Austin Walters saw the AI boom coming as an inventor on more than 650 patents. Now the IP Copilot founder explains how companies lose most of what their engineers invent. He spent two years at a major patent filer before anyone told him there was a patent program, built a large language model before transformers, and wrote a 2018 blog post saying "it's over." He sells software to in-house patent teams, yet predicts that companies will end up with in-house or outside counsel but not both, and says they're missing 95% of their inventions. Host Eli Mazour pushes back on where AI drafting is headed. Clause 8 is brought to you by: IP Copilot (https://ipcopilot.ai/) - Discover, manage, and protect your IP. Junior (https://docs.junior.law) - Your legal team's documents, drafted in your team's style. Listen to the full conversation at VoiceOfIP.com Chapters 00:00:00 Introduction 00:01:24 From Researcher to Prolific Inventor 00:06:17 Why Patent Programs Miss Most Ideas 00:10:18 The Origin of IP Copilot 00:17:10 Disclosures and Engaging Busy Inventors 00:21:01 Patents Versus Trade Secrets 00:22:48 AI Drafting and the Future of Patent Work 00:28:21 AI at the USPTO 00:29:44 Guardrails, Workflows, and Smaller Models 00:37:22 Evaluating AI Tools 00:40:27 Where AI Is Heading 00:46:57 The IP Function Inside Companies 00:50:02 Prior Art, Customer Success, and Review Committees 00:58:06 Advice and a Look Ahead to 2030 This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.voiceofip.com
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
SPEAKER_01The big issue with all these corporations is the engineers are kind of doing their thing, and I think they're missing 95% of the patentable ideas. I told the AI, I removed all its tools, and I said, okay, now do a prior art search. Instead, what it did is it built a tool to access the web, and it was trying to do a rainbow attack to break through our system. I have an idea. My AI talks to your AI. We negotiate a patent. Okay, here's a patent. Next year, you may have AI invent more or as much as humans.
SPEAKER_00Welcome to Clause 8, the number one podcast on intellectual property, brought to you by IP Copilot and Junior.
SPEAKER_02Welcome back to Clause 8. I've been talking with today's guest, Austin Walters, for a long time about AI and where patent practice is headed. Long before ChadGPT, he's working in AI and machine learning at one of the world's largest patent filers and became an inventor on hundreds of patents himself. That gave him a front row seat to how innovation actually happens inside a huge organization. Those experiences eventually helped him lead to found IP Copilot. What happens when AI can actually help companies see and understand their own innovation? How should pan-professionals separate useful tools from the hype?
SPEAKER_02Why does Austin think pan-attorney AI agents may be talking directly to pan-office examiner AI agents sooner than most of us imagine? Welcome to Clause 8.
SPEAKER_01Yeah, no, thank you for having me.
SPEAKER_02Your story is pretty interesting because you are actually a pretty prolific inventor before you started IP Copilot. Can you tell the story of how that happened?
SPEAKER_01Yeah, so I started really like in college. I was part of a group in Caterpillar actually doing computer vision related work. I had the opportunity to do a PhD or go to Capital One. Ended up deciding Capital One and kind of going down the more applied research path. While I was at Capital One, I didn't know about the patent program at all. So we were just publishing stuff, going about kind of our normal thing. And I ended up doing a lot of neural network design on CUDA chips and stuff in college, but then also at Capital One.
SPEAKER_01And then there was a competition internally to do the new fraud detection model. And we took a stab at it using neural networks. Around that time, Capital One informed me finally that we had a patent program. I was like there for probably two years without knowing about it, just giving away the IP. After that, my team filed a few with the team and we kind of waited actually to see if like we were told there was bonuses and stuff. We're like, is this real? And so we ended up finding out, you know, it was real. And then we ended up filing, you know, hundreds over the course of...
SPEAKER_01Part of it was my team would work with different groups within Capital One, kind of like as like a spoke. And we would try to bring external research in and then share it with the organization and take internal research or internal products and try to find external research that could be helpful. And so I got to kind of sit at the center of a lot of the R&D efforts. And through that, we got, you know, hundreds of patents, primarily around neural networks, primarily training, as well as like kind of utilizations, obviously, in financial services.
SPEAKER_02I call it the time before ChatGPT and after ChatGPT, and this was before. What made you interested in machine learning and AI at that point?
SPEAKER_01Yeah, so, I mean, at the time, it was, you know, people really, there were no neural network frameworks. So it was all like CUDA by hand. Originally, I was doing like some project where I was trying to code drones to identify moving targets, for use of a better word. And, you know, I thought it was just the best way to actually solve the problem. And the compute at the time was not very good, but you can see how as compute was improving and the GPUs were improving, how this would very... Did ChadGPT surprise you?
SPEAKER_01No. So actually at Capital One, we built, I would call it a large language model, before Transformers. So we were able to basically synthetically replicate any data set. And we published this back in like 2016, 17, 18. And there were other companies actually interested in utilizing it. So when I saw the Transformer come out, I was like, it's really slow. There's actually better ways to do it. I still maintain there's better ways to do it and faster and more efficient ways to do it. But at least for large language models, we're kind of taking the brute force approach.
SPEAKER_01And it seems to be working so far. It just takes exponentially more compute and money. I think at some point in the next 10 years, they're going to take some of the other approaches that actually we... Patented or did research on. And I think you're going to end up getting far smaller models that are actually more accurate and faster.
SPEAKER_02And I guess did the reaction to ChatGPT surprise you? Like the world is changing and everything. This is the moment.
SPEAKER_01So I always told my team at the time when I was at Capital One that like, we're going to be the last ship out because like, I think it was like 2018. I even wrote like a blog post about it. I'm like, it's open. Because it was like clear that we were on that exponential curve. This is before ChatGPT, when the Transformer came out, it was clear that this is like a general purpose computing system. Same with before that recurrent neural networks. So I knew I had to start a business, like honestly, where I was like using AI, because I'm like, I think it's going to dominate, be dominant. And I have expertise in it.
SPEAKER_01And I was like, it was just clearly the time to try to make it happen.
SPEAKER_00There are a lot of AI tools being built for lawyers. What stands out about Junior is where it came from. Junior grew out of patent lawyers and law firm leaders who spent years thinking about how technology could make sophisticated legal work more efficient. It's backed by integrations with 11 patent offices, trademark data, federal case law and SEP standards. And its web platform, Universe, pairs that data with examiner history to help draft amendments and arguments based on what has actually worked.
SPEAKER_00Including on Section 101 rejections for a particular examiner. You can learn more at junior.law.
SPEAKER_02We'll get back to that, but I want to talk a little bit about your experience as an inventor. Did you know what a patent was when you were told about the patent program? Like, what was your...
SPEAKER_01I'm actually opposed to software patents. So like when I was at Capital One, we'd get bonuses and I'd like donate to the EFF to like try to get rid of software patents. Actually, one of our investors too, like he said, like, why should I invest with you? I don't like software patents. And I said, by the end of it, I don't think there'll be patents, which I can talk about. But like I think, you know, as things are going to change so dramatically, you're going to democratize it. We're like patents are going to become very cheap to produce. And then you're also going to be able to very quickly and easily be able to license them. And that would probably help everyone because everyone does innovation.
SPEAKER_01So I didn't know what they were. I was just I think most software people are kind of naturally inclined to avoid it.
SPEAKER_02From your perspective, being within the company, anything that stood out to you about what makes a patent program work or not, you know, not even specific to your company, but I guess kind of as an inventor?
SPEAKER_01In general, what I think is everyone does innovation constantly. And I think one of the biggest takeaways I had was pretty much everyone has at least one patent or one patentable idea every product. They have multiple every year. And the big issue with all these corporations is the engineers are kind of doing their thing. And management knows what's important, but they don't know what's patentable either. And so you have this disconnect between management, which knows strategy, engineers, which are actually creating the patentable work. And this is software, mechanical, kind of whatever.
SPEAKER_01And then you have the IP team just desperately trying to figure out and connect the two. And I think it's very inefficient, very complicated. 95% of the patentable ideas in most corporations.
SPEAKER_02You know, I think you mentioned your name is on hundreds of patents. What was it about, I guess, maybe you or the experience that kind of made you so engaged in the program?
SPEAKER_01So I guess naturally we were in applied research. So we were taking like a bunch of like cutting edge stuff and applying it in unique ways every couple of weeks because we would work with multiple teams at the same time. And we were like consultants helping all these different teams. And so because of that, I got to see like all the problems in like, you know, one of the largest corporations in the US. And I got to figure out all that. Like we were constantly trying to brainstorm. We would give people kind of like, here's how you can like build a framework to solve your problem. And then we would consult with them. So for me, it was just it was a fun problem solving in terms
SPEAKER_01of like how I identified like what's patentable and what's not. We would brainstorm all the ways that we can solve this problem and then pick the best ones. So a lot of times you're, you know, that's what you want to do during patenting. You'll you'll focus on the one main one and you have all these other branches like we should cover these two to make sure no one else can implement a solution to it that we don't cover.
SPEAKER_02Any pain points that you recognized, besides within the company which you described, there's a disconnect, but from an investor standpoint that you thought, I wish this was smoother?
SPEAKER_01I mean, hundreds of patents means like probably thousands of disclosures, right? So like that process was not fun, like documenting everything, writing it all out, planning it all out. I always joke like, you know, like my goal as an engineer is not to talk to attorneys. Like my goal is to engineer things. So although I'm friends with a lot of patent attorneys now and like I work with patent attorneys, like That was, I think, was the biggest pain point is it takes away and, like, our team actually got reprimanded repeatedly. Like, you're talking to attorneys too much. And I was like, well, that's part of our output.
SPEAKER_02You know, essentially, you said you got reprimanded. I guess, is that because there's a disconnect between the value of the patent program within a company? Is that kind of the...
SPEAKER_01Different people value it. So like my management, they're going to be graded on one thing, right? And then you're going to have like maybe the highest rung of management, like the VP level be graded on another thing. And so from the line manager, they're like, how much code did you produce? Not how many patents you got. And then, but the VP might actually be graded on that. So there's like that disconnect. And so any kind of organization kind of worries about promotion. And so I think that was the biggest issue. Strategically, I think typically the patent program Like has like CTO and like a bunch of buy-in
SPEAKER_01if they have a program. But yeah, I think that's the main disconnect is like the engineering departments are not aligned.
SPEAKER_02And so you're working in AI and you think like, we might not have a job soon. So you need to start a company. How did you come up with the idea for IP Copilot?
SPEAKER_01So before IP Copilot, I went to Read AI, which is a company that does, they record meetings. And one of the reasons I went there is I was like, with all that recording data, there's a bunch you can do with it. Before IP Copilot, I had a thing on... Basically on the web and what I would sell, I would sell it to like friends who were doing stock trading. And what we would do is identify people doing kind of like leaking trade secrets online, but it was public and they were anonymous. So the system would de-anonymize them, identify that they probably worked for like AMD would be an example.
SPEAKER_01And then they're talking about their new chip coming out soon. And we would send that signal to the investors. And that would be accidentally, I guess. They would be like bragging about it on like Hacker News or Reddit or Twitter or whatever. And our system would track all that. Well, in 2018 and 2020, they started locking it down. All the APIs started going closed source and kind of walled gardens. And I shut it down when COVID happened. And then fast forward a couple of years, I was like, hey, I'm at ReadAI. I knew I wanted to start a business. And I was like, you know, one of the biggest problems I had at a corporation was everyone's inventing
SPEAKER_01and there's no... Ideas being captured, really. You have these IP teams going out, you have one IP attorney for 10,000 engineers, and you're just talking to people, and you're never capturing all the IP that way. I was like, why don't I just apply the system I already built to identify trade secrets I shut down a couple years ago, but I can apply the same thing internally. That's kind of where IP Copac came from. We want to capture and understand the knowledge that's in a company and how it maps to value within a business.
SPEAKER_02So that's interesting. It's something you've told me many times, that really IP co-pilot is about capturing the ideas within a company and not necessarily of invention disclosures for patents. That's part of it, of course. But why is that important? Why does people have all kinds of ideas? Things happen. They file some patents. Like, why?
SPEAKER_01So the way I view IP is kind of different. I think, you know, the way I always describe it is it's what's in the person's head, right? And then you have IP assets, which are just derived from what's in a person's head. So what our system is always trying to do is understand what a person knows and And how those concepts evolve over time. It's also like a really interesting technical problem that I enjoy. But part of that, right, is then mapping those concepts to the unique value of the business. Because most businesses, almost all businesses are IP based in some way, even if it's marketing. And so it's trying to figure out.
SPEAKER_01What concepts and who's creating those concepts that link it back to the product that's being produced by the business? And that's kind of what makes, I think, IP Copilot unique is that we're not worried about drafting. We're not worried about all this other stuff. What I'm worried about is who in your business is the most valuable and what are they creating that is valuable?
SPEAKER_02What is the vision for how the IP team leverages that or the company?
SPEAKER_01Yeah, so I think in general, I think it's more than just the IP team, but it's like even like a CTO, it'd be useful to know these three people are producing the 80% of the value to my products. Those are my most valuable people. I want to make sure I'm, you know, There's other people are also providing value and they might like iterate on that idea. But how do I make them more productive? How do I like leverage them more and stuff like that? So in my view, it's more of a living knowledge graph of kind of what's happening in your business
SPEAKER_01and then trying to map that to the value and trying to figure out how you can leverage that more effectively. A few different applications, right? So I worked a little bit in data loss prevention. You would want to identify when a person's about to leave and whether they're taking their most important information with them. That's like trade secret protections, basically. So that would be like one of the use cases. The other use case would be, okay, well, I can see that these people are developing IP around this new product. Maybe we don't care about that product, but we still want to patent it because we see our competitor doing it. So it's about tracking that and figuring out how to protect it from the IP team's perspective.
SPEAKER_01Showing you the graph and showing you how things map is pretty much like, in my view, like most of an internal counsel's job, right? It's making sure you patent the right ideas because you have a fixed budget. Yeah. And so we're trying to help that.
SPEAKER_02And I guess what do the inputs have to be into your system to make that happen?
SPEAKER_01Yeah, yeah. So we plug into engineering documents primarily. So Slack, Jira, Confluence. Some people are doing like Zoom transcripts. We have one company that sends us just everything, all the transcripts and everything that they record. So it's really, it could be that, but some companies just give us data dumps. They're like, here's all the documents related to our given product. And we do like infringement monitoring and product monitoring too, where you might just give us our portfolio. And then we're basically doing the same thing, but we're mining your competitor's products and then mapping them back to your portfolio and saying like,
SPEAKER_01Here's the patents that are reading on your products. But it's all the same type of problem. It's comparing an idea to assets. And then if you don't have an asset, here's the assets you should get to protect those.
SPEAKER_02You know, I was recently actually talking to somebody, what is the number of patents that you need for a product? And a patent attorney I heard say, well, you need whatever you need to protect it because of the value of the product. You know, that's the number of patents that you need. But it's a hard thing to do, right? Do you put financial documents into that, financial projections? Is that part of that?
SPEAKER_01Yeah, you can change it. So the thing, this one is just being beta tested now. One of the things we do is you can give us a bunch of product documents and we'll break out all the subcomponents. And then we'll identify which components that we can't find on the market elsewhere. And then we'll highlight that and say, hey, you have like two patent assets on it, but you have all these other concepts that you need to file on to make the protection like, you know, robust around that given asset such that no one else can build around it.
SPEAKER_02Yeah.
SPEAKER_01And so I think that, like finding the choke points in your product is really what the goal should be in my mind.
SPEAKER_02You know, one of the interesting, I guess, you shared with me is, you know, sometimes within a company, and it happens, I guess, within an industry, you know, people are working on the same problem. And I think you kind of alluded to that.
SPEAKER_01Oh, yeah. There's so much waste. I mean, like a company like Capital One, there's probably, I mean, I was a part of several projects where we had eight teams working. Competing basically, which means you have eight duplicate efforts trying to produce something. I imagine that's what we know of, right? With 40,000 engineers, you're going to have, I don't know, 150, 200 different duplicate efforts for like very basic problems that everyone has. And being able to identify, sort that and say like, hey, these two things, you know, are similar.
SPEAKER_01Our system does that right now, but I think in general is a huge problem that's outside of the IP space. I mean, I still view it as an IP problem personally, but no one else seems to. Everyone else is like, that's an engineering problem.
SPEAKER_02Yeah. It's interesting because obviously, sometimes when I do the work and And sometimes the best prior art is prior art from that same company that's working on applying the pen application. And, you know, I actually had a funny story because I once was with an inventor and I was telling him, well, we have this prior art existing. And he's like, no, no, that's what my pen application is supposed to cover.
SPEAKER_01And I explained to him like, oh, well, then I don't know if we could get that.
SPEAKER_02How do you get people like that? And you said, you know, they want to do engineering. They don't want to be like emailing back and forth with patent attorneys and writing in their disclosure materials and so forth. How do you get inventors like that more involved? You know, the kind of your high value, really busy engineers.
SPEAKER_01You know, normally the disclosure form is where you lose the most people. Right. If you say fill out this form, I'm going to close the window. Right. As soon as you leave the room and just ignore that, because like. I'm guilty of this probably thousands of times where I had a half-rent disclosure. I'm like, you know what? I got to go do my work. You know what I mean? So I closed the thing. And when we did surveys, it takes 40 minutes on average for people to complete an invention disclosure if you have documentation. Yeah. What we did in our tool is we flipped the script. So now we take their documentation, we have it already, and we summarize it for them.
SPEAKER_01So all you have to do is confirm it. And so from an engineer's perspective, it's like, this is awesome. I don't need to talk to anybody. I just say yes. And then our system will generate follow-up questions, which can or cannot be used if you want to. And then you can answer those questions and make a more robust disclosure. But because it's tied into all the engineering documents, we get all that context. It's actually way better than what the inventor would ever provide you. I'm sure you're familiar with, like, you know, getting one or two sentence disclosures and, like, you have to deal with that. Or the reverse now, you get, like, an AI-generated one that's, like, 20 pages.
SPEAKER_01Well, that's what I was going to bring up.
SPEAKER_02Yeah.
SPEAKER_01And, like, our system does both where it just makes it concise. And it's like, here's the idea, here's the inventive concept, here we'll do a prior search, and you get a package that's all ready for you.
SPEAKER_02Well, I was going to say even I've had some clients who've told their inventors not to use AI because what ends up happening, the invention disclosure is like 30 pages.
SPEAKER_01And it could be inventing stuff too, which is even a bigger problem.
SPEAKER_02And then I guess the job of the patent attorney is then kind of to dig through that and figure out like what actually was the invention.
SPEAKER_01That's even harder, I'm sure, because now
SPEAKER_02it's not narrow. Before, I guess you'd have the conversation, the invention disclosure call and kind of figure that out. But now you're kind of sometimes... You know, it probably, that's where I might start to forget where his idea was within that document.
SPEAKER_01I think the other thing, too, to consider is just from an inventor's perspective, it's just like a better user experience to be like, hey, you have an awesome idea. Rather than being like, what idea do you have? You're going to get shower thoughts half the time or whatever they were driving to work. They'd be like, oh, I had this disclosure meeting. I knew I had to come up with something. It's just going to be more natural, if that makes sense.
SPEAKER_02And obviously, if a company has a good patent program, then they're, like you said, they're not going to file for every invention disclosure. They're going to kind of try to find a way to call through them. And obviously, if it takes time and effort, but also brings disappointment to the inventor who might have spent time on that invention disclosure material. It seems like every company should have something like that, a graph of all the innovation. I guess, what is the reaction that maybe you got in the beginning and you're getting now and has that kind of changed?
SPEAKER_01So I would say there's two reactions. A lot of times, I mean, the engineers are all like, this is awesome. But I think a lot of the patent attorneys, about half of them, you'll see hesitation. They're like, I don't know if we want to be spied on because that's what it sounds like. And to a degree, it is, right? We are going to be reading your technical documents, but it's technical documents. So it's honestly already owned by the company and everything else. So it's less of a reaction. I think very quickly, we get everyone on board. And then it's Getting the security team
SPEAKER_01to approve us accessing all their most sensitive details, organizing it, and then making it available is like, there's always a hesitation there. I would say like initially there was a big hesitation around AI as all these companies had like generative AI committees spin up. Now, I mean, we can get through, like before it would take sometimes eight months because the committee would be formed as we were trying to get through the process. Now it might take six weeks. So it really just, it's getting a lot easier, honestly.
SPEAKER_02You know, one of the interesting parts of your invention is you are looking at ideas. So they could be some for patents, some could be trade secrets. Any thoughts about, like, solving that problem? Because obviously that's a problem that IP teams navigate. Kind of how do you have both protections if possible and at the same time keep trade secrets as trade secrets?
SPEAKER_01I mean, you know, I think what's going to end up happening is patents are going to become cheaper and cheaper to produce over time effectively free.
SPEAKER_02Yeah.
SPEAKER_01And I think that'll happen much sooner than
SPEAKER_02a lot of people think.
SPEAKER_01And I think because of that, the flip side of that is also you'll be able to take a patent and you'll be able to have the AI generate around it as well. So I actually think patents, there's some question whether or not they'll be as valuable as trade secrets over time. I think also because you have the search capability, You can also enforce your patents more effectively. So I don't know where that's going exactly, to be blunt. I think from the trade seeker perspective, I think they're underutilized. And I think there's not a good system today to categorize, maintain, and analyze and keep track of them.
SPEAKER_01So that's where one of the things we're focusing on is we do have a trade secret component because I think that the biggest problem is just the organization of it. How do you know even what a trade secret... A lot of people are trying to enforce it after the fact. Where, you know, you'll see someone taking your ideas and then you're like, well, we had a program. It's like, well, kind of. You had some trainings. You told people not to take it. But now in our system, right, we can say this person, when they go to leave, they knew all of this. You can have them sign it. And so when they take it with them, like you can actually prove with documented evidence that they took it and they knew
SPEAKER_01about it before they took it.
SPEAKER_02Yeah. I guess it makes a difference. Like if you're a company with tens of thousands of employees, you're not going to spend the time for each one.
SPEAKER_01You don't care. Honestly, you can't. Also, do you really want to be suing your employees is another question.
SPEAKER_02Well, yeah, you wouldn't want to. I guess I picked up that you think patenting will take less and less time. Obviously, there's a lot of other tools to kind of make that happen. What do you think about that? I guess the state of the AI technology for that, the state of the tools and different approaches you've seen it.
SPEAKER_01Yeah, so we always viewed AI drafting as a solved problem. Like even when we started in 2023, we said it will be solved. I don't know when, I don't know how, but I think like a general purpose model will solve it just like it's going to solve programming. And so our thesis was, let's focus on the thing that can't be solved with these generative programs. And so that's where we spent all of our time. But I think it's at the point where even we ran some tests and we were able to generate a patent that I thought was, I was like, we should just file this. And so I think we're at that point where at the very least, provisionals can be easily filed.
SPEAKER_01I think by next year, I think you'll have, I mean, you're already seeing it. You're getting AI generated patents today. I think specialized tools though are going to become less relevant, at least related to the generation portion. I think specialized tools might make sense in more like some of the other components, right? Collaboration, things like that. I think that's where that would be. Those would make sense. But I think in general, like, ChadGPT is pretty good at making a patent, like, or Claude, if you've... Yeah.
SPEAKER_02Well, I guess I have someone, at least for now, have someone with different experience. But, you know, to me, it's interesting.
SPEAKER_01There's obviously... Oh, it depends on the industry and all this other stuff. You know, I'm just saying, you can see the difference, though, between, like, 2023 and 2021. Oh, for sure. Things are getting better.
SPEAKER_02I mean, even with the patent drafting tools, you know, one of the problems that you do have, or sorry, patent prosecution, whatever, there's still a lot of data that you need to gather to do the task. And then you need predictable output, right? Somewhat.
SPEAKER_01I would argue like it depends on what you're trying to get because there's a lot of companies that just want to file some patents. You know what I mean? They just want a certain number. I think like you'll have probably two tiers. I mean, you already have two tiers, right? You have like the cheap tier. Sometimes there's three tiers for certain companies. I think like the lower tier patents are already commoditized, right? You're already outsourcing it to somebody you don't know. And so what's the difference between if they're using AI or not using AI? Like you have no, from your view, any difference.
SPEAKER_02Yeah, no, that's interesting. I think even before chat GPT, but pre-chat GPT times, the market was going two different directions. One was, let me figure out a way to make it as cheap as possible. And AI is obviously a way to facilitate that.
SPEAKER_01Yeah, and that's just going to drive it down. It's just dull.
SPEAKER_02But if your goal is to kind of figure out to get something valuable and worthwhile, then I think somebody still has to make that decision, right? And the question is, like, do you press that button internally or externally and make that call?
SPEAKER_01So my thesis is that both are going to happen. So you're going to have companies that realize they can outsource most of their IP efforts. So they're going to get rid of their in-house counsel and just have the law firm do it for really cheap. Or like, you know, you'll be like seconded or whatever it is. You're not going to have like the two sided thing, though, where you have like the in-house counsel and a bunch of external. I think that'll probably go away or like shrink. I think you'll have a lot more just, you know, that's my law firm and I just outsource it. Or you'll have the in-house counsel doing everything. And I think one of those two will happen just because a lot of the automation allows it to be centralized to just one person or two people
SPEAKER_01that can manage it.
SPEAKER_02And what do you think needs to happen? Like, you know, you said you drafted a pretty good application using LM, but do you think it's like the key is really good prompts, the right type of technology that's a good fit for that? Initial disclosure that you start with has to be good. Like, what do you think right now, as we sit today, that kind of makes it work?
SPEAKER_01I mean, I think it's with kind of everything, you know? So like our system, the entire system is designed to capture all that information you need. So I'm using it in our, like, we have an internal system that's connected to everything. And I'm like, okay, now draft a patent. It's like, well, it has like, you know, tens of pages of documentation about this. It has images, it has diagrams, like it doesn't need to do any of that. It just generates it all. It's like puts it all together and formats it. And then we're just like, OK, now check the MPEP, make sure it's all like accurate, you know. And so we, you know, it's pretty straightforward. It took like an hour for it to generate.
SPEAKER_01You know, is it the best application? No. But like, is it okay for a provisional? For sure. And it's probably, it might be good enough to actually just file it. So like, is it good enough to be like, stand up? It's like, well, you don't even know if an attorney drafts it, if it's necessarily good enough, like you hope it is, but there's always like, you know, edge cases. So I think there's two problems. I think it's that you don't know what the quality is unless like you're an IP attorney. So like I sent that draft off to an IP attorney to get reviewed, right? Yeah. And I think that's going to happen regardless. You're still going to want some sort of,
SPEAKER_01like as a CEO or as in-house counsel, you're going to want to outsource your risk to somebody else to handle that. Be like, well, they didn't do a good job then or whatever it is. And so I think that natural human instinct to avoid... Risk will continue. And I think maybe that's kind of saving grace potentially. But I think in general, like I think of it from a programming perspective, like I haven't opened up my IDE, which does the development in like a year. So I'm assuming anything generative, like content wise, that'll happen.
SPEAKER_01And we've probably all generated videos or images or songs and we could all see how good it's getting. Yeah.
SPEAKER_02I listened to Pandora, and then there was this one artist, I liked the music, and I couldn't find anything online. And my wife burst a bubble and said it was AI-generated music.
SPEAKER_01I live in Nashville area, and it was three years ago, I showed it to one of my friends who was a musician. He immediately bought a farm and started raising sheep. I mean, it's, yeah, it's very good.
SPEAKER_02You know, the other part of using AI, you and I were discussing before we started recording, is obviously the examination of patents. And, you know, the USPTO has been thinking about it. They've been kind of public with it. Where do you think that will go?
SPEAKER_01Well, I think, you know, from the start of kind of like IP Copilot and really any of these things, like I think we're going to go into a registration process effectively because it's going to be so quick to get IP or so quick to generate patents. If the U.S. Patent Office takes 18 months to prosecute it or to like even like, you know, start prosecuting it, depending how it works or examining it. Sorry. And then the prosecution process takes like whatever, five years. Like the point is, is AI is going to make everything faster. And so there's no point.
SPEAKER_01By the time, like if I was to publish, like I say, my patent goes live in 18 months, but I don't get protections for five years. Well, my company might be out of business. Because what they can do is another company can come in, use AI to generate everything that I just described perfectly for them, and then go to market with it before my patent becomes enforceable. And then I go out of business, and then what? So I think for patents to be valuable, they're going to have to figure out how to cut that time down to a couple of months to get a patent over the next few years.
SPEAKER_01I don't know exactly how they're going to do it. I think they're probably going to use a heavy amount of AI personally. But I think they have to do something related to that.
SPEAKER_02You know, the thing is, like, AI is good with making something seem like it's compelling and well-written and makes sense. But if somebody really knows the stuff, they might be able to identify kind of a shortcomings. And somebody who's generating that output, if they don't really understand what they're doing, they're more likely to generate output that probably misses the mark. Of course. Yeah. I mean, what is the key to providing AI as a tool with making sure it's leveraged in the right way?
SPEAKER_01So I always view these LLMs as the least efficient way to solve a problem. I mean, it's effective, like, you know, if I had, I used to joke, like, I would be more effective just having a room of interns and probably cheaper, you know, because the interns are going to be able to, like, it's a lot about your prompting, right? If I tell the interns solve this problem, they're going to, like, not do really silly things. Just as an example, you know, we saw, like, a hugging face get hacked. Well, we had a similar incident in our system, which I think I posted even on, you know, LinkedIn. Where I told the AI, I removed all its tools and I said, okay, now do a prior art search.
SPEAKER_01And I wanted it to like manually, like, just tell me like, hey, I don't have access to the tools. Instead, what it did is it built a tool to access the web and then went to our website. It identified the users because it had that tool still. And it was trying to do a rainbow attack to break through our system and guess the passwords into our system. So we're gonna do a prior art search on the website. And I was like, oh, well, that was shocking. So I turned off like the ability for it to like be in like that agent mode. But what I realized. How did you catch that? I was watching it. I mean, we were like, I was like, why is this taking so long?
SPEAKER_01Like just to tell me like, no. Yeah. And so we were watching and I was like, where is this going to go? And I watched it do the whole thing. And that was like December of 2025. And so I was like, this is clearly a problem. And in general, the approach we've always taken is use the LLM as kind of a brute force approach. Then once you know what you need to do, you don't need to use reasoning mode or any of these other systems. I don't know if you're familiar, but like large languages have these reasoning modes. They use a ton of tokens. You can actually just make it a very simple prompt once you know what you're looking for.
SPEAKER_01And so you kind of solidify a workflow. And then once it's solidified, it's very quick to easily analyze massive amounts of data. And I think that's where you can strain the problem and you can actually monitor errors. We have additional tooling that doesn't use large language models to check all of our systems. And I think that's also important. So for instance, If it cites something, is that actually in the text? Like you don't need a large language model to check that. You can just check that with something called regex. There's a lot of other approaches. And I think formalizing those workflows is a way to reduce risk. That's interesting.
SPEAKER_01Yeah. It's kind of like putting guardrails. I don't think you should be letting the large language model kind of go off and do its own thing most of the time, because you don't know where it's going to go. And it's honestly not going to be very efficient. And it also is based on, like, I don't think the AIs have gotten really too much smarter necessarily. What they've gotten better about, I mean, smarter is relative, I guess. What they've gotten better about is understanding what you're prompting them to do. And I think that's actually like the biggest like lift that we've been seeing over the last couple of years. Obviously, like context windows and all this other stuff, but that's the biggest one in my mind.
SPEAKER_00Companies are constantly innovating, but capturing those ideas and deciding which are worth protecting can be difficult. IP Copilot uses AI to uncover potential inventions in the tools teams already use. It connects those ideas, invention disclosure, That's what makes IP Copilot such a natural fit for Clause 8 this season. Discover, manage, and protect your IP at ipcopilot.ai. [music]
SPEAKER_02You know, obviously before there were, again, before ChedGPT, there were tools available to kind of make things more efficient and easy. Like, you know, simple things like checking antecedent basis for like 112 rejections or whatever. You know, you don't need LLM to do that. That's kind of like, oh, yes, you should check the article. And it's interesting now because obviously now there are tools that use LLM to do that. Yeah, it's just like an expensive way to do the same thing you could have done very cheap. How do you build those workflows? You know, how should the panel office think about building those kind of Workflows or like,
SPEAKER_02what's a good way to pack a prize?
SPEAKER_01Yeah. So, I mean, I don't know, like, I still think this is like a software engineering problem, ultimately, I guess, because maybe I'm a software engineer at heart, but, you know, in general, I think what we, you can create workflows like in, A system called like Zapier or any of these kinds of things where you just have like, you know, inputs, outputs, and you want to test each one of those. So in, I don't know if like Cloud or OpenAI or any of these systems necessarily have that exactly, but you can effectively run API call scripts that just go through and they like try each one and they have checks around them. I think that's honestly all you need to do.
SPEAKER_01And you can have like a programming, it's like a Python script. But like any kind of like scripting language, you can just say, here's the prompt. Here's what I'm looking for. Here's the checks I'm doing around it. And you just go down the list. And that way it doesn't think for itself and just who knows what it's going to do. Kind of a little bit more predictability
SPEAKER_02with that. Yeah, exactly. You know, you mentioned, I thought it was interesting, you mentioned you don't think, obviously, the conversation is more and more computing power for LLMs and because of transformers. But you think it's actually going to go direction towards smaller models?
SPEAKER_01Well, everyone's concerned about the compute. And I'm like, look, if that becomes a constraint, people are just going to innovate around it. You know, like that's not. That's not a real problem. We're doing it because that's the most efficient way to gain 1% accuracy. These models are not gaining huge leaps and bounds based on their training size or anything else. Actually, most of that has to do with the harness and the way that they're being put together. To understand the prompts or to get access to more tools. I think the more efficient way of doing it, I mean, there's a lot of different ways.
SPEAKER_01Just for example, there's certain models now that instead of using 16 bytes or 16 bits for every kind of compute layer, they're using four bits. And so you need to use four times less memory. Well, there's one-bit models too. And ironically, the one-bit model is equivalent to GPT 3.5 already, like a 27 billion one-bit model, which is like four gigabytes or something. So I can have it on my phone. It's equivalent to GPT 3.5 now. But give it four years, it's going to be as efficient as our best models today, but be four gigabytes.
SPEAKER_01And so what we're seeing is the frontier models are getting bigger and bigger. But like just behind them about 18 months, these small models you can run on your laptop are matching, you know, maybe six months or a year behind the frontier models. And they're much more efficient. So at some point, I think what's going to happen is intelligence, quote unquote, will become saturated and there's not going to be huge performance gains. And we're kind of actually, in my opinion, seeing that like there's not a huge performance gain between these models so much as just configuring the tool use. You know, there's pretty public.
SPEAKER_02Some companies discuss about developing their own models. Then they move away from that. Then they go back to that. When you hear stories like that, what are you kind of thinking is going on or...
SPEAKER_01So our thesis was that you should only ever just use the general models and then build in the tools and the scripts and the workflows you want. And so we actually have always just used the best open source models we can get access to that run efficiently. And we find that even like the accuracy in our model would often beat like OpenAI's model. And like, by the way, like there's no difference except for, you know, we have these workflows with additional checks around them to more accurately determine things. So we can use any LLM in our system. Because we have like tooling,
SPEAKER_01we have data, we have other things that help structure the whole thing. So it doesn't really matter how smart the LLM is.
SPEAKER_02You know, obviously we, I love the phrase you said, but LLM is the least efficient way to do things. But there's so many AI tools out there, you know, just putting aside the pen practice stuff, you know, for everything you mentioned, for coding, obviously, for marketing, for that. Let's say one of them, how do you evaluate if a tool is actually worthwhile for the problem? Like how do you kind of, What advice would you give?
SPEAKER_01Yeah. So it's an interesting question because we've come across people who try to build like copies of our tool. And it's always funny because we're like, okay, well, we'll stay in touch. And then like half of them or more come back within like six months because the reality is, so like software as a service, I don't know if there's like in the news, they've been saying like, that's going to go away. It's like, not really. And the reason I don't think so is like, you're not a good designer. Like as a regular person, You're only thinking about your narrow scope of problem, but there's a long tail of potential issues that you don't even know exist.
SPEAKER_01There's a bunch of unknown unknowns. And what all these companies are doing is they're solving for the unknown unknowns all the time. And so when you run into it, you want your tool to be able to handle it or recognize that there's a problem or whatever it is. I think in terms of assessing whether or not a tool is good or not, is whether or not, for one, can you just use ChadGBT to do it right off the bat? I think there's a fair argument. If I can give it a prompt and it does it, I don't need your tool. But I think it's more about your workflow and how cost-effective you can get it.
SPEAKER_01But I do think that it's becoming very saturated,
SPEAKER_02as you're noting, because it's also very easy to create a bunch of tools.
SPEAKER_01And so I know in the AI drafting space, there's what, like 100 companies? We literally have a channel that's just AI drafting tools that's just like every day is like a new one. And they're all getting funding. And we're like watching it. We're like, okay, but, you know, I'm just looking at the ChatGPT getting better every month. And I'm like, I'm just going to wait. Yeah, I think that the reason you would want a new tool would be a workflow or a collaboration of some kind.
SPEAKER_02You know, go back to our discussion about, you know, how the patent offices and practitioners are going to use AI. I think you mentioned And if someone jokingly or serious, but at a certain point, the AI tools are going to talk to each other.
SPEAKER_01I mean, they already are, right? But yeah, no, I mean, I think with the next couple of years, it might be China, it might be Korea, it might be the US, but whoever it is, you're just going to have a system where like, I have an idea, my AI talks to your AI, we negotiate a patent, and then you're just like, okay, here's a patent. One of the questions I have, or one of the theses I have, is why is the USPTO, if they're concerned about formatting, not just drafting the patents for us? If I know the claims and I have the evidence,
SPEAKER_01I think the biggest problem that's going to happen is proving that a human actually created the patent. And so I think the quickest way to solve that problem in my mind is I think you're going to have to get to the interview process as fast as you can or something to validate it, right? And so it's going to be me providing evidence. It's going to be me providing maybe an interview. I don't think the drafting of the patent will not be the important part. I think it'll be the process of confirming that that's actually you creating it.
SPEAKER_02I think there's been a lot of predictions about relating to AI, like, I don't know, three years ago, Everybody's not going to have a job in three years.
SPEAKER_01I think there's always like an overestimate and an underestimate, right?
SPEAKER_02But I guess going back to when CHPT came out, you already kind of knew this was coming to where we are today. Like, what do you think kind of the progression has been?
SPEAKER_01I mean, I said within seven years in 2023. And the reason I say that is like, that's just like a decent, like it takes about seven years to like for to transition like an industry in general through technology. But I don't know where it's going to end up. I think that's actually the hardest part is like a lot of technologies, it's very, it's almost like you can predict it. Like even sitting here, I'm like, I don't know if AI is going to just like keep progressing or like if it's going to get estimated cheaper, like what other uses can it be used for? I mean, one of the big things I always think about is like, you know, which people don't talk about is like,
SPEAKER_01we're going to have robots in all of our homes like fairly quickly. Like within five years. And the reason being is I know they're probably going to be able to ramp up to like 10,000 robots. Well, those 10,000 robots could build more robots. And so you're going to end up with like exponential growth. And it's like, how quick is that curve? 10,000 to a million or is it 10,000 to 100,000 the next year? It's just how fast can you produce it? I think similar with AI, it's how fast that technology trend is. I think in terms of like where we are on the progression, I think AI drafting is definitely coming.
SPEAKER_01And like, I think we all know it. I think a lot of us, including myself, like coding, it's like very hard for me to like look at it and be like, I spent 15 years learning how to do this. And like, I'm going to be doing something else. I'm still right now helping the AI. I'm still like, but I think we're going to have to transition into a different way of thinking about it. We're like, we're not generating content anymore, but we're doing content. Some other augmentation to that. Well, you know, it's not my original thought, but obviously there's a thought like the AI is going to create more work, you know,
SPEAKER_02more patterns. Maybe there's more attorneys and more.
SPEAKER_01Well, it's a joke, right? You get that email to generate the AI-generated email, and then you just summarize it on the other side. I'm like, can't guarantee that's a thing. Like we do it in code. The coder will generate a bunch of code, and then I just ask the AI to review it for me and give me a summary of the key parts I need to review, right? Yeah.
SPEAKER_02Same thing. Is there anything I guess you thought AI a problem that it would already solve that you are surprised that hasn't been addressed?
SPEAKER_01I'm surprised they're still using these super huge models. And I'm surprised that they're still trying to just like brute force it into the trillions and trillions of parameters when it's marginal gain, because most of the utility I think is very simple. Like just being able to ask the AI, like summarize this email is like a large portion of that utility. So that's the only thing I guess I would be surprised, like I'm really surprised by, because then they're spending billions You know, trillions of dollars building these data centers that they probably don't need.
SPEAKER_02You know, obviously you mentioned some working with AI and then it takes you like an hour to do something. And that's something we have to get used to, right? Because computers got pretty quick. So we kind of expect things to be instantaneous. And now using AI, sometimes it takes time for it to think things through.
SPEAKER_01Yeah. Yeah. That's a weird one because I think I talked about it in another talk. I did the equivalent of 3,000 hours of human labor and three hours on my phone the other day because I was just like, review this, review this, create this PowerPoint. I was just going through my checklist and I was just sending off to different AIs to do everything on my checklist. I was, by the way, driving while I was just talking this. I get to where I'm going and then I reviewed it all for an hour. And like normally like creating like a financial model or whatever it would be,
SPEAKER_01it would be like, you know, that takes days. And it's like, that's a hundred hours of work, right? Like in whatever it is, and it's just done. I think the trick is trying to do things in parallel because that way you're not like effectively bored. It's like, what would you normally have just outsourced to? Like we're all becoming managers of AI basically. And so like as a manager, my job is just outsource things as fast as I can to other people. And I think the same thing people should do with AI.
SPEAKER_02You know, this is not my original idea either, but during a workday, the average professional makes 150 transitions during a day between different things. And it's those transitions that kind of actually slow them down. At least that's the theory goes. And the question is like, how do you use AI in a way that you're not constantly building up your own context, if that makes sense?
SPEAKER_01Yeah. Ask the AI to organize your day so you don't context switch. Yeah, I mean, that's an interesting question. I don't know. I do think that's actually one of the challenges that we're going to have, like just in general as humans, is like trying to figure out because you could do so much. And I've talked to a lot of people and they're like, we're doing so much more work. I think to your point, it might not actually be more work necessarily, right? You still are working the same amount per day, but your transitions between them is exhausting. You constantly have to think about it. So I do think that's something people are facing. Yeah.
SPEAKER_02You know, obviously there were like kind of layoffs tied to AIs, but then there was like Uber saying, well, we haven't found any productivity gains. So it's hard to make sense of it all, I guess.
SPEAKER_01I think, well, the way I describe it is when you're in the singularity, which is like arguably where we are, right? It's almost like a black hole. And the closer you are to that event horizon, like where things are like constantly changing, you're going to experience things much more rapidly than other people. And so I like to like equate it to the Amish. It's like the Amish, they haven't changed in, what, 150 years? So between now and 150 years, they've progressed very little. But if you think about how fast technology is progressing, you'd be equivalent to the Amish today,
SPEAKER_01like 10 years from now, if you did not transition. Because that's how quick technology progresses. And I think a lot of people like Uber may be just further down that chain and not experiencing the latest technology, frankly. Or their analysis started two years ago, right? Well, it's like, yeah, two years ago, it wasn't very good. But like six months ago, it was really good. Yeah. Like I have some friends who are like, I'm the cutting
SPEAKER_02edge of AI. They're like, from the beginning, like I'm using clutch, GPT all in one and, you know, kind of saving time. But talking to them recently, I guess they're like, well, kind of like what you were saying, like a financial model that might have taken you, I know, a hundred hours, takes you an hour. And then they're like, well, it takes an hour. It's not as good as I would have come up with, but it's good enough.
SPEAKER_01Well, that's the thing. Is it good enough, right? Yeah. To be fair, the financial model I did create, I really did have to redo. So like from scratch, I was like, that wasn't good. So I just had to spend the eight hours to like fix it or whatever it was, you know. So I hear you. But like the other things were good enough. And I think the thing is, you don't know if you don't try. So you constantly are trying and just eventually it becomes good enough.
SPEAKER_02Yeah, I want to kind of bridge the gap. Do you want to discuss between what an IP co-pilot does, does this diagram of all the ideas, innovation within company, and, you know, drafting maybe chat GPT or Claude can do these days to some extent without any issues. What's the next step, I guess, you see kind of where AI can be helpful within?
SPEAKER_01So I think the big thing is that people think of the IP program as like a separate thing under the legal department. I think that's really weird. Personally, like it should be like almost inside your CTE under your technology division. And I think the big thing is the IP programs like looking over the technology division and constantly trying to pull out ideas. But I think what's going to end up happening is they're going to merge where you're going to have like a layer of insights over your like I call it like business intelligence over your entire system. Right. And you can use that a lot of different ways. I can say, like, build me the best team to solve this problem.
SPEAKER_01Or maybe I need to hire for these three technology categories because I did whatever. And I need these additional capabilities. I mean, part of it might be when you go to do layoffs, which bound to happen, Don't lay off like a core portion of your team, which like that happened repeatedly where I watched them and then they all had to get rehired back at like double the price. So I think there's like a bunch of uses outside of like classic IP that there's a ton of value. I also think cybersecurity is going to become a big deal.
SPEAKER_01Like who's the inventor, right? Is it the AI or is it the human is like going to become a problem? But then like also what if it becomes misaligned? Like, well, we're already tracking the alignment of how your ideas and concepts relate to the business. So I can start seeing if the AI becomes misaligned. I can start seeing if the human starts working on their side project at work, right? I think all of that actually relates and is like, in my view, should be part of the IP function, even though today it's not.
SPEAKER_02You know, you talk to a lot of patent attorneys and a lot of in-house teams. Is there anything that sticks out to you about what's like on top of mind when it comes to AI and patent practice?
SPEAKER_01I mean, there's obviously fear like it's going to take our jobs. I think you should lean into it personally and just be like, yeah, like, okay, let's try to make it automate my job because then you have more value to do other things. And you're also going to be more productive, right? And so in the meantime, you can make a lot more money potentially by getting through more patents. That's one thing I've just seen. I think the other thing that's interesting from my perspective is there's a lot of people are trying to skill up and use AI in different ways. And it's interesting to see how each person approaches it,
SPEAKER_01but they all end up coming to the same conclusion, which is like, okay, we can use it for drafting, but then they don't think about all the other things they can use it for. I think expanding the scope of potential uses is also going to be like an interesting component because I think a lot of people kind of put blinders on and they're really focused on like, how do I get my work done faster? You know, I always say like, you know, what an IP team cares about? It's like, okay, well, they don't want to look bad to their boss, like just like everybody else. And they want to do to get home on time. That's what most people want to do. And so if you can do that and you don't make them look bad and you can get them
SPEAKER_01at home on time, you're successful. And I think that's honestly where a lot of these tools go wrong and everything else. It's like, that's what you should be focused on.
SPEAKER_02You know, I'm curious about like the, it seems like a big part of your interaction with the patent process is you guys help generate these invention disclosures or maybe even track how often the inventors just accept what you guys generate.
SPEAKER_01It's interesting because we get like a really weird mixed bag where like people expect it to be perfect. And it's like, I know you miss your, like when you're doing searches too. So I think like, you know, I think there's, there's like an interesting, I don't want to say conundrum, but like challenge there. One of the things that you guys do is you do a
SPEAKER_02prior search.
SPEAKER_01It's like the base of our whole system, right? So we need to know, does this idea exist on the market in the patent space or on the web? We do NPL searches and everything else. Does it exist in video? Whatever it is. And then we also have to say, does it exist internally? And it's not being disclosed publicly. And so we're constantly running these searches to try to understand the mapping between the two. And the challenge to me is always, are we getting... We might miss something that people expect. It's like, well, did you actually review the prior art? And a lot of times they say no. I'm like, well, maybe it's better.
SPEAKER_01You don't actually know. And I think a lot of people have these preconceived concepts that they're already reviewing. The thing is, though, I think the biggest thing is the AI will hone in on a different thing that it thinks is unique.
SPEAKER_01So the search will actually be for something different than what you anticipate. And that's where I was saying it's a prompting problem. Like you got to make sure you're prompting it for the right thing to look for. But yeah, no, I totally see where you're coming from on that one.
SPEAKER_02Yeah, I'm curious, like the companies that your clients that have had most success with embracing your tool, like what did they do to kind of make it successful?
SPEAKER_01We had one person who just integrated with everything and then didn't even talk to the inventors. He just had them, he's like, you have a patent now. And he sent them the patent. It was like, please sign it. So that was he embraced it all the way, basically, is in my mind where it's like that's what I would be the dream. Right. I just as an inventor, I'm like, here, I would love to give you this patent. Can you sign it for me? It's like, OK, that's awesome. I think from other people who've like seen it work well is they integrate into their workflow. So they'll be using it to talk with outside law firms. They'll be using it to review the prior art and everything else and using it.
SPEAKER_01Primarily to collect the disclosure from the inventor. I think the other thing is people need to tune things a lot. For instance, in our system, you can tune the business value. What are you looking for as a business? We guess right up front. We have a pretty good education. We look at your 10K, your SEC filings, and all this other stuff. But it doesn't necessarily tell us what you as an IP individual is looking for. So what we found works best is we meet with one of our customers every week or every other week for the first two, three months, and we tune it.
SPEAKER_01And then they get really good results. It works decently out of the box, but once you tune it, you don't want to filter through a thousand ideas. You want to filter through the five that are the best ideas. And so it depends for each department, though, kind of what they're looking for.
SPEAKER_02You know, I'm curious, you said, you know, one of your clients just presents them with documents to sign for the patent application.
SPEAKER_01I presume that he also says, is this accurate?
SPEAKER_02It's interesting because obviously it's a balance because one of the things, and I've discussed many times, you know, the engineer's time is super valuable. So you want to find a way to kind of spend the least amount of time as you want. But at the same time, you want them to be bought in to what you're doing. Like, for example, your first encounter with the patent program you didn't know existed. Any thoughts about that, you know, with your product or without your products about the importance of having your engineers and inventors as champions for your patent program? And how do you use a tool like yours maybe to facilitate that?
SPEAKER_01I mean, I think in general, it depends on like the ethos of your company. I think some companies like, for me, right, it would be great if you just presented me with a patent and it was like, here's your patent, right? I'd say, here's the things I would add, right? And then you can add them or whatever. But that would be, I think, ideal in my scenario. I think for other corporations, you need the inventors to get bought in much deeper for whatever reason. Maybe they're on the patent committee or whatever. You want to build a patent committee system. You want to get really high-end patents that require technology, domain, expertise, etc.
SPEAKER_01I think certain things to really help with that is to have them feel involved in the process by creating patent committees or something like that. In our tool, you can have different committees and you can have patents auto-route and then people upvote or downvote and comment on them. And so we have seen a few people just create like custom committees and they'll just get auto routed and reviewed and then only move forward. And so effectively the patent IP team is like automating, like they don't even do like the initial review, the patent committee does the review. And so it's like the engineers feeling like almost like an internal like publication process, right?
SPEAKER_01Like they're all involved in it.
SPEAKER_02Yeah. And you know, that's also interesting obviously because your tool in some ways provides a recommendation, right?
SPEAKER_01Yeah, you don't have to show the, most people don't show the inventors. Okay, yeah.
SPEAKER_02So it would be interesting to kind of correlate that. I don't know if you'll ever have an opportunity to do that, to see what your tool versus the human evaluators.
SPEAKER_01We do keep track of that. We roughly match what humans end up reviewing. I think it depends on how strict. Some companies just have people who are just going to reject everything because they're like, you know, you know that person in the room, they're like, I'm the smartest one. I think that happens, you know, but I think in general, when we like look at distributions, we end up matching the distribution fairly closely, you know, plus minus 5% kind of deal. But that's kind of, it'll be very similar. And actually our scoring system was based on a person who gave us a bunch of disclosures. We then mapped our scoring system to their distribution.
SPEAKER_02So that way we knew we were at least matching them and we checked it with others and they all roughly, Well, that's interesting, obviously, if you have multiple panel review committees and one of them is out of bounds or whatever.
SPEAKER_01You'll have that one person who's just like, none of these ideas are good.
SPEAKER_02But it'd be good to have that kind of data to kind of show that, yeah. Any other, I guess, uses of your tools that kind of stand out to you that you think like, hey, this is a cool... It's something that I guess companies or clients have found like appreciation for.
SPEAKER_01Yeah, so my favorite component is honestly something we call Idea Atlas. So it takes all your internal knowledge and then it maps it to a 2D space, but then it compares it to your competitors. And then we have a tool where you can highlight it and ask questions about it. So you can say like, what are we developing in this area? And so each one of these dots represent either an idea Or concept internally, a disclosure or other patents. And you can see the white space where there's like, no one is like operating in, but you can file in. And then you can also click a button and we'll show you the emerging areas. So things that have changed over the last six to 18 months where people are starting
SPEAKER_01to file more or file less.
SPEAKER_02Within the company or just in general?
SPEAKER_01It could do inside the company or external. And then we'll summarize it. We'll say like, you know, we see like company X is now filing a bunch of stuff related to like kitchen appliances or something. And then here's the kitchen appliances. Here's the inventors. And, you know, with that information, you could start poaching those people if you want to, right? Like there's a bunch of use cases, but I think you can make, the easiest thing to do is that you can get budget with it. So you can make the argument like, hey, we need to patent all these ideas right here because we can see the competitors moving in that area. And it's like, it's something that like, No one even thinks about it, but our tool just does, and it's kind of cool.
SPEAKER_02Obviously, one of the difficult things about prior art searching, especially in a fast-moving field, let's say, like, AI, is there's a delay sometimes. Is there a way to kind of compensate for that, to kind of say, like, well, there's nothing right on it, but it's getting crowded over here probably.
SPEAKER_01Yeah, so our system predicts where the directionality of the concepts is moving. So we can actually say, like, there's no guarantee it continues. We can say, this is the direction it's moving. You should, they'll be here at some point. Yeah. You should file on those, right?
SPEAKER_02So interesting, the other way around, yeah.
SPEAKER_01But you can track that, right? There's an 18-month delay, so you're going to have to project it out a little bit further. And like I said, there's ambiguity, but you can definitely do that.
SPEAKER_02Well, I want to kind of get your thoughts. You know, you told me in 2018, you were sitting at work, you were telling your team, like, we want to have our jobs.
SPEAKER_01Yeah.
SPEAKER_02We got to be ready. What if there's somebody else kind of listening to this, let's say a patent attorney or not a patent attorney, and they're thinking like, all right, I want to find a way to solve some kind of problem that general AI or LLM can just be used to solve easily, kind of like Austin did. What advice would you give, like something that's worthwhile to focus on?
SPEAKER_01That's a good question. I think the biggest thing where like IP attorneys and like LLMs and stuff can provide a lot of value. Is looking through documents and summarizing documents. So if you're looking at like what an IP attorney can apply it to, it's like, you know, you could now analyze things you never could have analyzed before. So if you're trying to build a case, you can say, read all the documentation from that company. Sure, it takes five days. Who cares? I mean, that would have taken you 500 years. Like, you know what I mean? So like, you'll get that results back and you're able to actually analyze it and make better determinations. I think you got to constantly push yourself and push it beyond where you think is reasonable.
SPEAKER_01Because it will surprise you. And so I think that if I'm an IP attorney, that's what I'd be trying to do is just constantly push the bounds on what's possible. I think I would also try to get more into litigation or get more into where it's people-to-people interactions because I think content generation is going to go away at some point. I don't know when, but it will.
SPEAKER_02What about, like, I think you said 2030, I think that's your call. That's going to be the big difference. You know, let's say we are, you've kind of hinted at it, but like in 2030, we're sitting here, which I guess at this point is only four years away or three years. What will things look like, I guess, in the spam space?
SPEAKER_01Yeah, so I think my personal opinion is I think AI drafting is already taking off. I don't know what percentage it is. I'm guessing it's 20% of people are using AI in some way. I'm guessing in the next year it'll be 50%. I'm guessing the year after that it's going to be like 70%. It's just going to keep growing. I think at some point, though, it'll be a one-button click. And once it becomes a one button click, it like doesn't matter if you like your patent attorney or anything else. It's like no business is going to justify spending $30,000 on a patent. They're just not. They're going to be like, yeah, spend the 200 bucks or whatever it is.
SPEAKER_01And they're just going to risk it. Now, I think there's going to be obviously like those boutique firms that survive, but I think it's going to shrink down to a small number. I think other services you could provide, though, like, for instance, like going in and interviewing inventors, like that's not going to go away. So I think we're going to shift where people are kind of operating. But I don't think people are going to be drafting patents. I think it's going to be shifting more to the operational and customer service kind of space where you're talking to the USPTO, doing interviews, you're talking to the vendors doing interviews,
SPEAKER_01you're doing the customer relationship to keep money. I think that's where you're going to spend more of your time. I think IP is going to grow rapidly across everybody because it's going to be much easier to enforce.
SPEAKER_02What about like in the in-house perspective? How do you think in-house attorneys, if they're outsourcing, pressing the button for that application, you know, what do you think?
SPEAKER_01I think they're going to insource pressing the button, to be honest. But they're probably not going to insource like talking to the inventors because you can't. And you need to be able to expand or shrink that. And they also don't, like I said, don't want the liability. So I think what they're going to do is send drafts. I think we're going to collapse a lot of in-house counsel and outside counsel. So you're going to have a lot more companies that have outside counsel that maybe have in-house counsel today.
SPEAKER_01Or you're going to have a lot of companies that have in-house counsel and no outside counsel. And I think that's how it's going to split.
SPEAKER_02You obviously talked to a lot of in-house teams putting the drafting stuff aside. You know, they have your idea graph. You know, maybe they'll work closer with CTO office. What do you think that will kind of look like, the IP function within a company's?
SPEAKER_01I mean, I think a lot of things are going to collapse internally, right? So, I mean, it might be like the CTO is managing the IP function to a degree. You know, I think maybe the IP council is doing something else. Or like, you know, they have different job descriptions. Yeah, I mean, I hope that that's what happens, right, for my business, obviously. But like, I think in general, it's going to be very easy to obtain an understanding of people's businesses. We're seeing that with other AIs too, right? I mean, ChadGPT, you can plug in. I think you're going to have agents though, acting independently in organizations.
SPEAKER_01And I think that's going to become a bigger, bigger thing. I think like next year, You may have AI invent more or as much as humans inside of large corporations that have deployed AI. And that's kind of wild to me. I'm like, I don't even know how you track that. That's going to be a whole separate thing. We have a system in our thing that tracks, are you an agent or are you human? But we're seeing even in our own system, AI writes dramatically more than we do. So we'll write one sentence, then it takes all the context and builds us a huge document.
SPEAKER_01AI is producing like 90% of our content internally.
SPEAKER_02I think that's another problem. The companies are trying to, going back to what we were discussing, figure out, actually measure the value of the AI or whatever, the tokens that they're using. I guess it's not a very simple problem to solve.
SPEAKER_01We host our own AI, which is kind of nice because we're kind of unbound by tokens. It's more just like, how much GPU capacity do you have? So things get slower, but... Works for us. Because we're spending like, you know, one hundredth the cost of what you will for some of these other AI systems.
SPEAKER_02Any final thoughts as you think about AI practice, where things going?
SPEAKER_01I mean, I think it's really exciting personally, like from an engineering perspective, I think like there's going to be a lot of changes and I think people just have to be willing and able to adapt. Like I've adapted to engineering changing, right? My whole team has. I think it's fairly straightforward to do so, but you just have to like accept what's happening and then figure out how to make yourself more productive in that environment. So I don't think like anyone's losing their job necessarily. I think you said it before. It's like people are actually going to probably work more and they're going to work harder because they have to change what they're doing.
SPEAKER_02Well, Austin, thank you for sharing your wisdom and your insights. And we should have that conversation in 2030 to find out where things land.
SPEAKER_00Thanks for listening to Clause 8. Subscribe at voiceofip.com for more IP conversations and find extra content on the Clause 8 YouTube channel. See you next time.
Perry Saidman passed away on October 15, 2026, less than a month after we recorded this conversation. Perry built the market for design patents. His win in Avia v. LA Gear, a sneaker knockoff case, helped prove design...
When Coke Morgan Stewart became acting director of the USPTO, much of the patent world knew little about her. In this episode of Clause 8, Stewart, now President and CEO of the Council for Innovation Promotion (C4IP),...
To open the season, host Eli Mazour brings together three journalists who cover intellectual property for a living: Olivia Rafferty of IP Fray, Nisha Shetty of IAM, and Inbar Preiss of MLex. They discuss what happens...
Kevin Mack, the new president of Via Licensing Alliance, joins Eli for the Clause 8 season finale. Kevin talks about Via's plans to build its next patent pools around "de facto" standards - technology the market adopted...
Bruce Lehman, head of USPTO from 1993 to 1998, joins Clause 8 for a wide-ranging conversation about the modern IP system, the internet boom, and why the AI era may require a new copyright response from Congress. Lehman...