Subscribe
Share
Share
Embed
The NeuralPod AI Podcast
The NeuralPod is all about deep diving into technical machine learning topics and showing its real-world impact, bridging the gap between AI-first companies and adoption.
We chat with subject matter experts across reinforcement learning, recommender systems, deep learning and generative AI. As well as chatting with business leaders, VCs, operations and strategy experts.
Who are NeuralRec.ai? NeuralRec is a recruitment staffing agency. We build niche ML teams and represent some of the globe's best ML talent.
Chris: So, uh, Mical, uh, welcome
to the, uh, the neuro pod today.
How, how are you doing?
Michal Klos: Thank you.
Thank you for having me.
I'm good.
And yeah, busy with, with a
lot of things to build, but
you're happy to chat about it.
Happy to discuss and share some ideas.
Chris: Yes, thanks for,
for making the time.
Uh, I think we first got to know each
other, uh, when you went to share chat.
Before that you've, you've had quite a
good, uh, career in games development
and also spent three years as a, as a
software engineer at ance and doing some
cool real time machine learning there.
And, um, you know, now, now at Lex, do
you wanna just give the, the people a,
uh, uh, at, at more of an introduction
into your career and, and how it's shaped
into you now running a legal AI startup
in Lex, and as you say, bit busy building.
Michal Klos: So, as you
mentioned, uh, I did video games.
Uh, I started my career, but I think
a bit step before that I actually
learned programming in, uh, high school.
So, uh, really back then I was doing
sorts of developments, freelancing.
That time I was already quite fluent
with PH, p and c plus plus doing
different sort of projects, which
already started with visual basic and
very thin book, but practical examples
that was laying around on my house.
But pretty, at some point I realized
that there is certain power coming
with the computing or like having
the skill to like fill the blanks
computer screen with some impressive
maybe computations, maybe graphics and
because I was gaming a lot at that time.
But then the first very first
application to me as video games and
graphics, was at that time developing
at very fast speed or like over the.
Three, four years there were coming
different generations, graphic cards
and different new technologies involved.
and over the like decade, the
graphic situated so fast and that
was like very tangible of the, like,
innovation in that field that you
could, that you could experience.
And that was moving really the
imagination that I had at the time.
So I thought I pretty much every
studio had to create something,
some technical selling point to sell
their title, to allow to experience
the gamers with that they couldn't
experience in any other title.
And that was, yeah, that was
very inspiring them to try to
build something like this, uh,
at this or the other studio.
So that was what got me into video
games and that also helped to, uh.
Have this performance oriented
mindset because the games still this
day have to process a lot and there
is a lot of organizations going on.
And actually after this in my workshop,
this what, what help me to also, uh, I'd
be hired by Biden when there is a data of
data in the, the, the, the top companies,
that also requires some skill to write
that code that is pretty performant at
scale and, and optimize the systems.
Chris: Yeah, and talk to me
about it sounds really, um.
Interesting.
By the way.
Talk, talk to me about, uh, bike dance
and know, you know, a lot of people, uh,
speak highly of the engineers that work
at, at bike dance and, um, you know, under
the hood, um, you know, it's known for
if real time machine learning, et cetera.
Uh, you know what, what, what did
you learn there and, uh, how, how did
you find there being hands-on there?
Michal Klos: Good question
and good memories.
So actually I had a chance to work with
some really brilliant engineers there.
we worked with the Mountain View
office, uh, and then later I
worked also with Beijing office.
So first, uh, area was the Intelligence
Creations department, and it was
about, uh, creating the libraries
that would process graphics.
Or with CPU or with GPU or with
any machine learning models.
And that would be then
combined in the stickers.
So essentially we were attaching the and
nose and errors of the dog or cat to human
faces for sake of the content creation.
or some more maybe impre less or more
impressive filters that will come be
combined, for example, for the, and then
it was very famous in vi unicorn face.
So the morph of person would morphed
into unicorn very seamlessly,
very perfectly on, on the device.
And yeah, I must say it was very, I
like that environment much more than
actual video games that I worked before
because of some product way, uh, thinking
about some research way thinking and
the architecture that we had and, yeah.
As far as the architecture of the
video game engine was very similar, but
there it was more, more subtle, more
flexible and more tailored to, with the
same set of libraries, run off on the
very restricted, uh, uh, mobile device
capabilities with even with the inference
in, in place operating with both with
GPO and also being deployed on the Yeah.
Uh, cloud Linux machine.
So with, even with no GPO at all.
So that was very interesting.
And yet simple architecture, so adapting
the, let's say, current preparatory game
engine would be massive undertaking to,
to them to such, uh, such application
than to comparing, comparing to the
technology they've built in, in, uh,
within a couple of years inside Biden's.
Chris: And, um, yeah, they're, they're
on to, uh, share chat, which is not,
not quite as a gigantic scale as, um,
TikTok, but still huge in their own right.
Um, you know, my understanding of share
chat there, so essentially a lot more kind
of zero to one projects and, you know, how
did you go from probably having every tool
under the sun to, uh, you know, having
to be a bit more resourceful, um, at
share chat and, um, you know, what, what
was it like working on machine learning
there and kind of lessons learned there?
Michal Klos: so essentially after working
on some graphics and filters and later
on, also on some cloud deployments at
high scale, was called by church that,
hey, like there is this Indian, uh,
social media company aspiring to compete
with the other big tech companies and.
Yeah, like if I need somebody that
would be able to help them build
similar things, who has seen how it
would work on the big tech companies?
So I've been speaking
with two departments.
So one was like the AI for the recommender
system, the other one for the graphics,
and actually this of the graphics niche
was the first choice to help them.
The initial idea was to optimize
the conveyor, uh, effect engine
something that would run with
more effects, uh, on mobile.
So to help their content creation
to extend the funnel and convert
more users or have more content.
Because more content means
more audience, more users, more
users means more revenue, and.
In that sense, uh, I was building
couple of, uh, extensions there.
The idea was to also be able
to deploy these, uh, modules on
the, on the cloud the in India
has big problem of the coverage.
So let's say China, US, they
can target high, tier devices,
let's say iPhones or such.
at that time, the low end devices
were in, in the west maybe was 20%.
Still big, but not us that big.
Yeah.
When covering, but in India there are
a lot of, uh, the low tier devices.
one of the workaround that, Hey,
how about we try at least compute
something on the, on the cloud,
on the servers for the clients?
So maybe that was the driven, uh,
uh, adaptation of what, how we
can work around with technology.
So that was the first thing.
And after, uh, having a couple of
projects, actually the department
was closed and I had, I was lucky to
transition to the recommender system
and I was optimizing the, the cost
on the recommender system pipeline.
Chris: Okay.
And that brings us onto X ai, which
is a legal tech AI company that you
founded with some of your peers,
which we'll, we'll really get into
in the next part of, um, the podcast.
But in, in terms of, um, you know,
machine learning fundamentals and, you
know, having worked at some excellent
companies like Ance and Share Chat,
um, you know, what, what kind of, um,
fundamentals really helped you and
are transferable across the different
companies where, you know, teams and
businesses are, are sold differently?
Michal Klos: That's a good one.
Many people would think about
machine learning as an action.
Like, Hey, we train a model and
then we'll get some cool value,
or we can offer something.
But if you would ask any machine
learning engineer, they'll tell you
that, yeah, like training is cool,
but 80% even of the time they spend
on data and as in any engine or any
compute, like garbage in, garbage out.
So then a lot of thinking
and focus come into data.
Do we have the data?
Oh, yeah.
That's sometimes very problem.
So if somebody has like thousands
of, of rows in database, that seems
to be very a challenge at its own
to train something on such data.
has a lot of data, then is better.
But then this question, okay,
what is the quality of that data?
Is consistent?
Is, does it have any holes or what not?
That needs to be done.
Cleaned, understood, prepared.
And then also when training you, you
can try in, uh, like different methods,
good for given problem, but, uh, some
of the fundamental, uh, optimizations
can be even in the future, engineering.
So again, how do we can prepare the data
so this, even the simple grad descent
or any other deep dive on that descent
that is best suited for a problem,
how that, uh, actually optimizes to
the way that we would like to achieve,
like where we can not run something
and get good result, but how we can
actually get this result that gives us
the, the value that our company needs.
If it's something that we have, let's
say, in the segmentation for the images
or if it's something for, recommendations.
So predicting what, what, uh,
next post could be clicked
or engaged the most because.
Yeah, like now we have many of the
shelf, uh, methods, so algorithms,
so sometimes I didn't, wouldn't
even need to train something to get
the decent quality of the result.
So for example, I could go to the
hack phase and try this or that
to see if as a proof of concept
this may work or may not work.
But in the, uh, at scale, it's
another problem that sometimes people
specialize so much and they are
chasing the numbers in so much detail.
1% is a big, big number or big impact
on the organization or on the results.
we want to optimize down to,
like, again of the 0.001,
sometimes matter.
This is where we have to put a
lot more huge resources to go
from, let's say, 90 to 90.01,
rather than to go from 50 to 51 or 55.
And again, understanding the data helps
to see what is possible, what is or
what method we need to apply together.
Chris: This may be a silly question,
but do you think everybody does that in
your career, starts with the database
or are people trying to maybe implement
the uh, a fancy model without it?
You know, I'm potentially
not done this step before.
The step
Michal Klos: And you should ask,
like for example, when I was
helping all like spring startups
in, uh, in eu, like before I joined
Chris: I.
Michal Klos: ai, I've seen that
in every capital in EU there was
startup accelerator and every
each of vendor was drafting, like
contract drafting for legal startup.
Every single one.
And when I came across Lex ai, they
actually seem like they don't have this
drafting, at least not yet, but they are
gathering the data And there was some
also examples like, like or done by human.
So that was also sensible, uh,
baseline for any other generation or
evaluation that we could then derive.
And that was one of the reasons that,
hey, like somebody does this differently
or has approached like step, uh,
before actually jumping into training
or drafting or generating anything.
And that is one of the
fundamentals that, that we have.
And later on we even, uh, came
into partnership with one company
that was doing the drafting because
they had some no harm on the legal
side, they have the legal license.
But if you draft today with
the current regulation.
Like, how do you ensure that this
drafting will be up to date within
half year later for your clients?
So you need, again, the up to date
data from somewhere to, to actually
make your service sensible and,
and again, uh, giving the value.
Chris: Super interesting.
Yeah, we'll get, get onto that.
Uh, Charlie super keen to explore it,
but in, in terms of, um, moving on
to kind of next part, I know Yeah.
Just something I picked up on, um,
when we first chat, chatting a while
about you've obviously got your MBA.
Uh, congratulations on that, by the way.
Um, and also, you know, reached a, a quite
a high level in, in machine learning.
I think that mix a business acumen
and being able to zoom into technical
models and then maybe zoom out
how, how, what, what's the business
impact is, um, yeah, say unique.
What, what made you pick your
MBA and do you think a formal,
um, business qualification
like that has, has helped you?
Now you're build in Lex,
Michal Klos: Good question.
So I could like explain a lot when I
started, why I would do it, nowadays I
just used to say because of the curiosity
yeah, like I've seen many things happen in
startups, different startups in different
places, and uh, as somebody building
things, as an engineering I was okay.
I'm solution oriented.
I want to create the best solution I can.
Then I had a lot of sometimes
doubts why something works
or is managed the way it is.
So yeah, I think for engineers it's,
uh, like we are solution, uh, oriented.
We think with solutions
you tell about something.
We, we are used to okay, thinking
how to your problem or what is the
problem that you're telling us about?
And when, uh, have thinking about the
business, more about what I may have to
offer or like, like look what somebody is
dealing with and, and is there anything
that we can help with this sort of area.
So if you ask me what, what MGA will
help engineer, the answer is with
pretty much nothing because it's more of
this shift in perspective or thinking.
Many times I've seen young people deciding
on the studies and when connecting with
the, uh, other managers and directors
that I've met during the MBA course
was like, yeah, first it would be
good usually to have some, uh, uh,
some concrete skills, whatever it is.
It's maybe like accounting,
low engineering, anything that
helps first to build like the
personal proficiency in trust.
But I am very good at giving domain
and, and I'm a good specialist expert.
And then usually any kind of
management, uh, qualifications help
them to extend that into more people
or managing projects on or whatnot.
And when engineering, uh, uh,
candidates ask, okay, like, I have the.
Study to, to pick.
So I will pick computer science.
So I will, I assume that, uh, the
instruction for my career is that I'll
finish the studies, I will get a good,
uh, role in some reputable company and
then I will like continue my career.
And that's the, that's the instruction
with also with if, if good company
probably comes good paycheck that will
allow to support the family or live
comfortably, which is in many cases true.
And yeah.
So even if I would ask you, what,
what do I need the business degree
actually, so then it's not as simple
that if I finish the business degree,
then I get the management job because
then again, or it builds on some, some
prerequisites it is, uh, some kind of, I.
that I put energy into being a good
partner to business or to the venture or
something like this, me think, okay, what
I can offer somebody, like what we can
Chris: Hmm.
Michal Klos: for somebody before
jumping into solution phase.
Chris: Nice.
Yeah, it seems, it seems like
it's, uh, helped you become more
customer orientated and, um, yeah.
Closer to what the problem is.
It, well, the solution, what
you're trying to actually solve.
So Cool.
And, uh, yeah, me moving on to
kind of more, um, the startup
environment and l ai, I feel like
in a last year, um, certainly from a
recruitment perspective, it's, it's
become cool to join a startup again.
Um, you know, lots of amazing big
tech folks like yourself have, um,
you know, who have, have been working
on large scale systems like TikTok
or Meta or, uh, Spotify, um, are now
deciding that, um, you know, they want
to get up in, in the start of game.
Do you.
What, what's one big, uh, mind shift
set shift engineers or, or leaders
have to make when moving into a
startup from a large tech company?
Michal Klos: It can be quite
different, but apparently even at
Biden's, you can imagine big company,
still, after 12 years, themselves a
startup and inside I would say that.
There was this also high momentum, uh, and
of tooling and environment that made it
enjoyable to, or proficient to work with
when compared to maybe some corporations.
and the bigger the organization, usually
it comes with some processes which
are there in place to prevent some,
uh, errors or prevent reputational
loss or prevent client loss.
There are also more, uh, middle
management layers, but also help to
get this whole thing together, focus,
align on the goals, align on the client
or market, uh, But then we have this
uh, another player, this organization.
So with startup, yeah.
Like I was, when I work
with our organization, I've
seen that somebody takes.
Two months to release something, or
even two months to kick up the project.
was laughing that, yeah, in a startup
I would do two pivots in that the same
time because of how, how, uh, fast or
adaptable the, the process needs to be.
So also it's interesting to see
the, uh, original agile manifesto
that was done 25 years ago compared
to the agile processes that are
currently in the big corporations.
And yeah, like it's about a lot
of, like a lot of exercise into
how we can do what we do, but, or.
Maybe not faster, as in how we can click
maybe more times per second than we
used to do, but how we can cut corners
sometimes more to get straight to the
point of what we offer, or straight
to the point what somebody needs, or
straight to the point of some kind
of prototype which we can discover if
what we think is what somebody needs,
on that get the feedback faster.
How we can straight, like, straight to the
feedback that would help us to discover
the obvious that some company established
companies already are at the obvious that
when somebody sees that and declares that
it's, yeah, it seems obvious, but when
doing like navigating the ambiguity of
some business problems, some clients, many
com like many customers, how to deliver
the, the val, uh, value to them navigating
this market fit is like trying to like,
think a lot of things, but trying to.
I trying to notice what is this obvious
thing that we need do to deliver this
valuable service to people that they
say, oh yes, I needed that exactly this.
Chris: Okay.
Um, yeah, that, that,
that's super interesting.
Um, and moving on to kind of l ai,
and please correct me if I'm wrong,
the legal monitoring and, uh, regu
regulatory, um, complexity using ai.
Um, you know, what, what inspired
the idea there and kind of what, what
core problems are you trying to solve?
I feel, I feel like, um, you know,
gen AI and the legal tech space are,
are, are primed to work together.
So, um, yeah.
What, talk us through
what you're building.
Michal Klos: Yeah, actually company
was uh, created before I joined in,
so I am kind of late funder before
me there was some other technical
funders and the vision for the
company was to tell different,
as in more LinkedIn for lawyers.
So it was supposed to gather different law
professionals and them stream of the legal
updates and help them to gather, like
create community around discussion about
what this or that law means actually.
And then of course, have some
model training, uh, based on the,
on the content on the platform.
But, uh, over the time it was refined
or like after having some first uh,
it distilled into the of making the
laboratory changes easy to digest
for lawyers because, uh, many legal
teams are still using the same office
tools as they were using 20 years ago.
And many processors
still in the spreadsheet.
The idea is that we could create
this legal radar over what
comes in and what comes next.
So, uh, so these people could spare some
time for the creative work actually.
Then just tedious document, uh, swapping
and document work that is, yeah.
Sometimes people with a lot of
experience experts needs to do
this, you know, like manual document
crafting and so on, can, AI can help.
And for me, I usually tell a
story that what we do is imagine
somebody has to sign off that.
Yeah.
Like the, they are clear for some audit or
they're clear for some business assignment
based on the regulations that are in
place for nuke the regulations coming.
And this sign of usually
means responsibility.
So they probably fit like
twice uh, going with it.
And then imagine you have to do it.
Yeah, it's evening on on Thursday
and you like to sign off, but
you maybe are not sure yet.
Yeah.
Maybe something new came up.
Maybe you can double check.
you could then yeah, like
browse number of pages from,
uh, authorities that has, that.
Have the 200 pages per each
regulation just to double check and
digest if there isn't any query.
Or you could log into l ai filter
by recency, see if there's any or
any news there any summary about it.
that would be probably more like
four pages just to review if you
feel confident, sign off and go
spend evening with your children.
Chris: Um, okay.
And yeah, you touched on, hit at,
at, at the, the, the start of your
point there that, uh, you know,
people still in E Excel sheet sheets
and potentially, um, you know,
still some filing cabinets hanging
around in, in certain legal offices.
Uh, do you feel, um, this is
potentially not a just problem with
ai, do you feel there's a, you know,
a, a resistance to people wanting
to adopt these, these systems in, in
heavily regulated environments like
the, the legal space and, um, yeah.
What, what's your experience there?
Michal Klos: Yeah, so it's the
resistance there is, there is always
resistance because, uh, if you're
experiencing enough, then you'll
notice that every person at some
point just is resistant to any change.
And that's totally natural.
and as I left, like especially in, in
Germany, people are very used to some,
some habits and don't, don't like changes
and that makes it even, uh, harder.
So I usually like to visualize that.
Yeah, like in, in California we have
some, uh, very cool visionaries,
uh, changing the world, like grand
projects that trying to solve
in very broad way, almost
like the tech utopia.
And at the other extreme of the
spectrum, we have the European, or
even better German, some regulatory
teams that would like to see the
world black and white and continue as
they used to do it for last 30 years.
The problem is, uh, like the
world is changing anyway.
And, and if you have a company, you
know, that, that, that if you don't
adapt to what happens in the world,
you might go out of the business.
And yet, uh, experts are usually so much
focused on what they do and how to do it
and more effective that sometimes they
don't really have that time to notice that
how things, how much changed over time
and how much the risk has, like pressure
or stress has, uh, accumulated over time.
that's why then, uh, like if you ask
somebody to do something one day faster,
will think how to improve the process, but
there needs to be some bigger eye-opener.
Like, hey, like can you
do it one month faster?
Like, how can you break from
some habits because we have
something new that may help you to.
Work a bit differently or to adapt to
a bit more maybe convenient, more work.
And this is the, not only working with
people and like noticing what they
are thinking, how they are working,
uh, it's about noticing the workflows.
So how, if, if there is a team, how
do they usually work with some legal
updates and how we can fit our service
towards their workflow rather than
trade in this like, uh, Silicon Valley
style, Hey, we have this new mbus
and it will solve all your problems.
Please use it.
Yeah, it is so, it's so cool.
It's so awesome.
So, and, and everybody has it.
So this
Chris: Yeah.
Yeah.
Really well put.
Michal Klos: And
Chris: Yeah.
Michal Klos: on top of that, even
if we have people working in some
workflows, again, on the surface you
may not notice, but if somebody is
experiencing enough, they start to notice
communication also between these people.
And the politics sometimes.
Yeah.
So sometimes something doesn't make
sense, but in terms of that organization,
some politics make it work this way.
Even that's maybe not the most
optimal way, but the way it is.
like we say, see now AI call.
But if we see, even now the lesson or
most recent trend is agents also cool,
but if we jump to agents too soon, it may
look cool at the demo and on LinkedIn.
But then if you realize the usual
company, organization workflow, this
workflow needs to be meticulously
mapped and laid out in terms of
something that automation can before.
Yeah, throwing agent on it
and it will figure it out.
Because then it will be another
layer of mass and, and it
will be hard to, uh, operate.
And I think recently Walmart had this
use case and they did agent everything
and then they a step back to automate
the workflow that can be su supervised.
Chris: Interesting.
I think that segues nicely onto, um,
you know, some of the ML systems that
are powering the Lex AI platform.
Do you wanna just talk us
through what's going on?
Going on under the hood?
Michal Klos: Uh, yeah.
So, uh, since it's very lean on smart
startup, it's some, like, some people
expect AI to be chatbot and we indeed
do have chatbot, but I wouldn't focus
so much on chatting feature much.
Um, think about more as a database.
So more like Bloomberg for legal
monitoring, because if we accumulate
a lot of data, a lot of regulation, a
lot of updates, then the next thing we
would like to do is to search efficiently
through these updates and search and
optimize the search for the legal domain
with, for example, how the legal people
would filter this data to navigate
to what they are using the mouse of
what they are looking for, easiest.
And then on top of course, we have this
ai, uh, layer for example, you could with
also with, with, uh, profession search.
Uh, we can, uh, information and then
glue with AI into some constructive
assembly that will answer the questions
that, that people might have or find,
help them find the in one place, help
them find, uh, useful information
and from number of jurisdictions.
So, yeah, so, uh, from the more technical
point for me, yeah, I don't want to
have the headache of managing database
because I have a lot of things going on.
So I have database that
can grow into megabytes.
So this is one thing.
Uh, of course there is the vector
search enabled and, and, and,
and, uh, embedding storage.
that's combined together.
So we have, uh, convenient and fast,
uh, full text search also compared
with the, the semantic search.
So, again, or by words or by vectors.
And that sits as our no, like,
like fun, fundamental for any other
logic that we can, uh, do on top.
And if we look at the
like content journey.
Of course we have hundreds of sources
we watch, and this is, uh, automated.
So we have, yeah, like we, maybe
not yet agents, but I've looked very
carefully how it was done first manually,
uh, by people with legal knowledge.
And then that was implemented
automated workflow that runs for
each of the authorities we have.
And then that comes for some content
understanding, then some metadata, uh,
prediction based on the historic content.
then with that, we have candidates
to be, uh, released, published
on, on our, uh, platform.
And there is also human in the loop.
if anybody ask us, okay, cool, but is
this just some hallucinated gen AI slope?
It's not because it has
to pass the gate of the.
Of the, our legal team that, that has
good legal domain knowledge, so then it
Chris: Yeah, that, that.
Michal Klos: Yeah, and I've,
I've, uh, talked with some other
people doing similar systems.
The, there usually is spot with
human in the loop, especially in
the legal world, because Yeah.
How, how else you can, uh, we can
verify the, the, the, the quality.
So that comes,
Chris: Exactly that was going,
Michal Klos: yeah.
Chris: sorry.
Please go.
Michal Klos: we have the platform, we
can search it, but also we have quite,
uh, notification system built in.
So if we have any, uh, people watching
the, or like monitoring the news, then
they usually want to be notified about
what they feel relevant, so, or they can.
Like set some filtering on what they
want to receive, or we are, for example,
experimenting with some very, uh, very
simple recommender system also inside.
And examples that I could see here
is like, there is one, uh, query that
is like multi-stage that I could do
something almost similar as the elastic
surgery ranking is in the US version.
So it, it works quite well
out of the, out of the box.
So, so pretty much if you have any,
any, uh, homegrown content system,
uh, elastic search saves you a ton,
but for several reasons, uh, I'm.
that as a baseline and then,
or for special use cases.
And then our
database is postgrad and
I'm building on top of that.
A lot of those are like, rebuilding the
features that some other more specialized
database could offer because of the,
the size and easier way to manage the,
the, the volume of the data growing.
Chris: That's interesting.
And um, yeah, you, you, you kind of
touched it on it there in terms of
hallucinations, you know, anywhere where
AI is summarizing text, I think it's now
people say, oh, what about hallucinations?
Um, so you're saying, you know,
human in the loop, that's the
way you've kind of solved it.
Um, because I imagine any hallucination
in the legal world is potentially fatal.
Michal Klos: Yeah.
I would say there is, uh, two extremes.
Yeah.
So we have this AI.
Which is sort of statistical way of
digesting the text and transferring
one text to the other text.
And then on the other side we have
lawyers who see world black and white.
And on the one side we have like one
to a hundred percent confidence or like
probability that something is something.
One the side, we have the
black and white, good or bad.
And if bad, uh, then you have a problem.
You have big problem,
an escalation probably.
So this is where the workflow is first
thing, and this is where, uh, like
understanding how to translate or what
is good enough comes into play because
that what also happens in other areas.
So say we do 60% correctness, then the,
uh, most of the consultants in the US are
happy and they can tell a story already.
But if we come to lawyers, and
especially in Germany, then sometimes,
yeah, you have one, like 90%, 95.
Yeah.
It's, it's getting somewhere.
I would say that that's why I
sometimes love that LMS would
never happen in Germany because
unless it's 99, 9, 9%, yeah.
Then, then it
Chris: Yeah.
Michal Klos: considered, uh, ready
and it'll take much more long time.
Chris: And what do you think the
biggest challenge is building systems
in legal tech space right now?
Michal Klos: of course, uh, I would
say it's not about creating the best
foundation model, perhaps, unless you
go to these people and ask them what
they need or how they work or, or.
What is the smallest step that we
can pilot so they gain trust that
okay, this automation helps them,
uh, and it's not replacing them.
'cause there are still
some discussions like that.
we are not replacing anybody,
but we help want to help them
to do more of the creative work.
And this is where, working with the
actual consultants that help these
companies, uh, uh, comes into play.
this is one of the, I think one of the
biggest, uh, wins we had in the recent
time is that we partner with Deloitte.
And Deloitte is one of the big
four, uh, consulting companies.
That if anybody has risk and because
businesses, uh, are above managing
the risk, if anybody has any risk and
they want to move that somewhere or
pay somebody a lot of money to, to
manage that risk, that would be many,
uh, companies, Deloitte for example,
and they have some methodologies
and they have tools that they.
To a various standard in working
with, uh, legal compliance.
And this is what we are together now,
uh, working with to see what workflows
do they have and how we cannot only
digitalize some of these workflows.
So it's faster and, and many people can
seamlessly work on it, but also how to
give instant insight based on the data
that we already have in the platform.
So,
Chris: Yeah.
Okay.
And obviously delights, like
you say, a big four consultancy,
amazing achievement by the way.
Congratulations to, to
you on the team on that.
I know you obviously worked
incredibly hard on that.
Um, but what, you know, would it, would
this, would Lex be suitable for small
in-house teams as well and, and smaller,
uh, I guess law firms as well as,
uh, bigger organizations like Deloit?
Michal Klos: Uh, so Deloitte is
more of, uh, like knowhow partner.
And of course they would like to serve
more companies of growing low complexity.
That's why they need a tool and they
recognize, let's say, as a good tool
for legal monitoring, and this is how
they can, uh, their services better.
And this is how, uh, like usually
big companies turn to Deloitte.
So this is where we can see like
the full overview of what is, like
the palette of the legal problems
that they are dealing with.
But I would say, uh, any, any size of the
company would be good if they want to try
or if they want to use, feel big enough to
have a legal team that needs to watch if
the business is compliant with services.
So that's sometimes, that's usually
like more than 300 or more headcount
at the, at the headquarters.
more like small and medium businesses.
to like some maybe big discounts change
in Germany that also work with us, because
if we think about the biggest, then there
is the biggest competition on the, let's
say, from Harvard AI or, or such such
companies with, with much more ambition.
Chris: Nice.
Yes.
Okay.
Uh, I've got some friends in
the, the legal tech space.
Um, uh, I'm gonna mention
it to them for sure.
Um, let's talk about product market fit.
I know, um, you know, it's not
just a problem specific to,
to the legal tech space, how
legal AI space many people know.
And, uh, finding it can
be a, a real challenge.
And we've talked about journey,
the journey from exploration
to research to development of
a product briefly in the past.
Um, can you walk us through your
approach to finding, uh, product
market fit and or caveat that with,
uh, totally understand there's, there's
no silver bullet for these things.
Michal Klos: that's true.
So many people look around,
especially if that's some white
Combinator, like rock stars.
They usually have success,
so they don't tell.
Yeah, we did it that way.
And then many people like to follow, okay,
if you do this, then you'll get success.
But it turns out it's, they had different
sort of factors and environment that
what they did, what is actually the
story after the fact that it worked out.
uh, pretty much it's always, I
think about the, the termination.
So what I, for example, on the MBA, uh,
studies, what I thought, okay, I will now
know how to do a valet business plan or
startup pitch for, for example, for vc.
Uh, so yes and no.
So first I've learned that yeah,
I have enough tools or knowledge
to craft the pitch deck and
business case that would not fail.
So let's say I'm, uh, it's 50% better
than maybe of the ba it's 50% of the base
of the grade, maybe that you would get.
But then anyway, the, the
rest of the 50% is the team.
And so somebody might see, okay, do they
have enough, uh, they, are they enough
smart or clever to, to find some market,
some business case to, or like some angle
to the market that would make them better
than any other, trying the same, uh, idea.
And is this team good enough
to make it work, to cooperate?
And then if market changes adapt
and spot, what is their next step?
They actually wouldn't know at
the time of creating the pitch.
And I think this is also the part
of the, uh, at the feedback and
when approaching people, approaching
companies, I trained very closely on
that feedback and then seeing like,
okay, uh, I have maybe some vision.
But what is out there in
the wild that happens?
And I think the good case
could be in the hugging face.
So they had one vision, but because
they were out there in the wild with
their company already operating,
uh, this is where they spot,
okay, people behave differently.
They see some value, but in the
different thing that we offer.
let's say I'm doing the coffee
shop, I'm serving coffee.
I always dreamed that was serving coffee,
but for some reason people like this
machine that I use for, for coffee,
I would like to do it in their homes.
So now for me, the big decision, okay,
do I keep the vision and try to force it
or I will maybe try saying the machines.
So this is, uh.
Another good use case was, let's
say, can we offer something before
it's built even and see and gouge the
interest and see the feedback better.
I, again, it's all about how we can
cut shorter to the point of getting to
see interest or see the feedback from
and, and optimize for that niche and,
and capturing that niche and seeing
how far we can go how people, uh,
how valuable people, uh, think it is.
Chris: Yeah.
And you, you kind of touched on it there
in terms of, um, you know, listening to
feedback, um, you know, how do you balance
that framework from what is in reality?
Um, you know, constant feedback
that you potentially getting a
may feel messy or unpredictable.
You know, how, how do you balance
what the books are telling you to,
what, what's, uh, actually going on?
And you've touched on it before.
You might have pivoted two times
already, you know, how do you, how
do you know, and again, I appreciate
there's no silver bullet, but, um, how
do you balance all the moving variables?
Michal Klos: there is never a good answer
and it's usually like, okay, it's best
effort to try to get as close as possible.
if we have any feedback, um,
from the customer, this is
like the gold, the gold thing.
So first thing we do, the
feedback from customer.
Like we, I would even say there are
a lot of effort into actually getting
the feedback, with tools of analytics
or with anything if, if there is
enough signal that we can capture.
But, uh, fortunately in the remote
environment, we don't have possibility
just like watch over the shoulder how
somebody works and notice some things.
So usually it comes from to the meeting
with clients and like hearing some.
uh, what they see, what they like,
what they dislike, telling what, how
they thought something would be obvious
for them, whether they would click
or what information would expect.
And when working with special
legal teams, I, you, you wouldn't
get a lot something doesn't work.
And then, uh, there has to be way
to see, okay, what is expected?
It's because with only the, like the not
working, uh, signal, it's hard to you
know, like, okay, we can explore and we
can go for okay, explore them, evaluate
the, the, the best, potential candidate.
But that takes time.
That, that is some kind of framework
and we could jump straight at
something that we can show,
okay, is this example you like?
And then, then get.
Get feedback on that,
then we can move further.
And many times I had some ideas
or some vision how something might
work because it makes sense for me.
But then when, uh, confronting that
with client or customer, like any
kind of, okay, like, okay, I need
to differently because this is
what somebody else is expecting.
Chris: Yeah,
yeah.
It's a, it's amazing like we're
creating all these, uh, you know, really
advanced ML models and systems and,
um, getting in front of someone face
to face, a person to person is still,
um, one of the most powerful tools
out there in developing a, a really
great product, which is obviously free.
Michal Klos: I had the cases
like, can we maybe gouge some
answers that our chatbot is giving
That was, uh, for example, preparation
for, if we meet with client, can we get
a gist of what is their experience from,
from, uh, features including the ai,
and can we get some idea, okay, is it
mostly answering the question or maybe
we have some, uh, misses there and maybe
we can get some idea with analytics,
what topics they are asking about.
then under, yeah, like
understanding the behavior.
And if, like, if in anything that
anybody is building, understanding
the behavior is first things probably
somebody is, is maximizing to get
the signal from the behavior and
from the data for the iteration.
And then I would notice that, yeah,
somebody would ask the, ah, chat bot to
maybe enable notifications because I, I
would never thought about building that.
But somebody is asking about,
Hey, can the chat already.
Like push some buttons for me in, in the
tool that they are using, which was kind
of hilarious, are if they failed that
maybe they have constructed very, very,
very complicated prompt and even another
lawyer could not really know how to answer
that complex question because there is
just too many, too many threats in that.
So again it comes, okay, what, what
is interesting, like what people think
that AI would for them, and even the
AI perception is much different since
JGPT release because from the market fit
navigation, we, there was one anecdote
that Lex AI had no chatbot yet integrated.
couldn't log in and chat about
legal, like please find me some
legal regulation and answer if
there was some penalty, let's say.
So we had the.
Metadata part in AI and the
summarization for summaries in in ai.
But that was more of the back office.
And when my co-founder went to the
meeting, he's a service guy, he met with
some other serious lawyers and they have
very high voice about like, can you show
us what, what we can do with Lex ai?
He showed how it lacks ai and
they have very, very question.
It's Lex with AI in the name show US ai.
And he was a bit baffled,
like, like, like we have here,
like this is our like service.
We have some AI summaries on.
And they, no, no, no.
Show us the ai.
And they were expecting the
perception of the chat with the
AI chat bot Since chat, GPT.
So after hearing that, I
was like, yeah, good point.
if that's silly, sometimes we have to make
something just so people can see the proof
that we have, have this under the hood.
And what I did was on a Friday, I had
that, uh, that story I pulled some of the
like examples that I've seen about Thes.
by Monday I said, Hey, if anybody ask
about the ai, here is the internal
chatbot that I combined with some
latest news and show them does this and
this, this should, uh, be good answer.
Just so yeah, we have this ai and
again, we'll need to build more based
on what somebody would like to ask
it, because then it's not that easy.
But this would be something that
we can then already and build
this ship while, while we need to
sail, based on, on the feedback.
Chris: Yeah,
I think it's that you've.
Testament to how you're embracing,
uh, life as a startup founder.
And, uh, you know, you're getting
feedback on the Friday and you're, you're
shipping the feature by by the Monday.
Uh, that, that's, um, that's great.
And I, I, I guess the next part will be
how, how, uh, people are defining AI and,
um, a GI, uh, that's seen a lot out there
and people arguing on technicalities
of what, of what things means.
So, um, cool.
Okay.
And, uh, yeah, just moving on to
the last couple of parts about, um,
Lex, and let's move on to culture.
I know, um, you know, there's a
classic mime out there that says,
you know, don't, don't hire Google
engineers because they'll over,
they're over engineer everything.
Um, you know what, what,
what truth is behind that?
And do you think there's,
Michal Klos: So
Chris: you know,
Michal Klos: co-founders.
Chris: it should be about
shipping over perfection.
Michal Klos: this meme and this, like,
this is a strong statement that could
put on LinkedIn to general discussions.
my co-founders would follow it, they
wouldn't partner with me basically.
Yeah.
So if they were to take this
advice, but it's, I think the Google
engineers are the most type of the
minds and, and brilliant, yeah.
Like software engineers out there.
And I can, couple of my friends are
indeed working at Google and they are
pretty good at, at, at algorithms,
math and, and, uh, many other things.
But I think could imagine like why
somebody might be a good engineer
or what makes them, so sometimes
it's just because somebody Yeah.
Is brilliant in terms of the STEAM
subjects and they can navigate that and
they can, that that's their element.
But sometimes it's also that, yeah,
somebody might be, like in civil
engineering, they would like to build
the biggest canon that can shoot, uh,
into the low orbit, the satellites.
And that's their ambition.
They would like to build this,
their project they like from heart,
they love, and then they look
for opportunities to build that.
And then they may be so much focus on
the solution that they don't really
consider other aspects or, or like, they
like to build the perfect system and
many other aspect might not come into
their mind first without, without, uh,
So this might be the same
case for the, uh, big tech.
Many times we have the perfect
opportunities to craft be two systems,
like the most performance systems
or that you can nerd snipe at, hey,
like, we have the best, this or that.
And there are a lot of tools out of other
brilliant colleagues to help with that.
But then, yeah, like at these companies,
optimizing something by, by again, but
fraction of percent is a big impact.
But at startup, many times we
are optimizing for existing even.
Yeah.
So, and there is other parts to the, I
think this, uh, big tech, uh, culture
that from the business perspective, from
even the finance perspective, if we,
we prefer to the tools or the software
that we built internally, it's like
not built, uh, not built here syndrome
because, uh, many times, if anybody.
Usually take a decision to
build or buy a big, big take.
When it's comes to scaling to millions
or even uh, billions of people,
licensing would be pretty, pretty much.
So even spending three months on
building something in house sounds
like a better deal than buying it
and then facing the risk of very
high, very high li licensing cost.
So for that reason or for the promotional
reason, like if I find this like need
pick on some edge cases and that will
allow me to build something that works
better for our fraction of percent
improvement and impact, then I will
get the promotion and, and this will
be something that's not my evidence.
That what, why, my brilliance
needs to be rewarded.
And for these also organizational
problems, this is why
these people optimize for.
being brilliant, building beautiful
systems, which up the environment that
somebody will put, it's optimizing
for revenue or optimizing for runaway,
which means existing, or how fast
we can build something that maybe
is not yet perfect or just exist.
And we can, as a prototype, put in
front of somebody to learn more before
we will, uh, prepare for scalability
on, on a scale of thousands of so many
times somebody will argue that yes,
we, we need to spend a couple of more
months to prepare for the next phase
of scaling the number of, of people.
But what if these people,
these users will never come?
And that's the under risk.
I, everybody would like to be
com uh, comfortable that it, the
system wouldn't break, but if it.
If, if anybody wouldn't predict that
they have a scalpel system and then
many users, then this is the, the
pain of growth and, and patching it
and then making scaling as you go,
as you have these outages or whatnot.
So this is the, the growing pain
that not always you can predict.
Not always.
It is a good idea even
to prepare for sometimes.
And this is what I would
say the, the happy accident.
If it happens, we, we can
really guard for everything,
uh, when building something new.
And we want to be, uh, faster of reasons.
Chris: Yeah, makes total sense.
You know, as you said, um, at the start of
the podcast, can lab mindset and building
things quickly within, um, a startup.
So important.
And, um, yeah, not always
time for that, for perfection.
Um, okay, I'm moving on to the
kind of last, uh, question in
terms of the, the legal tech.
Um, you know, what.
As we kind of said, it can, well, it
potentially can be conservative at,
at, at times the legal tech industry
for, for obvious and right reasons.
Um, what, what's your thoughts on
where the sector's headed with AI
in the next kind of, uh, six to two,
uh, six months to, to two years?
And I know, uh, with how fast, uh,
technology is changing in AI that
no one can really predict this,
um, which makes it part of the fun.
Let, let's see, uh, how right or
or wrong we were in six months.
Michal Klos: So as mentioned, uh, I
think that one point as we have many
passwords coming over and over, let's say
from the visionaries, uh, in the Silicon
Valley, it's not easy to adopt them.
Start fast at the, let's say,
consumer corporation level.
Like of course, it's for individual easy
to sign on to, to the charge GPT, but
it's not easy to convert the workflow.
So again, as with preparing data
for any kind of training here,
we need to prepare the workflow
for the, for the, for the AI use.
So I think there will be not that much,
uh, uh, not that much in terms of ai.
So hopefully the tools will gradually get
better and there will be more adoption.
But I would also focus more on
the people side, so how these
people will behave using theses.
there I hope there won't be,
not much fear about, okay, is
this AI going to replace me?
And rather it would be more, okay, I
now more time for something else than.
document rewrite, or I will
get but much better searching
thanks to integration of ai.
also this depends on the region.
I think so us, yeah.
Usually the, the law,
uh, works differently.
So there is more, uh, more
places for spectacle cases
because of the precedent's law.
Uh, so maybe there could be some more
news that somebody used new AI set
up to find the, the work, like to do
the work of maybe hundreds of interns
just by looking for the documents and
that that was already done somewhere.
So probably there will be more cases
like this, maybe more spectacular,
more, more to write story about.
think in Europe it will be more
gradual in terms of, uh, like
through layers and layers of EU than
the, uh, national, uh, and maybe.
Nearby jurisdictions and hopefully
there will be more, uh, adoption there.
Another topic that is, that is
Chris: Hmm.
Michal Klos: but very, uh, very like work
heavy is, for example, internal policies.
So internal policies like our thousands
of in front documents that describe
what, how things should work or what are
the standard operating procedures and
updating this manually is a lot of work.
Hopefully there.
Also, this could be done easier and
faster without breaking any privacy.
So I hope also anybody using the
services will not, will not see any
data breaches, let's say has some trend
documents leaked this is another very
high concern, especially in Germany.
Chris: Cool.
Okay.
Uh, that's pretty much it for
the, the legal, um, tech section.
Thanks so much for sharing Lex AI
and, and what it is you're building.
Um, really appreciate that.
In, in terms of just the last couple
of questions, it j just a bit of fun
in terms of, um, sharing a bit of
knowledge as a founder, obviously.
Time is very precious to you.
Are there any AI tools, agents, or
workflows that help you, uh, keep
you really productive at the moment?
Michal Klos: Many good engineers stepping
into startup note also getting burn out.
So I think that that's
the first challenge.
Even if best AI system, it's
will be better not to burn out or
not to suffer from overworking.
That may happen if there
are too many things.
And, and it's pretty overwhelming,
uh, I would say at times.
So the prioritization is the first
part, and then having the, like
matching the energy, uh, with the
workload and there is pretty no, no
again, uh, best way for everybody.
So finding everyone's, uh, good
energy distribution of what day or
how to, uh, prepare for marathon
with, with good energy, like with
good sleep, with good, with good life
habits is something very beneficial.
That actually, yeah, for
me it works, uh, till now.
So I try to manage, for example, working
hours where, when I have the energy
for it or like when I have the like
best like for thinking or, or whatnot.
And yeah, for me that's evening usually
so, so contrary to maybe some, uh,
Amazon, Amazon guides or whatnot.
Chris: It's a really good point.
And one what?
Yeah, it's a really good point and one
that I wasn't expecting you to, to say,
but as a, as a solo founder myself, I
can relate and, um, you know, knowing
when to work and when not to is, um, the
key to longevity in, in the startup gig.
But I think there's so much, uh,
information out there in terms of,
you know, startups, founders in San
Francisco working seven days a week.
And, uh, like you say, I think you
put it perfectly, it's about what
works for you and, um, you know, I've
certainly worked weak as before, but
our value taking time out and think
it, we should normalize that more, uh,
rather than the the ground culture.
Michal Klos: I've seen there,
there is, uh, 9, 9, 6 in China.
It, you can do it, but everybody
does it and nobody is ahead actually.
So it can fall short and uh, like if
nobody in Silicon Valley does it, cool
if, depending if you are young, if you
have nothing to spare, you like spending
being productive if you like it, do it.
But I also have seen some solar funders
in other, uh, industries that would do
maybe like three days of hackathon mode
and then two days of, uh, uh, relax
of of, of getting the energy back.
And then again, like intense days.
So it works differently for anybody and.
For example, yeah, if you can
imagine I don't drink coffee.
Um, because yes, but then I realized
the way I usually drink my beverage
is, uh, very similar to Moderat.
then again, like helps to
break from some, some cycles.
And with that comes the time, yeah.
How to, how to manage the time.
And usually if you can, uh, delegate some
things the time, that's usually good.
But for example, I've never seen good
meeting summarizer from transcription
transcriptions that would work for me
because that was making apparently more
text that I need to digest and understand
and this heavier workload for, for me
to comprehend, um, among many things.
but there comes
Chris: Hmm.
Michal Klos: tools.
True, I do use some ai, uh, even charge
PT sometimes to do some research.
it's searching over the web pages much
faster that I can, that I would do
sometimes back in day to open like 10
tabs and then digest what is there.
So it's good for like surface research
or it's good sometimes to, like, if
we have some meeting, uh, notes or
whatnot, then structure it in more,
uh, for example, product driven way.
So have some extension of me that could
debrief product management for one or
the other part just to get the structure
good and so we can see if we don't
have any gaps that we need to discuss.
So having this assistance
sometimes beneficial.
And I use also the, uh, coding
tools, so it's not entirely like, uh.
Coding.
But, uh, as I with my friends,
is more of the, again, more
stretched strategic thinking.
So I need, I know what I need to build in
what steps, and then I can see, okay, if
this can help to maybe fast forward to the
next step that I have on that checklist.
it's not really like missing all the,
Chris: Yeah.
Michal Klos: all the, authentication
on other house because with enough
knowledge we know how to build things.
And sometimes it's like per programming,
sometimes it's, uh, trade off with
freelancer, if I were to give that
task to team member or somebody and
get the code that I, with some, some
fixes or some, some, uh, stage scalping
that I could do and, and refine it.
Sometimes it also is rubbish, so not,
I can accept every, every change that
I get and I, I think already somebody
noticed that if you try coding with
prompts, pretty much first answer
will be always rubbish that you need
to reject or, or tweak or whatnot.
But it, if the context is good, if
managing what we need to give, uh, any
kind of coding assistance, what kind of
samples on the parts of project to digest,
sometimes it can actually work quite well.
And I think even to extent that, you
know, like I have a friend that is
also managing IT team and he realized
that, yeah, he would, uh, move from
senior engineers to junior engineers and
use these coding tools in, in inside.
Apparently he got better result and he
built something in half yard than he
would for two years with some external
agency he had, again, uh, better
integration site, better feedback
inside than, uh, yeah, like lengthy
sessions, feedback session with customer
success with somebody external that
doesn't have that internal incentive.
So apparently in that case, like
Chris: Nice.
Michal Klos: may see, okay, like
junior or AI assistant and, and,
and debate which one is better.
The, the takeaway is both because,
uh, we need an operator that can
adapt and we need a machine that can
help them to, to operate, uh, faster.
Chris: Oh, uh, yeah, it doesn't sound
too dissimilar from a, a found a
friend, um, of mine, a a little bit
different, but one of his processes
now for managing his team is, you
know, go and put it free, free coding.
Assistance our agents and then ask
him to review it and not before.
Um, so yeah.
Um, okay.
And, and in terms of, um, you know,
a book or, or an idea or a podcast
that's not the neural pod that may
have shaped, um, your approach to
leadership and innovation is any kind of
papers or, or books as I say that have
really, um, helped, uh, refine how you
think or how you approach the world?
Michal Klos: so when we think
about the, uh, it, science three
years ago is already, feels pretty
much old, but I still think, uh,
Chris: Yeah.
Michal Klos: reads are pretty very much
up to date because on the technical side.
Books may be outdated over the one year
even, but on the people side, I think
the mentality or some organization
behaviors are still update, uh, even
in the matter of a hundred years.
So I'm fascinated how much many
parts of the medical man with from
seventies still are up to date.
How the, for example, nerd leadership
works still after 50 years.
Next time that you can, yeah.
That you can score the.
Respect points by solving something hard
and, and people who are into engineering
will really, they will appreciate it.
And then still over the time I come
back to Joe Polsky with the stories he
has from Microsoft from two thousands
and still many of these games I think
are very up to date until today and
over time more of the creation of
something interesting and managing
or like noticing things as a manager.
Uh, I think there are, the
creativity in by at Cat close to
my heart because I like the story.
I appreciate, uh, Ed's contribution
to the graphics of within
my, my interest in graphics.
And that was very well explained how they
also not only wanted to build things,
but also how they, how they wanted
to manage the relationships with the.
Oh, we have steep jobs, uh, like, uh,
it's debut, uh, if people like them or
not, but, but, uh, he, as the manager
of the Pixar division had to deal with
jobs and then he also had to navigate the
ship with much bigger, think, like very
big players like Disney and how to try
not to compete with them, but to create
something with them and then create the
movies and very start really, really small
as a very small player, uh, in that case.
Chris: Nice.
Okay.
And, and final question, um, you know,
outside of legal tech, which I know
you're super passionate about, um, you
know, what AI and and ML trends are
you, are you most, um, excited for?
And, and just to tag a, a mini question
onto that, uh, where do you think the next
big breakthrough in, in AI will come from?
Michal Klos: that's
really hard to tell that.
I noticed is that like finally even a
recent change was that people accused
recently Chad GPT five to be less polite.
What I found that GPT five was, yeah,
it's quite okay, but the four was
so sugar coating that it was hard
to stand and I think it was too nice
Chris: too nice for you.
Michal Klos: I think there is some trend
probably it's not, not different than
any other service actually, because
if you notice some social media.
They for revenue, which means
they optimize for the content you
like, and then, uh, they put you
in your bubble of the information
you like to get to stay there.
And I'm afraid that this will
be the route for the, like, mass
market AI models, like open ai.
If they want to be liked by everybody
and be popular, they might be
optimized the same way as we've seen
the recommender engines for social
media have been optimized over.
Uh, so is one of the, I think,
uh, for me, negative trends Yeah.
For maybe social media positive
trends because they know how
to optimize it for revenue.
So this is one thing.
Another thing is that, uh,
we, we've seen one wave with ml,
we've seen Next Wave with like GPT.
think that's.
may be experiencing some
compound of many smaller waves.
And I'm really, uh, like I'm
really, interested why we don't
have yet so much video, John Wave.
Because when I look at the, our
Bidens, our Google models or even
the not so recent, uh, thing that
was the Chinese manufacturer,
uh, I can't recall at the moment, but
they would create, uh, short video that
could be worthy of the movie release
of like Pixar short and probably with
John AI and the quality they are getting
approaching to, that's pretty striking.
So this is something that, uh, I don't
know why it's so silent and if you
Chris: Yeah.
Michal Klos: it more, yeah.
Chris: Do.
Would,
would, would you go and see, I mean going
a bit off topic here, but would you go
and see a media that's completely AI
generated, even if the story was good?
Um, and I guess you might go and see
it for the novelty at first, but then
I dunno, surely there's something
to be said about going to see one of
your favorite actors, um, in a movie.
Michal Klos: true.
I think this, uh, border is not
yet so distant and black and
white again, because we already,
with the, this is the, like, the
Chris: Mm-hmm.
Michal Klos: that I see
and maybe next disruption.
Uh, already we've seen, uh, artists
complaining that JGPT and other similar
models were stealing the artwork
because they just crowded all the
internet, including the not licensed
work of arts from, from artists,
and they can replicate the style.
And we already seen that, uh, uh, like
the AI gen is applying first to the
fields that have less error factor.
So let's say if something fails that,
or they are not a hundred percent
correct, that's not that severe.
we've had El Musk predicting,
like promising the card
next each next three years.
But because of the criticality of the,
like, it's let weapon and, and this
lethality, if anybody, anything breaks the
responsibility for that is hard question.
And nobody wants to ask it yet.
Like, like, like to that yet solve it.
But we have that with text Take ai
slap for LinkedIn, nobody bats an eye.
We have that for generating images.
Because if there's number of pixels
are out in the body curves, we
have that for AI code because that
can be easily verified with one
test if that runs and compiles.
And for the graphics again.
we have so many v uh, visual
effects studios around the world.
And unfortunately in that
business, it's about the cost
of production to get it lower.
And a lot of that work was
already outsourced to Asia,
like India or Philippines.
And this could be the next place where if
we use the AI tools, this workload, this
workforce may not need to be as many.
So maybe if we will not even realize,
we already have seen some actors
that we liked, but they were done
entire cg like, like some, uh.
passed away that were revived with cg.
But if we add some AI tooling
Chris: Yeah.
Michal Klos: compositing or, like masking,
uh, there is, I think much better, uh,
visual effects style, uh, jargon for it.
But these are the very labor consuming
areas that probably will be first
introduced to AI or done AI to generate,
by hand, but, but by the automation.
And probably gradually you'll
see more and more such effects.
And of course it won't be how, uh, like
the, the full gen AI movie was show off
that, hey, we, we can do it in production.
That probably will still adopt with
some person And uh, yeah, we might
not even realize that hell of that was
actually from Gen AI and we liked it.
Chris: Yeah,
it is a really interesting pine.
And I think, uh, it'll be super
interesting just to bring you back to
the, the legal space where the, the
copyright line falls and a lot of this
stuff, I know Anthropic got the legal
case right now with the, all of the
book sites eaten up into its models.
And, um, yeah, time will, will, uh,
tell, I guess if we're watching, uh,
AI generated movies this time next
year with, uh, Tom Cruise in on.
Uh,
yeah.
But, um, that, that's it for today.
Mikel, thank you so much for coming
on and sharing your knowledge.
I, I've really enjoyed speaking to you
today and, and getting to know you, you
more and, um, Lex and some of the amazing
things that you're doing over there.
So thank you for, for joining us.
Michal Klos: hope there will be
some interesting insights from
that, that anybody can learn.
yeah.
Looking forward to, yeah, your series
of podcasts to, to next guests.
Chris: Thank you, Mikel.
Michal Klos: Thank you.
Chris: Cheers.
Thanks so much.