Latest / AI for Founders with Ryan Estes / Legal AI: Why Lawyers Are Finally Free to Think

Transcript

Legal AI: Why Lawyers Are Finally Free to Think · AI for Founders with Ryan Estes

0:00
Are lawyers about to become obsolete or are they about to
0:03
become 10X more powerful than ever before?
0:06
What if the real problem isn't bad lawyers, but broken systems,
0:09
bloated tools, and 10 different tabs just to write one contract?
0:13
AI isn't replacing attorneys, it's exposing how inefficient
0:16
the entire legal system has been.
0:18
Right now, most legal work is chaos.
0:20
Fragmented tools, endless documents, missed context,
0:24
hidden risks. Now imagine this.
0:26
You upload 100,000 pages of case files and instead of searching,
0:30
you talk strategy. The system maps relationships,
0:33
detects contradictions, run simulations and tells you this
0:37
argument wins. This one loses.
0:39
Here's why. Not guesses, not vibes.
0:41
Auditable reasoning. That means one hour of work can
0:44
produce what used to take 10 or even 20.
0:47
So the real shift isn't AI replacing lawyers, it's lawyers
0:50
finally doing what they were trained to do.
0:52
Thinking. So if your product still
0:54
requires humans to clean up bad outputs, you're building
0:57
dependency. But if you eliminate friction,
1:00
you unlock entirely new markets. More cases handled, more people
1:03
served, more access to justice. So are you building something
1:07
that replaces effort or something that expands what's
1:10
possible for humans? Welcome to AI for founders.
1:13
I'm Ryan Estes. This is for founders who want to
1:15
build smarter, faster, and actually win with AI, not just
1:19
bolted on. Today, you're going to learn how
1:21
elite builders are eliminating complexity, scaling output, and
1:24
creating unfair advantages in legacy industries.
1:28
If you're serious about playing this game, go to
1:30
aiforfounders.co and get in the room where the real builders
1:34
are. We send leverage.
1:35
One e-mail could change your entire road map.
1:37
And if this hit, drop a review like it's a term sheet.
1:40
One sentence, high conviction. Let's go.
1:48
Thank you so much for joining me.
1:49
Welcome to the podcast. Thank you for having me, Ryan.
1:52
Absolutely, my pleasure. Very excited to talk about your
1:56
project because you know from the, in the heroes section of
1:59
the website, it says do more legal work in less time with
2:02
fewer tools. And you know, I'm not a legal
2:05
professional, so if I can limit my legal work, I'd love to do
2:09
it. So I would love for you to give
2:11
us an introduction to your, your tool, kind of what you do and
2:15
who you do it for. Absolutely.
2:19
So let me give you a real quick rundown of Iris and exactly why
2:23
we created it the way we did. Essentially, when you go into
2:26
legal, one of the things you see is first, there are a lot of
2:32
point solutions trying to sell like, oh, we do a Word doc
2:35
plugin. Oh, we're going to host
2:38
documents for you, we're going to do this for you.
2:40
And each of these does one very niche little thing and
2:46
essentially lawyers are forced to buy ten of these point
2:50
solutions to do their work more efficiently.
2:52
And if you're a big law firm, that can work really well for
2:55
you. But if you're a small and medium
2:57
and medium sized shop, it just you just get overwhelmed.
3:00
And what we started to see with ChatGPT coming out of generative
3:04
AI is that this problem became 10X worse because suddenly you
3:08
could create like a very niche little product, which is
3:10
essentially just a system prompt around one of these models and
3:13
say this is we've built an AI for transactions, We've built an
3:16
AI for misery law. Like it got people just started
3:21
getting extremely overwhelmed with lawyers.
3:23
And that my Co founder was a former big low lawyer himself.
3:26
He tried a lot with these tools. He'd experimented a lot with
3:29
them. And he was constantly inundated
3:31
with outreach with these tools that in fact did not do what
3:36
they should have because there was just these little point
3:38
solutions. So what we built was an entire
3:41
end to end legal platform that allowed people to basically, if
3:44
there's any kind of legal work you have, we have mathematically
3:49
guaranteed algorithms, some of which are open sourced that
3:52
basically adapts to your work and is able to organize your
3:56
data better. And in in essence just give you
3:59
higher quality legal work without charging you an arm and
4:02
a leg for it. Because the other problem is
4:04
that a lot of these guys charge you a lot of money for very
4:07
little value. 100%, you know, and what's been like maybe
4:11
holding attorneys back from AI is just the hallucination
4:15
conundrum. You know, a year ago, anybody
4:18
that was going to working with generative tools would be like,
4:20
oh man, this is just quoting things completely wrong.
4:23
How do you feel like we've come as far as the large LLMS and
4:27
maybe how does Irish kind of approach, you know, legal
4:31
writing, which obviously cannot have hallucinations?
4:34
Absolutely. So when it comes to
4:36
hallucinations, the most important thing for anybody to
4:40
remember is just in the fact that hallucinations really come
4:44
from multiple categories. You can have hallucinations
4:48
within your data where you have you say, oh, there's a court,
4:55
there's a case that cites this that's useful where the case
4:58
doesn't exist and the most common kind of hallucination
5:01
that everybody talks about. But we also, Ryan, have a lot of
5:06
hallucinations around. So we have a lot of
5:13
hallucinations around where an AI will say, oh, this court case
5:17
said something and on the surface it looks right, but when
5:21
you actually start digging into it, it was not.
5:24
It's not applicable for you for many various reasons.
5:26
It might be a different jurisdiction where your
5:28
jurisdiction contradicts it or it's a completely different
5:31
domain where suddenly it's no longer like if it was something
5:38
in medical and you're doing something in finance, there
5:41
might not be a carry over. So you have all of these issues
5:45
that often pop up with hallucinations.
5:48
And this is a much more insidious form or a third kind
5:50
of hallucination is really comes from context.
5:54
And this is where I just really shines is that if you have very
6:00
large sets of doc inputs and this has been done over many,
6:04
many years, for instance, you're, you're, you're reviewing
6:08
the doc documents with litigation or an arbitration
6:11
where you've gone back and forth between multiple years, You have
6:15
expert testimonies, you have expert testimonies countering
6:18
it, you've cross examinations, etcetera.
6:21
Very unlikely that or or pretty much never that your documents
6:26
are organized in a way where any I would read document number one
6:29
and say, OK, I should now read document #2 to that's this point
6:33
to answer this question. Now I should read 303 makes this
6:36
claim, but document 2 or a court law hub that came out actually
6:42
would apply this conditionally or something else.
6:45
So these kinds of context management and reasoning is
6:48
where also creates a lot of issues and these creates all
6:52
kind of hallucinations that typically don't get flagged
6:55
because or are hard to that's for because if you're just
6:59
trying to use an LLM for everything and you're just doing
7:01
vector search, you will miss a lot of this.
7:03
So the leave the lawyer just sees a bad end product and they
7:09
call it a hallucination. But behind the scenes there are
7:11
multiple things. And what we have done is we have
7:14
systematically by breaking down every possible source of error,
7:19
we are able to build subsystems that capture every individual
7:22
aspect of hallucinations. And then it's about constantly
7:27
improving that each subsystem with the feedback we get when
7:30
the edge cases we're able to catch.
7:32
And that allows us to become much, much more reliable than
7:35
any other legal AI competitor out there.
7:38
Amazing, you know, it's a contact is really key,
7:41
especially here and if there's a higher consequences probably
7:44
illegal than anywhere else. When you're thinking about the
7:47
the kind of work an attorney does on average, what percentage
7:52
of kind of this kind of administrative writing work is
7:55
it replacing right? Are you replacing 8090%?
7:59
Is it 50%? How much are the how much review
8:03
should do you recommend that attorneys really spend with the
8:06
documents they're they're going in and out of iris with?
8:10
So I think again here it's helpful to see it as nuanced
8:14
because it's not 50% of all of that work, it's 100% of some of
8:19
their work then maybe 10% of something else.
8:23
The ratio we're often quoted is we are hiring anywhere from
8:27
three to five times better improvement to 1% get actually
8:30
quoted as a 20% twenty time better improvement.
8:34
So for every one hour they were putting into the platform, they
8:37
were getting 20 hours of actual benefit.
8:43
So and a lot of this again comes down to what it is that Iris
8:47
does. You know, right now, a lot of
8:50
lawyers work flows will either go into doing things manually if
8:54
they're a small or medium shop, or they have to if they're a big
8:57
law shop and they have a lot of tools, they have to go to the
9:00
document management sync system, pull out a few documents.
9:04
So they have to upload that into something like ChatGPT.
9:07
But then suddenly they realize, hey, I can't upload everything
9:10
into that because of confidentiality requirements or
9:14
these files are just too big, ChatGPT can't read it or one of
9:16
the big issues that you think these systems have solved.
9:20
But having this OCR like the documents are have handwritten
9:24
annotations, they have all of these and think like very big
9:28
competitors like Harvey, I often struggle to do this because
9:32
there needs to be, there are very special systems designed
9:35
for this. So then they have to scan these,
9:39
do these and they spend a lot of their time doing all of this.
9:42
So they don't actually spend their time doing a lot of legal
9:45
analysis and legal work. So one of the big benefits of
9:48
Iris is we take everything out of their hands where it just you
9:52
just upload as many documents as you want.
9:54
Our own system is going to acknowledge graphic OCR, build
9:58
the relationships, build the entity maps, and then all you
10:01
have to do is really discuss your legal strategy.
10:03
You tell it, hey, I'm thinking of drafting it this way.
10:06
What do you think? It will run 50-60 different
10:09
simulations, identify what should work, what won't work
10:12
where, where are the risks likely to be.
10:15
You get a very transparent and auditable thinking trace where
10:18
it will both site where it's coming up with its answers from
10:22
and tell you, This is why I thought of this.
10:24
This is why I thought of that, so you're able to interact with
10:27
it. And because of this, the
10:30
secondary benefit is that even if you disagree with something
10:32
or you're want to push back with something 8, you're because of
10:38
this very auditable thinking trace, you're able to get a much
10:40
deeper and richer answer. And all of that really needs to
10:44
a much superior work product, even one hour compared to other
10:49
products where you might have to go spend 6710 hours doing
10:52
something and still not get a valuable final.
10:57
It's kind of bringing the the whole tool suite into one
11:00
container, yes. And really the more important
11:07
thing to think about here since a lot of your audiences likely
11:10
founders, is it's not that we said, let me take this tool
11:14
here, let me take that tool there because from an
11:17
engineering perspective that doesn't work.
11:20
You know, what we have to do is we have to rebuild ground up to
11:24
ensure that integrations that will miss out to ensure that
11:27
when people upload new documents into their DMS systems or into
11:30
Iris, our systems automatically configure the new relationships
11:34
that are able to catch new information or often lawyers
11:38
will tell you things in path and like, hey, I've I've argued this
11:41
kind of case before. In my opinion of the current
11:45
judges that are sitting or the current place, I'm arguing and
11:48
probably not very friendly to this.
11:49
So we should tweak our arguments.
11:52
So these kinds of memory features where the AI has been
11:55
intelligently re plan on the fly, that's not something that
12:00
other people are able to pull off.
12:02
And you only get this through deep software engineering
12:05
integrations. It's not something where you can
12:07
just try to chain together a bunch of APIs who calls and call
12:12
it a day. Amazing.
12:14
And you know, kind of looking through your material, you know,
12:18
this is good for litigation attorneys, insurance, defense,
12:21
corporate and contracts, employment law.
12:23
Is there are there any kind of like legal areas that Iris would
12:27
not be good for? So far we haven't found any.
12:32
Well, we've tried there's obviously things were better or
12:35
worse at and we're always trying to improve.
12:37
But for instance, we have users all over the world that have
12:41
tried the product and loved it. We have users from every kind of
12:44
law that has tried and loved it. And really the crux of this is
12:49
entirely in our geometry, the way we represent information,
12:53
the way we're able to reason through things, some of which
12:55
we're open source so people can verify these claims themselves
12:58
and look at the way we research, look at the way we organize
13:01
embedding. All of these lead to much better
13:04
outcomes. Yeah, nice.
13:07
You know, and I'm thinking this, I'm not an attorney, but I have
13:11
an attorney and is this something I, and I want to make
13:14
the recommendation be like, hey, you should go check out Iris.
13:17
Could I then say this is going to 5X your output, so maybe you
13:20
should reduce my billable hours by 5X?
13:24
Or is this going to basically give me 5X the insurance that
13:29
like, hey, I know my attorney is, is equipped with the tools
13:32
that like is going above and beyond now maybe what he was
13:36
capable of before? Absolutely both.
13:40
I think depending. Like we've had both use cases.
13:44
Some people, you're just drafting simple documents, you
13:47
just want things done quickly. There's no reason that an NDA
13:51
should take you more than 10 minutes of drafting and actually
13:57
2 minutes of drafting and then maybe half an hour of
13:59
verification. There's no reason should take
14:01
you any more than that. And with our platform like I'm
14:06
really giving you the higher higher level edge cases, what we
14:09
see is people draft an NDA in one shot and it's very good.
14:12
So they're done there, right there.
14:14
And in other times you're, you're preparing for a
14:17
deposition, you're preparing for something else.
14:19
You just want very good quality responses.
14:22
You have 100 GB document dump based on all of the legal
14:26
context, which is a real like our lawyers have dumped that
14:31
under the platform. And you just need to understand
14:33
what to do and you need to push something out quickly because
14:36
your judge hearing is in a day or two.
14:39
It will give you orders of magnitude better performance
14:43
than anybody else. And obviously I'm, you know,
14:45
talking a big game. Every founder comes here and
14:48
talk, says they're changing the world.
14:51
We're so confident in our product that we've made a free
14:53
sign up for anybody. So you don't have to take my
14:55
word for it. You can go try the platform out
14:57
yourself and you can do it yourself.
15:00
There's a reason we always win in our head to head demos versus
15:03
every legal AI competitor on the planet.
15:07
We have, we have. And there's a reason we have
15:09
never lost a customer. We've never lost an enterprise
15:12
customer ever. Because once you actually start
15:15
using the platform, you will see the differences both in the
15:17
speed and the quality. Yeah, how are attorneys doing?
15:22
You know, the, the word for, you know, a year has been a is
15:25
coming for our jobs. It's coming for attorneys jobs
15:27
first. Now, of course, there's probably
15:29
like, you know, if, if you're in the courtroom, that's never
15:32
going to be replaced by a robot, I don't think.
15:35
But there's a lot of areas where maybe it is.
15:37
It just makes sense. If you're like drafting wills or
15:39
you're doing repetitive kind of agreements, There's always
15:42
nuance where people need to need a, a personal touch.
15:45
Not to mention, just if you're in need of an attorney, you're
15:48
usually in some kind of emotional distress and it's nice
15:51
to speak to a professional. But how are attorneys doing?
15:54
Are they worried that Iris is going to turn into a legal robot
15:57
and their days are numbered? Not at all.
16:00
It's really the other way around.
16:03
Because our platform does so much work so quickly.
16:06
You're able to take on work that was just never possible before.
16:10
Yeah, and that really goes back to why we're releasing Iris
16:13
Light. That goes back to why we're
16:15
doing so much for other stuff, why we made the platform free to
16:18
begin with. You know, it's a very big risk,
16:21
especially for a start up at our stage to be making things free
16:25
and to be open sourcing all of our work.
16:26
So or open sourcing so much because you're always worried
16:30
like, hey, a well funded competitor could copy, you could
16:32
do this. But we truly believe in that
16:35
legal has to be democratized. Right now, courts are
16:38
backlogged. Tons of people need access to
16:41
justice and they don't get it. And our platform lets you get
16:44
this. So On the contrary, the presence
16:46
of a good AI platform like Iris actually allows your eternity
16:50
to, I'm sorry, your attorney to do much more work, much faster
16:55
work, higher quality of work, and thus take on many more legal
16:59
cases than than they would have otherwise.
17:01
It just, it does change the billable hourly model.
17:05
You can no longer bill for hourly rates.
17:09
But what you can definitely do is use the AI to be more
17:12
efficient and then more people will come to you.
17:16
More people can use your services and that completely
17:19
changes your game. Yeah, amazing.
17:22
And and two really kind of big call outs that stood out for me
17:25
about IRS is the unlimited context.
17:27
So it's basically the token maxers are out there.
17:30
They can just go crazy, you know, with documents and upload
17:33
it and it it reads every single one independently and also zero
17:39
data retained. So this this is basically saying
17:42
that it's not getting uploaded to the major LLMS. So are you
17:46
basically creating a MCP server locally for for every client
17:50
where all of their information is stored locally?
17:54
So I think in both cases slight misunderstandings, but largely
18:00
you have the right picture. First, with regards to infinite
18:05
context, we it is processing all these documents.
18:09
It's not, it's not independent in the sense.
18:12
And there's actually a big problem with traditional
18:14
techniques like vector search where they will, if you use
18:17
something like hardware, you can upload as many documents as they
18:19
want, but they're just simple rag wrappers, which is going to
18:22
chunk it and then hope that your prompt has enough things to have
18:25
cosine similarity matches with Iris.
18:29
We actually link across documents, entities,
18:31
propositions, assertions, all of these things.
18:33
If there's anything that could be possibly contradictory, we
18:36
update, We have like a self updating graph.
18:39
And all of this basically ensures that even when like for
18:43
any question, we're, we're getting a truly comprehensive
18:46
insight into your knowledge base.
18:47
We're not giving you a proposition or a case law
18:50
argument strategy that is actually going to be that that
18:55
might be countered by something else in the document set that
18:58
you missed. Or you know, if you have a
19:01
massive 1000 page document that is out of context window for any
19:06
other land and there's like a proposition CA section C codes
19:11
something in section BA. Lot of the times drag search
19:14
will miss that. So we're just able to catch all
19:16
of these things by mixing context into each other.
19:20
The second aspect of this is 0 retention.
19:24
What that means is that we, we have special partnerships with
19:28
all the LLM providers that they will never train on your data.
19:30
They don't even retain your data.
19:32
So we don't have to do anything fancy or with the MCPS,
19:36
etcetera. I could sit here and pretend
19:38
like we do a these very sophisticated solution there.
19:42
But really we just spoke to all the LLM providers.
19:44
We have a very good relationship with them all and because they
19:48
respect the work we do, they don't retain your data, they
19:51
don't touch it. And what this allows us to do is
19:53
it simplifies the architecture quite a bit, which means it's
19:56
cheaper for us to run, it's cheaper for you as the user.
19:59
Otherwise, we would be having twenty other headaches and we'd
20:02
have to pass those calls on to our customers. 100% and that,
20:06
that creates its own leaks problem internally.
20:08
There's almost, there's always a trade off, you know, but you
20:11
know, if, if you have assurances and guardrails in place, I mean,
20:15
it's the, the cost benefit analysis gets pretty easily.
20:19
And I, I'd love to talk about cost too, and kind of how you
20:22
approach pricing, you know, because now with, with AI
20:25
implementation, there's all kinds of different pricing
20:28
models, whether it's by token or outcome based pricing.
20:31
A lot of people are doing kind of traditional sats and buy
20:34
seats. And it looks kind of like that's
20:36
kind of how you have it. You kind of have an ABC pricing
20:39
model where you have one seat, come on, come on in and give it
20:42
a shot. And then once you love it, then
20:44
you can bring on more people into the, into the, the, the
20:48
program. So how did you approach pricing
20:51
with kind of the different pricing options available for
20:53
you? Yes, that's a great question.
20:59
So really I think this also goes on to product start.
21:05
We wanted to keep things as simple to begin with because a
21:09
lot of the issues that with small and medium size firms you
21:12
have is that there's a very big difficulty with understanding
21:18
where like just how to use the platform, how to how it what it
21:23
would cost them. Because with legal tools you are
21:26
like all kinds of random pricing models I saw per word per page,
21:31
like somebody was try charging them for word outputted.
21:36
There's it was just a mess. And the first thing we wanted to
21:38
do was ensure that anybody that's coming on our platform
21:41
knows that we're not trying to squeeze every dollar out of
21:43
them. So simplify that tremendously.
21:47
Now that we've simplified that tremendously, we can start
21:50
adding some usage based pricing for the future.
21:53
So we don't intend like we didn't start out this way, but
21:58
because our context management is so good, people started using
22:01
us as a document management system, which means people are
22:03
just dumping everything on there.
22:05
So adding some prices for storage as in adding some, we
22:09
have some special modules for specialized work that we can
22:12
charge extra for where we will introduce some token based
22:16
pricing or something simple. The key inside there is just we
22:21
never want to make our monetization a bigger priority
22:25
than ease of use for our users because that will not really
22:29
benefit anybody. We want to keep things as easy
22:32
for our users to understand as possible and then do
22:36
monetization second. So we're and I think we've done
22:40
a pretty good job with it. We're still running a pretty
22:42
successful business at the end of the day because of all the AI
22:46
and the optimizations we have done.
22:49
It's not as though that puts us in the red or we lose money on
22:52
it. It's just a matter of ensuring
22:56
that everybody gets pricing like fair, transparent pricing.
23:00
Love it. You know, doing research on you.
23:03
There's a couple of things that popped up I have to ask about.
23:06
What is the chocolate milk cult? So that is the AI research
23:13
community that we've built. We have over about a million and
23:17
a half people that we reach every month, mainly through my
23:21
newsletter on sub stack called AI made simple, although there
23:24
are other platforms as well. So there again, the goal is to
23:29
democratize access to cutting edge research.
23:31
I think if you look at the AI landscape in commentary
23:35
landscape, you, you tend to have people that just like make money
23:39
by telling you, oh, these are the five tools you can buy.
23:41
These are the 10 prompt tax I've learned.
23:44
This is the 20 stuff I've learned or they make very simple
23:47
things like this, how you build a neural network and this is
23:50
that's all. Or you have people that just
23:52
speak to pH DS with this is exactly like this is the most
23:56
recent AI research paper. But if you were an actual
23:58
builder or an investor or a policy maker or anything like
24:01
that, who needs to understand the cutting edge, but like you
24:05
don't have the resources to invest into your own AI lab to
24:09
be like, OK, like this paper has released a different
24:13
quantization method. How does that change our
24:15
financials? We can we do that for you.
24:19
We just research a bunch of stuff.
24:21
We dig into a lot of the papers. As mentioned, we have amazing
24:25
relationships with a lot of major AI labs with whom we share
24:29
data and they share data back with us.
24:31
And that allows us to know exactly what is the most
24:35
impactful information in cutting edge AI and how can you as a
24:39
random builder without any major resources, go start building
24:45
from it tomorrow. Because ultimately, technology
24:48
only benefits humanity. So it can benefit everybody.
24:52
It shouldn't be a tool that just benefits a few extremely rich
24:56
people. But I think to me that's a very
24:59
to be sad if this is, well, the contribution that tech had to
25:03
the world. This episode is brought to you
25:06
by a dangerous idea. It's called reading,
25:09
specifically the AI for Founders newsletter. 47,000 Founders
25:13
strong now. And it's the good stuff.
25:16
No top ten AI tools that will change your life written by
25:19
someone who doesn't ship product.
25:20
It's just real breakdowns of what's actually working in AI,
25:24
who's raised, who failed, who pivoted, and which founders
25:27
quietly built $1,000,000 in ARR while everyone else was arguing
25:31
on X. It's written for people building
25:34
with AI every week. You get tactical frameworks,
25:36
unfair advantages, breakdowns of real companies, and the
25:39
occasional existential question about whether your startup is
25:42
intelligent or just, well, branded.
25:44
It's free, It's sharp. It might mildly offend your
25:47
current workflow. Go to aiforfounders.co and
25:50
subscribe before your competitor does, because the only thing
25:53
worse than being replaced by AI is being replaced by a founder
25:57
who reads. That's the newsletter at AI for
25:59
founders.co. Now back to the show.
26:04
This episode is brought to you by Moburst, the AI first growth
26:08
engine trusted by iconic startups like Calm, Robin Hood,
26:11
Reddit, Uber and Google. Moburst is a digital, full
26:16
service, mobile first marketing agency that helps companies to
26:19
scale and become category leaders.
26:21
Whether you're finding product market fit or ready to
26:24
accelerate growth, Moburst delivers creative, data-driven
26:28
solutions that help with scale and efficiency and impact.
26:32
At Moburst, we're not just guessing what works, we're
26:35
predicting it. As an AI first agency, Moburst
26:38
uses cutting edge tools to analyze user behavior, adapt
26:42
creative in real time, and cover hidden growth opportunities.
26:46
The result? Marketing that's faster,
26:48
smarter, and built to convert from early stage startups to
26:52
global brands. We've helped hundreds of
26:54
companies overcome complex challenges, unlock new
26:57
audiences, and optimize every metric that matters.
27:01
At MO Burst, we believe that creative thinking, advanced
27:04
technology, and data are the core drivers of success at
27:07
scale. If you're serious about scaling
27:09
your business with precision and purpose, it's time to talk to MO
27:13
Burst. Visit moburst.com and let's
27:16
build something category defining together.
27:21
Yeah, but honestly, obviously you've found like that the
27:26
message is really resonated. You know, your your sub stack
27:30
has 250,000 Subs. You know, the chocolate milk
27:33
cult is spreading chocolate milk AI over the world.
27:37
Tell me a little bit of how you've been able to leverage
27:40
that audience you've created to help Iris grow.
27:44
And also, you know how you went about like, building that
27:47
audience? So for helping Iris grow,
27:51
honestly, it hasn't been as impactful as you think because
27:55
my work is BPI research and engineering.
27:58
So it's not like lawyers sit there and read, are interested
28:01
in reading about what are the maths, what is the math behind
28:04
attention and how does that impact your financial
28:06
statements? You know, they're not sitting
28:08
there looking for GPU pricing and all of that.
28:12
Where it has helped us tremendously.
28:15
He is in the fact that we are able to do all of this research
28:19
like we open source something called latent space reasoning.
28:22
And I can tell you right now, one year later, you guys are
28:27
going to come back to the spot pass here, the readers are going
28:29
to come back. And then the 1st place we heard
28:32
of it, it's what all the future labs are betting on with AIDS
28:37
because it has a lot of structural advantages that don't
28:39
exist anywhere else. We were able to work on it,
28:42
pioneer it and we open sourced a very, very cheap, like a very
28:46
lightweight version of it that anybody could run in test.
28:49
And already in that you'll see that we create some emergent
28:52
properties that don't exist in traditional auto regression.
28:55
Just the way LLMS work normally. What we open sourced creates new
29:00
abilities that don't exist in them.
29:02
And you can't just get that through prompting or training
29:05
that that kind of stuff has been extremely helpful in getting us
29:10
the partnership. So we we talk to best the best
29:13
engineers at NVIDIA and my newsletter.
29:15
I had guys from Zebu AI come on a DeepSeek a GLM like all them.
29:22
Sorry, min Max GLM is the Zebu's model.
29:25
But we're also talking to guys in deep mind.
29:27
I'm going to I've been invited to Google next a week from now
29:30
or two weeks from now. Sales force has invited me like
29:34
we're able to build these very strong technical partnerships so
29:37
that everything we build like we know for a fact, we're not just
29:41
relying on external stuff, like we're not relying on people to
29:47
tell us what to build. You know, we're not traveling
29:49
the Internet to be like, oh, what is the next optimization?
29:52
We, we are the ones publishing that stuff and that we get
29:56
extremely long term modes for us from a product perspective. 100%
30:01
you know, and it's actually very rare, you know, there everyone's
30:04
building with AI, but you really by building kind of a personal
30:07
brand, personal relationship first, then you have this
30:11
unbelievable kind of insider connections because there's,
30:14
there's fair exchange. You know, you, they get access
30:16
to your audience, you get access to them that influences your
30:20
product. That being said, you know, you
30:22
could have probably pointed your AI expertise in any category of
30:26
business with that kind of support.
30:29
Why did you choose legal specifically?
30:32
What I mentioned earlier, democratization, I think in a
30:36
world where you have technology that can make things happen
30:40
faster, it is extreme. It is unconscionable that you
30:45
can have people that are scared to get access to a legal help in
30:49
like I know around New York City there's people who are like,
30:54
hey, my landlord doesn't has been doesn't do their stuff,
30:58
messed up the building and now they're overcharging us on rent
31:01
and we have nowhere to go because we don't have resources.
31:04
In India, where I'm from, you have you have a 10 year backlog
31:09
in civil laws, civil cases, which means if you went to court
31:13
tomorrow to say that my employer didn't pay me, and this happens
31:16
much more frequently than you think or my employer stole wages
31:19
from me, you will not even get a high sitting for 10 years in the
31:23
future. And you know, you see that you
31:27
think, oh, that's an India specific problem, but that
31:29
exists everywhere. How many times are you forced to
31:32
eat shit because you just dare to go into the law system
31:36
because you don't have the money?
31:38
With my newsletter, I often get like, I hear that I have that
31:40
all the time where sometimes if I'm critical of companies, I get
31:44
lawsuits tests where it's like, oh, we'll sue you for we'll sue
31:50
you for taking money from somebody, which I've never taken
31:53
money. It's an absolutely absurd claim.
31:55
But they say I have an agenda for writing against them or
31:58
something. And in all of these cases, like
32:01
the reason this stuff can happen is because fundamentally our
32:05
legal system has become very inaccessible to people.
32:08
So when I what I Co found that this was something we both
32:11
agreed on that there is no reason this stuff has to take
32:13
this long and it there's no reason it has to be so expensive
32:17
and this is what partnered up to there to change.
32:22
It's beautiful, man. Power to the people.
32:23
I'm all about it. Was there a specific instance
32:27
that kind of pointed this out to you explicitly?
32:31
Is there like a Joker to your Batman?
32:36
I mean, the example I brought up from India, my, it's something
32:40
I've seen very first hand. You know, you have a lot of
32:44
farmers that were exploited because they were oversourced.
32:50
They were, they get overcharged for pesticides, etcetera, and
32:53
then they have to go into debt. And if it doesn't work out, then
32:56
you have loan sharks coming after them.
32:58
So it's a pretty grim situation overall.
33:00
So the inaccessibility to legal is something that I've kind of
33:04
felt for a while, even when I was indirectly working on this.
33:08
You know, if you look at a lot of my work early on, I would
33:11
often like try to point out research that was trying to do
33:14
this and improve this. But what we realized very early
33:17
on is that a lot of the people that do legal AI research, at
33:20
least pre 2023, were not building parts.
33:24
It was just there's a database here that we have trained on and
33:27
I got my PhD and that's it. Or I built this as a project but
33:32
it was not going into meaningful progress with my Co founder.
33:37
The thing that changed is that I had somebody that we compared
33:42
with and that really that is what led to the creation virus.
33:48
Amazing. I love that.
33:50
All right, we're ready to enter the Quantum Round.
33:53
Are you ready to launch? Absolutely.
33:56
All right, aside from Iris, what experiments are you running with
34:01
AI? What are you playing around
34:02
with? So I think there's so much
34:06
across a few different areas of computer use are very important
34:10
to see like what can be done, what can't be done.
34:12
We're we're testing security stuff extensively.
34:15
I think we spoke a little bit about that, but people don't
34:19
actually understand how many security vulnerability is being
34:22
shipped, shipped into software right now.
34:24
And this is leading at software systems because the AI is AI,
34:29
people are using AI to just open port requests and do this and
34:32
it's creating and extreme. So we're just profiling security
34:36
risks. Hardware is something we're very
34:38
interested in. So just looking at how hardware
34:41
is changing etcetera tends to my work tends to be a bit more
34:45
technical and cutting edge. It's not really we're trying
34:49
these tools or doing this. Although in the computer use
34:52
front, there's some clipping software I'm trying for videos
34:56
and just some automation stuff. Yeah.
35:00
You know, Speaking of vulnerabilities, you know,
35:02
obviously that big announcement from Anthropic about mythos, you
35:05
know, finding and exploiting 0 day events and like all kinds of
35:09
open. Source.
35:11
Coding. What's your take on that?
35:15
I think with the mythos specifically, I think it's hard
35:19
to know how much of that is real, how much of that is hype
35:23
because they did, they didn't get so much information out.
35:26
It's, it's actually really hard for any serious person to know
35:29
what they've actually built. That being said, I would not be
35:33
surprised if it had done something.
35:36
Truth be told, I don't think it's a it's because of their
35:39
model. I think more entropic has been
35:42
experimenting with harnesses like Lord Code, etcetera.
35:45
And one of the benefits of harnesses is that you can start
35:47
to see where things fail more frequently because it's just
35:50
going to take actions by itself. So I think that is likely
35:53
creating a lot of like you're starting to identify their, you
35:58
know, gaps in software. Because The Dirty secret about
36:01
software is that so much of this is like build either by free
36:05
maintainers who don't get paid so they they're overworked and
36:09
sometimes just pass things along.
36:11
Or it's like, even if you have perfectly built modules,
36:16
combining the modules is very difficult.
36:18
So like people will combine them in sloppy ways and then build
36:23
systems on top and top of that. And that'll just create all
36:27
kinds of issues that harnesses are now able to find because
36:30
they're, they're just running like 100 actions and reading all
36:33
the logs where humans could not. So I'd be less.
36:38
I think it's more exciting than it is running because really
36:42
it's the perfect way to start identifying where our systems
36:45
are weak. I would not be as worried about
36:48
it because people like a lot of this is often hype.
36:53
Yeah, it is kind of unfortunate that like that is become kind of
36:57
the marketing plan for the large LLMS is basically every six
37:02
months, let's try and really scare everybody to try and
37:05
create a a regulatory mode around what we're building and
37:08
nobody can have it and like it, it does.
37:10
It's kind of becomes kind of like crying wolf, you know?
37:14
Absolutely. I think you said it perfectly.
37:17
I think that is the issue that we always have it.
37:22
One of the biggest issues you see is that with a lot of these
37:26
systems now, they have no way to differentiate themselves.
37:28
So they just make up these most random metrics that will say
37:32
like, oh, this system works for 8 hours straight and didn't stop
37:36
working, or the system did something else, which is like
37:40
they're just trying to create some kind of a story around it
37:43
because like it's very hard otherwise to sell.
37:46
And Iris has that issue too. Like every time we go to sell,
37:49
we heard about 10 competitors we've never heard of before.
37:53
But everybody's trying to get ratio of this.
37:55
So I do get the temptation. I just think it's a bit sad that
38:02
to in order to sell things, we're completely like destroying
38:06
the discourse around tech, destroying the ability to have
38:10
good faith conversation around what's working and what's not.
38:13
Yeah. And how do you kind of approach
38:16
moats, let's say in general, because you're very public,
38:19
you're you build in public, people know who you are, know
38:21
what you're working on. And so they they can just copy
38:24
it. They could vibe code something
38:25
together in a couple of weeks, some chat TPT wrapper, you see
38:28
them pop up, they're trying to move in on your market.
38:30
Like what are you doing to make Iris like a defensible business?
38:37
So that's a good question. And actually I wrote about this
38:39
recently and funny enough, I had an outreach from our Business
38:45
School that they might be or somebody there that wants to use
38:49
this in their course. We where I'm essentially just
38:53
wrote about this problem because there's not a new question in
38:55
tech. Yeah, why?
38:58
Like why can somebody not copy you?
39:01
And I think people don't understand how difficult it is
39:04
to copy software, especially from different communication
39:08
paradigms. So if I was a rapper and you're
39:11
a rapper and you built a vector UX around something, it's easy
39:15
for me to copy that and vice versa because we have the same
39:18
paradigm. And this is what I mentioned a
39:20
little bit starting into our call.
39:23
It's not just about building features, it's about building a
39:26
system that interplays with it. Well, you know, why could GitHub
39:32
Copilot not copy cursor and why could neither of them copies
39:36
cloud code properly, even though each, if you look at the place
39:41
in the market, had exceptional brand name, exceptional money,
39:45
was it? It's just because every software
39:47
is built with certain assumptions around what what it
39:51
does, what it doesn't do, how it operates, how it moves data, how
39:54
it processes that what is valuable, what should not be
39:57
done by it and should be like left to the user.
40:00
And often like these are entire life philosophies.
40:04
So you can't just. Cut out one part of it and put
40:08
it into another because you'll, you'll find all kinds of
40:10
migration conflicts. And that's one of the biggest
40:13
modes we have is just, we have built everything ground up.
40:17
We are like the first truly agentic system in legal.
40:20
We are the first truly non wrapper solution in legal, at
40:25
least in the new generation of stuff, because we from day one,
40:28
in fact, even from before I started Iris, we were we
40:33
identified fine tuning doesn't work.
40:35
We've I was writing about how vector search is limited in
40:38
2022, which is before people even knew were like rag as a
40:43
word didn't exist. Then when I wrote down why
40:45
vector search was limited. And that's not to say it doesn't
40:50
work, but we did identify some structural issues and why you
40:53
had to work around that to improve it.
40:56
So when it comes to stuff like this, we built from the from day
41:00
one, keeping in mind every, all these limitations.
41:05
And we do build a lot in public, but there's a lot we don't
41:08
share, you know, that's the collaboration with the labs.
41:10
We do this like I, we just don't have the time to share
41:13
everything. And that allows us to build
41:16
technically a much superior product that loops on that
41:20
grows. Like having these relationships,
41:23
that deep mind means I know what's going to happen with the
41:26
models before they're released. So we know we can see like how
41:30
we change our URL policies. Is this worth focusing on?
41:33
How could we improve? Secondly, there's a context mode
41:36
to it. The more you use a platform to
41:38
build it to be self organizing and self learning, the more it
41:42
will learn your preferences. It's not going to train on your
41:44
data. We're very strict about that.
41:46
But it will learn that Ryan likes these kinds of answers
41:49
more. Ryan likes to argue in this
41:51
format. Ryan's best practices state that
41:55
this form, this format should be followed when writing an e-mail.
41:58
This should be the memo. This should be and all of these
42:01
little quality of life improvements as you, as you
42:04
interact with the legal AI and you say, hey, for this case, I
42:07
prefer this strategy. It's annotating all of that.
42:10
So it enriches your data set, which means suddenly you cannot
42:13
leave because if you leave, you lose a lot of embedded
42:17
intelligence. So both in the technical side
42:19
and the product side, we create these native AI native modes
42:23
that just don't exist in today's technology.
42:27
Yeah, it's brilliant. You know, the, the legal world
42:30
like really lends itself to, you know, AI as a with a tool set
42:36
very well. You know, like the listeners of
42:38
this podcast love ideas because we all chase shiny things all
42:41
day long. You know.
42:42
Is there an another category that if not legal, because you
42:46
guys pretty much have that category buttoned up, is there
42:49
another one that maybe is primed and and ready for something
42:52
similar, a way to democratize maybe a high barrier of entry,
42:58
highly complicated, highly nuanced industry.
43:01
I think you just answered your own question.
43:03
Every industry like that is ready to be democratized.
43:07
Yes, there is. And I think not just that low,
43:12
often low barrier to work industries or things that don't
43:16
have existing solutions, they can also be democratized by
43:18
bringing something to them. You know, if, if you have, I
43:23
know a person that's building AI automation for modeling agencies
43:28
and like, you know, modeling as an industry is extremely
43:31
outdated technically so, but her system will handle a lot of that
43:36
for the agencies. And in doing so, we'll reduce a
43:40
lot of manual work, allow reduce scheduling conflicts for models
43:43
because some they're often have to go through that or other
43:46
kinds of issues. You can use stuff to bring it to
43:51
new industries. I think what people don't
43:54
realize is you can just try stuff and it will likely work
43:58
what kinds of things you try. Like some of these optimizations
44:02
we're building would likely not work in other industries because
44:05
law has a high paying capacity. A lot of other industries don't.
44:09
So you don't want to build like a very expensive optimization
44:14
for the NTH percentile, NTH percentile performance in a
44:18
field where it might not work. But many, many cases, you don't
44:22
need to worry about the 99.9999% accuracy. 99 is good or 80% is
44:28
good. And if you can deliver 80%,
44:30
where the world gets 0 percent or where it's 20% normally.
44:35
And if you if you put in a lot of work, you've got 100%.
44:38
But where you can people do that, this kind of stuff, it's
44:42
right for disruption. And there's so much capital
44:44
flowing around looking for problems to solve.
44:47
There's a lot of people who are asking like, hey, can, how can
44:50
you make my life easier? I think it's the perfect
44:52
opportunity right now. Yeah, yeah.
44:55
You know, you you can just build things.
44:57
And right now it's easier than ever.
44:58
So it almost comes down to what you should not build.
45:01
You know, when I think about you and Iris, I can tell you have a
45:04
passion for democratize and legal.
45:07
You know, do you think it's it's it's safe to trust that passion?
45:12
Is that where founders should really kind of do a gut check
45:15
and say, like, what's the thing that really puts a fire in my
45:18
belly and go for that? You have different kinds of
45:22
founders, so I can't speak to them all.
45:24
I can just say that the chances are if you're trying to do
45:28
something great and you're trying to disrupt something, you
45:30
really want to change something, or you think you're going to be
45:33
working in a field with a lot of competition, then the simple
45:38
truth of the matter is that if you're not super duper
45:43
passionate about it, if you're not going to want to work on
45:46
things, then what will happen is somebody will build what you
45:52
wanted. And nothing sucks more than
45:55
seeing somebody else succeed of an idea you had and you're like,
45:58
oh, I knew it's work, but it didn't.
46:01
So like either save yourself the heartache of not going through
46:04
that or just be ready to put in a lot of it.
46:09
And sometimes it's passion, sometimes it's desperation.
46:11
Like you can have multiple reasons for doing this, but
46:14
unless you have a reason to absolutely have to win this
46:17
field, just move so quickly that like, I find myself feeling
46:21
outdated every few weeks because all of some assumptions are had
46:25
about what technology can or cannot do just no longer applies
46:29
or something I thought worked really well certainly breaks.
46:34
So from that perspective, I think it's always good to have
46:39
some reason to want to do things, whether that's passion,
46:43
whether that's something else. But on the flip side, I do think
46:47
there's a lot of people that just want to try things.
46:49
You know, they want to build a site project to try things and
46:53
maybe turn it into a hobby hustle, just see what works or
46:57
doesn't. And that's absolutely fine.
46:59
Like I think if you're interested in building
47:03
something, you shouldn't let like I think you shouldn't let
47:08
the scary stories about entrepreneurship, eating all
47:11
their days and all of that to try because it's OK to just try
47:14
things, to try things and then realizing it's not for you.
47:17
And AI makes that very easy. Hey man, brother, I love it.
47:22
That's a great place to resolve here.
47:24
Now, if you're an attorney, go to iris dot AI.
47:28
It's IRYS dot AI. It's in the show notes right
47:31
now, wherever you listen to this show.
47:33
And if you have an attorney, send it to him.
47:35
Maybe you get a discount on that bill.
47:38
If you want to connect with Davanch, then just Google
47:43
chocolate milk cult leader. That's the best way to do it
47:47
because he's #1 and I can't recommend his sub stack highly
47:50
enough. That's incredible, and
47:53
congratulations on the success of that.
47:54
I know how hard it is building up a newsletter.
47:57
You know, I've been working on mine about a year and feel
47:59
really good about where we're at.
48:00
So that's incredible. Any other place where people can
48:04
connect with you personally online.
48:07
I leave all of my social media links that you can use to
48:10
connect me. LinkedIn, Instagram, any of
48:13
those are good. I think I try to be as
48:17
accessible as possible because it's very important to like if
48:23
we're trying to build things, democratize access to building
48:26
will be very weird if I was just not available.
48:28
And I think if people are going to reach out to me, I think it's
48:31
I the least I can do is give them a good reply back.
48:36
So any of those platforms will work.
48:39
That stack itself has a group chat and DL features now, so you
48:42
can use that instead if you don't want to expose personal
48:45
information to me, all things work.
48:49
Incredible. Well, I really appreciate the
48:50
conversation. This was fantastic.
48:52
Thank you so much. Thank you for having me, Ryan.
48:56
You are about to spend anywhere from 100,000 to $300,000 on your
49:00
kids college education, and that's not a small bet.
49:03
And right now there's a very real chance that they graduate
49:06
and can't find a job. Not because they weren't smart
49:08
enough, not because they didn't work hard enough, but because
49:11
the job market changed and nobody told them.
49:14
Employers are hiring AI native candidates, people who don't
49:17
just know their field, but know how to work with AI at a
49:20
professional level. And most colleges are not
49:22
teaching this, not even close. ainativestudent.com is the
49:27
program that fills that gap. It takes high school and college
49:30
students and teaches them how to actually think and work with AI
49:33
the way top companies expect on day one.
49:36
So your investment in their degree actually pays off.
49:39
You're already paying for tuition, housing, textbooks.
49:42
This is the one thing that actually makes it all worth it.
49:44
Go to ainativestudent.com because a diploma is a starting
49:48
point. Being AI native is what gets
49:50
them hired. That's AI native student.com.