three of them. Um so that's kind of the minority of the companies in the space that are doing this this way.

SPEAKER_01 7:16

Yeah. Our models, um, I didn't mention this before. Our models are really small, so they run on a model of 250 to 500 kilobytes in size. So they only need a CPU to run.

SPEAKER_00 7:24

Yeah, I was about to say, but the little bit of background research that I did did say that you can run this all on the CPU on the edge. And you because one of the things that like struck me about well, the risk, like the supply chain risk of also trying to have a GPU on your uh like the the hardware requirements for like the stuff that we're used to, and then like competing with people using GPUs for all kinds of other AI training loads and data centers. And again, this is not a super well thought or researched thing, but like my I don't know, my spy my Nikki sense says that that's substituting one risk for another. Um that I don't I don't know the implications of that, but like to me, I don't know that most manufacturers want to be worrying about now having to have a GPU out on their floor, like in the controller of the robot. Yep. Um yeah. So maybe tell me a little bit more about that, that differentiation between because there's like physical AI is what everybody's saying now. Like I feel like I didn't get much of a chance to walk the floor at uh automate, but the little bit that I did, it was like physical AI is here. This is the year of the physical AI. And I think that's probably true and why everybody was exhibiting that, but like to the untrained person that doesn't know much about AI but knows about industrial automation.

unknown 8:42

Yeah.

SPEAKER_00 8:42

Somebody's telling you that now you can use AI to pick parts from a bin without having to orient them. Right? Like that's the that's the application they care about.

SPEAKER_01 8:51

Can you everyone can palotize? Yeah, yeah. That was like an automate thing. It was like everyone can palletize, everyone can pick. Uh it's still there's still the issue around localization, though, when they're picking because the camera, because you know, when cameras get too close, they have a hard time, you know, they have a hard time localizing.

SPEAKER_00 9:07

But but it's like to a controls person, the the differences in how you accomplish this with AI are not necessarily immediately apparent because the marketing is just we do it with AI. And it's like great. But there's a lot, there's a lot of, and this is why I've been kind of like cautiously interested and skeptical in the physical AI it claims over the years, because I used to work in um AI, quote unquote, mostly machine learning, but at the time where neural networks were starting to become part of architectures in enterprise AI, and I was doing it for forecasting demand and demand planning um in in retail applications, but basically, you know, sort of the big data uh applications at the time. And there is just so much that I didn't know until I started working in that area about like just the nuance in different types of AI and how they work. Like you were talking about backpropagation, right? And it's like if you don't know the difference between different types of models and how they work and how much they need to train, are they deep, are they wide, are they both? Are they, you know, it's it's all about and and just like thinking of a machine, it's like it's not one component. It's how you put the components together, how they work. What process are you trying to accomplish? There's no like silver bullet model for anything. Um so I don't know. I would posit that most of our audience is not AI native or or hasn't been working in that area for a long time. Um I'm sure like automated with Brian Heater is where you're gonna get the audience that like knows your stuff in and out, knows all the no one knows our stuff in and out, startups, like all that kind of stuff. My audience is like the I don't know, the the controls engineers, the plant engineers, the maintenance checks, like the people that are gonna have to work with this stuff once the management decides, I'm gonna stick this in your plant. Yep. So yeah, just with that in mind. From that angle, a little bit like what's the secret sauce that you guys have there and how is it

Skills Based Control With Goal Functions

SPEAKER_00 11:10

better?

SPEAKER_01 11:10

So most AI today, um, like like I said, like 99.9% of AI today that you're gonna see uses back propagation. And then there's a layer on top of it of how back propagation is basically you tell the model it knows it, and it uses a uh a reward function and weights, various weights is it and uh it's throwing a bunch of different things at it to try and figure out how to get to the end result that it wants, right? Essentially, um, or that you're trying to get to. Um, and so what it ends up doing is creating a ton of model bloat. And then when it's close enough to what you need, um ideally like you've covered all the edge cases, um, you can deploy it. But then sometimes like with these high context, high context type tasks, you need a ton of data for them. Um and the haptics become an issue, physics, kinematics become an issue, and that you need really, really high fidelity data for all of these. Um and so you end up your model just ends up bloating and getting very, very big. Um you end up having to have, you know, um, uh, like you said, a bunch of GPUs next to it in order to run the model, which creates latency in the performance. Um, but if you're talking about diffusion models, which are teleoperation-based models, meaning a human, human tells the model what to do. Um, and when there's an edge case, you go back, you have the human teach it that edge case, and then it goes back into the model, right? Um, that is how, or uh deep reinforcement learning, again, is kind of what I was describing earlier, or reinforcement learning, uh generative models, all of it kind of falls into this backpropagation-based bucket. Okay. Um it's like 99.9% of what you're seeing out there. Um what we're doing, it has a reward or it has a goal function versus a reward function, and we're developing skills versus task-specific applications. And so, meaning we could take a deformable object insertion, um, think a rubber part, right, um, that you're inserting, and we could take that same model and um on robot do training some training on it, or do a couple minutes of simulation training and then put it on the robot and have a robot kind of it understands what it's trying to do, and it can actually go and figure out how to now take a plug and insert it, or a connector and insert it. Um, and we took that and we added some more context to it, some a little bit more simulation training. And now it can do some wire harness and cable manipulation, some lightweight. We're still doing building that out. Um, but it doesn't need to know all the potential edge cases. What it needs to know is the parameters or boundaries. Think like bumpers in a bowling alley, right? Like, but we put up like, think like for us, it's 20-foot walls because it can never go outside it because it doesn't have that data to go outside it. Um, but it has uh kind of the boundaries in which it's allowed to operate, and it can adapt within those boundaries to continue operating to get to its goal. And so it will never do anything that's uh not like outside of what it's been trained to do be or and outside of its goal function. And it, but it will continue operating within the parameters of its goal and to perform the goal that it is designed to do. In this case, like one of our uh models we deployed early with BP, we have some papers on it, so I can talk about it, uh, was to manipulate a high pressure gas valve. Think um edges could be worn off, uh corroded. Uh you might need it could be rusty, so you might need different torque. It could be covered in oil or grease. Um, they never knew the condition of the valve, and they needed to consistently turn that valve uh, you know, a certain number of degrees. Uh it's high, like three out of every thousand tasks according to OCI, there's result in an injury. So high injury rate task, perfect for automation, right? Supported by unions, yay. Um, and so uh we took that same model and put it on bolts. Um, it's the exact same model kind of transfers really nicely. Um, and we can cover a bunch of different shapes and sizes of bolts because the goal function of it is essentially the same task. And then the perception model, so the vision model that's associated with that is telling us what part it is that then set the parameters around like how much it needs to grip and et cetera. Yeah. Um so it's it's we kind of play in the control model side, so non-vision, so all the physics and kind of kinematic side of it. Um, close your eyes and try and do something. That's kind of where we are, and that's kind of the functions that that we do. So um we have some training videos that we show people um around uh the rubber part insertion that I was talking about, where you can actually see it feeling its way through to do an insertion. And I describe that as like trying to plug in a plug behind your bed. That's essentially what the model's doing, and that's to account for errors in movement or lighting changes and things like that, where it can kind of feel it feel its way through doing a task. So you can you create a lot of robustness in what you're performing on top of vision. Um so we complement existing models that people have designed around vision um to add robustness to them for a lot of the more edge cases that you run into and not have to collect all that high fidelity data around that. That makes sense.

SPEAKER_00 16:03

Yeah, absolutely. So you're that's kind of your part of it. And then you have you said partners for the perception. So those would be vision companies.

SPEAKER_01 16:11

Yep.

SPEAKER_00 16:11

Um kind of sorry.

SPEAKER_01 16:13

Oh, go ahead. I was gonna say any vision model we can work with, um, really. So um, and like there, there's a lot of vision that's been solved, um, but it still needs there's a lot of data that's required for it. We kind of reduce that amount of data um is one of the major benefits that uh people like and we could retrain for new parts like in a matter of hours on robot um all on edge. So it's really good for high security environments um and where IP is protected because we can also train in representative parts and then transfer it on site uh to the actual parts. Um so we work really well in the defense sector. It's kind of where we've been uh our bread and butter industries, but we also work in uh so defense, aerospace automotive in our big industries.

SPEAKER_00 16:56

So is this being deployed on any existing robots to make them better or to make them more flexible? Or are you working exclusively in new installations?

SPEAKER_01 17:06

Um we have primarily been greenfield um in uh greenfield applications, things that you can't do today because candidly, if you can do it with the existing automation techniques, you should. Um it like we're not trying to point. Yeah, yeah. It's uh we're we're not trying to reinvent the wheel. Um, we're trying to unlock applications that you can't do with existing tools because it's either there's too many edge cases, um, maybe it's too high mix, so you can't justify the ROI. Um, the latency becomes a required, like because cycle time and latency is a really big issue in robotics, um especially when you look at a manufacturing setting. It's why you've seen all the humanoid companies actually moving over to consumer. Um, is they don't they can't perform like if if I'm having a robot fold my laundry, I don't care if it takes five hours to fold my laundry as long as I don't have to do it. But if I'm in a factory and I have to have them fold something, they better be able to hit the human times and um and so uh that's kind of like a little bit of where we've been we've been playing because our our models can update in in milliseconds. And so they're they can add kind of that that cycle time speed and robustness and reliability that's needed because we can handle those edge cases.

SPEAKER_00 18:22

So I was just recording an episode earlier today um with a gentleman that they uh help technicians. Um they do digitization of documentations for electrical panels and like maintenance instructions and different things like that. And a big part of what why they do what they do is also like safety, because if you have outdated documentation and the the panel like design or the the drawings are wrong, like or the lockout tag out procedure wasn't done right, like people get injured in the in these. And we were just talking about like the humanoids, and I was like, and at what point does like do you actually have then a robot it go into the potentially dangerous situation where like if somebody gets blown up, it's gonna be them. And we were like just kind of talking about how none of us really know like these timelines and whether or not humanoids are gonna be the thing. But like when you're talking about the those rusty valves and like inserting, and that just made me think of that. Like those are cases where you're not gonna have general purpose automation or not general, like specific automation stuff built for that, because it's not a thing that gets done every day, all day, every day, it needs to be done fast. It's more of like a you need a certain type of manipulation, you need a certain amount of fluency and understanding and what is the task is being done. And so we were like, hey, everybody, you know, like you can go into the trades and technicians, we need more of them. Like they're aging out, people aren't going into these fields. You can now, if you're uh a UIUX designer or uh just a you know software programmer, like good luck if you're not architecting and using AI to like be a team, right? That is no longer like a safe employment choice. Like you're you're gonna be employed forever if you know how to code. No. Um, and then the question to me is like, can you be employed for and I think the answer to me is like, there's nothing that tells you you can be employed forever. You have to keep learning and you have to keep 100%.

SPEAKER_01 20:15

Yeah.

Cobots Today And The Humanoid Gap

SPEAKER_00 20:16

But do you see your models and the ability now to do some of these tasks that haven't been feasible to do in the past? Are any of them in the humanoid form factor? Or are we more so talking like cobots or or what's the hard part you guys are?

SPEAKER_01 20:30

100%, 100% cobots. Um, we've been we've been on five degree freedom, six degree freedom, and seven-degree freedom arms. Um, primarily, we did have some conversations at Automate for like a 27 degree of freedom hand. Um and Matt is very excited to get to start working on that kind of stuff. But candidly, um, we have been primarily focused on there's a so there's a this gap right now that we're seeing, right? So, like, okay, let's say the future is humanoids. Great. Like, okay, human form factor, everyone wants humanoids because hey, I can watch a video, a YouTube size video video data set, right? Watch a video, train the model using teleoperations and diffusion models, right? Um, cool. I get that that's where the internet side data set is, but physics, you can't get the physics from those videos, right? Like there's this, there's this gap that exists in the market in order to do it. So there's, you know, companies like Hapley that are trying to capture haptics, but like how uh a rubber part like squishes, like how squishy it is, like that's it's very hard information. You can't get that information from video, and it's it's you can't simulate it. It's just it can't be simulated, or how like stiff a harness is because when it was manufactured, one person used more tape than another person when wrapping the harness and when they were manufacturing it. So slightly differently, like all those kind of things you you can't simulate. None of the simulators, like simulation platforms today, actually. Can fit simulate physics very well. And so you really have to capture, and it's high fidelity data you need in order to do it if you're trying to use traditional automation. Yeah. Um, and so like you really need something that updates in in real time in order to perform those. Like that it's that is our thesis and our strong belief system around it. So we've had a lot of people ask us to like use it to train existing models to kind of fill that gap. Um like to capture that data. And it's like, yeah, we could do that, but like, why would you just create model bloat that's unnecessary? Yeah. Like you can have something that will literally just keep updating in real time and knows what it's trying to do, as long as the task that it's trying to do maintains the same is the same, then it's not you're gonna, it's gonna continue performing the task. Um, like you can even add up to that, um, the kinematics in the robot, or excuse me, the um the harmonics in the robot. So if something gets off, it can adjust to that because it works like your brain does. Um like if we hurt our elbow and our arm, we're gonna continue, uh, you know, we can figure out how to do something. It's the same kind of uh learning process. And so um, like the coming back to the humanoid conversation, um like they're not ready, let's be honest. Like even the the three-finger grippers are uh, I think Google licensed the the three-finger gripper it designed to another company to start doing testing industrial like industrial testing on it last year, uh maybe the year before. Um, and even that's still, you know, getting to a point where it's ready for industrial use where it can reliably perform, you know, a task 8,000 times in a day over several years, right? Like you need that level of reliability and testing um in order to put it into a manufacturing setting, um, let alone the runtime, like how long it can run for and not need to be recharged, because God knows you don't want to leave a cord laying around in a facility, like that's something someone could trip over, let alone the robot if you have food versus you know, um, it's on rollers. Um so you know, there's we have our thesis, like that there is a use for humanoid robots. I a lot of people are saying the manufacturing setting is going to be one of the first areas that it goes into. I think it'll be healthcare. Okay, the human um form factor, like even just collecting data and observing data and like serving as a support for nurses would make, I think makes a little more sense than it does to me than it does in a manufacturing study because of latency, performance, um like the actual environment is not like yes, you have to walk around, but like have you seen those things walk around? They're not like they're not super stable. Um it's like, why put why are you putting it on legs? Like, there's no point in putting it on legs. And then there's stairs, and like, you know, and then if you're looking at using it to train traditional model, like using traditional models that are back propagation, like say there's a bump in the floor, like and it needs to go somewhere else. Like there, yeah, you can do the vision model, but like how is it gonna, how is it gonna adapt? Like you need you need something that can adapt in real time. Like I like that's just a strong belief. Like, I don't know everything I'm gonna need to go and need to know. And as you were saying about jobs, like you need to continue learning. Like, why shouldn't our models like continue learning, which is kind of one of the trends that I saw at Automate 2 was um continuous learning. Um the new

Continuous Learning Data Ownership And IP

SPEAKER_01 25:19

trend.

SPEAKER_00 25:19

Um continuous learning, let's talk a little bit about that. One of the hesitations I've heard from end users on any of these AI control things is also like who owns the data? Where is it going? Does it have to go to the cloud? And if I come up with a new task or a motion path or something or other that is valuable to me, um, does your model take that and then train other robots on that? What where does that IP live? Does it belong to me? Does it belong to you? If I choose to stop paying your subscription, does my robot stop working? Like, can you tell me a little bit about kind of the the ownership economics and how that relationship works between you guys as the technology provider and an end user? And do you work with systems integrators or do you implement this directly at the end user? Like, how does that relationship work? Because we all know that this is like you don't just buy one of these things and throw it in and expect it to work for the next 30 years without any involvement from no goddamn.

SPEAKER_01 26:16

No, not at all. Um so the way our models work, we consider it like an 80-20 rule. So um 80 80% generalized, um, and then 20% specialized for the specific task that it's performing. Because it's a skill. Again, it we're not doing um general models, like that's not our world. We're we're skills. Um and the continuous learning side, I mean, I'm sure like the reason they're doing that is because uh there's a general perception that you need production level data in order to do production level tasks.

SPEAKER_00 26:51

Yeah.

SPEAKER_01 26:51

Um, we don't. Okay. So we again, it's why we do really well in that kind of these high security, high IP environments. We just don't need it. Um we trade on simulated data, representative data, and we deploy uh in facility on their specific um data. Now, there is there are some things that we learn from our customers around performing the task and like the challenges if they've tried to use traditional automation, and to a certain extent that could be considered IP, but um the tasks we're doing are our general tasks before we go uh and and uh specialize it for a particular or generalized skill that before we specialize it for a particular task. So we don't really run into that issue like other companies do. Um there uh like the reason everyone looks at it data is because VCs, I like I mentioned my background is in venture, and VCs look at data as the moat, right? Like they always have. Um and so because it makes the stickiness really high on a piece uh on a piece of technology. Um like imagine trying to switch off of your sales for as a salesperson or your sales force or your ERP or your vendor management system. Like once you're on, like it's like pulling teeth to get me off of that thing because especially as an enterprise organization, like training your team is 50% of the cost and getting them to adjust it. Um and so that's just you know, from a from that perspective, that's why there's always this battle kind of over IP, and it's actually tanked a lot of companies from what I've seen in in the past, um, because they can't scale up the solution to other customers. Yeah. Um the customer ends up owning the IP because it's IP data.

SPEAKER_00 28:32

Yeah.

SPEAKER_01 28:33

Um our models, we have background IP that we have patents on. So that's you know, that's beneficial. Um, we have a paper coming out on IEEE this week this year that's kind of some light, some early background um pieces of our IP, but doesn't actually give up give away the IP, but it kind of explains what we're doing. Okay. Um a little bit more about what we're doing. But um the we're not collecting any customer data, we don't store any customer data at all, runs on edge. So nothing relieves the facility, so it's only secure as your facility is. Okay. Um, and so we don't really run in into that issue at all. Um, but it is an issue we are some of our partners run into. Um, and so we talked to them about that, but uh they kind of they've been kind of going with the 80-20 rule as well, where it's like as much as you can generalize as a skill set, and then it can learn on the customer's IP um for the actual

Off The Shelf Hardware Integration Strategy

SPEAKER_01 29:23

task.

SPEAKER_00 29:23

Do you guys have specific uh partners for the hardware, or can it be whatever customer chooses to use?

SPEAKER_01 29:30

So we a lot of again, this comes into another VC issue. Uh VCs want you to do a vertically integrated solution for reliability and because there's a lot of costs when it comes to integration. Um uh and when you're specializing for a particular task that you're performing, um, you should be able to go vertical, right? Like, but the problem is your customers um are familiar with certain types of technology, and you were talking about who the people listening to this podcast are. Like, they don't want to learn a new piece of technology. No one should have to. If you need to shut something down and fix it, you want something you know how to fix. Like, and so if you have a vertically integrated solution with new technology, like we saw a bunch of stuff at Automate that is brand new technology, like uh they design their own end effector, they design their own controller, there's like different, different new types of hardware. Um and if you require someone to go and learn a new piece of hardware and how it operates and works and they have to troubleshoot it because it they need to fix it, they're just gonna rip that sucker out and not want to adopt it. So you need to use things that people know how to use. Um, and so we try and go with everything off the shelf. Um, they've tested, like getting AI into production, I call it like it's like a drug discovery process. Like it mirrors a drug discovery process much, much more than it does a traditional software. Um because there's, you know, there's the all the RD work to actually build the AI. Then you go into like lab testing, right? And then you go into like your animal testing, your human trials, right? Like there's all the usability, there's the reliability. Like, where are all the edge cases? And um a lot of people don't realize that. And so that's kind of where people want these vertically integrated solutions because like once you've done the full end-to-end stack, you know that it's like you as an automation company or an autonomy company can reliably perform. But the end user adopted adoption is almost more important than like, yes, it's the robustness gets you past to the end user adoption, but they're never going to use the damn thing if it if it doesn't work the way they understand and they aren't gonna want to learn a new piece of software or a new piece of hardware and software for like a small piece of their entire line. Um and so we've gone with everything off the shelf. And we work with um all industrial grade robots. Um we're not trying to use special end effectors or grippers, we try and use generalized end effectors when we can. Um, we try and use built-in force torque when we can. We are actually really surprised the UR force torque was as sensitive as it was. Um we start with cobots and then move on to the industrial robots. Um and uh for the industrial grade robots, a lot of it, if we can use them, has to do with uh what kind of uh sensors we can apply to it for the tasks that we're performing because we do need certain levels of sensitivity. But again, we can find things that are off the shelf and that our customers are familiar with and have tested.

SPEAKER_00 32:17

Really, really interesting. Okay. I now wish we had a ton more time, but it sounds like we're gonna we're I'm gonna be staying in touch and uh paying attention to the stuff that you guys are gonna be putting out.

Automate Takeaways And Real Adoption

SPEAKER_00 32:29

And then I'm assuming we will see you again in automate next year.

SPEAKER_01 32:32

Oh, 100%. Yeah. Well, uh, we're there every year. We usually have a small booth in the back, but we're usually pretty busy. Um, but like this year, like compared to last year, oh my God, so many.

SPEAKER_00 32:43

This year was, I would say uh everyone that I've talked to, like this was like the best automate ever, which it does seem to get better every year. So that's not a huge surprise, but just uh the like my booth was so busy. Yeah. And almost everyone that I've talked to was just like they're I mean, people are just they're they're hungry, they're searching, they're prepared. They're not just like aimlessly walking the floor and asking random questions, but like these are people coming to you knowing that they want to see what you've got. Um and that's that's great. Like I feel like our industry, granted, like the economics and everything, there's gonna be probably possibly some contraction in certain areas next year. And you know, there's there's always automation touches many things, right? And certain things are probably gonna be going down a little bit as much as other things are gonna be up. But the the economics also say then is if you're a plant owner, right, or if you're a PPG or you're a manufacturing company that may be seeing a little bit of a slowdown next year, that's also exactly the time that you need to look inward a little bit. And any extra capacity that you have or any extra potential plan downtime that you can have over busier years, that is the time also to invest, right? So that orders pick back up, you have capacity, you can make better product than your competitor, you can do so less expensively, you can be more. I mean, you you're if if you're slow at all, this is also your opportunity to tackle how the heck do we face how do we hire when things or do we have to hire as many people again when things pick up back up? Right. The labor force situation um is only getting worse. And now that we're competing with the centers for the skilled labor, uh, you know, just everybody this is the time. And and this is also, I guess, the time, like you said, this is the year of physical air where I think more of it is leaving the demo POC stages, um, and not just super early adopters can actually get some value out of spending some serious time considering these, and maybe things that weren't automatable in the past from an ROI perspective, like you said, either tasks that weren't possible or they were too high mixed, um, low volume, like the bar for those projects is getting lower and lower every year. But I haven't seen like a big leap shift over the last like five years. I'd say it's kind of like, yeah, you see the same stuff, and it's the demos are getting more impressive, but they're still mostly demos.

SPEAKER_01 35:03

They're still all in cages. Yeah. We're trying, we, we our ours is all fixureless and out of cages. So it's like you can come and mess with we we like to mess with our robots. Um it's like put them to the challenge in real world environments.

SPEAKER_00 35:17

Um I really wish that Courtney had been here for this conversation. She was like, man, so we finished up our conference last week, and then I'm I guess I'm the only one that's flexible enough that was like, Yeah, I'll take some uh some interviews today, but Allie's feeling sick, and then Courtney was like, Well, my first day back to work after a week off, like there's just no way that we can make it happen. Um, but she would probably want to talk to you all about the the mechanics and the torque sensors and um because she's yeah, working on a bunch of stuff that she can't talk about either because of NDAs and whatnot. But I think this is the first time also she's not just taken an off-the-shelf robot and made it do something. She is now actually having to work with off the shelf with like robot hardware and and making it do stuff rather than the robot company's controller and just making that do stuff. So she's she's thinking on all kinds of different levels right now of robot. She used to when she was just doing, or not just, but she was doing training for UR and uh, you know, kind of more in that traditional uh let's take a cobot that's established that has some standard software and like you know, make it do a thing.

SPEAKER_01 36:27

Yeah. I was gonna say that's a good conversation for my co-founder, Harne. He uh used to teach robotics at Duke University and he actually worked on the first Roomba. Um, the model for the first Roomba, which is pretty cool.

SPEAKER_00 36:37

Um well, maybe maybe, and I say this a lot, I probably more often than I should, but I'm like, we can we can have a follow-up episode. Uh one schedules allows, but um, yeah, it it's I would I would go on forever, but I think um coming up on we're a little over half an hour here, Whitney. If there's anything that you that I didn't ask or we didn't cover, what what should people know about you guys um that should be included in this conversation before we sign off with our traditional where can people find you? What should they look for? Um and you can skip straight to that if if that's all we got.

Career Advice For The AI Shift

SPEAKER_01 37:13

No, um, I mean, there's a lot of trends in robotics. I would the big thing I would say is a lot of people are worried about uh robots taking their jobs. I would not be worried about that. Um, the jobs that robots are doing today are probably jobs people don't want all that much. Um and the it's you know, you're talking about software engineering earlier and that kind of going away. Um, the thing I've seen is it's more uh software engineers are turning into managers and supervisors of the AI. And I see it being very similar on the robotic side where you can do a lot more with less. But um we on the job growth side, um I actually have seen from yes, there's layoffs that are happening, but it's usually um managers weirdly, uh, but it's people without the technical uh chops to actually be able to work with AI. So it's just the sooner people can adopt it, the better. I have a 16-year-old niece um who wants to get a degree in business. And I was talking to her and I was like, she's like, oh no, not AI, because there's schools been pushing them, you know, not to use AI and actually learn critical thinking and things like that. Um, I know in college they'll push her hard, but um I was just like, sweetheart, you like, I love you, but you need to learn how to use AI. Like you need to know what it's capable of and what it's not capable of, and what to use it for and what not to use it for. Um, and the same thing goes in automation. Like you were talking about the physical AI world. It's still relatively new as a field in reality, like 20 um 15 was really when it came about. And so when you look at that, there's really only a few people um that when we're working with them that we actually look to because who understand robotics and AI. And knowing that, like as people who are in in in factories, like learning that how AI works and the robot works, or how the AI and the robot work, it's a really specialized skill. And you you will have a career for a very long period of time uh into the future if you understand those two things, because there's very few people who actually have taken the time to learn about them. Um, and so that's one thing I would just like I'm just encouraging people to do because a lot of people don't understand the all the requirements that are needed to even automate a task. Um, because we talk to them and it's like, what do you want to automate? And they can't tell us all the challenges that they've they've tried on why it can't be automated. Um, and so if that that's who we're always looking to talk to, um, is someone who actually has that kind of experience who can tell us those things. And so anyone who's looking, you know, at their career, um, I'd encourage more people to do that. Uh because we can like what I was telling my daughter.

SPEAKER_00 39:47

So my daughter's eight, and and like maybe two years ago or three years ago, it's like, you know, you don't ask kids seriously what do you want to be when you grow up. I I was have hated that question. I also think kids don't know enough about what's actually out there in employment.

SPEAKER_01 39:59

I wanted to be a singer and I I have the worst voice ever.

SPEAKER_00 40:02

So let's I wanted to be a professional skier um at some point. I I found I found a survey on the internet that I filled out when I was in like fifth grade. Um, and that's what I said I wanted to do, apparently. Um I did use ski when I was a kid, but I was always pretty bad at it. I'm not good at sports. Um and I and yeah, I wanted to be a lawyer at some point, I wanted to be an anesthesiologist, like just all over the map. Um, and then I ended up with just like, well, I never know what I'm gonna do, and I'm just gonna do something that probably facilitates some ongoing learning. And so I went to business school, and then I was like, I like to do technical stuff. And like, I personally have always found interest in combining a couple different things. And so, like what you said, knowing about AI and knowing about robotics, it's like it's one thing to know about AI, but then you got to know how to apply it. And so, like, I have a background in machine vision, and then when I was working in AI in 2016, there's a lot of people doing computer vision. And those things kind of sound similar, but they're absolutely not because one relies on actually in the physical world getting a picture and then making a decision based on said picture. And the other one involved getting a bunch of pictures from the internet and trying to say, is it a cat or a dog? And like both things are useful, but they're entirely two different things. And then, like, what value do you get out of knowing if this is a cat or a dog? Like, that's that's a piece of value that then needs to go into something that produces something else. And so, like somebody knowing about, I remember there was a guy that like he came up with a more explainable neural net and was trying to find like a commercial purpose for that. And then maybe I talked to him like two years later, and he's like, Oh, I came up with this neural net that requires less training data. And like it's so valuable because, and I'm like, Well, you want me to sell it to who? And it's like people that are you doing it already, but they could do it better. And just like you said, it's like if you're already doing it with something that works, like there's not a lot of cases where shaving a little bit of time off of the training is going to be worth your while, like substituting the models or changing it or adding to it, or like, I'm like, dude, what like not just like, oh, you came up with the cool novel thing, but like what the heck does it solve?

SPEAKER_01 42:18

Yeah.

SPEAKER_00 42:18

Anyway, this is a very long-winded way for me to come back to my point about my daughter, but she wanted to be a YouTuber. And I was like, great, learn a skill and then make YouTube videos about you learning that skill so that if one day, like if nobody watches your YouTubes, you've learned a skill that's worth something. But if you put together learning a valuable skill and then you create an audience on YouTube teaching that skill, then yes, you've put like one plus one and gotten more than two out of it. Um, because you're you're kind of like taking two disciplines and finding a way to marry them in a way that creates more value. And I think with the whole AI thing, it's like, what can you apply AI to that you know that not everybody else knows? And then that makes you so valuable in the job market. It's not just you know some general skills, but you like know a general skill and then also a, and I would call at this point like AI should be a general skill that everybody should learn how to use.

SPEAKER_01 43:17

100%.

SPEAKER_00 43:17

How do you apply it to your area of expertise or your area of interest? Right. And I think that's what creates like that future proofing of you have a career that although it will keep changing, it will constantly keep changing, you will have the skill set to keep changing with it and being the person that is managing that process rather than the one being managed out.

unknown 43:36

Correct.

SPEAKER_00 43:37

Um yeah, and I love that you made that point. And then I will, Courtney will probably listen to this later, or if she doesn't, I'll tell her like, this is why you should feel so good about what you know and what you're doing. Cause she's using AI in robotics and she knows, like, she's taught the kinematics behind robotics for so long. She knows everything about how to implement a robotics project. And having been a systems integrator, also, she knows the vision. She knows the motors, she knows everything that goes with it. And like that's super valuable to be able to put together. Um, I was at an event in Phoenix three weeks ago. There was a couple college girls that came and they were they're studying robotics. Uh, and I think that they're maybe like a year away from graduating, like Arizona State or something like that. And we were talking about AI on the panel, and the girl next to me, she was like, AI, like, why? And I was like, she's like, How are you guys? Are you using it? And I was like, Oh, I know, you know, like I know people that are using it to program robots and stuff like that. And she's like, Oh, no, really? Like, I don't want to use AI to program the robot. She's like, I like writing the robot programming, like that's what I like. And and I'm like, girl, you're the one in college and you're all over here being Miss Naysayer about AI, like, really? And I was like, no, sorry to tell you. And this will take longer to travel from the West Coast down into where you are, right? And it depends on what level manufacturer you decide to get a job at, like, how long it's gonna take for them to be like, girl, you better be using Claude to program this robot or whatever, right? Is gonna be the the model de jour in that particular enterprise. That's that's um, but she's like about to graduate college and having that same mentality, like, I don't want to touch the AI, and and and I shouldn't be because I was like, Yeah, no, they're looking at like your budget's not just your time and and your money anymore. It's like your tokens and like they don't even have keyboards anymore. They're whisper, you know, they've got whisper microphones at their desks. There's there's like all she was like, no, no, no, I don't want this. And I'm like, oh, I feel like you're gonna be in for such a rude awakening. And it may be, it may be in your first, it may not be in your first job. Um, but like, ah, ah, next next three to five years, like if you're not talking about using AI to do whatever job you're doing, I I feel like that's the that's the risky card.

SPEAKER_01 45:51

Yeah, there's also a you were talking about language earlier, um, and AI language. Uh, there is a big language issue in AI. It's like a PR issue plus a language issue because so much of what's coming out is around is is highly, highly academic. Like even I like my founders, I I am not an engineer. Um, and my founders, I had to sit down with them. I was like, I'm a kindergartner. Talk to me. Like, explain this to me. I was like drawing photos of like how it actually works to really understand it. I'm like, okay, you've defined, like, you've redesigned the perceptron. What is the perceptron? I was like, oh, it's the node. Okay, it's the node. Okay, so what does that mean? And it's like, no, it's like each node is uh is now its own little mini model that where the adaptivity lives. And it was just like, okay, why don't you say that? Like, you know, like explain it to me. Um, but there's there's a really big language issue and and PR issue uh around AI. Like I talked to my parents, I talked to my nieces and my nephews, and I like being in AI, and my husband's also in AI, like being in this field, it's just like the the gap is it just really we really need to do a better job um explaining things to people and the benefit to them.

SPEAKER_00 47:03

Um and stop fear-mongering uh in reality because it's not and like like how we stop in the this train whether you want to or not, or even whether we should or not. Like it's it's kind of pointless to debate it. It's it's how do we harness it? How do we then guardwheel it? Like, how do we make sure it actually works for all of us? And we all have to know about it for it to work for everyone. And if you're gonna stick your head in the sand and hope it goes away, then yeah, it will you know serve the people that care to get involved. Um anyway, we're getting on a much bigger philosophical discussion here at the at the end hour, um, or the end of the No, that's okay. I appreciate it. And and uh I one of the things I love about doing this podcast is I get to meet people like you um and have these conversations and then hopefully have follow-ups to them uh eventually.

Where To See Thought Forge

SPEAKER_00 47:52

So when when is the next time or place that uh people can see you guys exhibiting or doing something?

SPEAKER_01 48:00

Um we're not doing IMTS this year, uh, because that time of year is just our busiest time of year. So it'll probably be automate next year. We have a lot of like smaller conferences we do that are like like smaller groups, like industry conferences. Um but like the big one is probably gonna be automate next year, um, to be honest with you.

SPEAKER_00 48:18

Um and I think that's where the majority of our audience is probably gonna get a chance to come see you guys again.

SPEAKER_01 48:23

So uh we have been invited to exhibit a cut some of the larger robotics companies booths, so it'll be easier to find us if you don't know about us. We get we have customers a lot of the times who are like, why have we not heard of you? And it's just like because we spread via word of mouth. Um honestly. Uh, and that's kind of how we we've managed our growth because robotics automation and robotics is very much a survival game.

SPEAKER_00 48:44

Um it is, and that's the other thing is like we were talking about the hardware stuff, and I'm sorry to interrupt you. Um my ADHD is is speaking now.

SPEAKER_01 48:52

I have it too, don't worry. I try I almost interrupted you too.

SPEAKER_00 48:56

But that uh like the novel hardware stuff, it's enough, like it's enough to manage running some novel software stuff. But the other thing is like this industry is about longevity and survival, and and it doesn't matter how cool your thing is or what kind of problems it solves, if you're not around five years from now to support the thing that your POC customers put in their factory and they've now got a dead robot or whatever, and there's obviously enough of those stories, there's enough of those graveyard companies, there is a lot of skepticism and just also a lot of like for good reason, I guess, conservatism in going with the company that you if you don't know if they're gonna be around or not, because you're going out on a limb to invest in something that may give you a huge competitive advantage, or it may just have been, you know, throwing a bunch of money down the drain because two years from now, and it may have nothing to do with you as the customer, right? Like if you're if your VC decides not to give you a follow-up round, you're dead.

SPEAKER_01 49:52

Um you haven't gotten customer traction. Yeah.

SPEAKER_00 49:55

Um, or you don't have customers, usually those two things coincide, right? Yeah.

SPEAKER_01 49:59

Well, there are so much out there that smoke and mirrors. We hear it from our customers all the time.

SPEAKER_00 50:04

Yeah. And I love talking to companies that are like it's not the the PR isn't just about, oh, we raised X like a huge amount of money and we just came out of stealth, and like nobody knows who we are or what we do, but we somehow are worth like you know, all these billions of dollars. And like that is an area of technology that I'm not so familiar with because I've always worked for uh like the AI startup that I worked for back in 2016 was bootstrapped and they still they're you know they're you've got to build a real business in robotics and automation and autonomy, like you have to.

SPEAKER_01 50:36

There's not anything.

SPEAKER_00 50:37

It's true, and like your customers are gonna they take that into account when they look at what you have is is can I trust that you're gonna be around? Um because these these are these are long, long-term relationships. And like I said, you know, factories want to invest in equipment that they can keep around for 30 years. And I don't know that's like a hard thing to like coupled next to kind of consumer and tech, like nobody really thinks in those terms anymore, those long terms, but like that's what PLCs are. That's why they're still around, is because they they've been humming on those lines for 20 years or whatever, and like to change the cat the way that that factories think about spending money from that to oh, we're gonna do a new thing every three to five years. Like it just doesn't work that way.

SPEAKER_01 51:18

No, it does not.

SPEAKER_00 51:19

So, anyway, thank you so much for the time. I'm gonna uh I'm gonna cut this off. I'm gonna cut myself off now. Uh, thank you for those of you listening. Stay tuned for more of our automate series where we're gonna catch up with all the stuff we missed at Automate this year. We're still never gonna get to all of it, not even if we continued this series all the way until Automate next year, but um it'll just kind of an ongoing conversation, and we will see you in Vegas. And I will very much like to try to make some time to come see you. Um, because as we both know, if if it's not on the schedule, it's probably not gonna happen because automate that busy. And so, those of you, if you're listening to this and this, you know, captures any part of your interest or imagination, or you think a friend works at a plant or maybe interested in something like this, like ping them about it now and put this on your show calendar or whatever. Um, figure out a way to start following these guys and pay attention because yeah, this is um to me anyway, one of the more interesting and most likely to like add value to some operations in the short term from what I've seen of the of the AI stuff. So I I appreciate it. And I'm really glad I I got a chance to find out about you guys.

SPEAKER_01 52:30

Awesome. Thank you. I really appreciate the opportunity.

SPEAKER_00 52:33

Um and where should give us give us the the quick like where should people go and find you, connect with you on LinkedIn or your team

Contact Info And Final Signoff

SPEAKER_00 52:40

or go to your website?

SPEAKER_01 52:42

Yeah, so uh my name again, my name is Whitney. My email is Whitney at thoughtforge.ai. Um, if you're trying to sell me something, I won't respond. But if you're reaching out to to talk to me as a uh as a potential customer, would love to chat. Uh the two uh two areas again that we're we're working on uh actively are insertion and kind of complex manipulation type tasks, high mix low volume. Um doesn't need to be high mixed low volume, but more complex. You can't use traditional automation today. Um and uh our website is thoughtforge.ai. So thought and then forge, like forgeymetal.ai.

SPEAKER_00 53:15

Yeah, perfect. We'll make sure to put a link to that in the show notes. And then yeah, let's see how many um AI generated messages you get about the context of this conversation in the subject line and in the body. Oh, Whitney, saw you on the podcast recently. I love the point that you made about blah, blah, blah, blah, blah. I would like to sell you X. I feel like that's like half of my inbox now.

SPEAKER_01 53:36

I know. People like don't realize that you can tell when it's AI. It's it's almost like you want a typo in there, so you know it's not AI.

SPEAKER_00 53:43

Yeah. I also like, I just I seriously know that you do not have the time to go like listen to a full-on hour-long conversation of everybody that you're cold outreaching to email. So there's just like very little likelihood that that's actually the case. Um anyway, good and good and bad AI applications. Uh, we should start like doing some sort of running tally, ladies, uh, about what what we think about good and bad uses of AI. All right. Have a wonderful rest of your evening, Whitney. Thank you so much. Bye. And bye. Thank you for listening to Automation Ladies. If you like our content and you want to stay in touch, please connect with us on LinkedIn, follow the show page, subscribe to our YouTube channel, and you can send us a message or a copy on our website, automationladies.io. We look forward to getting to know you.