In 2017, Google published a research paper called “Attention Is All You Need.” It quietly introduced the transformer—the architecture behind today’s AI boom.
They didn’t patent it.
Fast forward a few years, and that single decision helped unlock an entire industry… but also left billions in potential defensibility on the table.
That tension—between speed and ownership, innovation and protection—is exactly where Dina Blikshteyn operates.
The Founder’s Blind Spot: You’re Moving Fast… But Leaking Value
Most founders don’t ignore intellectual property.
They just delay it.
You’re building, pitching, hiring, shipping. IP feels like something you’ll “get to later”—right after product-market fit, right after your next round, right after things slow down (which they never do).
Here’s the problem:
IP has a clock. And it starts ticking the moment you talk.
Dina Blikshteyn, Partner at Haynes Boone and co-chair of its AI practice, sees this pattern constantly. Founders disclose their ideas in pitch meetings, conferences, or demos—only to realize later that they’ve accidentally disqualified themselves from global patent protection.
In the U.S., you get a one-year grace period after public disclosure.
In most of the world? You get zero.
The startup doesn’t die instantly. But something more subtle happens:
your “crown jewels” are no longer defensible.
The Great Misconception: “We Don’t Need Patents”
There’s a popular narrative in startup land right now:
“Speed is the only moat.”
It sounds right. It feels right. And in some cases—it is right.
But it’s also incomplete.
Dina breaks founders into two camps:
Those who think everything they build is novel
Those who think nothing is worth protecting
Both are usually wrong.
The real game isn’t patenting everything. It’s identifying where your leverage actually lives.
If you’re building on top of large language models, your moat probably isn’t the model itself. It’s:
How you structure inputs and outputs
How your system orchestrates multiple components
The specific workflows or outcomes you enable
In other words:
not the engine—but the system around it.
That’s where smart IP strategy starts.
The Trade Secret Trap
Many founders default to a seemingly clever workaround:
“We’ll just keep everything a trade secret.”
On paper, it sounds efficient. No legal fees. No filings. No waiting.
In reality, it’s fragile.
Why?
Because startups are inherently leaky systems:
You pitch investors
You hire (and lose) employees
Your team publishes research
Your product reveals behavior
All it takes is one leak—and the “secret” is gone.
Even worse: someone else can independently file a patent on a similar idea.
Now you’re not just unprotected.
You’re potentially blocked.
AI Changes Everything… and Nothing
At first glance, AI feels like it breaks the entire IP system.
Models evolve every few months.
Products iterate weekly.
Entire categories appear and disappear in a year.
So why bother with patents that last 20 years?
Here’s the paradox Dina highlights:
Most patents aren’t valuable when they’re young.
They become valuable when the world catches up.
Some of the most powerful patents being enforced today were filed over a decade ago. If written well, they don’t lock onto a specific implementation—they cover future variations.
That’s the real art:
Writing claims that survive the next wave of technology.
The Real Battlefield: Not Models, But Systems
There’s a quiet shift happening in AI.
Early on, the focus was on foundation models—who builds them, who owns them.
Now, the battleground is moving up the stack.
Founders are asking:
Can I patent something if I’m “just” building on OpenAI?
Is my product defensible if the model is interchangeable?
Dina’s answer is nuanced:
You don’t patent the model.
You patent how you use it.
Think of it like this:
Everyone has access to the same Lego blocks.
But not everyone builds the same machine.
The invention isn’t the brick.
It’s the configuration.
Regulation: A Moving Target (and a Strategic Variable)
While founders are figuring out product and IP, governments are trying to catch up.
The U.S. is currently taking a relatively hands-off approach—prioritizing innovation and speed.
Europe has gone the opposite direction, introducing the EU AI Act with structured risk categories and compliance layers.
For startups, this creates a strange dynamic:
Move fast in the U.S., but with uncertainty
Move cautiously in Europe, but with clarity
And looming over all of it is one unresolved question:
Is training AI on copyrighted data… legal?
There are dozens of active cases right now. The outcome could reshape how value flows across the entire AI ecosystem—from creators to model providers to startups.
If history is any guide (think Napster → Spotify), we’re likely heading toward a new economic model, not just a legal ruling.
The Hidden Advantage: AI for Lawyers
Ironically, AI isn’t just changing what gets protected—it’s changing how protection itself works.
Tools like Solve Intelligence are automating the “busy work” of legal drafting, allowing attorneys to focus on strategy and claim design.
The result?
Faster filings
Better quality patents
More leverage for founders who use them early
In a way, lawyers are becoming product managers—designing protection systems instead of just documents.
So What Should Founders Actually Do?
If you strip everything down, Dina’s advice is surprisingly practical:
Don’t wait. If you’re talking about it publicly, you should be thinking about IP.
Focus on the core. One or two strong patents beat ten weak ones.
Think globally, act selectively. File where it matters for your business.
Balance protection types. Patents, trade secrets, and copyrights all play different roles.
Work with experts who understand your tech. Not all legal advice is created equal.
Zooming Out: The Real Game
Every technological shift creates a window.
For a brief moment, the rules are unclear. The system hasn’t caught up. The boundaries are still soft.
That’s when the biggest advantages are created.
We’re in that window now with AI.
The founders who win won’t just build faster.
They’ll lock in their advantages while the system is still forming.
Because in the end, innovation isn’t just about creating something new.
It’s about making sure it’s still yours when everyone else arrives.
👂🎧 Watch, listen, and follow on your favorite platform: https://tr.ee/S2ayrbx_fL
🙏 Join the conversation on your favorite social network: https://linktr.ee/theignitepodcast
Chapters:
00:01 Introduction to Dina Blikshteyn and AI + IP background
00:42 Dina’s origin story: engineering, Wall Street, and law
01:47 How technical background shapes IP and AI legal work
02:23 Key challenges in patenting AI and emerging tech
03:03 Why startups delay IP—and the consequences
03:59 Patent basics: filing vs protection timeline
05:04 International patent strategy for startups
06:26 Common founder misconceptions about patentability
07:41 Patent vs trade secrets vs trademarks vs copyright
08:19 Risks of NDAs and IP leakage in startups
09:35 Publishing vs protecting IP in AI research
10:53 Patent surprises and broad claims in emerging tech
12:05 OpenAI, patents, and shifting strategies in AI
14:21 Comparing AI to past platform shifts
15:07 Patent enforcement and proving infringement
17:39 Litigation, settlements, and patent dispute dynamics
19:18 Famous patent cases and startup vs big tech battles
21:41 Lessons for startups from major IP cases
22:45 How AI tools are changing patent workflows
24:24 Are moats dead? Rethinking defensibility in AI
25:19 What AI startups should actually patent
26:25 Patent lifespan vs fast-moving tech cycles
27:54 Open source vs proprietary IP strategies
29:24 Evolution of AI regulation (US vs states vs EU)
31:23 How regulation impacts innovation and startups
34:04 AI governance frameworks (NIST, ISO)
35:04 Future of AI regulation and legal landscape
36:01 AI copyright lawsuits and fair use debate
38:01 Derivative works, copyright, and AI-generated content
40:22 Implications for creators and content economics
41:03 Rapid fire: AI and IP misconceptions
45:14 Closing thoughts and future of AI + law
Transcript
Brian Bell (00:01:17): Hey, everyone. (00:01:17): Welcome back to the Ignite podcast. (00:01:19): Today, we’re thrilled to have Dina Blikstein on the mic. (00:01:21): She is a partner in the Intellectual Property Practice Group at Haines Boone in New (00:01:25): York, (00:01:26): co-chair of the firm’s artificial intelligence practice, (00:01:29): an expert at the intersection of patent law and emerging tech, (00:01:32): and a frequent speaker and writer on AI, (00:01:34): patent eligibility, (00:01:35): and tech regulation. (00:01:36): Dina brings both technical expertise from her early career developing (00:01:40): high-frequency trading systems and seasoned legal perspectives shaping how (00:01:44): innovators protect and deploy their technologies.
Dina Blikshteyn (00:01:47): Thanks, Brian. (00:01:49): It’s always great to be here and do a podcast.
Brian Bell (00:01:51): Yeah. (00:01:51): So I’d love to get your origin story. (00:01:53): What’s your background?
Dina Blikshteyn (00:01:54): So my background is in computer and electrical engineering. (00:01:57): So I was a computer geek. (00:01:59): I did a lot of computer engineering, computer science work. (00:02:03): Ended up on Wall Street and then decided after spending about five years there that (00:02:09): I needed more human interaction and (00:02:12): And went to law school at night. (00:02:13): So then lo and behold, (00:02:15): you know, (00:02:15): we can fast forward 15 years later and I am still doing technology, (00:02:20): but on the legal angle. (00:02:22): And now I’m also, you know, throw AI into the mix. (00:02:24): So extremely math heavy. (00:02:26): So now it’s computer engineering, (00:02:28): computer science, (00:02:30): AI, (00:02:31): and...
Brian Bell (00:02:31): Yeah, (00:02:31): so you very much live at the intersection of law and IP, (00:02:35): but also like AI and technology. (00:02:37): Yeah, very interesting. (00:02:37): I had the same realization 20 years ago when I worked on Wall Street, (00:02:40): you know, (00:02:40): staring at a spreadsheet and analysis all day, (00:02:44): every day was not for me. (00:02:46): I enjoyed the work, but not 80 hours a week or whatever. (00:02:48): How do you think your technical background shapes your approach to IP and later AI?
Dina Blikshteyn (00:02:53): Quite a bit, (00:02:54): actually, (00:02:54): because with the technical background and then also working in the industry, (00:02:59): I know how technical systems work on a perspective of a developer. (00:03:04): So when I talk with inventors about either helping them obtain patents or (00:03:09): invalidating patents, (00:03:10): all that background comes into play. (00:03:12): So it’s a lot less theoretical and more hands-on. (00:03:16): And then applying that to the legal concepts.
Brian Bell (00:03:19): What are some challenges you faced when working on patents and cutting edge tech fields?
Dina Blikshteyn (00:03:25): Challenges when I help inventors get patents?
Brian Bell (00:03:28): Yeah. (00:03:28): What are some of these challenges around, let’s just take AI. (00:03:32): We’re an early stage venture capital firm. (00:03:33): We’re backing, you know, around a hundred early stage firms a year. (00:03:38): As founders approach us, you know, how should we counsel them? (00:03:41): Obviously get them in touch with you, (00:03:42): but what are some of the challenges that they face, (00:03:45): you know, (00:03:46): when they’re trying to protect their IP?
Dina Blikshteyn (00:03:48): So, you know, the biggest challenge that I’m seeing is IP is not a priority, right? (00:03:54): And neither it should be when you’re just a startup, but that comes with consequences, right? (00:03:59): So in US, (00:04:00): you have a one-year grace period to file a patent application from an earliest (00:04:04): disclosure. (00:04:05): And other countries, it is an absolute bar. (00:04:07): Essentially, you have to file for a patent application before any disclosure. (00:04:11): So at the time the startups actually figure out they need a patent and want to file (00:04:16): a poor patent, (00:04:17): it’s just too late. (00:04:18): And the patent protection is not available to protect their crown jewels. (00:04:22): Now they can file on improvements, (00:04:24): but more often than not, (00:04:25): you want to protect the core technology. (00:04:27): And sometimes it’s just from a legal perspective, it’s too late.
Brian Bell (00:04:30): Maybe you could define for the audience what the notices versus the actual patent. (00:04:34): There’s a lot of entrepreneurs that listen and probably don’t know the difference.
Dina Blikshteyn (00:04:38): All right, (00:04:38): so when you file a patent application, (00:04:41): right, (00:04:41): you essentially get a date from the patent office of when the patent application is (00:04:46): filed, (00:04:47): right? (00:04:47): It takes about three years to have the patent issued, (00:04:50): sometimes a bit more, (00:04:51): sometimes a bit less. (00:04:53): But your protection starts from the filing date. (00:04:57): Now, (00:04:57): if you’re a startup going 100 miles an hour and trying to develop technology, (00:05:02): you will be talking to different VCs or you will be talking about it publicly or (00:05:07): submitting submissions to different conferences that outline your technology. (00:05:11): All of that is public disclosure. (00:05:13): And that starts the clock for filing the patent application. (00:05:17): So in the US, you have one year to file from the earliest disclosure that you make. (00:05:22): And that’s without the NDA. (00:05:23): And then the rest of the world, you can’t do any disclosure.
Brian Bell (00:05:26): Yeah, (00:05:26): maybe tease that out a little bit, (00:05:28): you know, (00:05:29): because a lot of the startups we back, (00:05:31): obviously, (00:05:31): are U.S. (00:05:31): based. (00:05:32): How should they be thinking about the international filings, if at all?
Dina Blikshteyn (00:05:35): So you want to file where A, you can protect the technology and B, where you’re doing business. (00:05:41): You don’t need to file in every single country in the world. (00:05:44): And we recommend against those types of filings. (00:05:47): So if you’re in the tech space, typically it’s U.S., Europe. (00:05:51): And then you pick several countries in Asia. (00:05:54): It can be China, India, South Korea, for example, sometimes Japan. (00:05:58): But it always goes back to the business case, right? (00:06:01): Where will people use your technology and where it will be sold? (00:06:06): And also how much of an appetite you think you’ll be protecting it worldwide. (00:06:10): Now, I know if you’re a VC, you’re looking at it the opposite way, right? (00:06:14): Where is the technology being protected? (00:06:16): when you’re looking at different companies. (00:06:18): So it’s always a calculus, (00:06:20): but it’s a business calculus more than a legal calculus of where a company should (00:06:25): be filing versus what would be the long-term benefit from those, (00:06:29): right? (00:06:29): And you don’t need to file on everything, right? (00:06:31): You don’t need to file on every single technological improvement, (00:06:34): especially for startups, (00:06:36): like our recommendation, (00:06:37): one to two patents. (00:06:39): You can start in US and then you can decide whether you want to file globally down (00:06:43): the road and make sure that’s directed at the crown jewels of the technology.
Brian Bell (00:06:48): So you probably talked to lots of founders who think they have really unique IP, (00:06:52): but upon review, (00:06:53): they don’t. (00:06:54): Tell us some of the common traps or pitfalls or ways that founders actually don’t (00:06:59): have a patentable IP that they think they do.
Dina Blikshteyn (00:07:02): Right. (00:07:03): Okay. (00:07:03): So there’s usually two groups of people that one group thinks that everything’s (00:07:09): novel and the other group that thinks that nothing is novel and we’re going to file (00:07:13): a patent. (00:07:14): For if you’re money conscious, especially in the beginning, you can always do your own. (00:07:20): search to see what’s out there and the easiest and free way to do it is to do (00:07:26): google patents and then you can kind of see what’s already out there something to (00:07:30): remember is that most patent applications don’t get published right away there’s (00:07:36): about an 18 month lag unless a patent issues before that so you’re a bit behind the (00:07:43): technology when you do when you do those searches (00:07:45): But at the same time, (00:07:46): doing a search can give you a fairly quick sense of what is out there (00:07:51): technology-wise.
Brian Bell (00:07:52): So how should a startup think about patent strategy versus other forms of (00:07:55): protection like trade secrets or trademarks?
Dina Blikshteyn (00:07:59): Okay. (00:07:59): So there are four types of protections, right? (00:08:03): Then you mentioned three of them and the last one is copyright. (00:08:06): And they all work together, right? (00:08:09): So a lot of startups think, well, we’ll just do everything a trade secret. (00:08:13): Well, (00:08:13): It doesn’t work for a number of reasons. (00:08:17): One, when you shop around, you are doing the disclosure under the NDA. (00:08:21): And then you want a risk of someone saying, (00:08:23): well, (00:08:23): I’m going to file a patent application on your technology.
Brian Bell (00:08:27): So tell us more about that. (00:08:28): There’s a disclosure under NDA. (00:08:30): What does that mean?
Dina Blikshteyn (00:08:31): Disclosure under an NDA is when you sign a document. (00:08:34): It’s essentially a non-disclosure agreement where if I’m a startup, (00:08:38): I’ll say, (00:08:39): I will tell you about my invention and you agree not to disclose it outside of our (00:08:44): conversation, (00:08:45): right? (00:08:45): Does it always work practically? (00:08:47): Maybe yes, maybe no. (00:08:49): Right. (00:08:49): You can also have instances where you have developers leave. (00:08:52): Right. (00:08:53): And while they can’t take the source code, (00:08:55): because that would be covered under an NDA as well, (00:08:58): or non-compete agreement, (00:08:59): they can still take the know-how and try to build their own product. (00:09:03): Right. (00:09:05): And then the other one, (00:09:06): and this is always true in the AI space, (00:09:09): since it’s such a hot and emerging field, (00:09:11): a lot of talent likes to publish at conferences. (00:09:14): And what do they publish on? (00:09:16): Whatever they are inventing, whatever they’re coding, whatever their research is in. (00:09:21): So while you as a company or startup think that something can be a trade secret, if you have a (00:09:26): paper that publish something the same or similar, (00:09:30): that is the genie out of the bottle that you can.
Brian Bell (00:09:32): Yeah. (00:09:32): So like if I’m a AI researcher at Google, (00:09:35): you know, (00:09:36): name your favorite company and, (00:09:37): you know, (00:09:37): I really, (00:09:38): I’m on the research team. (00:09:39): I’m not in the applied engineering team actually shipping products. (00:09:42): I’m doing like primary research. (00:09:44): There’s this kind of push that I, you know, if I’m a PhD, I want to publish. (00:09:47): What are some considerations for organizations as they scale? (00:09:51): to kind of maybe say, (00:09:52): hey, (00:09:52): researcher, (00:09:53): maybe don’t publish that in archive or at the next, (00:09:57): you know, (00:09:57): AI conference. (00:09:58): We’re going to, how do you, how should organizations think about that?
Dina Blikshteyn (00:10:01): You know, it’s always a balance between, you know, small, medium and large organizations. (00:10:06): And if you have, if it’s a large organization, there’s typically a program in place (00:10:12): Where they say, (00:10:12): yes, (00:10:13): go ahead, (00:10:13): publish, (00:10:14): but we will file for a patent application the day before you do the submission. (00:10:18): That way, the technology is protected. (00:10:21): And if you’re a researcher wanting to publish, (00:10:23): you can go ahead and submit your paper for the publication. (00:10:27): If you’re looking at the startups or you’re looking at medium-sized companies, (00:10:30): it becomes more problematic because at that point, (00:10:33): they have to make a business decision. (00:10:35): Do we submit or not submit or do we file for the patent application or not?
Brian Bell (00:10:40): Are there any, to the degree that you could speak about, any interesting stories? (00:10:44): I mean, (00:10:45): you could change the names to protect the innocent, (00:10:47): but are there any interesting stories where you were surprised or where you’re (00:10:52): like, (00:10:52): okay, (00:10:52): this is like a shoo-in, (00:10:54): this is definitely a protectable IP and it wasn’t or vice versa? (00:10:57): You didn’t think it was a protectable IP and it ended up being?
Dina Blikshteyn (00:11:00): I’ve seen a lot, (00:11:01): a lot of cases where I didn’t think something would get patented, (00:11:05): and lo and behold, (00:11:06): it was. (00:11:07): And not only was it patented, the protection was also extremely broad. (00:11:12): So that’s the surprises you have, especially in the tech space. (00:11:15): Right now, (00:11:16): in the healthcare and AI, (00:11:17): I expect to see a lot of those types of claims there as well. (00:11:21): Because you have to remember, (00:11:22): every time there is a new technology, (00:11:25): the patent office examiners have to come up to speed on the technology. (00:11:29): And while they’re coming up to speed, (00:11:31): that’s really your chance to file for patent applications and get broad claims, (00:11:36): right? (00:11:37): Because then once the field gets saturated, the claims will become lengthier and more narrow.
Brian Bell (00:11:42): Yeah, (00:11:43): I mean, (00:11:43): the most famous example in the current platform shifts that we have is the (00:11:47): Transformer architecture out of Google, (00:11:49): which was published in 2017, (00:11:51): right? (00:11:52): But since they publish it, it’s not protected. (00:11:54): Yeah. (00:11:55): I bet Google, (00:11:56): do you think Google’s like, (00:11:56): oh man, (00:11:57): maybe we should have, (00:11:58): you know, (00:11:59): instead of publishing that research, (00:12:00): we should have just, (00:12:01): you know, (00:12:02): protected it with a patent?
Dina Blikshteyn (00:12:04): I have a counter story to that. (00:12:06): So this is one with, it’s publicly known with OpenAI. (00:12:11): So when OpenAI became public and the HIGPT came out, (00:12:15): It was all about, (00:12:16): well, (00:12:16): we don’t need patents because all this technology has existed long before. (00:12:21): And then long behold, after ChatGPT became popular, (00:12:25): they started filing for patent applications on improvement and they were getting patents. (00:12:30): And they were also filing under track one, (00:12:32): which essentially for track one, (00:12:35): you pay an exorbitant fee in thousands of dollars and you can jump to the front of (00:12:39): the line at the patent office. (00:12:41): So you can-
Brian Bell (00:12:42): So they have like the Disney fast pass for patents where you pay some (00:12:46): extra money-
Dina Blikshteyn (00:12:47): Exactly. (00:12:47): And the patent office will review you next. (00:12:49): Exactly. (00:12:50): So you can pay for-
Brian Bell (00:12:50): Yeah. (00:12:51): Then it’s where you can get a number of patents. (00:12:54): pursue that route. (00:12:56): So what I started seeing is all of a sudden you have open AI who went from the (00:13:00): mantra, (00:13:00): we don’t need patents to filing a lot of patent applications in the LLM space. (00:13:06): How has this platform shift? (00:13:08): I mean, obviously you started 15 years ago in law, you said. (00:13:12): So you kind of started during the kind of the mobile and cloud kind of boom. (00:13:16): How is this AI platform shift the same or different from kind of the previous (00:13:22): platform shifts that we’ve lived through?
Dina Blikshteyn (00:13:24): You know, it’s, it’s, I can’t say it’s different, but we are on this cups. (00:13:29): There’s a lot of innovation going on. (00:13:31): So let’s, (00:13:32): a lot of companies are just filing for a lot, (00:13:34): a lot of patents just to have their piece of the pie. (00:13:38): And then, so I’ve seen that a lot in the beginning around 2018. (00:13:42): picking around 2022, (00:13:44): where you can actually file for a patent application and get claims that were (00:13:48): fairly broad from the patent office. (00:13:50): Right now, (00:13:50): you’re sort of seeing the field being saturated a bit, (00:13:54): but now you’re seeing innovation based on technology types. (00:13:58): So, (00:13:58): for example, (00:13:58): AI and healthcare is becoming extremely popular, (00:14:02): and I see a lot of filings in that space.
Brian Bell (00:14:04): Makes sense. (00:14:05): So I get my patent. (00:14:07): I’m scaling, revenue is going up and to the right, everybody’s happy. (00:14:11): But then I notice three, four, five competitors crop up. (00:14:15): How do I inspect and prove that they’ve somehow stolen my IP? (00:14:21): I’m, you know, I’m looking at it and like, this is, this feels the same as what I’m doing. (00:14:26): Like it’s, it’s, you know, same input gets the same output. (00:14:30): There must be something in that black box that I’m not seeing that I’m curious or suspicious of. (00:14:34): Right. (00:14:35): So how do you prove that that’s from the outside looking in?
Dina Blikshteyn (00:14:37): It’s, it’s pretty hard, right? (00:14:38): It’s not always easy. (00:14:40): You have to look at the claims of your patent. (00:14:42): And again, (00:14:43): the claims, (00:14:43): not the entire specification, (00:14:45): because claims is what is where the protection is. (00:14:49): And you can, (00:14:50): search online and you can see if the claims map to their technology. (00:14:54): Now, (00:14:54): if you’re sure, (00:14:56): you can always file a complaint with the courts and that would open up discovery (00:15:00): where you can actually go into their systems and see how they work.
Brian Bell (00:15:04): So is it an impartial third party that does that investigation? (00:15:07): It would have to be, right? (00:15:08): Because if company A is accusing company B, (00:15:10): it’s not like company A’s engineers, (00:15:12): the plaintiffs are going to get access to the defendant’s systems, (00:15:16): right?
Dina Blikshteyn (00:15:17): So the plaintiffs may not get access, (00:15:19): but their attorneys may get access and they may not be able to disclose to their (00:15:23): client what they’ve been looking at or they hire an expert to do that. (00:15:27): And the expert can tell the attorneys, yes, there is infringement. (00:15:30): No, there is no infringement. (00:15:32): And then the attorney can take that to the client.
Brian Bell (00:15:34): Really interesting. (00:15:35): Yeah. (00:15:35): So the attorneys get involved, (00:15:36): but then you guys have to, (00:15:38): you know, (00:15:39): have somebody like you on staff that understands code, (00:15:41): understands the space, (00:15:43): but also you can hire experts in discovery to, (00:15:46): to sort of make that determination.
Dina Blikshteyn (00:15:48): Exactly. (00:15:49): And, and, you know, you don’t always pursue the litigation route, right? (00:15:53): Sometimes until there is infringement, (00:15:55): sometimes it’s your former employee who opens up their own company. (00:15:59): So at that point, (00:16:01): you can kind of see that they’re all doing almost the same thing and that may be (00:16:05): covered by your patent. (00:16:06): So sometimes it’s just a communication saying, (00:16:09): I believe you’re infringing my patent and you can send the system the system (00:16:13): letter. (00:16:13): Would that always work? (00:16:14): Probably not. (00:16:15): But you also have to do a calculus. (00:16:17): How much money is at stake?
Brian Bell (00:16:19): Yeah. (00:16:19): So what percentage of patent cases end up in like a settlement? (00:16:24): Like what are some of the outcomes and rough statistics around this space? (00:16:28): This is completely fascinating. (00:16:29): I have no idea.
Dina Blikshteyn (00:16:30): Right. (00:16:31): So it’s not an easy question to answer because there’s a lot more complaints filed than trials. (00:16:37): So someone gave a statistic that it’s about 5% of cases actually make it to trial (00:16:43): and most of them settled before that. (00:16:45): Right now, the question is, when will they settle? (00:16:48): Will they settle in the beginning? (00:16:50): Will they settle closer to the trial? (00:16:52): That depends on the discovery that can take, (00:16:55): whether there is actual infringement or what else is going on in the case. (00:16:58): There’s also a parallel proceedings called post-grant proceedings in front of the (00:17:03): patent office, (00:17:05): right? (00:17:05): So say if I have a patent and I sue you for patent infringement, (00:17:10): you can file a petition at the patent office saying that that (00:17:15): patent is invalid, right? (00:17:17): And now the proceedings from the patent office, they’re very quick. (00:17:21): They’re about a year from institution to the final decision. (00:17:25): So roughly 18 months from start to finish if you count in the preparation time. (00:17:32): So now you can use the post grant proceedings to force settlement and litigation. (00:17:39): So there’s a lot of different strategies that you can take.
Brian Bell (00:17:41): That’s fascinating. (00:17:42): Are there any famous examples where it made a company successful or broke a company (00:17:48): where a company was like, (00:17:49): wow, (00:17:49): we got to shut down or eventually just kind of petered out because there was a case (00:17:55): decided against them that was existential. (00:17:57): I don’t want to talk about them.
Dina Blikshteyn (00:17:59): Publicly well-known ones, you know, like the Amazon buy now button comes to mind, right? (00:18:03): They had the buy now button. (00:18:04): They, they put a patent on it. (00:18:06): Other people tried to copy it. (00:18:07): And I think Amazon sued them and won. (00:18:09): Right. (00:18:10): There’s also an eye for eye case, (00:18:12): which is actually a few, (00:18:14): that one I can probably talk about because it’s a, (00:18:17): It’s more than a decade old, right? (00:18:19): Where it was also post-grant proceeding that I think ended up in a $400 million (00:18:26): verdict in district courts, (00:18:27): right? (00:18:27): So that is an example of a startup taking on Microsoft, right? (00:18:32): And Microsoft did not want to settle.
Brian Bell (00:18:33): Wow, they took it all the way to trial. (00:18:35): Took it all the way to trial.
Dina Blikshteyn (00:18:37): And that was a $400 million verdict in 2010, 2011.
Brian Bell (00:18:42): What was the technology that was in dispute?
Dina Blikshteyn (00:18:45): It was the Word document technology, like XML type technology.
Brian Bell (00:18:50): Right, right. (00:18:51): And what was the startup in this case?
Dina Blikshteyn (00:18:53): i4i.
Brian Bell (00:18:53): i4i, I don’t recall them. (00:18:55): And maybe that’s indicative of this $400 million ruling kind of put them out of (00:18:59): business, (00:18:59): basically.
Dina Blikshteyn (00:19:00): Well, no, no, they won. (00:19:02): They won against Microsoft.
Brian Bell (00:19:03): They won. (00:19:04): Wow. (00:19:04): So what were the particulars of that case where they were able to argue and win a (00:19:09): settlement or win a ruling?
Dina Blikshteyn (00:19:11): So, I mean, Microsoft alleged that the patent was invalid and not infringed, right? (00:19:17): And they did not want to settle. (00:19:18): They took it all the way to trial. (00:19:20): And I guess they were wrong.
Brian Bell (00:19:22): That’s crazy. (00:19:22): So from that situation, (00:19:25): how can, (00:19:25): you know, (00:19:26): what are some of the lessons from that case that startups can take away and apply (00:19:32): as they’re developing new IP?
Dina Blikshteyn (00:19:33): So you should really talk to an attorney. (00:19:37): or a patent attorney when you, (00:19:39): it doesn’t have to be one, (00:19:40): it can be several, (00:19:41): but some people understand the technology. (00:19:43): Because a lot of what I’m seeing on my end, (00:19:46): especially with startups, (00:19:47): they can say, (00:19:48): well, (00:19:48): we can write our own patent application, (00:19:50): but they don’t understand.
Brian Bell (00:19:52): I’ll just plug it in.
Dina Blikshteyn (00:19:53): Exactly. (00:19:54): And ChatGPT will give you claims. (00:19:56): Would it give you defensible claims? (00:19:58): Probably not. (00:19:59): So that’s when the skill comes in, (00:20:00): to make sure you have quality claims that you can get through the Patent Office. (00:20:05): Because there’s a huge difference if you look at a claim, a claim that’s 150 words, (00:20:10): or that’s 500 words. (00:20:11): It’s a lot more narrow. (00:20:12): Narrow claims are harder to infringe.
Brian Bell (00:20:14): Yeah, and so AI is changing how this is done. (00:20:16): You recently joined the advisory board for Solve Intelligence, which is a portfolio company. (00:20:21): It’s how we got in touch. (00:20:22): What is it about their approach or about AI that’s kind of changing how this work (00:20:26): is done and maybe accelerating patents and IP?
Dina Blikshteyn (00:20:30): Solve is really a cutting edge of what they do. (00:20:34): And the way Solve works, it minimizes the busy work that an attorney needs to do. (00:20:40): And it actually lets us focus on the legal work, like how to draft the best possible claims. (00:20:46): Do the claims have support in the specification? (00:20:49): Like, (00:20:50): What are different claim types that we can do and what can we direct it on? (00:20:55): Because remember, say you have a system with a client and a server, right? (00:20:58): You can write claims directed at the server and do it directed at the client, (00:21:04): and you can do it directed at both, (00:21:05): right? (00:21:06): Now, (00:21:06): keep in mind, (00:21:07): there’s probably two different entities who own the client and the server, (00:21:11): right? (00:21:11): So those claims may not necessarily be the best ones for infringement, (00:21:17): especially if you’re looking to enforce it against one party. (00:21:20): So with Solve, (00:21:22): Solve can help me draft spec while I’m focusing my legal expertise on the claims to (00:21:28): make sure good quality claims are written and those get through the patent office.
Brian Bell (00:21:33): Interesting. (00:21:34): So it’s accelerating the legal work that can actually get done, (00:21:38): which is what any good technology does. (00:21:39): It’s a leverage.
Dina Blikshteyn (00:21:41): It’s a leverage to minimize the busy work and focus on the high value legal work. (00:21:47): It’s like the same thing as asking a question out of any chat bot. (00:21:52): Depending on the type of the question and the type of the prompt, (00:21:55): you may get a good output or bad output.
Brian Bell (00:21:57): Right, right. (00:21:58): So there’s this meme or idea floating around in tech circles that there are no moats. (00:22:03): Execution and speed is your only moat. (00:22:05): Would you agree with that or are there still moats given Java AI?
Dina Blikshteyn (00:22:09): I think there are still modes. (00:22:10): It’s just you need to figure out what they are and where the boundary is.
Brian Bell (00:22:14): Yeah, (00:22:15): so there’s lots of startups out there, (00:22:17): you know, (00:22:17): at the application layer, (00:22:19): you know, (00:22:19): with a platform shift, (00:22:20): like large language models have been, (00:22:22): you know, (00:22:23): there’s this disparagement where, (00:22:25): oh, (00:22:25): you’re just a chat GPT rapper, (00:22:27): you know, (00:22:28): you’re just sitting on top of their tech and you don’t actually add any value. (00:22:32): How do you respond to all the application startups out there that are kind of (00:22:36): building on top of foundational models? (00:22:39): What should they look for? (00:22:40): What are some areas that are commonly patentable?
Dina Blikshteyn (00:22:43): Okay, (00:22:43): so your question, (00:22:45): if I convert it to patent terms, (00:22:47): it’s patenting the LLM model itself or is patenting the system as a whole? (00:22:52):
Brian Bell (00:22:52): Yeah, it’s probably the system that sits on top of the LLM, right? (00:22:54): Because the LLM is, you can swap it in and out like a database, right?
Dina Blikshteyn (00:22:58): Right, right. (00:22:59): I mean, unless you’re the LLM provider and you’re actually looking how to make the LLM faster. (00:23:04): So if you’re an LLM provider, that’s where you’ll be filing your patent applications. (00:23:08): If you are looking at the entire system, what we would recommend is (00:23:12): filing a specific components, right? (00:23:15): How do you have an application? (00:23:17): How is it different from whatever has been filed? (00:23:20): Are there any improvements that you’re doing on the input or the output? (00:23:24): Like what is your final output product, right? (00:23:27): Can I file a patent application on that? (00:23:30): So you’re looking at the entire system, (00:23:33): you’re figuring out where the value is, (00:23:35): and you’re trying to direct the patent applications on that value.
Brian Bell (00:23:40): What are some famous, because patents, they last for a couple decades typically, right? (00:23:45): About 20 years?
Dina Blikshteyn (00:23:46): 20 years from filing.
Brian Bell (00:23:48): From filing, not from awarding, but from filing. (00:23:51): 20 years from filing for utility patents.
Dina Blikshteyn (00:23:55): Yeah.
Brian Bell (00:23:56): Which is interesting. (00:23:56): I think in pharmaceuticals, you develop a compound, some new drug. (00:24:09): The 20 years is really lucrative. (00:24:12): you know, the semaglutide GLP-1 patent, like that can be very lucrative. (00:24:17): But with technology, technology moves so fast, right? (00:24:19): There’s every year or two, there’s some new model, some new platform. (00:24:25): How do technology companies think about this?
Dina Blikshteyn (00:24:28): I see where you’re going with this, Brian. (00:24:31): But I think if you look at what the age of the patents, (00:24:35): when they’re being asserted, (00:24:37): they’re all over 10 years old. (00:24:39): So even though the technology has changing, (00:24:42): If the patents are drafted well, they would apply to that future technology. (00:24:47): That’s why if you look at big companies, (00:24:48): they also file multiple patent applications in different fields and on different (00:24:53): technologies because they won’t assert all of them, (00:24:55): but they’ll assert maybe a handful.
Brian Bell (00:24:57): I see a lot of startups wrestle with closed source versus open source. (00:25:01): How do you advise clients, (00:25:03): if you have advised clients in this area, (00:25:05): to think through going open source versus closed source?
Dina Blikshteyn (00:25:08): Yeah, and it’s a catch-22 for startups because they do like to go open source. (00:25:13): You can still use open source where the end product is something that’s inventive. (00:25:18): You’re not getting a patent on open source. (00:25:20): You’re getting a patent on your entire system or a portion of that system. (00:25:25): So you can still use open source and have something be inventive and patent. (00:25:31): Right. (00:25:31): It’s tricky though. (00:25:31): Like if somebody submits a pull request into a repo and the license says, (00:25:36): now I own that, (00:25:37): you know, (00:25:37): I’m the company and now I own that. (00:25:39): It’s tricky legal waters, right?
Brian Bell (00:25:41): It is tricky, but you won’t be getting a patent on that request. (00:25:46): You’re getting a patent on something that’s specific to you and to your technology (00:25:50): and to your area. (00:25:51): Because the way I look at it, open source, it’s a tool of how to make your invention work. (00:25:57): And you have to combine multiple functions of that open source to achieve a result. (00:26:02): So that combination may be.
Dina Blikshteyn (00:26:05): How has, (00:26:06): you know, (00:26:08): AI regulation and law in this area kind of evolved over the last five or 10 years?
Brian Bell (00:26:13): It’s evolving and it doesn’t stop to evolve. (00:26:16): Right. (00:26:16): So especially with this administration, they’re very pro AI innovation. (00:26:22): So what they’re doing is minimizing legislation on the federal level with a Trump’s (00:26:28): executive order. (00:26:30): And essentially, (00:26:31): they’re saying develop AI as quickly as possible for national security reasons. (00:26:36): Right, because some of the states were trying to develop their own laws. (00:26:41): So you got to do this in Colorado, (00:26:43): you got to do this in New York, (00:26:44): you got to do this here and there.
Dina Blikshteyn (00:26:45): You’re correct. (00:26:46): So the states are trying to step in. (00:26:48): And if you look at the previous administration, (00:26:50): there was a lot of proposed state legislation that has never been enacted because (00:26:57): there was an administration change. (00:26:59): So if you’re talking about the Colorado AI Act, (00:27:02): so that was, (00:27:03): I think it was like 2024, (00:27:04): I think it was May 2024, (00:27:06): it was enacted. (00:27:07): It was supposed to go into effect of January 1st. (00:27:10): Now it’s delayed until June. (00:27:13): And this administration has been very forward in its thinking that states should (00:27:19): not legislate AI on a state level. (00:27:22): It should be done on the federal level. (00:27:24): And at the same time, (00:27:26): this administration wants AI to develop with as little oversight as possible.
Brian Bell (00:27:32): Now compare that to Europe. (00:27:34): You had an EU AI Act that they’ve been forced with different levels for different (00:27:40): types of systems with high risk, (00:27:42): low risk, (00:27:42): medium risk. (00:27:44): And then all of a sudden you have startups who have trouble innovating in Europe (00:27:50): because they don’t know which of those risk categories apply or if they apply. (00:27:56): So there’s a school of thought that the EU AI Act actually is stifling innovation, (00:28:03): especially for the smaller and medium companies.
Dina Blikshteyn (00:28:05): Yeah, it’s kind of interesting, right? (00:28:07): Because some governments will try to, (00:28:09): it’s very unusual that a government kind of steps in and tries to regulate (00:28:12): something before it’s harmful. (00:28:14): Usually regulations are more of a reaction to harm already inflicted, (00:28:19): like the smog in California, (00:28:22): right? (00:28:22): You had all the regulations around carb and clean emissions and stuff like that, (00:28:28): which was a reaction to all the smog sitting in LA and stuff and other places in (00:28:33): California. (00:28:33): It’s really interesting that states would try to go, (00:28:35): oh, (00:28:36): AI could be really powerful and harmful, (00:28:38): so we’ll try to get ahead of it. (00:28:40): That’s kind of an unusual stance I don’t usually see from governments.
Brian Bell (00:28:44): I think it depends on the state, (00:28:47): because we have the same issue with the privacy law, (00:28:49): because the EU has a really firm legislation on the privacy law. (00:28:56): The GDPR 10 years ago. (00:28:58): And in the United States, there’s no federal privacy law. (00:29:01): So you have states like California who are trying to regulate it on the state (00:29:05): level, (00:29:05): and some states do so more so than others. (00:29:08): And it seems like AI was going in that direction until this administration has been (00:29:14): actively trying to discourage that.
Dina Blikshteyn (00:29:17): You remember my big, beautiful bill, because who could forget that? (00:29:22): As part of that bill, (00:29:24): they wanted to impose a 10-year moratorium on the states for AI legislation. (00:29:29): Now that failed. (00:29:31): But what if you start reading Trump’s executive orders and subsequent orders? (00:29:36): What he’s saying is if the states would try to legislate AI in a way that we don’t (00:29:43): agree, (00:29:43): that would impact the state funding.
Brian Bell (00:29:45): Interesting. (00:29:45): So yeah, (00:29:46): this administration, (00:29:47): it’s all for, (00:29:49): you know, (00:29:49): developing AI with as, (00:29:51): I don’t want to say as little oversight as possible, (00:29:54): but they seems like they want to do, (00:29:56): they want to give companies a chance to develop AI and then try to legislate it. (00:30:00): Right. (00:30:00): But then again, (00:30:01): if you are a startup that just keep in mind, (00:30:04): there are frameworks that are in place, (00:30:07): right? (00:30:07): Such as NIST and ISO that have that, (00:30:10): uh, (00:30:11): that have the risk analysis, (00:30:13): the risk assessment frameworks for AI. (00:30:15): that companies still use. (00:30:17): It’s just at this point, instead of maybe seem as more optional than mandatory.
Dina Blikshteyn (00:30:23): Yeah, so you threw out a couple acronyms there that the audience may not be aware of. (00:30:28): Maybe you could define NIST and ISO. (00:30:30): I always forget when this stands for. (00:30:32): What is the spirit of that? (00:30:34): How does it matter to founders?
Brian Bell (00:30:35): Okay, so NIST, it’s a risk assessment framework for AI, right? (00:30:39): And the other one is ISO 42001. (00:30:43): And it’s all about AI governance. (00:30:45): What is a high-risk system? (00:30:46): What is a low-risk system? (00:30:49): How do you implement a risk framework so your AI is deemed safe and ethical? (00:30:56): And how do you apply proper AI governance, (00:30:58): depending on whether you develop your own AI tools or whether you use third-party (00:31:03): AI tools?
Dina Blikshteyn (00:31:04): You know, we talked about the last five or 10 years of changes. (00:31:06): How do you see it kind of unfolding over the next, (00:31:08): you know, (00:31:09): remainder of the decade, (00:31:10): over the next four or five years?
Brian Bell (00:31:11): In U.S., I think the next two years, there’s going to be a lot of innovation happening. (00:31:16): After that, it will depend on whether there is or isn’t a change in administration. (00:31:20): So we may get, (00:31:21): if there is a change in administration, (00:31:24): we may move more towards pro-AI governance. (00:31:28): Then go back to, (00:31:29): I don’t want to say EU AI Act, (00:31:31): but something that’s similar and moving more in that direction. (00:31:34): If there isn’t, (00:31:35): there’ll be continuing a lot of AI innovation with not as much federal oversight. (00:31:42): I think what’s going to happen is something that’s happening with the copyright (00:31:45): law, (00:31:46): where it just has to go through courts. (00:31:48): And at some point, (00:31:49): when we get a Supreme Court case to see what can be done about AI and copyright, (00:31:54): and if that is infringement or not. (00:31:56): But at this point, it’s just weaving its way through courts.
Brian Bell (00:32:01): Yeah, (00:32:01): kind of the most famous right now with this current paradigm shift is probably (00:32:05): OpenAI versus New York Times or New York Times versus OpenAI. (00:32:08): I think there’s about like 39 different cases that are found. (00:32:12): Mostly in California and New York, some of them are consolidated. (00:32:16): But there is a question of whether training AI model on copyrighted data or not copyrighted.
Dina Blikshteyn (00:32:23): data is fair use and not and different judges apply though the four factors (00:32:28): differently how do you think if and when it makes it all the way to the supreme (00:32:31): court if you had a crystal ball how do you think the fair use kind of comes down (00:32:35): with with ai models
Brian Bell (00:32:36): it will be fact-specific. (00:32:38): Because, you know, the problem is not fair use or using your books on AI, right? (00:32:45): It’s the fact that all these LLM models, (00:32:47): they can create something that’s extremely similar to artists’ work. (00:32:53): And then the value of the entire work just crashes.
Dina Blikshteyn (00:32:56): Tell us more.
Brian Bell (00:32:57): Okay. (00:32:57): So if I’m an author or if I write songs as a company right now, you can buy my song or my book, (00:33:06): And because that’s valid, that’s not infringement. (00:33:09): And then you can use it to train AI model. (00:33:13): And now what a lot of AI providers are doing, (00:33:17): they’re putting, (00:33:18): I don’t want to say it’s firewalls, (00:33:19): but essentially they’re putting, (00:33:22): I guess, (00:33:22): I see firewalls on the inputs and the outputs. (00:33:25): So you can get content that’s the same as my work. (00:33:28): So they’ll change a word here and there. (00:33:31): So that’s not infringement. (00:33:34): But at the same time, (00:33:35): once trained, (00:33:36): that model can still produce work that’s very similar to my own. (00:33:39): That’s in my style, but that uses different works, right? (00:33:42): So there is no... A university would train a human being to read a bunch of books. (00:33:47): And, you know, you could argue that a PhD is just regurgitating other people’s ideas. (00:33:52):
Dina Blikshteyn (00:33:52): You’re right. (00:33:53): You absolutely can. (00:33:54): But now you look at the speed, right? (00:33:57): How fast can a PhD replicate my work as opposed to how fast an LLM can do it? (00:34:02): Right. (00:34:02): And at which point, a value of my work would just plummet exponentially.
Brian Bell (00:34:06): Yeah, that’s really fascinating. (00:34:08): So I think it’s going to be the same thing as what happened with Napster. (00:34:12): There’s going to be a change in the industry, (00:34:15): right, (00:34:16): on how work are getting evaluated and where the value is. (00:34:20): Because I remember we went from CDs with the music to now individual songs that can (00:34:25): be downloaded to your playlist. (00:34:26): I think at some point with AI and copyright, we’ll see the same type of change.
Dina Blikshteyn (00:34:31): Right. (00:34:32): I think it’s a little different because, you know, I ingest a book and train on it. (00:34:36): There are ideas in that book that I can then summarize, right? (00:34:40): Or regurgitate. (00:34:42): And so you think ultimately, (00:34:43): we’re going to have to develop a system to compensate the authors of every book (00:34:48): ever written that’s still like under copyright?
Brian Bell (00:34:51): Potentially, right? (00:34:52): Assuming, but you have to remember for copyright infringement, (00:34:56): at least non-derivative copyright infringement. (00:34:59): It has to be the same thing. (00:35:01): It’s the same output.
Dina Blikshteyn (00:35:03): So there’s derivative works and non-derivative works. (00:35:06): Maybe you could like, what’s the difference?
Brian Bell (00:35:08): So the derivative works are the works that can be different from the actual work, (00:35:14): hence it’s a derivative. (00:35:15): So I’ll give you an example with a Mickey Mouse. (00:35:18): Right. (00:35:18): Recently, Disney changed how Mickey Mouse and Minnie Mouse, how they all look. (00:35:22): Because the copyright on the Mickey Mouse has about to expire if it hasn’t already. (00:35:27): So that second slightly different Mickey Mouse, Minnie Mouse, that’s the derivative.
Dina Blikshteyn (00:35:32): And that’s copyrightable?
Brian Bell (00:35:34): It’s copyright. (00:35:35): You can copyright derivative works for like another 70 plus years.
Dina Blikshteyn (00:35:39): How do you think that applies to books and AI and articles in New York Times now?
Brian Bell (00:35:44): Well, that’s right. (00:35:46): So that’s, (00:35:47): that’s sort of the issue because if, (00:35:49): if you have a vote base going through the AI model and you have the AI model (00:35:56): provider, (00:35:56): that’s modifying the content, (00:35:58): right. (00:35:59): It’s not direct infringement anymore because the input is different from the output. (00:36:03): Right. (00:36:03): The question becomes, can it be a derivative?
Dina Blikshteyn (00:36:05): And if the court rules that it is a derivative work, (00:36:08): then the IP copyright flows back to the original?
Brian Bell (00:36:12): Potentially, yes, but it will be fact-specific. (00:36:15): I think if you start looking at it just in terms of policy, (00:36:20): it’s not the fact that AI can infringe one work. (00:36:24): It’s that the value of the author’s work will become extremely low just because AI (00:36:31): can replicate the author’s style. (00:36:36): which may not fall under copyright infringement or under derivative infringement.
Brian Bell (00:36:40): Yeah. (00:36:40): That’s like, (00:36:41): like almost arguing that like a book review or a Wikipedia page is, (00:36:45): you know, (00:36:46): infringing on my copyright of like be as an author. (00:36:49): Right. (00:36:50): But not really, right? (00:36:52): Because it’s just writing a summary of the book. (00:36:54): Your book report doesn’t infringe on my copyright necessarily. (00:36:58): But if you write a whole other book, that’s effectively the same book, just rephrased.
Dina Blikshteyn (00:37:03): It can be a different book. (00:37:04): You can just be in my style, right? (00:37:07): Or think of it as songs, right? (00:37:09): If you have, (00:37:10): you know, (00:37:10): all of a sudden you can have 100 songs that sound like Taylor Swift, (00:37:13): but they’re not Taylor Swift. (00:37:15): Right. (00:37:15): I’m sure Taylor Swift would have to have something to say about it.
Brian Bell (00:37:18): Yeah. (00:37:18): There’s a famous song example recently that was like the Robin Thicke blurred line song. (00:37:24): I think the Marvin Gaye estate sued them and said, no, this is, you’re copying Marvin Gaye. (00:37:31): But really, (00:37:31): I think, (00:37:32): I think the court kind of ruled in favor of Robin Thicke and his production team. (00:37:36): They’re like, no, it’s not, it’s not enough. (00:37:39): It’s not similar enough to infringe on the copyright.
Dina Blikshteyn (00:37:42): Right. (00:37:43): I think the issue with LLM is just they can do it so much faster. (00:37:49): Right. (00:37:50): So whereas you can come up with one song in the month, maybe. (00:37:53): Right. (00:37:54): All of a sudden you can come up with a hundred in a day. (00:37:56): Right. (00:37:57): And you’re seeing this in music right now with AI-generated music. (00:38:00): You can generate hundreds of songs per day, (00:38:03): potentially, (00:38:04): that sound just like a number one song right now.
Brian Bell (00:38:06): Exactly.
Dina Blikshteyn (00:38:07): Yeah. (00:38:07): So that’s where the underlying problem is. (00:38:09): I think you see it more in music than in books now, (00:38:12): but it’ll go into the books, (00:38:15): it’ll go into paintings and into the artwork. (00:38:18): It’ll go into everywhere. (00:38:19): That’s why I’m saying, from my perspective, I think there’ll be a shift. (00:38:23): on how business in the art space is being conducted and how artists ultimately get compensated.
Brian Bell (00:38:30): It’s a really interesting thing that we’re teasing out here, (00:38:33): which is the legal professions and our construct of legality and copyright, (00:38:39): how it kind of shifts with the times. (00:38:42): Right. (00:38:42): Because I think what you’re describing is, (00:38:44): you know, (00:38:44): yeah, (00:38:45): if I made one song in the 80s, (00:38:47): that sounded like Michael Jackson’s like, (00:38:48): whatever. (00:38:49): Great. (00:38:49): But now because you can take Taylor Swift to make 100 songs that sound like her (00:38:53): instantaneously, (00:38:54): the legal apparatus and frameworks need to adapt to that reality.
Dina Blikshteyn (00:38:58): Exactly. (00:38:59): Yes. (00:38:59): And that’s the argument that Taylor Swift’s IP lawyers are making.
Brian Bell (00:39:03): Probably. (00:39:04): I mean, right now it’s almost with AI, it’s like trying to have the legal framework adapt to AI. (00:39:11): And sometimes it works, sometimes it doesn’t like, just like with the copyright and fair use. (00:39:16): Right. (00:39:16): So it ultimately either be a Supreme court decision or, (00:39:19): you know, (00:39:20): the federal government decided they would have to legislate.
Dina Blikshteyn (00:39:24): At the federal level, (00:39:26): because otherwise you’re going to have state sporadic legislation that is born to (00:39:30): anyone any good.
Brian Bell (00:39:31): Right. (00:39:31): Which would be the next step. (00:39:32): If it gets all the way to the Supreme Court and I get a ruling I don’t like, (00:39:36): well, (00:39:36): then I go turn around and go to my lawmakers, (00:39:39): my elected officials. (00:39:40): I say, hey, this is not fair.
Dina Blikshteyn (00:39:41): Exactly. (00:39:42): But that takes time.
Brian Bell (00:39:43): On the other hand, you can also make an argument, you know, playing the devil’s advocate here. (00:39:48): that you need to let technology develop to figure out where the issues are. (00:39:53): And you can’t stifle innovation with too much legislation because then people would (00:39:57): just go to other countries to do the same thing.
Brian Bell (00:40:00): Really fascinating. (00:40:00): Well, let’s wrap up with some rapid fire questions. (00:40:02): What’s a misconception about AI and IP you’d like to debunk?
Dina Blikshteyn (00:40:06): That IP does not apply to AI. (00:40:09): It does. (00:40:09): There’s a lot of AI patent applications that are being filed and a lot of AI (00:40:14): patents that are being issued by the patent.
Brian Bell (00:40:16): What regulation or policy development do you think would most help startups building AI today?
Dina Blikshteyn (00:40:21): I would look at NIST or ISO frameworks for AI governance. (00:40:26): Otherwise, the way the environment is right now, you can develop AI, just develop it.
Brian Bell (00:40:33): What’s the most surprising trend you’ve seen around founders and IP strategy?
Dina Blikshteyn (00:40:37): The resistance to IP.
Brian Bell (00:40:39): Yeah, not a lot of...
Dina Blikshteyn (00:40:40): You know, (00:40:41): startups want to file for patents and when they realize that they do, (00:40:44): it may be too late.
Brian Bell (00:40:45): What’s a bit of advice you give every early stage founder about working with legal counsel?
Dina Blikshteyn (00:40:50): Shop around to find a legal counsel that you like, has expertise and that you can trust.
Brian Bell (00:40:56): Where do you think the next wave of AI innovation is coming from?
Dina Blikshteyn (00:40:58): I think it would be in healthcare. (00:41:01): There is a lot of AI movement in startups that applies to healthcare. (00:41:07): to help develop drugs, particularly in drug discovery. (00:41:11): So I think we’ll be seeing a lot of innovation there.
Brian Bell (00:41:13): Yeah, that’s really exciting. (00:41:14): And healthcare and medicine tend to lag, you know, five or 10 years. (00:41:17): Other industries are just, there’s a lot more regulation there.
Dina Blikshteyn (00:41:20): You’re absolutely right, Brian. (00:41:21): But at the same time, (00:41:22): it also takes, (00:41:24): you know, (00:41:24): five to 10 years to go through all the clinical trials and find out the drug that (00:41:28): you like. (00:41:29): Right. (00:41:30): And AI can speed that process up.
Brian Bell (00:41:33): Yeah, very exciting. (00:41:34): Reading and hearing about, (00:41:35): you know, (00:41:35): basically models of humans where they can test every compound all at once, (00:41:39): you know, (00:41:40): against like human simulations and kind of know exactly how that compound will be, (00:41:44): you know, (00:41:44): digested and will impact your systems.
Dina Blikshteyn (00:41:47): Exactly, right. (00:41:48): It’s things up. (00:41:49): I mean, the problem you often run into there is that those compounds may be creatable or not. (00:41:57): So you have to find a compound that you can create in the lab and then apply.
Brian Bell (00:42:02): Right, right. (00:42:03): What’s a piece of tech or legal writing that changed how you think about your work?
Dina Blikshteyn (00:42:08): It’s not. (00:42:08): So I like Brian Garner, right? (00:42:11): Just for legal writing, active legal writing. (00:42:15): I guess if you’re thinking of it from a technical perspective, (00:42:18): non-boring legal writing that doesn’t want you to sleep and that sounds like you’re (00:42:23): reading a statute, (00:42:24): I would say... (00:42:25): For all startups and also, (00:42:28): you know, (00:42:29): attorneys who are just starting out read how Brian Garner writes.
Brian Bell (00:42:33): If you could change one thing about how the legal community engages with AI (00:42:37): developers, (00:42:38): what would it be?
Dina Blikshteyn (00:42:38): I think the legal community needs to come up to speed on math and AI systems. (00:42:44): So they actually know what they’re talking about instead of just using the buzzwords.
Brian Bell (00:42:49): What’s a part of your work that you wish more founders understood deeply?
Dina Blikshteyn (00:42:52): That attorneys, (00:42:54): there’s some attorneys who actually know how AI works and we don’t just focus on (00:43:01): legal. (00:43:02): We’re also looking for the business case. (00:43:04): which I think where miscommunication often lies because there’s this perception (00:43:08): that attorneys are a bunch of naysayers telling you what you can’t do. (00:43:12): We really are there to guide you to make sure you don’t get into legal trouble and (00:43:17): at the same time also expand your business.
Brian Bell (00:43:19): Yeah, (00:43:19): that’s always the funniest part of working with lawyers and attorneys is how (00:43:23): cautious they are. (00:43:25): They’re trying to keep you out of legal hot water. (00:43:26): That is like your primary function.
Dina Blikshteyn (00:43:28): It’s like... (00:43:29): Yeah, no, I agree with you because that is our job, right? (00:43:32): But at the same time, you can legally stifle the company to non-existent.
Brian Bell (00:43:38): Right, right. (00:43:38): So looking forward five years, (00:43:39): what’s a development in AI law, (00:43:43): this area you hope to see realized?
Dina Blikshteyn (00:43:44): I think more law firms will be using AI and you can kind of tell that what’s solved (00:43:51): and how fast they’re growing. (00:43:52): I think the value actually will be, the AI would help rid attorneys of the busy work. (00:43:58): and help them actually focus on the legal work. (00:44:00): Right. (00:44:01): It’s almost as if everybody now with AI has a team of people working under them, (00:44:06): whether you’re, (00:44:07): you know, (00:44:07): a paralegal or associate level person all the way up to a senior partner. (00:44:11): Now everybody kind of has with AI kind of these agentic systems that help them get (00:44:15): their work done. (00:44:16): Right. (00:44:16): And they’re great, right? (00:44:18): Especially if, when you start looking at a number of these systems, they are fantastic. (00:44:24): But at the same time, (00:44:25): and I think all attorneys need to know how AI works and also reviewing the results (00:44:32): because there’s tons of cases all over the country. (00:44:35): with AI creating fake case law and inciting those in motions and then those being (00:44:42): caught by judges and the other parties, (00:44:45): right? (00:44:45): So there’s a lot of traps that you can fall into, (00:44:49): but at the same time, (00:44:50): using the AI the right way and verifying what the AI product is a huge step for (00:44:57): attorneys.
Brian Bell (00:44:58): Thanks for coming on, Dana. (00:44:58): I learned a ton. (00:44:59): So glad I have this podcast because I get to talk to experts like you and learn something new.
Dina Blikshteyn (00:45:04): Thanks. (00:45:04): Thanks again for the time.
Brian Bell (00:45:05): All right.
Dina Blikshteyn (00:45:06): Thanks, Brian. (00:45:07): And thank you for having me.
