May 6, 2025 • 48 mins
https://www.genspark.ai/ https://genaimeetup.com/ Follow the podcast: https://podcast.genaimeetup.com/
Join hosts Shishank and Mark as they dive deep into the world of generative AI agents with Lenzoy Lin, Engineering Lead at Gens Park. Discover how this rapidly growing startup is revolutionizing productivity through their suite of AI agents - from their groundbreaking phone call agent to deep research tools and slide creation capabilities. Learn how Gens Park has grown from $10M to $22M in ARR in just one month, their approach to building reliable AI systems, and get a glimpse into the future of human-AI collaboration. Whether you're a tech enthusiast, entrepreneur, or AI professional, this episode offers valuable insights into one of 2025's most promising AI startups.
In this episode:
Lenzoy Lin's journey from Google to leading Gens Park's engineering team How Gens Park's mixture of agents approach solves complex tasks Behind the scenes of their phone call agent, deep research tools, and slide creation capabilities The technical challenges of building reliable AI agents at scale Gens Park's position in the competitive AI landscape and future roadmap
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
(01:00:00):
Hello, everybody. We have
a very exciting day today.
We are joined with Lenjoy Lin from
GenSpark, which is an exciting new
startup working on generalist AI agents
and advanced specific vertical agents,
like building slides, doing deep
research, and tons of
(01:00:20):
other really cool products.
But before we get started, this is the
Generative AI Meetup podcast with your
hosts. I'm Shashank, and I have Mark
Whitney, who's joining remotely while
he's traveling across the world.(...) And
we also have Lenjoy Lin,
the Eng lead from GenSpark.
So GenSpark is a really exciting company.
(01:00:43):
2025 has been founded as the year of
agents. And we are so excited to get a
company that is working on a truly
generalist AI agent and has a couple
focuses on specific AI agents. And
they've gained a lot of traction in the
last few months. They are apparently the
(01:01:04):
fastest growing startup. If you look at
ARR growth in the last month, they jumped
from 10 million to 22 million in ARR in
just a month with the launch of their
exciting new agents. They've raised, I
think, close to 100 million in the most
recent Series A funding, bringing their
valuation upwards of 500 million.
(01:01:26):
And they're in a really exciting space
with lots of other competitors like
Proplexity, which have massive, massive
unicorn valuation. So they have a really
exciting future ahead of them.
With that, let me ask our guest, Lenjoy
Lin, to maybe briefly introduce himself.
Give us a little bit
about your background.
(01:01:46):
You were in big tech for a long time.
You were an Uber TL engineering manager
at TikTok, Uber, Google, Houzz, and also
ByteDance,(...) leading ads and so many
other really interesting projects.(...)
So can you tell us a little bit about
yourself and your journey to GenSpark?
(01:02:08):
Yeah, thank you for the pleasure, Wenche.
I started my career at Google. So I spent
almost 10 years, 10 very happy years in
Google. And built something very exciting
in search and also ads area.
Later, I left Google for Pinterest.
(01:02:30):
At that moment, Pinterest was a startup.
So I got an opportunity to grow with the
company, grow with the team. And later I
left Pinterest to join another startup in
the house. And then I joined TikTok in
the US,(...) grow with the TikTok US
(01:02:50):
team, from very small to very big. So
because I did, now I'm doing something
even bigger. So it's a GenSpark. So we
want to solve the gap between human
creativity and tech execution(...) by the
agentic engine. So
(01:03:11):
what's GenSpark is doing?
Okay, very cool.(...) So you mentioned
that you are making the agentic engine.
(...) Can you maybe describe in your own
words, what GenSpark is and what the core
kind of problem you're solving that with
(01:03:33):
GenSpark, that existing
tools don't necessarily solve?
Yeah, we want to empower users by
democratizing knowledges and the tools.
So then we can solve the pain point, the
way you first mentioned. So in the real
(01:03:54):
world, tasks are often like multi-step,
requires switching across apps and tools
and data sources.(...) And FIM is very
powerful at the end of today. And then it
will become more and more powerful.
However, this pain point is not easy to
solve. And also users need to get
(01:04:16):
consistent and trustworthy
and the reliable data at the outcome. So
GenSpark solution in high level is try to
combine large-range model, which is a
foundation, and also the tools.(...)
Because large-range model need to talk to
the digital world. So including for
(01:04:37):
example, search and the gentle video, et
cetera.(...) And going forward,
agent can think, can act. So then in
GenSpark, we recently launched the phone
call. It can actually like interact with
people in the real
world, not only digital world.
(01:04:58):
So on top of that, we
also have the data sets.
So then we can make it more controllable
and reliable. Yeah, build these three
things, like a model and the toolings and
the data sets. Then we can build the
(01:05:18):
different AI experience in the future.
Yeah, very cool. So you mentioned the
phone call agent. That is a really cool
agent that I saw on your website. So my
understanding is that with the phone call
agent, you can go and then make a
restaurant reservation,
reserve hotel, something like that. One
(01:05:40):
question that I kind of had was, it
sounds incredibly technically complex,
and also difficult to test. So how do you
test that? Do you simulate calls or are
you actually calling restaurants? Do you
do something else?
Like how does that work?
We do both.
Yeah, we cannot send every single task to
(01:06:02):
the real restaurant. It's very, very
spammy. Sometimes it's not working well.
And we build our own testing environment
to make the call. To make sure that it's
working. And of course, before the
launch, we need to do the real test. Then
we actually do the phone call. And after
launch, we also collect the data to make
(01:06:24):
sure the user experience is good. And
then of course, there are some gaps we
need to improve. We need to improve. Then
on top of that, we gradually improve it.
Okay, very cool. I'm curious, who is kind
of the main type of business that you're
calling? Is it hotels? Is it restaurants?
(01:06:47):
Is it somebody else?
What do you use it for?
Right now, restaurant reservation is a
typical one because it's relatively
straightforward compared with other
tools. And then the task is clear because
it makes the revision and the corrector
(01:07:08):
transfer uses the requirement to the
language which restaurant can understand.
And then we can easily, we can actually
easy to avail the task
is successful or not.
That kind of reminds me of one of the
early demos at Google where they had a, I
(01:07:30):
forget what the name of the product was
called, which allowed you to call a
restaurant for you. And I think
eventually they incorporated some of
those voice features into the Pixel
phone, which screens spam calls for you,
maybe waits in the phone tree for you and
helps you navigate like your customer
(01:07:53):
service. If you call Comcast, you just
have to wait for a long time. Do you have
any plans to simplify those
kinds of use cases for users?
Maybe for the future.(...) Yeah, the case
you mentioned is in Google IO, I think in
2017 around,(...) I just remember the
concrete year. So then we will also
(01:08:14):
inspire by that moment and then we'll
make it as a real product in the launch
with your very exciting.
Nice.
So that sounds really exciting.
Maybe before jumping into some of the
other specific advanced agents,
can you explain your mixture of agents
(01:08:35):
approach? Like how much of the GenSpark
advantage is in maybe your secret sauce
versus just the underlying models just
getting better every month? Because I
think cursor's growth partially, I think,
is just attributed to cloth 3.7 on it
coming out and the underlying models
getting better at coding.
(01:08:56):
So I'm curious how much of your core
engineering is focused on building
tooling on top of these LMs?
For user-facing product, it depends on
the scenario. So because we built the
foundation for the mixture agent and then
depends on which task and how competitive
(01:09:18):
it is. And then for more, the task is
more complicated, requiring planning, et
cetera. Then mixture agent will be better
or like you need to check multiple
different sources and then to make the
result more reliable, then mixture agent
will be more suitable for these kind of
(01:09:39):
cases. Then meaning
that way user mixture agent
depends on the actual task we need.
Is there any particular tasks that you
find that this agent excels with with the
mixture of agents? And I wanna kind of
(01:09:59):
get into, on your website, it seemed like
your flagship agent was the super agent.
(...) Is there any
particular tasks that excels at?
I think there's a lot.(...) So if some
simple questions, so I think every model
will give you a similar answer. But if
(01:10:19):
you need to verify some information and
to organize some
information from multiple sources,
most of the time I think we need a
mixture agent to get the reliable
information, particularly for some like
marketer research.(...) So agent need to
(01:10:40):
pass lots of information. And at that
moment, the power of this
agent will be very clear.
Very cool, yeah. I gotta say that I tried
out your slide making agent and I gave it
just our websites. I wanted to make
something for our meetup. Actually, I was
(01:11:01):
hoping to use at the meetup on Thursday,
I wanted to be able to make a slideshow
for it. So I just gave our website and I
put it into your slide making agent. And
it actually came out with some really
good slides. I only gave it the website
and it was able to parse the website and
then put it into a nice four slideshow
(01:11:22):
presentation that was really succinct.
And it even added some information that
we didn't say. So on the slides it said,
oh, South Bay Gen AI, the South Bay's
most active meetup. And I was like, wow,
that's a really good way of putting it. I
wouldn't have thought it that way. So
kudos to you, because I think that the
agent was like incredibly good.
(01:11:42):
Glad to hear that.
Yeah, it's funny because I did the exact
same thing. So I'm actually switching
teams at Google. I'll be working on and,
I'm not sure if I can say, but the AI
Labs team. So working on AI agents that
can do really cool stuff for users. And I
wanted to say like a goodbye to my
current team. So I just put some
(01:12:03):
information into GenSpark, had to create
like a big presentation about my past, my
background, just through
a bunch of information.
And it made the entire slide presentation
in a couple of minutes, which is just
fantastic. I think the product seems
really polished too, apart from the
actual LLMs. It seems like you have a
(01:12:24):
visual editing tools where you can like
drag and drop, or like edit
the text, edit the images.
How much of your engineering effort is
spent on the product, like the software
engineering versus like the ML part,
building the agents or fine tuning the
models, if you guys do that.
(01:12:44):
With the growth of the capability of the
large language model, I think the
borderline between traditional engineer
versus like the ML
engineer will be blurred.
So meaning that we are aiming for
something in the future. We build the
(01:13:06):
foundation, we build our tech stack, will
be more suitable for the live language
model. That's why like we're able to
launch quite a few products this short
time, and then we combine, we try to
combine the traditional engineering work,
something like InFibre, pipeline,
whatever, it's the same. And also how to
(01:13:28):
unleash the power of the live language
model and then,(...) and we
can adjust them accordingly.
Once the live language model that
capability become better and better.
That's awesome to hear. I think
especially with the rise of these coding
IDs, maybe I'm sure your team uses some
(01:13:49):
of these cursor, Windsor
for maybe like cloud code.
So let's move on to one of your other
prominent product, which is the agentic
deep research. So this was a really cool
feature that I first started using with
the Gemini models.(...) And then later,
Chachi PT introduced one. I think our
(01:14:09):
perplexity also has one, Grok has one.
How is the GenSpark agentic deep research
different from the other competitors?
Before mentioning the GenSpark deep
research, I'd like to
mention another agent,
we launched last Q3. It's the factor
(01:14:32):
check. So at that moment,
we call it a cost check.
And behind the thing, this agent
and the deep research we launch later,
they share a lot of like a tech stack
because we prepare that for a long time.
And then we just pick up the right moment
to launch the deep research after tuning,
tuning the overall, we call it.
(01:14:56):
So for the cross-check or factor check,
that's the world first asyncholized
agent. I think these are one of the
innovation because if you have a team,
you ask a question to them. If they get
back to you, maybe tomorrow, I think
sometimes it's always very good. But for
(01:15:20):
agent, if it can return with your result
within about one hour, it will be much
better than interacting with the team. So
our invitation for that is about 10 to 30
minutes. So people cannot wait in front
(01:15:40):
of the computer for 30 minutes, just
waiting for the result. That's why we
make it asyncholized. That's why we built
the first world, we're the first
asyncholized agent.(...) And behind the
thing, agent will check all the website
and find the right information and mix
them together and make it more
(01:16:04):
human-friendly representative.(...) And
based on that, we found that the deep
research is a good opportunity.
And we launched the deep research agent
actually earlier than OpenAI,
earlier than a chat dbt feature because
we prepared that for a long time. And for
(01:16:26):
that one, we can imagine that
fundamentally, agent needs time to
consume and organize the information.
That's why we try to leverage the
computer power and then try to make it
more helpful and then
do the right balance.
So I saw that fact check agent that you
(01:16:48):
had and it looks really interesting. So I
am a bit curious on how you actually
verify what is true, especially for
things that are a little bit difficult to
fact check. So there's a lot of things
maybe in history or maybe something with
nutrition advice where a lot of people
may disagree. So some people might say,
(01:17:09):
oh, high carb diet is better, low carb
diet is better. What types of things do
you do to say like, hey, this is the
actual source of truth. Do you do
research, do you read articles? Do you
actually do independent research? What
does that actually look
like for the fact check agent?
Yeah, you're asking a very great question
and also a very hard question. So I think
(01:17:31):
even for human beings, it's very hard to
verify some factor. So human are relying
on the information on the web or on the
library or some other
literature just found.
So for the agent are similar. If the data
is accessible to human,
(01:17:54):
if it's digitized, it can be accessible
for agent. Then first part of the
question, we should be able to find the
relevant information.
That's the first accountability.
Second one, how to do the cross-check
thing for human being.(...) So A, Tokyo
(01:18:15):
something, B, Tokyo something. If they
are connected to each other, we need to
have some way to judge.(...) Then we have
the mixture agent.(...) And then one
agent will challenge another agent.
Then we can do something more readable.
(...) Third one,(...) in reality,
sometimes the factor or the truth is from
(01:18:37):
small group of people because the
majority of people
are wrong. For that one,
I would say it's hard to verify.(...) But
if we make something more reliable,
we put the source along with opinion or
the fact together, then some expert can
challenge it or can look into the actual
(01:19:00):
source, the original source
to verify it's a fact or not.
Lastly,
for some open-ended
question or its pure opinion,
we train the agent to let the user know
this is just an opinion.
We don't know like, I mean, two on the
(01:19:20):
fourth. Then we try to give the both
information to let him know. So then try
to make people say like, sometimes the
agent cannot know. So just let people
know what's so far agent can get.
Yeah, that seems like a really hard
(01:19:41):
problem to solve. Like you said, even for
humans, this is really challenging.
And I feel like it's a big responsibility
because now if we are replacing the
actual source of truth and this is where
people go instead of a search engine and
they trust the LLMs with whatever
response it gives us, I feel like it's a
(01:20:03):
big weight to carry and a big
responsibility to entrust companies with.
One really cool example that I remember
is, I don't know if you've ever used
consensus.app.(...) It's like an AI
powered scientific research tool which
allows you to search
across academic papers,
(01:20:25):
research literature, and then you can ask
it questions. And then it looks through
all of the academic literature and
compares the question that you're trying
to ask and gives you, okay, there's five
papers that are saying it's bad, but
there's like 10 papers that are saying
it's good. So maybe it's probably good
because of these facts.
And I really like how they kind of
(01:20:47):
explain that.(...) So I'm curious if you
guys present the sources that you use to
give your answer or your opinion,(...)
kind of like perplexity or consensus,
so people can then dig into the sources
and make their own opinion later.
Yeah, so that's how we designed this
(01:21:07):
product(...) because we just give people
the raw information, raw source, which we
can find in the internet. And also, you
also mentioned a very good point, like
the data source is so important.(...) And
that's why is a purpose I mentioned that
our foundation is a LAM
(01:21:28):
plus tooling plus dataset.
Reliable dataset is so important. And
then we need to spend lots of effort to
make sure dataset is reliable. And it's a
really hard problem for the industry.
Maybe going forward,
(...) it will be even harder.
So you mentioned a reliable datasets. Can
(01:21:50):
you share some of the datasets that you
use to train your LMs? Like are you using
internet data? Are you using Wikipedia?
Are you using your own dataset that you
created from somewhere? I know that
datasets are kind of the gold thing. I
think data is one of the most important
things for everything. Can you maybe
(01:22:10):
share some of the
places you get your data?
For the, yes.(...) For the public data, I
think it's acceptable for
everyone. I think it's not the,
and depends on the area. So for some
particular industry data, we need to
purchase some data.(...) So that's why
(01:22:33):
it's more like authentic data. It's more
valuable data. The other one, we have
user. User will do some search with an
e-caching with agent. Based on that, we
can also accumulate data by kind. I think
for these kind of things.
(01:22:55):
Okay, very cool. So I want to bring the
conversation back to the agentic deep
research that we mentioned before. And I
think that you kind of already answered
the question a little bit to say that
sort of the secret sauce and the main
value they provide over maybe other tools
is the ability to do the asynchronous
(01:23:16):
fact checking and having high quality
data. But I was wondering if there was
any other things that kind of set
GenSpark apart for the agentic deep
research where people would want to use
you over some of the other AI research
tools like Consensus, like Google search,
like Perplexity. Like what
kind of sets GenSpark apart?
(01:23:37):
I think it's the infinity risk.
So it's hard to tell because by time,
(...) everybody is working on data. And
for us, we spend a lot of effort on the
data and then we do the cross-check or
whatever. I think by time,
after we are accumulating
(01:24:00):
more and more data, and then it will make
it apart. The second
one, like for tooling part,
for cross-check, the tooling part is
relatively simple compared with our other
agents.(...) I think for that one,
by time, it will be more and more
efficient and more powerful.
(01:24:22):
I think for what make us very different
is not only the deep research. So like
the other agents will make us more
different. For data part, that's one hour
single source(...) data plus
misagent for the deep research.
So it seems like you talked a lot about
(01:24:44):
data, the asynchronous pipelines.
What are some good use cases for this
kind of tooling? It seems like you all
have a massive user base right now, at
least like massive growth.
What do you see people using your product
for and what are your
favorite use cases for these agents?
(01:25:07):
For me, I use deep search and cross-check
for some of the, I mean, use it a lot.
So for example, China news, there's
something straightforward. And also some
China something, some academic,
(01:25:27):
or like something about history.
We need to,
it's not the rumor, it's like some
history fact.(...) I think the
cross-check will tell you the conclusion
as well as some
sources. And for user side,
according to current understanding,
(01:25:49):
there's a huge range of use cases. It's
hard to summarize,
summarize what's the thing.
And again, so the same way are very
different from others is the agent,
the other agents. So not only the deep
research. So the recent launch for the
(01:26:10):
secret agent, people can use that for
the, for a generator video or summarize
the news and or even like use the South
Park to gen, to report the recent news.
So these kind of very creative things is
more important than the,(...) I will say
(01:26:31):
the traditional like deep research.
Yeah, of course, for the, as you
mentioned that is a website,
good website or good slides,
the format aside is very beautiful for
the new agent tools. And if we want to
get the reliable information, then behind
(01:26:52):
the scene, we use the deep research and
the cross-check as enabled to do the
foundation work. So for that the point is
you can imagine that this is a
productivity tools for user and for us.
After the cloud information, then we can
represent the content to
more like user-friendly.
(01:27:14):
Yeah, that's actually really interesting
that you mentioned about the video
generation. I'm curious what types of
videos people are creating. Are they
creating entire YouTube videos from
scratch? Are they just creating B-roll?
What types of things are people making
with the video generation?
Right now we cannot create long form
(01:27:36):
video. So only like a five second or 10
seconds like that. So people can do some
work around to like build a bunch of the
video clips then to combine them
together. And also people can create a
website. So with multiple video over
there. So these kinds of things can
(01:27:57):
speak all the information representative.
This is also good away. So not only the
pure video. And video plus audio. And
this is another thing,
say like a South Park
for news example.
(01:28:17):
If only video, only script is not like a,
fun enough, right? So create audio, put
them together. Then that's making the
information more
attractive, more user-friendly.
So I'm curious,(...) what kind of video
models do you use? Or is it similar to
your other approach where you use a
(01:28:38):
mixture of video models? And what kind of
orchestration do you do on
top of the existing models?
For video it's different from the mixture
agent for chart or for color check.
Because right now no easy way to, I mean
for single video, I mean leveraging
(01:28:59):
multiple model to combine video together.
(...) So if we go to the website, go to
the video generation, we list a bunch of
the video generation models.
These are the tools we are used for
genetic video. And then depends on which
style. We integrate most of the top video
(01:29:22):
generation tools already.(...) However,
is a real one, depends on your scenario.
Some video, some model is good at making
some funny video or make it more real.
The other one is more about,
like I mean the style is quite different.
We choose, we played with video a lot.
(01:29:44):
That's why we have some good science
about which kind of problem, which kind
of task we want to call which model.
We want to make a
limited number of the choice
to show user, then user can give
feedback. If we change too many videos
for user, it's not a good experience.
(01:30:09):
So that's interesting that you mentioned
about routing to a particular agent or
model. What, I'm just kind of curious,
and you don't have to answer this if you
don't want to, but I'm sort of curious as
to what types of heuristics that you use
to figure out which
agent or model to route to.
Oh, that's very complicated. So some of
(01:30:30):
those things are based on the past
experience. So we know, we have some
rough idea about which model is good at
which scenario for the video, same for
the image. And based on that, we combine
the things with the rules and also with
the larger model. The model will help to
(01:30:50):
the planning. So we ruin the model based
on the solution together. Then we can
pick up one. And sometimes we also made a
mistake. Then based on some actual
results, if we think it's not good, so we
have some good way to come back and then
iterate our rule or our code. Yeah,
(01:31:11):
that's by time. We want to make the
product right.(...) Lots of iteration,
lots of testing, lots of work behind.
Yeah, I guess I have a follow-up question
regarding how you build the heuristics.
(...) You mentioned a lot of different
strategies, but maybe high level.(...) Do
you rely more on manual heuristics where
(01:31:31):
you observe this model is better at one
task and another model is better at
another task? Or do you try to, you know,
train a model to learn that automatically
at scale with lots of
data and lots of examples?
Do both.
So first of all, in the beginning, of
course, you need to rely on people.(...)
Then by time, we need to build some good
(01:31:55):
strategy, have some
internal feedback look.
At least we know something, I mean,
partially know something good or bad, and
then we can iterate, iterate in the proof
and then some way to measure. So just
something high level.(...) Lots of
detailed work behind. I think this is a
very committed performance to solve. And
(01:32:16):
then it's hard to tell if it's
well-solved or not, but we just make it
product right, make it more practical.
Very cool.
So kind of switching the topic a little
bit,(...) I saw you have another agent
that we haven't really touched on yet,
which is the agentic data tables. Can you
maybe briefly describe
what that offers users?
(01:32:39):
If you try some query,
you'll see like, for example,
if you want to understand the top,
like top brand in some industry, and then
the ideal format, the outcome for the
result is the table. And then you can
even define a few dimension.
(01:33:02):
So then the table is something,
the good way to view the overview, and
also we put the source of the data.(...)
Then you can do the check by yourself.
(...) That's one thing.
If no data, it will table. Another
example is like,(...) for example, who is
the key players in Trump team? There is a
(01:33:27):
presidential election.
So there's nothing about data, but it can
give you the people name, and then their
role, and all the other information. So
meaning that table is very powerful.
People can present the data in a very
abstract and brief way. People can easily
to get information and they're easier to
(01:33:47):
compare. So this is the thing we built in
last Q4. Behind the tech stack, it's
still similar with the CodeShack and the
Deep Research. That's why we are able to
launch it and iterate them together.
(01:34:08):
Maybe taking a step back a little bit, I
think we covered most of the agents that
you guys offer. Most of the advanced
agents, we talked a little bit about the
basic agents,(...) some of
the mixture of agent systems.
If we take a step back,
what are some of the
biggest challenges your team faces
in building these agents in a reliable,
(01:34:28):
trustworthy way and at scale? Is it like
the conceptual challenges where you have
to create these ideas from scratch, or is
it just like technical software
engineering challenges where you have to
integrate a lot of different tools and
make sure the evals are right at scale?
Challenges, that's a lot. Because if we
(01:34:52):
want to build the product right,
not only demo, it's
actually solve the problem.
There's lots of detail thing we need to
handle. Lots of common cases we need to
debug, we need to cover, we need to have
some other solution.
For mixture agent,
(01:35:13):
the architecture is a little bit
complicated. So how to combine them
together, that's one typical sense.
I'd like to mention another one, another
different challenge as well, to give you
better sense about the kind of challenge.
Say for example,(...) the model's context
(01:35:34):
window,(...) there's a limit over there.
And by time, it will grow.
And in practice, if we solve some very
complicated problem,
the window is still a limit.
Then we can either optimize the data, try
to fit the window, or like we just leave
(01:35:55):
it as is, or some other work around. And
then we can predict the content window
will be longer. Then how to leverage the
end resource to the right place.
So I think this is another challenge we
need to face. So by time, that's to be
more practical. So make things work, and
then make it the
future tech stack scalable.
(01:36:22):
That's a really interesting point. You
mentioned the context window might be
small now, but it
might grow in the future.
So how do you, as a high level product
planning(...) for the roadmap for the
future,(...) how much of your effort is
spent solving the problems and
limitations of the models today versus
(01:36:43):
just waiting for the models to get
better, waiting for them to get better
reasoning skills or longer
context window and so on?
In practice, we do both, depends.(...)
Yeah, because there's no thing for answer
to say yes or no, or like what's the
right way.(...) We just need to be more
(01:37:04):
adaptive, and then be
more careful about the sound.
Sound good, I mean sounds
important to technically see.
So to answer your question,
we have lots of different products, and
they're different. That's why like we
(01:37:26):
depend on which product. Some products we
need to optimize the data, no matter
what. And some product we just leave the
window short for now, and then we imagine
in the near future, it will be big, and
then we don't need to see end of the
organization too early.(...) So
(01:37:47):
combination, combination of
both, depends on the product.
So I'm curious, you mentioned that there
was a lot of the issues around the
limited context window as well. Can you
maybe share some of the techniques that
you use to deal with it, and assuming
that the models, I mean, we all think
(01:38:09):
that the models will get better, but I'm
curious what kind of specific techniques
you're able to use to deal with the
limited context window. So I know one
technique that people will use is they
will summarize lots of the previous
conversation, and then throw that into
the bottom of the context. They'll use
RAG, I don't know, what kinds of things
(01:38:30):
do you use to kind of handle
the limited context window?
Yeah, we're doing something similar.
Yeah, of course the summarizer would have
all just a slow sound of data, or like we
have some judges, the data is already not
that important or whatever.
When I mentioned the content window is
(01:38:51):
like, it's just one example we need to
deal with, and how to make the tech stack
more scalable for the future, because the
model will improve,(...) and the other
part will improve as well. That's why
like, pick up as a writer, try our best
to pick up the right tech stack, how to
(01:39:13):
have the good engineer and the good model
can work together. So that's the big
question the engineering team is facing.
Yeah, and of course, content
window is just one example.
Yeah, speaking of the engineering team,
it seems like you guys are working on a
lot,(...) a lot of different products, a
(01:39:33):
lot of different challenges.
How big is this team? How many people are
working on it right now? And how many
people are focused on like the research
part? How many are focused on the
engineering product? And are you looking
to hire more people
in any specific areas?
We are a very smart team right now. So
the total company is around 20 people. Oh
(01:39:56):
my God. Many people are engineer.
So then, yeah, we have some people
working with, I mean, very good at
modeling part. They have lots of
experience.(...) And lots of stack, stack
means like back end, the modeling and the
front end or whatever. Because it is a
smart team. Everybody can do everything.
(01:40:19):
Are you trying to hire for any specific
open roles right now?
We are not a,(...) like I said, we have a
lot of people, but we
are slowly, slowly hiring.
That's, yeah. So we talked a lot about
(01:40:41):
all the different
products, the company size.
I had a couple of follow-up questions for
the future.(...) But before that,(...)
how do you see yourself in this crowded
space?(...) Because we have a lot of
competition from China. I think the first
generalist AI agent was Manus.
(01:41:03):
We have a lot of new specific vertical
agents. I think the industry leaders like
Hanma, for example, they
have their own AI slides agent.
There is competition from deep research
agents from like the fan companies.
How do you see the position of like
smaller startups like
Genspark in this massive landscape?
(01:41:26):
I think the market is huge. And then
going forward, the AI experience will be
very different. So then it's very, yeah,
it's very competitive. Yeah, of course. I
think the marketer is even bigger.
If you compare the agents, I think it's
(01:41:49):
very early stage.(...) For us, it's about
only 10% for the launch already,
according to our product vision. So
that's why we are very confident to build
something that's bigger. And if we are
talking about some existing bigger names,
like Google or Chai GPT,(...) of course
(01:42:09):
they are very good. They are very, very
big already.(...) Google is more or less
like cut to the search engine and the
search box we have. This is because the
user experience itself.
Chai GPT is a very, very great product
for the chart interface. For us, we have
a more complicated scenario. So I can
(01:42:32):
take a slice AI, for example.
This interaction cannot be simply sold by
either search box or the chart box.(...)
So that's why we feel like going forward,
the scenario is different. We want to
empower every user, empower everyone in
(01:42:54):
their daily work and daily life. That's
why the scenario is more complicated. So
then let's worry about the user case
forever. Marketing is huge. In the end,
the future AI experience will be quite
different. Your home
without the answer question.
No, I absolutely agree.
I think it's a big market and it's still
so early that I don't think we know what
(01:43:14):
the future will look like. And it's
really exciting to see startups because I
think in a smaller company, you're able
to move much faster.
I mean, I work at Google and I love
massive companies because they are able
to do things at scale, but we also move a
lot more slowly. We're not able to
iterate as quickly. And I'm really
excited that startups like yours exist
(01:43:36):
because it is really exciting to use the
AI Slides agent today and share it with
my friends and colleagues.
But I am curious, there
are other agentic companies.
Do you foresee interacting with other
agents outside of your company? Because I
think maybe we can see a future with
(01:43:57):
multiple different kinds of agents and
agents interacting with each other
without human intervention.
Do you have any plans or are you working
on features to use like the, I think
Google announced a
protocol for agent to agents,
some standard tooling and like the MCP
protocol to allow other
(01:44:19):
people to connect to your agents.
I would say the depends on user scenario.
So we are consumer facing AI agent.(...)
So for daily life, daily work, if it's a
scenario where we want to build something
valuable to empower users.
(01:44:42):
Do you have any plans for business use
cases, B2B products?
I would say, so the answer is not now.
But going forward as the
growth of the AI product,
the bottom line between work and life and
(01:45:03):
the internal scenario will be blah.
So that's why hard to tell B2B if we're
enterprise facing but for
user scenario, it will be shared.
Very interesting, yeah.
So I don't know if you can see my video
anymore. I think I lost it. But anyway, I
still hear you. Okay, that's fine.(...)
(01:45:24):
So one kind of question that I have is
following the $60 million of funding,
(...) taking this money, what kind of are
your key priorities for your growth? And
then if you can share maybe a little bit
about your product roadmap.
So we are doing the BOD,
(01:45:48):
product and lead growth.
We don't do content distribution or
marketing as of today. So all the growth
are from organic growth. Then building
our products is the top priority. So as I
mentioned before, right now, according to
our product vision, only 10% progress at
(01:46:11):
this moment. So the more resource will be
put to build a useful
valuable product for users.
Are there any specific new advanced
agents that you're looking forward to
having in the next few
months or year or two?
(01:46:33):
For us, we are going to launch another
one.(...) Stay tuned, it's about one to
two weeks, within one week, we're going
to have another one launch.
That's really exciting, the speed at
which your startup is iterating, not in
the order of like years or months,
everything is happening in weeks and
days. That's incredible.
(01:46:56):
Yeah, I mean, I think that's everything
we had. I really appreciate
your time and thoughtful answers.
To all of our listeners, check out
GenSpark. It is really easy to use. And
the free plan offers plenty of usage.
The paid plan is fantastic. It offers you
higher limits and I think more capable
(01:47:18):
models too, if I'm correct.
More quotas.
More quotas, yeah. More credits.
It's like a free model. You show people
how good this thing is and people just
want to use more. That's kind of what I
felt with the AI Slides agent.
Yeah, again, thank you so much. But I
wanted to leave a couple of minutes for
(01:47:38):
you. If there's anything you want to
announce or anything you want to share,
(...) anything you'd
like to tell our audience.
For more information, so people can
follow our x.com. In just search,
GenSpark AI, so you can find the account.
So most the latest information, we
(01:47:59):
usually post to our Twitter account.
That's awesome. Yeah, we'll share the
links in the podcast notes and you guys
can follow the Twitter account, the x
account. And also if you're here in the
Bay Area this week on May 8th, we'll be
having GenSpark and a couple other
(01:48:21):
agentic startups too.(...) A fully
integrated coding ID and a browser agent.
So it'll be like a full on agentic
product demo and a tech talk show in Palo
Alto from 6 to 9 p.m. And yeah, I'm
excited to see you all here if you can
make it. And thank you so much again,
(01:48:42):
Linjoy, for joining
us, this was fantastic.
Thank you for having me. Looking forward
to the meetup. I will show you more
videos. Awesome. And
the demos over there.
Excited to see you in person.
All right, thank you everyone.
Bye-bye.
Hello, everybody. We have
a very exciting day today.
We are joined with Lenjoy Lin from
GenSpark, which is an exciting new
startup working on generalist AI agents
and advanced specific vertical agents,
like building slides, doing deep
research, and tons of
(01:00:20):
other really cool products.
But before we get started, this is the
Generative AI Meetup podcast with your
hosts. I'm Shashank, and I have Mark
Whitney, who's joining remotely while
he's traveling across the world.(...) And
we also have Lenjoy Lin,
the Eng lead from GenSpark.
So GenSpark is a really exciting company.
(01:00:43):
2025 has been founded as the year of
agents. And we are so excited to get a
company that is working on a truly
generalist AI agent and has a couple
focuses on specific AI agents. And
they've gained a lot of traction in the
last few months. They are apparently the
(01:01:04):
fastest growing startup. If you look at
ARR growth in the last month, they jumped
from 10 million to 22 million in ARR in
just a month with the launch of their
exciting new agents. They've raised, I
think, close to 100 million in the most
recent Series A funding, bringing their
valuation upwards of 500 million.
(01:01:26):
And they're in a really exciting space
with lots of other competitors like
Proplexity, which have massive, massive
unicorn valuation. So they have a really
exciting future ahead of them.
With that, let me ask our guest, Lenjoy
Lin, to maybe briefly introduce himself.
Give us a little bit
about your background.
(01:01:46):
You were in big tech for a long time.
You were an Uber TL engineering manager
at TikTok, Uber, Google, Houzz, and also
ByteDance,(...) leading ads and so many
other really interesting projects.(...)
So can you tell us a little bit about
yourself and your journey to GenSpark?
(01:02:08):
Yeah, thank you for the pleasure, Wenche.
I started my career at Google. So I spent
almost 10 years, 10 very happy years in
Google. And built something very exciting
in search and also ads area.
Later, I left Google for Pinterest.
(01:02:30):
At that moment, Pinterest was a startup.
So I got an opportunity to grow with the
company, grow with the team. And later I
left Pinterest to join another startup in
the house. And then I joined TikTok in
the US,(...) grow with the TikTok US
(01:02:50):
team, from very small to very big. So
because I did, now I'm doing something
even bigger. So it's a GenSpark. So we
want to solve the gap between human
creativity and tech execution(...) by the
agentic engine. So
(01:03:11):
what's GenSpark is doing?
Okay, very cool.(...) So you mentioned
that you are making the agentic engine.
(...) Can you maybe describe in your own
words, what GenSpark is and what the core
kind of problem you're solving that with
(01:03:33):
GenSpark, that existing
tools don't necessarily solve?
Yeah, we want to empower users by
democratizing knowledges and the tools.
So then we can solve the pain point, the
way you first mentioned. So in the real
(01:03:54):
world, tasks are often like multi-step,
requires switching across apps and tools
and data sources.(...) And FIM is very
powerful at the end of today. And then it
will become more and more powerful.
However, this pain point is not easy to
solve. And also users need to get
(01:04:16):
consistent and trustworthy
and the reliable data at the outcome. So
GenSpark solution in high level is try to
combine large-range model, which is a
foundation, and also the tools.(...)
Because large-range model need to talk to
the digital world. So including for
(01:04:37):
example, search and the gentle video, et
cetera.(...) And going forward,
agent can think, can act. So then in
GenSpark, we recently launched the phone
call. It can actually like interact with
people in the real
world, not only digital world.
(01:04:58):
So on top of that, we
also have the data sets.
So then we can make it more controllable
and reliable. Yeah, build these three
things, like a model and the toolings and
the data sets. Then we can build the
(01:05:18):
different AI experience in the future.
Yeah, very cool. So you mentioned the
phone call agent. That is a really cool
agent that I saw on your website. So my
understanding is that with the phone call
agent, you can go and then make a
restaurant reservation,
reserve hotel, something like that. One
(01:05:40):
question that I kind of had was, it
sounds incredibly technically complex,
and also difficult to test. So how do you
test that? Do you simulate calls or are
you actually calling restaurants? Do you
do something else?
Like how does that work?
We do both.
Yeah, we cannot send every single task to
(01:06:02):
the real restaurant. It's very, very
spammy. Sometimes it's not working well.
And we build our own testing environment
to make the call. To make sure that it's
working. And of course, before the
launch, we need to do the real test. Then
we actually do the phone call. And after
launch, we also collect the data to make
(01:06:24):
sure the user experience is good. And
then of course, there are some gaps we
need to improve. We need to improve. Then
on top of that, we gradually improve it.
Okay, very cool. I'm curious, who is kind
of the main type of business that you're
calling? Is it hotels? Is it restaurants?
(01:06:47):
Is it somebody else?
What do you use it for?
Right now, restaurant reservation is a
typical one because it's relatively
straightforward compared with other
tools. And then the task is clear because
it makes the revision and the corrector
(01:07:08):
transfer uses the requirement to the
language which restaurant can understand.
And then we can easily, we can actually
easy to avail the task
is successful or not.
That kind of reminds me of one of the
early demos at Google where they had a, I
(01:07:30):
forget what the name of the product was
called, which allowed you to call a
restaurant for you. And I think
eventually they incorporated some of
those voice features into the Pixel
phone, which screens spam calls for you,
maybe waits in the phone tree for you and
helps you navigate like your customer
(01:07:53):
service. If you call Comcast, you just
have to wait for a long time. Do you have
any plans to simplify those
kinds of use cases for users?
Maybe for the future.(...) Yeah, the case
you mentioned is in Google IO, I think in
2017 around,(...) I just remember the
concrete year. So then we will also
(01:08:14):
inspire by that moment and then we'll
make it as a real product in the launch
with your very exciting.
Nice.
So that sounds really exciting.
Maybe before jumping into some of the
other specific advanced agents,
can you explain your mixture of agents
(01:08:35):
approach? Like how much of the GenSpark
advantage is in maybe your secret sauce
versus just the underlying models just
getting better every month? Because I
think cursor's growth partially, I think,
is just attributed to cloth 3.7 on it
coming out and the underlying models
getting better at coding.
(01:08:56):
So I'm curious how much of your core
engineering is focused on building
tooling on top of these LMs?
For user-facing product, it depends on
the scenario. So because we built the
foundation for the mixture agent and then
depends on which task and how competitive
(01:09:18):
it is. And then for more, the task is
more complicated, requiring planning, et
cetera. Then mixture agent will be better
or like you need to check multiple
different sources and then to make the
result more reliable, then mixture agent
will be more suitable for these kind of
(01:09:39):
cases. Then meaning
that way user mixture agent
depends on the actual task we need.
Is there any particular tasks that you
find that this agent excels with with the
mixture of agents? And I wanna kind of
(01:09:59):
get into, on your website, it seemed like
your flagship agent was the super agent.
(...) Is there any
particular tasks that excels at?
I think there's a lot.(...) So if some
simple questions, so I think every model
will give you a similar answer. But if
(01:10:19):
you need to verify some information and
to organize some
information from multiple sources,
most of the time I think we need a
mixture agent to get the reliable
information, particularly for some like
marketer research.(...) So agent need to
(01:10:40):
pass lots of information. And at that
moment, the power of this
agent will be very clear.
Very cool, yeah. I gotta say that I tried
out your slide making agent and I gave it
just our websites. I wanted to make
something for our meetup. Actually, I was
(01:11:01):
hoping to use at the meetup on Thursday,
I wanted to be able to make a slideshow
for it. So I just gave our website and I
put it into your slide making agent. And
it actually came out with some really
good slides. I only gave it the website
and it was able to parse the website and
then put it into a nice four slideshow
(01:11:22):
presentation that was really succinct.
And it even added some information that
we didn't say. So on the slides it said,
oh, South Bay Gen AI, the South Bay's
most active meetup. And I was like, wow,
that's a really good way of putting it. I
wouldn't have thought it that way. So
kudos to you, because I think that the
agent was like incredibly good.
(01:11:42):
Glad to hear that.
Yeah, it's funny because I did the exact
same thing. So I'm actually switching
teams at Google. I'll be working on and,
I'm not sure if I can say, but the AI
Labs team. So working on AI agents that
can do really cool stuff for users. And I
wanted to say like a goodbye to my
current team. So I just put some
(01:12:03):
information into GenSpark, had to create
like a big presentation about my past, my
background, just through
a bunch of information.
And it made the entire slide presentation
in a couple of minutes, which is just
fantastic. I think the product seems
really polished too, apart from the
actual LLMs. It seems like you have a
(01:12:24):
visual editing tools where you can like
drag and drop, or like edit
the text, edit the images.
How much of your engineering effort is
spent on the product, like the software
engineering versus like the ML part,
building the agents or fine tuning the
models, if you guys do that.
(01:12:44):
With the growth of the capability of the
large language model, I think the
borderline between traditional engineer
versus like the ML
engineer will be blurred.
So meaning that we are aiming for
something in the future. We build the
(01:13:06):
foundation, we build our tech stack, will
be more suitable for the live language
model. That's why like we're able to
launch quite a few products this short
time, and then we combine, we try to
combine the traditional engineering work,
something like InFibre, pipeline,
whatever, it's the same. And also how to
(01:13:28):
unleash the power of the live language
model and then,(...) and we
can adjust them accordingly.
Once the live language model that
capability become better and better.
That's awesome to hear. I think
especially with the rise of these coding
IDs, maybe I'm sure your team uses some
(01:13:49):
of these cursor, Windsor
for maybe like cloud code.
So let's move on to one of your other
prominent product, which is the agentic
deep research. So this was a really cool
feature that I first started using with
the Gemini models.(...) And then later,
Chachi PT introduced one. I think our
(01:14:09):
perplexity also has one, Grok has one.
How is the GenSpark agentic deep research
different from the other competitors?
Before mentioning the GenSpark deep
research, I'd like to
mention another agent,
we launched last Q3. It's the factor
(01:14:32):
check. So at that moment,
we call it a cost check.
And behind the thing, this agent
and the deep research we launch later,
they share a lot of like a tech stack
because we prepare that for a long time.
And then we just pick up the right moment
to launch the deep research after tuning,
tuning the overall, we call it.
(01:14:56):
So for the cross-check or factor check,
that's the world first asyncholized
agent. I think these are one of the
innovation because if you have a team,
you ask a question to them. If they get
back to you, maybe tomorrow, I think
sometimes it's always very good. But for
(01:15:20):
agent, if it can return with your result
within about one hour, it will be much
better than interacting with the team. So
our invitation for that is about 10 to 30
minutes. So people cannot wait in front
(01:15:40):
of the computer for 30 minutes, just
waiting for the result. That's why we
make it asyncholized. That's why we built
the first world, we're the first
asyncholized agent.(...) And behind the
thing, agent will check all the website
and find the right information and mix
them together and make it more
(01:16:04):
human-friendly representative.(...) And
based on that, we found that the deep
research is a good opportunity.
And we launched the deep research agent
actually earlier than OpenAI,
earlier than a chat dbt feature because
we prepared that for a long time. And for
(01:16:26):
that one, we can imagine that
fundamentally, agent needs time to
consume and organize the information.
That's why we try to leverage the
computer power and then try to make it
more helpful and then
do the right balance.
So I saw that fact check agent that you
(01:16:48):
had and it looks really interesting. So I
am a bit curious on how you actually
verify what is true, especially for
things that are a little bit difficult to
fact check. So there's a lot of things
maybe in history or maybe something with
nutrition advice where a lot of people
may disagree. So some people might say,
(01:17:09):
oh, high carb diet is better, low carb
diet is better. What types of things do
you do to say like, hey, this is the
actual source of truth. Do you do
research, do you read articles? Do you
actually do independent research? What
does that actually look
like for the fact check agent?
Yeah, you're asking a very great question
and also a very hard question. So I think
(01:17:31):
even for human beings, it's very hard to
verify some factor. So human are relying
on the information on the web or on the
library or some other
literature just found.
So for the agent are similar. If the data
is accessible to human,
(01:17:54):
if it's digitized, it can be accessible
for agent. Then first part of the
question, we should be able to find the
relevant information.
That's the first accountability.
Second one, how to do the cross-check
thing for human being.(...) So A, Tokyo
(01:18:15):
something, B, Tokyo something. If they
are connected to each other, we need to
have some way to judge.(...) Then we have
the mixture agent.(...) And then one
agent will challenge another agent.
Then we can do something more readable.
(...) Third one,(...) in reality,
sometimes the factor or the truth is from
(01:18:37):
small group of people because the
majority of people
are wrong. For that one,
I would say it's hard to verify.(...) But
if we make something more reliable,
we put the source along with opinion or
the fact together, then some expert can
challenge it or can look into the actual
(01:19:00):
source, the original source
to verify it's a fact or not.
Lastly,
for some open-ended
question or its pure opinion,
we train the agent to let the user know
this is just an opinion.
We don't know like, I mean, two on the
(01:19:20):
fourth. Then we try to give the both
information to let him know. So then try
to make people say like, sometimes the
agent cannot know. So just let people
know what's so far agent can get.
Yeah, that seems like a really hard
(01:19:41):
problem to solve. Like you said, even for
humans, this is really challenging.
And I feel like it's a big responsibility
because now if we are replacing the
actual source of truth and this is where
people go instead of a search engine and
they trust the LLMs with whatever
response it gives us, I feel like it's a
(01:20:03):
big weight to carry and a big
responsibility to entrust companies with.
One really cool example that I remember
is, I don't know if you've ever used
consensus.app.(...) It's like an AI
powered scientific research tool which
allows you to search
across academic papers,
(01:20:25):
research literature, and then you can ask
it questions. And then it looks through
all of the academic literature and
compares the question that you're trying
to ask and gives you, okay, there's five
papers that are saying it's bad, but
there's like 10 papers that are saying
it's good. So maybe it's probably good
because of these facts.
And I really like how they kind of
(01:20:47):
explain that.(...) So I'm curious if you
guys present the sources that you use to
give your answer or your opinion,(...)
kind of like perplexity or consensus,
so people can then dig into the sources
and make their own opinion later.
Yeah, so that's how we designed this
(01:21:07):
product(...) because we just give people
the raw information, raw source, which we
can find in the internet. And also, you
also mentioned a very good point, like
the data source is so important.(...) And
that's why is a purpose I mentioned that
our foundation is a LAM
(01:21:28):
plus tooling plus dataset.
Reliable dataset is so important. And
then we need to spend lots of effort to
make sure dataset is reliable. And it's a
really hard problem for the industry.
Maybe going forward,
(...) it will be even harder.
So you mentioned a reliable datasets. Can
(01:21:50):
you share some of the datasets that you
use to train your LMs? Like are you using
internet data? Are you using Wikipedia?
Are you using your own dataset that you
created from somewhere? I know that
datasets are kind of the gold thing. I
think data is one of the most important
things for everything. Can you maybe
(01:22:10):
share some of the
places you get your data?
For the, yes.(...) For the public data, I
think it's acceptable for
everyone. I think it's not the,
and depends on the area. So for some
particular industry data, we need to
purchase some data.(...) So that's why
(01:22:33):
it's more like authentic data. It's more
valuable data. The other one, we have
user. User will do some search with an
e-caching with agent. Based on that, we
can also accumulate data by kind. I think
for these kind of things.
(01:22:55):
Okay, very cool. So I want to bring the
conversation back to the agentic deep
research that we mentioned before. And I
think that you kind of already answered
the question a little bit to say that
sort of the secret sauce and the main
value they provide over maybe other tools
is the ability to do the asynchronous
(01:23:16):
fact checking and having high quality
data. But I was wondering if there was
any other things that kind of set
GenSpark apart for the agentic deep
research where people would want to use
you over some of the other AI research
tools like Consensus, like Google search,
like Perplexity. Like what
kind of sets GenSpark apart?
(01:23:37):
I think it's the infinity risk.
So it's hard to tell because by time,
(...) everybody is working on data. And
for us, we spend a lot of effort on the
data and then we do the cross-check or
whatever. I think by time,
after we are accumulating
(01:24:00):
more and more data, and then it will make
it apart. The second
one, like for tooling part,
for cross-check, the tooling part is
relatively simple compared with our other
agents.(...) I think for that one,
by time, it will be more and more
efficient and more powerful.
(01:24:22):
I think for what make us very different
is not only the deep research. So like
the other agents will make us more
different. For data part, that's one hour
single source(...) data plus
misagent for the deep research.
So it seems like you talked a lot about
(01:24:44):
data, the asynchronous pipelines.
What are some good use cases for this
kind of tooling? It seems like you all
have a massive user base right now, at
least like massive growth.
What do you see people using your product
for and what are your
favorite use cases for these agents?
(01:25:07):
For me, I use deep search and cross-check
for some of the, I mean, use it a lot.
So for example, China news, there's
something straightforward. And also some
China something, some academic,
(01:25:27):
or like something about history.
We need to,
it's not the rumor, it's like some
history fact.(...) I think the
cross-check will tell you the conclusion
as well as some
sources. And for user side,
according to current understanding,
(01:25:49):
there's a huge range of use cases. It's
hard to summarize,
summarize what's the thing.
And again, so the same way are very
different from others is the agent,
the other agents. So not only the deep
research. So the recent launch for the
(01:26:10):
secret agent, people can use that for
the, for a generator video or summarize
the news and or even like use the South
Park to gen, to report the recent news.
So these kind of very creative things is
more important than the,(...) I will say
(01:26:31):
the traditional like deep research.
Yeah, of course, for the, as you
mentioned that is a website,
good website or good slides,
the format aside is very beautiful for
the new agent tools. And if we want to
get the reliable information, then behind
(01:26:52):
the scene, we use the deep research and
the cross-check as enabled to do the
foundation work. So for that the point is
you can imagine that this is a
productivity tools for user and for us.
After the cloud information, then we can
represent the content to
more like user-friendly.
(01:27:14):
Yeah, that's actually really interesting
that you mentioned about the video
generation. I'm curious what types of
videos people are creating. Are they
creating entire YouTube videos from
scratch? Are they just creating B-roll?
What types of things are people making
with the video generation?
Right now we cannot create long form
(01:27:36):
video. So only like a five second or 10
seconds like that. So people can do some
work around to like build a bunch of the
video clips then to combine them
together. And also people can create a
website. So with multiple video over
there. So these kinds of things can
(01:27:57):
speak all the information representative.
This is also good away. So not only the
pure video. And video plus audio. And
this is another thing,
say like a South Park
for news example.
(01:28:17):
If only video, only script is not like a,
fun enough, right? So create audio, put
them together. Then that's making the
information more
attractive, more user-friendly.
So I'm curious,(...) what kind of video
models do you use? Or is it similar to
your other approach where you use a
(01:28:38):
mixture of video models? And what kind of
orchestration do you do on
top of the existing models?
For video it's different from the mixture
agent for chart or for color check.
Because right now no easy way to, I mean
for single video, I mean leveraging
(01:28:59):
multiple model to combine video together.
(...) So if we go to the website, go to
the video generation, we list a bunch of
the video generation models.
These are the tools we are used for
genetic video. And then depends on which
style. We integrate most of the top video
(01:29:22):
generation tools already.(...) However,
is a real one, depends on your scenario.
Some video, some model is good at making
some funny video or make it more real.
The other one is more about,
like I mean the style is quite different.
We choose, we played with video a lot.
(01:29:44):
That's why we have some good science
about which kind of problem, which kind
of task we want to call which model.
We want to make a
limited number of the choice
to show user, then user can give
feedback. If we change too many videos
for user, it's not a good experience.
(01:30:09):
So that's interesting that you mentioned
about routing to a particular agent or
model. What, I'm just kind of curious,
and you don't have to answer this if you
don't want to, but I'm sort of curious as
to what types of heuristics that you use
to figure out which
agent or model to route to.
Oh, that's very complicated. So some of
(01:30:30):
those things are based on the past
experience. So we know, we have some
rough idea about which model is good at
which scenario for the video, same for
the image. And based on that, we combine
the things with the rules and also with
the larger model. The model will help to
(01:30:50):
the planning. So we ruin the model based
on the solution together. Then we can
pick up one. And sometimes we also made a
mistake. Then based on some actual
results, if we think it's not good, so we
have some good way to come back and then
iterate our rule or our code. Yeah,
(01:31:11):
that's by time. We want to make the
product right.(...) Lots of iteration,
lots of testing, lots of work behind.
Yeah, I guess I have a follow-up question
regarding how you build the heuristics.
(...) You mentioned a lot of different
strategies, but maybe high level.(...) Do
you rely more on manual heuristics where
(01:31:31):
you observe this model is better at one
task and another model is better at
another task? Or do you try to, you know,
train a model to learn that automatically
at scale with lots of
data and lots of examples?
Do both.
So first of all, in the beginning, of
course, you need to rely on people.(...)
Then by time, we need to build some good
(01:31:55):
strategy, have some
internal feedback look.
At least we know something, I mean,
partially know something good or bad, and
then we can iterate, iterate in the proof
and then some way to measure. So just
something high level.(...) Lots of
detailed work behind. I think this is a
very committed performance to solve. And
(01:32:16):
then it's hard to tell if it's
well-solved or not, but we just make it
product right, make it more practical.
Very cool.
So kind of switching the topic a little
bit,(...) I saw you have another agent
that we haven't really touched on yet,
which is the agentic data tables. Can you
maybe briefly describe
what that offers users?
(01:32:39):
If you try some query,
you'll see like, for example,
if you want to understand the top,
like top brand in some industry, and then
the ideal format, the outcome for the
result is the table. And then you can
even define a few dimension.
(01:33:02):
So then the table is something,
the good way to view the overview, and
also we put the source of the data.(...)
Then you can do the check by yourself.
(...) That's one thing.
If no data, it will table. Another
example is like,(...) for example, who is
the key players in Trump team? There is a
(01:33:27):
presidential election.
So there's nothing about data, but it can
give you the people name, and then their
role, and all the other information. So
meaning that table is very powerful.
People can present the data in a very
abstract and brief way. People can easily
to get information and they're easier to
(01:33:47):
compare. So this is the thing we built in
last Q4. Behind the tech stack, it's
still similar with the CodeShack and the
Deep Research. That's why we are able to
launch it and iterate them together.
(01:34:08):
Maybe taking a step back a little bit, I
think we covered most of the agents that
you guys offer. Most of the advanced
agents, we talked a little bit about the
basic agents,(...) some of
the mixture of agent systems.
If we take a step back,
what are some of the
biggest challenges your team faces
in building these agents in a reliable,
(01:34:28):
trustworthy way and at scale? Is it like
the conceptual challenges where you have
to create these ideas from scratch, or is
it just like technical software
engineering challenges where you have to
integrate a lot of different tools and
make sure the evals are right at scale?
Challenges, that's a lot. Because if we
(01:34:52):
want to build the product right,
not only demo, it's
actually solve the problem.
There's lots of detail thing we need to
handle. Lots of common cases we need to
debug, we need to cover, we need to have
some other solution.
For mixture agent,
(01:35:13):
the architecture is a little bit
complicated. So how to combine them
together, that's one typical sense.
I'd like to mention another one, another
different challenge as well, to give you
better sense about the kind of challenge.
Say for example,(...) the model's context
(01:35:34):
window,(...) there's a limit over there.
And by time, it will grow.
And in practice, if we solve some very
complicated problem,
the window is still a limit.
Then we can either optimize the data, try
to fit the window, or like we just leave
(01:35:55):
it as is, or some other work around. And
then we can predict the content window
will be longer. Then how to leverage the
end resource to the right place.
So I think this is another challenge we
need to face. So by time, that's to be
more practical. So make things work, and
then make it the
future tech stack scalable.
(01:36:22):
That's a really interesting point. You
mentioned the context window might be
small now, but it
might grow in the future.
So how do you, as a high level product
planning(...) for the roadmap for the
future,(...) how much of your effort is
spent solving the problems and
limitations of the models today versus
(01:36:43):
just waiting for the models to get
better, waiting for them to get better
reasoning skills or longer
context window and so on?
In practice, we do both, depends.(...)
Yeah, because there's no thing for answer
to say yes or no, or like what's the
right way.(...) We just need to be more
(01:37:04):
adaptive, and then be
more careful about the sound.
Sound good, I mean sounds
important to technically see.
So to answer your question,
we have lots of different products, and
they're different. That's why like we
(01:37:26):
depend on which product. Some products we
need to optimize the data, no matter
what. And some product we just leave the
window short for now, and then we imagine
in the near future, it will be big, and
then we don't need to see end of the
organization too early.(...) So
(01:37:47):
combination, combination of
both, depends on the product.
So I'm curious, you mentioned that there
was a lot of the issues around the
limited context window as well. Can you
maybe share some of the techniques that
you use to deal with it, and assuming
that the models, I mean, we all think
(01:38:09):
that the models will get better, but I'm
curious what kind of specific techniques
you're able to use to deal with the
limited context window. So I know one
technique that people will use is they
will summarize lots of the previous
conversation, and then throw that into
the bottom of the context. They'll use
RAG, I don't know, what kinds of things
(01:38:30):
do you use to kind of handle
the limited context window?
Yeah, we're doing something similar.
Yeah, of course the summarizer would have
all just a slow sound of data, or like we
have some judges, the data is already not
that important or whatever.
When I mentioned the content window is
(01:38:51):
like, it's just one example we need to
deal with, and how to make the tech stack
more scalable for the future, because the
model will improve,(...) and the other
part will improve as well. That's why
like, pick up as a writer, try our best
to pick up the right tech stack, how to
(01:39:13):
have the good engineer and the good model
can work together. So that's the big
question the engineering team is facing.
Yeah, and of course, content
window is just one example.
Yeah, speaking of the engineering team,
it seems like you guys are working on a
lot,(...) a lot of different products, a
(01:39:33):
lot of different challenges.
How big is this team? How many people are
working on it right now? And how many
people are focused on like the research
part? How many are focused on the
engineering product? And are you looking
to hire more people
in any specific areas?
We are a very smart team right now. So
the total company is around 20 people. Oh
(01:39:56):
my God. Many people are engineer.
So then, yeah, we have some people
working with, I mean, very good at
modeling part. They have lots of
experience.(...) And lots of stack, stack
means like back end, the modeling and the
front end or whatever. Because it is a
smart team. Everybody can do everything.
(01:40:19):
Are you trying to hire for any specific
open roles right now?
We are not a,(...) like I said, we have a
lot of people, but we
are slowly, slowly hiring.
That's, yeah. So we talked a lot about
(01:40:41):
all the different
products, the company size.
I had a couple of follow-up questions for
the future.(...) But before that,(...)
how do you see yourself in this crowded
space?(...) Because we have a lot of
competition from China. I think the first
generalist AI agent was Manus.
(01:41:03):
We have a lot of new specific vertical
agents. I think the industry leaders like
Hanma, for example, they
have their own AI slides agent.
There is competition from deep research
agents from like the fan companies.
How do you see the position of like
smaller startups like
Genspark in this massive landscape?
(01:41:26):
I think the market is huge. And then
going forward, the AI experience will be
very different. So then it's very, yeah,
it's very competitive. Yeah, of course. I
think the marketer is even bigger.
If you compare the agents, I think it's
(01:41:49):
very early stage.(...) For us, it's about
only 10% for the launch already,
according to our product vision. So
that's why we are very confident to build
something that's bigger. And if we are
talking about some existing bigger names,
like Google or Chai GPT,(...) of course
(01:42:09):
they are very good. They are very, very
big already.(...) Google is more or less
like cut to the search engine and the
search box we have. This is because the
user experience itself.
Chai GPT is a very, very great product
for the chart interface. For us, we have
a more complicated scenario. So I can
(01:42:32):
take a slice AI, for example.
This interaction cannot be simply sold by
either search box or the chart box.(...)
So that's why we feel like going forward,
the scenario is different. We want to
empower every user, empower everyone in
(01:42:54):
their daily work and daily life. That's
why the scenario is more complicated. So
then let's worry about the user case
forever. Marketing is huge. In the end,
the future AI experience will be quite
different. Your home
without the answer question.
No, I absolutely agree.
I think it's a big market and it's still
so early that I don't think we know what
(01:43:14):
the future will look like. And it's
really exciting to see startups because I
think in a smaller company, you're able
to move much faster.
I mean, I work at Google and I love
massive companies because they are able
to do things at scale, but we also move a
lot more slowly. We're not able to
iterate as quickly. And I'm really
excited that startups like yours exist
(01:43:36):
because it is really exciting to use the
AI Slides agent today and share it with
my friends and colleagues.
But I am curious, there
are other agentic companies.
Do you foresee interacting with other
agents outside of your company? Because I
think maybe we can see a future with
(01:43:57):
multiple different kinds of agents and
agents interacting with each other
without human intervention.
Do you have any plans or are you working
on features to use like the, I think
Google announced a
protocol for agent to agents,
some standard tooling and like the MCP
protocol to allow other
(01:44:19):
people to connect to your agents.
I would say the depends on user scenario.
So we are consumer facing AI agent.(...)
So for daily life, daily work, if it's a
scenario where we want to build something
valuable to empower users.
(01:44:42):
Do you have any plans for business use
cases, B2B products?
I would say, so the answer is not now.
But going forward as the
growth of the AI product,
the bottom line between work and life and
(01:45:03):
the internal scenario will be blah.
So that's why hard to tell B2B if we're
enterprise facing but for
user scenario, it will be shared.
Very interesting, yeah.
So I don't know if you can see my video
anymore. I think I lost it. But anyway, I
still hear you. Okay, that's fine.(...)
(01:45:24):
So one kind of question that I have is
following the $60 million of funding,
(...) taking this money, what kind of are
your key priorities for your growth? And
then if you can share maybe a little bit
about your product roadmap.
So we are doing the BOD,
(01:45:48):
product and lead growth.
We don't do content distribution or
marketing as of today. So all the growth
are from organic growth. Then building
our products is the top priority. So as I
mentioned before, right now, according to
our product vision, only 10% progress at
(01:46:11):
this moment. So the more resource will be
put to build a useful
valuable product for users.
Are there any specific new advanced
agents that you're looking forward to
having in the next few
months or year or two?
(01:46:33):
For us, we are going to launch another
one.(...) Stay tuned, it's about one to
two weeks, within one week, we're going
to have another one launch.
That's really exciting, the speed at
which your startup is iterating, not in
the order of like years or months,
everything is happening in weeks and
days. That's incredible.
(01:46:56):
Yeah, I mean, I think that's everything
we had. I really appreciate
your time and thoughtful answers.
To all of our listeners, check out
GenSpark. It is really easy to use. And
the free plan offers plenty of usage.
The paid plan is fantastic. It offers you
higher limits and I think more capable
(01:47:18):
models too, if I'm correct.
More quotas.
More quotas, yeah. More credits.
It's like a free model. You show people
how good this thing is and people just
want to use more. That's kind of what I
felt with the AI Slides agent.
Yeah, again, thank you so much. But I
wanted to leave a couple of minutes for
(01:47:38):
you. If there's anything you want to
announce or anything you want to share,
(...) anything you'd
like to tell our audience.
For more information, so people can
follow our x.com. In just search,
GenSpark AI, so you can find the account.
So most the latest information, we
(01:47:59):
usually post to our Twitter account.
That's awesome. Yeah, we'll share the
links in the podcast notes and you guys
can follow the Twitter account, the x
account. And also if you're here in the
Bay Area this week on May 8th, we'll be
having GenSpark and a couple other
(01:48:21):
agentic startups too.(...) A fully
integrated coding ID and a browser agent.
So it'll be like a full on agentic
product demo and a tech talk show in Palo
Alto from 6 to 9 p.m. And yeah, I'm
excited to see you all here if you can
make it. And thank you so much again,
(01:48:42):
Linjoy, for joining
us, this was fantastic.
Thank you for having me. Looking forward
to the meetup. I will show you more
videos. Awesome. And
the demos over there.
Excited to see you in person.
All right, thank you everyone.
Bye-bye.
Popular Podcasts
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
24/7 News: The Latest
The latest news in 4 minutes updated every hour, every day.
Therapy Gecko
An unlicensed lizard psychologist travels the universe talking to strangers about absolutely nothing. TO CALL THE GECKO: follow me on https://www.twitch.tv/lyleforever to get a notification for when I am taking calls. I am usually live Mondays, Wednesdays, and Fridays but lately a lot of other times too. I am a gecko.