Y Combinator Startup Podcast - AI and the Future of Law: The 10 Year "Overnight" Success Story | Main Function

Episode Date: December 28, 2023

Casetext started out in 2013 as a crowdsourced law library — a sort of “Wikipedia meets Reddit” for the law. Ten years later, Casetext is one of the biggest wins to date in AI, capable of turning weeks of arduous legal work into hours or minutes. Just months ago it was acquired for $650 million dollars. What happened between those two points? For this episode of Main Function, YC President Garry Tan sits down with Casetext co-founder Jake Heller to learn the real story of their 10-year “overnight” success: the 3 a.m. origin story, how the company evolved as fast as tech would allow, and the “magic demo” that helped turn Casetext into a rocket ship. Apply to Y Combinator: https://yc.link/MainFunction-apply Work at a Startup: https://yc.link/MainFunction-jobs

This is Jake Heller. He's the co-founder of Case Text, which sold for $650 million earlier this year. It's one of the mega wins in AI. And today, he's going to tell us how he did it and how he built something that reduces weeks of painstaking legal work down into just minutes. And why that turned into a million dollar contracts for his startup. Large language models are creating ridiculously huge opportunity, and today we're going to learn about Jake's story. Let's get started. At YC, one of the things I've learned is that it takes someone who understands the world from a very specific perspective. That person needs to be combined with people who can build incredible technology.

Jake's origins actually started in the legal world. I had actually a very early traditional legal career. I practiced law at a big law firm before that. I clerked for a federal circuit judge. I had the privilege of just for a summer working in Obama's White House as an intern in the White House counsel's office. So I was able to see some interesting sides of the practice of law. And it was also pretty apparent early on that it was an area that could be improved with technology pretty substantially. You know, it's not any one story, but there are many instances where it was like 2 a.m. or 3 a.m.

And I'm trying to just find one piece of information that might help with a case. And we're talking about like a piece of evidence that might swing a billion dollar lawsuit in one direction or another. Or for our pro bono work, a single legal case that might help somebody who might go to jail, not go to jail. So like literally life or death or making or breaking a business. and it was just so hard. And I would then go on my iPhone and look for like take out Thai food or something. And that was insanely easy. And so we knew that the technology to do relatively trivial stuff, like find nearby restaurants

or good reviews of them or what have you, was there. And it just felt like this big disconnect where things that really mattered at the end of the day, like whether or not somebody spends their life in prison, technology wasn't really helping there. Among the best founders, this feeling was. that Jake describes is very common. What does it feel like when you realize technology has made basic consumer actions like ordering my lunch very easy? But my work is so hard. We're still just a decade into this megatrend idea that great software and now AI has only made it into a very, very small percentage of the total pie of GDP, what the world needs. And like Jake mentions, sometimes it's very

important life or death things that are the largest markets. Case text was one of the companies I got to work with in my early days in YC in 2013. One of my notes from June 2013 about Jake, I said, impressed by this team, articulate, accomplished, and good at building software to boot. So case exchange considerably over the course over 10 years. What saved the same is we had a vision around applying some of the best technology to the legal profession to make it so that when lawyers are doing their critical work, oftentimes life-changing or saving a business or what have you, that they were getting the best access to the most modern technology. In the very early days, the best technology and the ones that we got most excited about were around, for example,

crowdsourcing and applying very early and rudimentary versions of natural language processing to help lawyers do legal work. They started off as a crowdsourced campaign. They started off, crowdsourced case law library where users would edit and annotate, then have other users upvote or downvote the annotations, like Wikipedia meets Reddit, but for the law. From 2013 to 2023 selling for $650 million, that's literally a 10-year overnight success. That was literally us. It took us basically 10 years to get to a place where we had a product that was truly, truly amazing. But we also made mistakes. There are times during that path. where we thought we had extreme product market fit,

where we thought everything was perfect. At YC, we have a diagram that describes the 10-year overnight success as the process. After the launch, after the adrenaline wears off, you go into the trough of sorrow and the wiggles of false hope. One early moment, we created a product that a number of enterprise firms seemed really excited about. And this is, you know, backward machine learning and artificial intelligence weren't nearly as powerful, but it could still do pretty incredible things for law firms. Like they can just drag and drop in a set of documents and it would read through them and say,

based on what I've read, here are the things you need to read next. Things you're probably missing that aren't into documents yet, like cases, regulation statutes that you'll do next. So we early on had a number of large law firms enterprise clients who were super stoked about it. And they started paying us $50,000, $100,000, $150,000 for client. And we got super excited about that. And so we thought, like, just keep on scaling. There's a lot of big law firms globally that just hire a bunch of salespeople and watch us grow. And of course, it turns out that not all law firms are the same and not all are ready to buy immediately and quickly for new technology.

Some are early adopters, some are not. It shows us true not just of law firms, but of all customers. So they started with larger clients and then found that they had exhausted that set of customers. But then focusing on smaller law firms, it worked again until it didn't. And then we also started focusing later on smaller law firms who saw extreme value in using artificial intelligence at the time, we didn't know as powerful of this today, and applying that to their workflow because they don't have enough people, right? So they're by definition, a small law firm. Not enough people, not have hours in the day. Anything they can make them

work faster was depreciated. And again, went after that market and saw this influx of thousands of customers. And we were celebrating like every day because we had 100 customers today, 400 customers a day, 1,000 customers in a day. And if, you were celebrating, you know, 100 customers a day, And it felt like, again, we were on fire. And then some of our marketing channels just stopped working as well. And we got the group of people who were responsive to the message. And it got harder and harder and harder to reach that next group and growth slowed again. But then as the business waxed and waned, they kept at it.

And then a breakthrough. So we were really lucky to get early access to GPD4 before it come out. Like the brief history there is that we've been working on like large language models. Actually, for the last five or six years at this point, as soon as the Burt paper came out, we saw immediate applications to law. They'd been experimenting with AI since 2016, but getting access to GPT3 and GPD4 from OpenAI was the moment that allowed them to create their mega product, co-counsel. This is what product market fit feels like.

With this most recent breakthrough of the state of the art in artificial intelligence, we're able to build a really incredible product. one that for our customers was life-changing. And we saw something that we've never seen before, where you start adding millions of dollars of revenue a month. You start working with those same clients from years ago who may have taken nine or 12 or 18 months

to make a decision because they're a large enterprise client to make a decision in a month. And I think that's when we knew we had something really, really, really special. And frankly, more important than even the numbers, it was just you can see it just, talking to one of our clients the way they would light up in a way that we've never seen before in any product we've built. I mean, we were almost certainly going to triple our revenue course of

this year. And that's after 10 years of building the revenue we had before, right? Again, there are different levels to this kind of product market fit. Founders are always asking us what product market feels like. The answer is when you hit it like that, you know. When we got the chance to work GP4, maybe about six or so months before it was publicly released, we immediately saw that This is different. For us, we saw the raw capabilities of this AI to do human level work at a superhuman speed. What that would ultimately mean for our customers

is something that's really life-changing, right? To make an AI assistant that lawyers can delegate complex legal tasks to and have they get done really fast and at the same level of quality and reliability as you'd expect from somebody quite good working for you. The key pattern we're seeing around the very best LLM-based startups is this, a golden demo that gets people to sign on the dotted line to become large dollar revenue customers immediately because people see the value right away.

We would look at real cases with real data in the past and show how you could have immediately caught the fraud. You know, upload all the emails and instant messages and so on from within this organization and ask questions like which of these emails might evidence potential fraud. And it immediately flags people kind of making jokes. In this case, it's about a case around Enron, you know, failed company that was kind of known in the early 2000s for rampant fraud. All their emails are online. And it would read through all of those and flag like, hey, here's an email where they're talking, and this is literal in like cookie monster talk as a joke about, you know, hiding assets and avoiding auditors and special purpose vehicles and so on.

And another email that sarcastically calls Kenneth Lay an honest man. And the fact that the AI could pick up on the sarcasm and highlight that to the lawyer to say, hey, this might be really important evidence, was an example. And taking a step back from that, the killer demo was doing that activity that you're reading a million documents, finding answers, and a few other, like maybe half dozen others doing research, reviewing contracts, et cetera. So at the end of the demo, you say, the last 10 or 15 minutes, I've done like, four or five days of work, if not much more. And lawyers kind of sat back in their chair and

said, okay, I get it. Like, I get how AI can help me do my work better. So for those watching, that's what good looks like, a magic demo that shows that you can have five days of work squished down into minutes. That's one pattern that is turning into a repeating one. For every large language model based startup, we see that it's going to win their market and build something huge. The Golden Magic Demo. One of the main reasons why Jake and Case Tech succeeded is actually their grit and determination. And also, they were constantly linking what they had to build to what customers wanted. Outcomes to customers.

And they did this masterfully for their customers, the lawyers. But it turns out there's a lot of work kind of going from the raw model to a product that lawyers can actually use. use, we can upload thousands or millions of documents, have to read all of them and tell you which ones are relevant and not relevant to what you're working on, or for it to automatically do research for you and find the right answers out of a billion pages of cases, rules, regulations, and statutes. That kind of scale and accuracy is really, really hard, at least still, you know, it's GP5, so I'll be obviated.

But, you know, we just kind of really, like, move fast building that product. And I think one of the things that never left us from the beginning days of YC, is just a velocity of product and a velocity of hearing customer feedback iterating on it. The legendary scientist Louis Pasteur once said, in fields of observation, chance favors only the prepared mind. And that's what case text did. They started originally as a crowdsourced user-generated content website. But because they were prepared from building software over many years for the same customers. The second large language models appeared, they were ready. And the crazy thing is, the impact of LLMs might just be getting started. So this is going to sound probably insane to

most people who hear it, but I think the GPT technology is underhyped. Because what we see in GPT4 specifically as a model, and I'm assuming the same will be true as new models come out in the state of their advances, is a machine that can, uh, and understand and write and logic to some degree at the level of a pretty good operating postgraduate, right? Maybe a young paralegal or associate at law firm or at McKinsey or what have you. And once you figure out how to use that technology to, with a high level of accuracy and a high level of scale, like review thousands of documents or millions of documents, you could

all of a sudden take things that were necessarily human processes. apply it at scale through technology. And that is just so insanely powerful. What we're seeing in legal is, for example, the California Innocence Project that would get hundreds or thousands of applications of people who are presently in jail who are looking to prove their innocence. And the applications are thick. It's like police reports that are very detailed and witness reports and trial transcripts and deposition transcripts and much else. And they would have to like personally read every single page of that individually. And they still have enough people, none of time.

So they have a four-year backlog to even evaluate cases. Right. And with technology like this, they can read over thousands of pages in minutes and accurately tell you all the details of the case and who's this and whatever, you know, you can cut down that four-year wait time to like one year or one month. That's like life-changing. People are waiting in jail to prove their innocence, right? And that's just like such a small example of what we think.

is going to happen in a very big way. When you compress the work of dozens or hundreds of people, sometimes the most drudgery-laden type of work, and take it away and give an equivalent or better result, that's magic. A lot of people are going to want that magic and the tooling and developer tools necessary to bring that magic to the world is one of the current mega opportunities out there for startups. There's going to be a really rich ecosystem of technologies that will see. support builders. And I think this is going to look a lot like the cloud. To me, GBT and other

models is like a base layer of capability, like cloud computing, that a lot of companies, if not all companies, will plug into some degree. And there'll be another layer of technologies that help you get the most out of it. Langchain's a great example. And then there's going to be an application layer built on top of that. And each one of these folks provide value and there's great businesses to build in each. What we found kind of at that last mile, right, of course, like GBT4 itself is really incredible, but how do you engineer it so that you can have thousands of users on the same time and it's reviewing millions of pages of text all at the same time? That's really hard. How do you make sure that it's not inaccurately saying something about a document with so-called

hallucinations? That's really hard. How do you test that what you're doing as you make alterations to your code or to the prompts you use when you're talking to the AI? doesn't produce the wrong output or hallucination, that's also really hard. We had to build a lot of that ourselves internally. Who knows, maybe someday some of those things will become products. But I also know that there's going to be a lot of really fantastic products that support companies, just like there were things that built on top of the cloud like Kuroku. So that's the good news.

What a time to be alive. Even though Jake has reached the top of one peak, he's already encouraging the next adventurer, you, to step up to the mountain. I think the thing I'd like to pitch right now is never been a better time to start a company. And, you know, while we're here, apply to YC and get the right advice to start off with. I probably sound a little bit crazy to saying this, but I think it's an underhyped moment. I mean, the more you dig deeper into this technology and what I can do and the kinds of problems that can solve for people, the more extreme amount of white space you'll likely see. And so I would recommend folks who are kind of sitting on the sidelines or playing around or hacking.

I think now is the moment to really consider getting going. And I think you'll be massively rewarded for working with this new technology and being on this ride. Case text is one of the biggest mega wins in AI. I am so proud to work with Jake early in his YC days, and I can't wait to see what his team does next. I can't wait to see what you do next, too. That's it for this time. I'll see you next time.