WEBVTT

00:00:00.000 --> 00:00:04.600
Hello, today we're chatting with Qiu Yuan, Jin Qiuyuan

00:00:06.200 --> 00:00:07.800
Sandy, you graduated from Yale

00:00:07.900 --> 00:00:09.600
Yes, yes, Yale undergrad

00:00:10.200 --> 00:00:10.900
And then

00:00:12.000 --> 00:00:14.600
Right, so what was your career path after that

00:00:15.700 --> 00:00:17.000
Yes, I'm a serial entrepreneur

00:00:17.000 --> 00:00:18.700
So after graduating from Yale, what happened next

00:00:18.700 --> 00:00:19.700
I first worked in consulting

00:00:19.700 --> 00:00:21.300
And then later transitioned into product management

00:00:21.600 --> 00:00:24.900
Before starting my own business, I worked as a product manager at Tantan

00:00:25.000 --> 00:00:26.300
In charge of growth and monetization

00:00:26.800 --> 00:00:29.300
After Tantan was acquired, I started my own business independently

00:00:29.500 --> 00:00:31.900
I built an online events platform

00:00:31.900 --> 00:00:33.100
Called VLAN Lingxianchang

00:00:33.600 --> 00:00:35.700
I previously raised two rounds of funding

00:00:35.900 --> 00:00:36.700
From Shouyuan and Sequoia

00:00:37.300 --> 00:00:38.400
So this platform of ours

00:00:38.400 --> 00:00:40.500
Mainly serves enterprises online

00:00:40.700 --> 00:00:42.500
For hosting virtual events

00:00:42.900 --> 00:00:46.700
Everything from online recruiting to corporate open days

00:00:47.100 --> 00:00:50.100
Yes, we ran this product for a little over four years

00:00:50.500 --> 00:00:51.800
And went through the whole cycle, so to speak

00:00:52.100 --> 00:00:54.700
During the pandemic we had a lot of

00:00:55.000 --> 00:00:55.800
Momentum

00:00:55.900 --> 00:00:58.200
After the pandemic, many events returned to being offline

00:00:58.400 --> 00:01:00.400
So last year I paused the company

00:01:00.400 --> 00:01:02.200
And this year I'm exploring new directions

00:01:03.700 --> 00:01:05.400
We've already been WeChat friends for many years

00:01:05.800 --> 00:01:07.700
We were online friends

00:01:07.700 --> 00:01:09.900
We used to be online friends, now we're friends in real life

00:01:11.200 --> 00:01:12.400
Back then it was at Yale

00:01:12.400 --> 00:01:14.500
What did you study, and how did you end up coming back to start a business in China

00:01:15.600 --> 00:01:17.800
Right, I studied economics as an undergrad

00:01:17.800 --> 00:01:22.400
At that time, China had this environment of mass entrepreneurship and innovation

00:01:22.700 --> 00:01:24.100
So back then it was quite emphasized

00:01:24.400 --> 00:01:25.900
Trying out entrepreneurship as a direction

00:01:26.300 --> 00:01:28.600
So after graduating I came back to look at various opportunities

00:01:29.600 --> 00:01:31.900
So how long did you work at Tantan

00:01:32.300 --> 00:01:35.000
I was at Tantan around 2018 to 2019

00:01:35.800 --> 00:01:37.500
So it happened that right after that the pandemic hit

00:01:38.100 --> 00:01:40.600
The virtual social space is booming

00:01:40.700 --> 00:01:41.900
Started working on virtual socializing

00:01:42.600 --> 00:01:46.000
I've actually always wanted to work on connecting people

00:01:46.300 --> 00:01:49.100
Right, so my career as a product manager

00:01:49.100 --> 00:01:50.000
Well, it's about social connection

00:01:50.100 --> 00:01:52.700
Then later I built a metaverse events platform

00:01:52.800 --> 00:01:56.800
It was also about creating, for event organizers and participants online,

00:01:57.100 --> 00:01:59.800
a connection and an immersive social environment

00:02:00.600 --> 00:02:02.500
So what were you mainly responsible for at Tantan

00:02:03.200 --> 00:02:05.500
At Tantan, I mainly focused on growth and monetization

00:02:05.600 --> 00:02:08.400
For monetization, I basically went through the whole zero-to-one phase

00:02:08.700 --> 00:02:12.000
We went from a very low monthly revenue level

00:02:12.000 --> 00:02:14.700
to a monthly revenue of about 100 million

00:02:14.900 --> 00:02:17.700
We built a lot of features and did a lot of

00:02:18.000 --> 00:02:18.700
A/B testing

00:02:18.900 --> 00:02:22.300
So I'm actually quite proud of that process

00:02:22.800 --> 00:02:25.000
So what do you think was the most important feature back then

00:02:25.500 --> 00:02:26.400
There were quite a few

00:02:26.700 --> 00:02:28.900
Right now in Tantan, paid features like

00:02:29.400 --> 00:02:30.400
seeing who likes me

00:02:30.700 --> 00:02:31.900
or Super Like, for example

00:02:32.300 --> 00:02:35.400
Each feature brought in roughly tens of millions of RMB in revenue

00:02:35.900 --> 00:02:38.000
as a monthly paid feature

00:02:38.600 --> 00:02:40.900
So was Super Like something Tantan pioneered

00:02:40.900 --> 00:02:42.600
or did Tinder already have something like that

00:02:43.100 --> 00:02:45.900
We were actually feeling our way across the river too

00:02:46.500 --> 00:02:50.100
We made a lot of innovations and adjustments based on Tinder

00:02:50.100 --> 00:02:52.600
Right, because Tinder was already a fairly large company by then

00:02:52.600 --> 00:02:53.900
and had fairly stable revenue

00:02:54.100 --> 00:02:58.200
so their monetization features and functions

00:02:58.400 --> 00:02:59.300
were relatively mature

00:02:59.500 --> 00:03:02.300
So at first we mainly borrowed from their paid features

00:03:02.500 --> 00:03:04.900
mainly localizing for China

00:03:04.900 --> 00:03:08.800
and doing some new design around bundling and payments

00:03:10.200 --> 00:03:12.200
So Tantan came out after Tinder

00:03:13.600 --> 00:03:14.500
Yes, that's right

00:03:17.100 --> 00:03:18.900
It was a copy-to-China product

00:03:19.500 --> 00:03:19.900
Right

00:03:20.500 --> 00:03:22.900
Back then a lot of platforms were copy-to-China

00:03:23.000 --> 00:03:28.200
My first tech job was actually at Didi

00:03:28.300 --> 00:03:30.600
At the time, it was also modeled after Uber, essentially copying Uber

00:03:30.800 --> 00:03:35.400
At the time, it actually happened to be right after Uber China's merger that I joined

00:03:35.500 --> 00:03:37.900
So at the time there was still a lot of 'copy to China' happening

00:03:38.300 --> 00:03:40.400
This kind of mobile internet entrepreneurial approach

00:03:40.900 --> 00:03:42.700
Tantan will help you develop a user model

00:03:42.900 --> 00:03:46.500
You'll roughly know in your mind what a user might do when entering this product

00:03:46.900 --> 00:03:47.600
Right

00:03:47.600 --> 00:03:49.900
It's essentially like a path of validation

00:03:50.200 --> 00:03:52.200
But in terms of payment, there might be a difference due to exchange rates

00:03:53.600 --> 00:03:54.700
I learned some user models

00:03:55.200 --> 00:03:59.100
Later when doing Tantan, these user models

00:03:59.400 --> 00:04:05.000
Did the user models you formed in your mind help you later when working on metaverse social products?

00:04:05.900 --> 00:04:07.100
This was indeed helpful

00:04:07.200 --> 00:04:11.600
Because actually a lot of paid feature design still revolves around human nature

00:04:12.100 --> 00:04:14.300
For example, the gender ratio is very imbalanced

00:04:14.400 --> 00:04:19.900
How to give male users more proactive rights to view female users

00:04:20.000 --> 00:04:25.300
Or let them show off their own strength/status

00:04:25.700 --> 00:04:29.100
Actually a lot of our paid features were designed around this

00:04:30.200 --> 00:04:34.800
From the perspective of our later entrepreneurship doing online social activities

00:04:35.100 --> 00:04:38.600
It actually also leveraged a need in human nature

00:04:38.700 --> 00:04:42.100
That is, under pandemic conditions

00:04:42.300 --> 00:04:44.500
Many people had a lot of video conferences

00:04:44.500 --> 00:04:46.100
Extremely exhausting

00:04:46.200 --> 00:04:53.300
And there was no sense of immersion or interactive way of engaging

00:04:53.600 --> 00:04:56.300
Online, we actually used a gamified experience

00:04:56.500 --> 00:05:00.000
To recreate that offline social experience

00:05:00.300 --> 00:05:02.300
Actually, everyone has their own avatar in it

00:05:02.600 --> 00:05:05.400
And when this little character walks up to other users

00:05:05.700 --> 00:05:07.700
It automatically generates an audio-video connection

00:05:08.100 --> 00:05:11.300
So that you'll feel like this person seems to be in the same space as you

00:05:11.700 --> 00:05:13.600
Creating this kind of social experience

00:05:13.900 --> 00:05:17.400
So this is actually also centered around a human need

00:05:17.800 --> 00:05:18.800
Building product design around it

00:05:19.400 --> 00:05:20.500
So is it one-on-one online

00:05:21.200 --> 00:05:22.200
It's one-to-many

00:05:22.400 --> 00:05:24.300
It mainly embodies a sense of space

00:05:24.500 --> 00:05:27.600
For example, right now we're in a lobby environment

00:05:27.700 --> 00:05:29.500
And there are different tables in there, right

00:05:29.600 --> 00:05:31.100
Over there they're discussing another topic

00:05:31.100 --> 00:05:33.600
And over here we're recording a video

00:05:33.800 --> 00:05:35.800
Then over there they might also be running a workshop

00:05:36.000 --> 00:05:39.600
Actually it's the same space that can give you an immersive experience

00:05:40.300 --> 00:05:42.200
This is something we've actually done online through a

00:05:42.800 --> 00:05:45.100
So-called gamified space

00:05:45.700 --> 00:05:48.300
And an immersive environment to help you create and recreate this

00:05:48.700 --> 00:05:51.800
So your launch event would feel like it's happening at an actual event venue

00:05:52.200 --> 00:05:55.300
And your online job fair would feel like a booth setup at school

00:05:55.500 --> 00:05:59.600
And you could actually walk around between different booths

00:06:00.000 --> 00:06:02.300
And after finishing a conversation at one booth

00:06:02.300 --> 00:06:04.100
You could also switch to another channel

00:06:04.300 --> 00:06:05.100
So they don't interfere with each other

00:06:05.100 --> 00:06:08.500
But at the same time you get a pretty smooth experience

00:06:09.200 --> 00:06:10.800
And I think this is quite fun

00:06:11.300 --> 00:06:12.400
It's really fun

00:06:12.400 --> 00:06:14.300
For us it's also quite well received

00:06:15.300 --> 00:06:19.100
It's a product that's quite popular with students and younger users

00:06:19.400 --> 00:06:20.800
Because we always say that Gen Z

00:06:21.100 --> 00:06:23.500
And the generations after them are all gaming natives

00:06:23.500 --> 00:06:25.500
So we actually used the approach they like

00:06:25.500 --> 00:06:29.800
To create a platform with more interactivity and entertainment value

00:06:30.600 --> 00:06:32.800
The key thing is last month I talked to a 19-year-old

00:06:32.800 --> 00:06:34.900
A young guy who came from Houston

00:06:35.400 --> 00:06:38.500
He almost wanted to do a similar idea

00:06:38.500 --> 00:06:41.900
His starting point was that he had previously attended an online conference

00:06:41.900 --> 00:06:42.400
Summoner

00:06:42.800 --> 00:06:44.900
And it turned out to be a group of three hundred people

00:06:44.900 --> 00:06:46.800
With some big shot lecturing to a crowd

00:06:46.800 --> 00:06:47.800
He was sitting there listening to a lecture

00:06:47.800 --> 00:06:48.800
Like Americans, you know

00:06:48.800 --> 00:06:51.300
Including myself, being from the new generation, I also find that annoying

00:06:51.300 --> 00:06:51.700
Right

00:06:51.700 --> 00:06:53.100
I came here wanting to interact with people

00:06:53.100 --> 00:06:54.900
And it turned into being lectured at

00:06:54.900 --> 00:06:57.500
He wanted to build an online platform kind of like this

00:06:57.500 --> 00:06:58.900
Where you go up and generate a stage

00:06:59.200 --> 00:07:01.200
And people can walk up close and chat one-on-one

00:07:01.200 --> 00:07:04.800
Yeah, at that time we were among the first, I think

00:07:04.900 --> 00:07:07.700
At least the earliest to launch in the Asia region

00:07:07.700 --> 00:07:07.900
Right

00:07:07.900 --> 00:07:10.300
So I think this demand still exists

00:07:10.300 --> 00:07:13.200
It's just that everyone may have been cooped up too long again because of the pandemic

00:07:13.200 --> 00:07:14.600
and returning to an offline way of socializing

00:07:15.700 --> 00:07:18.300
because I saw that kid playing Roblox

00:07:18.300 --> 00:07:20.600
there's really a lot of room

00:07:20.600 --> 00:07:20.900
right

00:07:20.900 --> 00:07:22.100
so what do you think then

00:07:22.100 --> 00:07:24.400
and that 19-year-old still felt

00:07:24.400 --> 00:07:25.900
that this need of his wasn't being met

00:07:25.900 --> 00:07:27.700
he thought this thing felt pretty futuristic

00:07:27.700 --> 00:07:28.700
what's your take on that

00:07:28.700 --> 00:07:32.700
this is essentially a disconnect between the user and the payer

00:07:32.800 --> 00:07:35.300
in other words, our paying customers

00:07:35.300 --> 00:07:36.300
are actually enterprises

00:07:36.300 --> 00:07:38.000
and within those enterprises, the decision-makers

00:07:38.000 --> 00:07:39.600
tend to be a bit older

00:07:40.400 --> 00:07:43.400
so when they're allocating budget and paying for this

00:07:43.400 --> 00:07:46.000
they may not fully understand the need

00:07:46.000 --> 00:07:48.800
so it's our users and participants

00:07:48.800 --> 00:07:50.500
everyone who tries it loves it right away

00:07:50.500 --> 00:07:52.700
NPS is basically above 80%

00:07:52.700 --> 00:07:55.600
so it's a platform that's very well received by participants

00:07:55.600 --> 00:07:59.000
but the ones paying are event organizers and enterprise clients

00:07:59.000 --> 00:08:01.900
so there's bound to be a certain lag

00:08:02.000 --> 00:08:03.500
so this is also one of the challenges

00:08:03.500 --> 00:08:05.100
this platform faces in terms of monetization

00:08:06.400 --> 00:08:08.500
but at the time, since enterprises were the ones paying

00:08:08.500 --> 00:08:10.300
it probably did drive quite a few enterprises away

00:08:10.300 --> 00:08:11.900
for us, it was actually mostly in British pounds

00:08:11.900 --> 00:08:13.900
because our first client was Ola

00:08:13.900 --> 00:08:16.600
and afterward, a lot of enterprise clients started reaching out to us proactively

00:08:16.600 --> 00:08:18.200
saying they wanted to host a similar event

00:08:18.200 --> 00:08:19.000
with brand exposure

00:08:19.000 --> 00:08:22.000
with a branded venue of their own

00:08:22.000 --> 00:08:22.200
right

00:08:22.200 --> 00:08:24.900
so this was also one of our selling points to enterprises

00:08:24.900 --> 00:08:27.000
basically we did both 2C and 2B

00:08:27.000 --> 00:08:31.500
for us, we had C-end participants and B-end organizers

00:08:31.600 --> 00:08:32.200
a dual-sided market

00:08:32.800 --> 00:08:35.600
so what did the C-end user side look like around that time

00:08:35.600 --> 00:08:38.200
we hosted about 3,000 events a year

00:08:38.200 --> 00:08:43.600
Then there might be dozens, hundreds, or even thousands per event, right?

00:08:43.600 --> 00:08:43.800
Right

00:08:43.800 --> 00:08:47.000
So it's a relatively large platform

00:08:47.000 --> 00:08:47.700
Right

00:08:47.700 --> 00:08:49.500
So that's it

00:08:49.500 --> 00:08:50.900
But it also has a problem

00:08:50.900 --> 00:08:54.400
That is, it may not necessarily be a particularly recurring model

00:08:54.400 --> 00:08:55.500
Because once an event is finished

00:08:55.500 --> 00:08:58.000
The next one might not happen until next quarter or the second half of the year

00:08:58.000 --> 00:09:00.600
So it doesn't have that great an automatic renewal

00:09:00.600 --> 00:09:00.800
Right

00:09:00.800 --> 00:09:03.100
So this is also a problem with this type of product

00:09:03.100 --> 00:09:05.000
That is, event platforms

00:09:05.000 --> 00:09:06.500
It has its novelty

00:09:06.500 --> 00:09:10.900
But it's probably hard to embed into a company's workflow

00:09:10.900 --> 00:09:12.700
To let them pay long-term

00:09:12.700 --> 00:09:16.800
So if kids nowadays really end up having a need for virtual socializing in the future

00:09:16.800 --> 00:09:19.700
And they start playing on it

00:09:19.700 --> 00:09:20.900
Well, that's certainly possible

00:09:20.900 --> 00:09:21.300
Right

00:09:21.300 --> 00:09:24.500
I think in the long run

00:09:24.500 --> 00:09:29.400
This is still a quite valuable product and platform

00:09:29.400 --> 00:09:32.700
So going forward, as the terminals change

00:09:32.700 --> 00:09:34.900
And as the user structure changes

00:09:34.900 --> 00:09:37.100
I think there might be new opportunities

00:09:37.100 --> 00:09:39.100
So what are your thoughts now

00:09:39.100 --> 00:09:42.000
In terms of AI, what do you want to do

00:09:42.000 --> 00:09:44.400
Right, because my background and interests

00:09:44.400 --> 00:09:47.900
Have always been in human behavior

00:09:47.900 --> 00:09:50.800
The intersection of human connection and technology

00:09:50.800 --> 00:09:51.500
Right

00:09:51.500 --> 00:09:55.100
And my previous experience has also developed along this main line

00:09:55.100 --> 00:09:56.500
The direction I want to pursue now

00:09:56.500 --> 00:10:04.000
Is actually to give AI some human intuitive judgment and some humanity in it

00:10:04.000 --> 00:10:05.500
Because right now when we look at a lot of AI execution

00:10:05.500 --> 00:10:08.000
It follows a workflow

00:10:08.000 --> 00:10:10.500
Follows a hard coded rule

00:10:10.500 --> 00:10:11.600
And then executes

00:10:11.600 --> 00:10:14.700
But if you ask it to make its own decisions and judgments

00:10:14.700 --> 00:10:17.900
Or to act based on context, right

00:10:17.900 --> 00:10:21.100
Then when making some of these professional choices

00:10:21.100 --> 00:10:23.500
it often produces quite large deviations

00:10:23.500 --> 00:10:25.200
possibly even worse than an intern

00:10:25.300 --> 00:10:28.800
so what we hope is to let AI cross over

00:10:28.800 --> 00:10:31.800
so it can be like a professional with certain experience

00:10:31.800 --> 00:10:34.400
like a seasoned professional

00:10:34.400 --> 00:10:38.100
and learn to do these professional tasks

00:10:38.100 --> 00:10:39.800
especially long-horizon ones

00:10:39.800 --> 00:10:41.200
and long-term ones

00:10:41.200 --> 00:10:44.500
that can be applied to actual work with economic value

00:10:44.500 --> 00:10:45.400
Right

00:10:45.400 --> 00:10:50.800
So basically we build a product platform

00:10:50.800 --> 00:10:54.500
to help these professionals

00:10:54.500 --> 00:10:55.400
share their knowledge

00:10:55.400 --> 00:10:58.100
share their workflow and some of their professional

00:10:58.100 --> 00:11:00.400
computer use operations on the computer

00:11:00.400 --> 00:11:05.000
and from these I extract some key effective behavioral information

00:11:05.000 --> 00:11:09.200
then teach AI how to do all kinds of work

00:11:09.200 --> 00:11:12.300
so that's our product positioning

00:11:12.300 --> 00:11:14.800
So you think it's not a technical problem

00:11:14.800 --> 00:11:16.000
it's a data problem

00:11:16.000 --> 00:11:17.300
It's actually a data problem

00:11:17.300 --> 00:11:20.500
because we've found that all the large model companies

00:11:20.500 --> 00:11:24.400
allocate about two-thirds of their manpower and resources

00:11:24.400 --> 00:11:28.000
to data collection, cleaning, labeling, and pre-training work

00:11:28.000 --> 00:11:31.400
so actually the quality of the data determines the quality of the training

00:11:31.400 --> 00:11:33.600
as the saying goes, garbage in garbage out

00:11:33.600 --> 00:11:35.900
I think data is extremely important

00:11:35.900 --> 00:11:40.200
Right now AI training has just run into this kind of data bottleneck

00:11:40.200 --> 00:11:42.800
because data available online

00:11:42.800 --> 00:11:46.100
as well as data that can be manually labeled

00:11:46.100 --> 00:11:47.500
has basically been exhausted

00:11:47.500 --> 00:11:48.900
so there's a large amount of real-world data

00:11:48.900 --> 00:11:52.800
that hasn't been effectively filtered and collected by anyone

00:11:52.800 --> 00:11:54.800
so this is a fairly big problem we want to solve

00:11:54.800 --> 00:11:58.500
that is, once there's large-scale human demonstration

00:11:58.500 --> 00:12:04.200
that can teach AI how to operate all kinds of manual trades and actual work processes

00:12:04.200 --> 00:12:08.500
it can actually greatly improve AI's applicability in real-world scenarios

00:12:08.500 --> 00:12:09.900
and its usability

00:12:09.900 --> 00:12:13.700
so this is one of our product goals

00:12:13.700 --> 00:12:17.100
It can be classified as part of the data labeling industry

00:12:17.100 --> 00:12:18.800
We don't want to call it the data labeling industry

00:12:18.800 --> 00:12:22.900
We are more of a behavioral intelligence layer

00:12:22.900 --> 00:12:26.300
Based on our data channels and real data

00:12:26.300 --> 00:12:29.000
We can build something like a behavioral model

00:12:29.000 --> 00:12:32.800
This model can make AI agents more efficient

00:12:32.800 --> 00:12:36.900
And execute certain types of tasks at a lower cost

00:12:36.900 --> 00:12:38.400
Compared to a traditional large model

00:12:38.400 --> 00:12:40.000
It may require fewer steps

00:12:40.000 --> 00:12:41.400
And fewer tokens

00:12:41.400 --> 00:12:45.600
And instead has domain-specific expertise and intuition in that field

00:12:45.600 --> 00:12:46.700
It can perform better

00:12:46.700 --> 00:12:49.100
So this has actually been validated in research

00:12:49.100 --> 00:12:52.400
Because now, companies like Amazon

00:12:52.400 --> 00:12:55.600
Their research team has done a shopping agent project

00:12:55.600 --> 00:12:58.400
Using traces from shoppers

00:12:58.400 --> 00:13:02.000
That is, real shoppers' behavioral trajectories

00:13:02.000 --> 00:13:06.600
To train an AI shopping agent

00:13:06.600 --> 00:13:08.600
A shopping agent

00:13:08.600 --> 00:13:11.100
The actual result compared to a base large model

00:13:11.100 --> 00:13:12.100
When letting it choose products

00:13:12.100 --> 00:13:14.100
And complete the entire shopping process

00:13:14.100 --> 00:13:15.500
Showed a twenty to thirty percent improvement

00:13:15.700 --> 00:13:19.100
So this has actually been validated in one specific domain

00:13:19.100 --> 00:13:21.500
So we want to, from the data acquisition side

00:13:21.500 --> 00:13:24.700
As well as from the infra product-level construction

00:13:24.700 --> 00:13:29.100
Both be able to produce a different result

00:13:29.100 --> 00:13:32.400
And help AI actually be deployed and executed effectively

00:13:33.700 --> 00:13:36.400
So do you want to sell it to large model companies

00:13:36.400 --> 00:13:38.200
To help them make their general-purpose models better

00:13:38.200 --> 00:13:40.500
Or sell it to specific customers

00:13:40.500 --> 00:13:43.900
To help them improve their own workflows

00:13:43.900 --> 00:13:45.500
There's a sequence to this

00:13:45.500 --> 00:13:47.300
Right now, large model companies

00:13:47.300 --> 00:13:52.200
Have a lot of demand to purchase this kind of data

00:13:52.200 --> 00:13:54.400
That can also bring us early revenue

00:13:54.400 --> 00:13:56.000
As for the future

00:13:56.000 --> 00:13:59.300
Our vision is still to become a behavioral layer

00:13:59.300 --> 00:14:03.200
That is, to have, on top of the base large model

00:14:03.200 --> 00:14:08.100
A layer that resembles behavior

00:14:08.100 --> 00:14:10.600
Enable agents to teach themselves

00:14:10.600 --> 00:14:12.000
Learn on their own

00:14:12.000 --> 00:14:12.900
From human demonstrations

00:14:12.900 --> 00:14:14.900
From the experience of experts

00:14:14.900 --> 00:14:18.000
It could serve agents

00:14:18.000 --> 00:14:19.400
This kind of expert marketplace

00:14:20.700 --> 00:14:23.300
I hadn't thought about serving agents either

00:14:23.300 --> 00:14:24.700
Right, because in the future

00:14:24.700 --> 00:14:30.700
Agents will actually go find a lot of learning content on their own

00:14:30.700 --> 00:14:31.600
Right, because currently agents

00:14:31.600 --> 00:14:35.500
You actually need to feed them a lot of context

00:14:35.500 --> 00:14:38.400
Tell them a lot of documentation and skills

00:14:38.400 --> 00:14:40.600
But later on agents will actually go find it themselves

00:14:40.800 --> 00:14:42.700
For things I don't understand or haven't done before

00:14:42.700 --> 00:14:44.800
How should I go about learning this

00:14:44.800 --> 00:14:48.100
Then you need a repository with a large number of human

00:14:48.100 --> 00:14:51.300
demonstrations and experience

00:14:51.300 --> 00:14:55.100
This repository can let AI learn from demonstrations

00:14:55.100 --> 00:14:56.800
And then it can teach itself

00:14:56.800 --> 00:14:59.400
To carry out these specific tasks

00:14:59.400 --> 00:15:01.000
Because actually every job has been done by someone

00:15:01.000 --> 00:15:02.000
It's just that this data

00:15:02.000 --> 00:15:03.600
hasn't been well preserved until now

00:15:06.100 --> 00:15:09.000
Let me think, so, do you think that

00:15:09.100 --> 00:15:12.400
although models, for example, when answering questions now

00:15:12.400 --> 00:15:15.600
have a lot of data from their training set

00:15:15.600 --> 00:15:18.900
but there are some things they search for live online

00:15:18.900 --> 00:15:22.000
and do you think that afterward, for agents

00:15:22.000 --> 00:15:25.000
the part where they actually do things will be like this too

00:15:25.000 --> 00:15:25.800
Yes, yes, exactly

00:15:25.800 --> 00:15:27.600
Because actually, just like now

00:15:27.600 --> 00:15:29.300
if you're a developer

00:15:29.300 --> 00:15:32.400
and you don't know whether a certain solution

00:15:32.400 --> 00:15:34.000
has an existing API and SDK

00:15:34.000 --> 00:15:35.600
you'd go search on Github yourself

00:15:35.600 --> 00:15:36.800
this is actually a similar logic

00:15:36.800 --> 00:15:40.300
in the end, for an agent to learn a piece of hammer-like knowledge

00:15:40.300 --> 00:15:42.000
how a person carries out a specific task

00:15:42.000 --> 00:15:44.000
it will also go find the corresponding reference

00:15:44.000 --> 00:15:45.300
So we hope to provide something like this

00:15:45.300 --> 00:15:46.800
It's equivalent to saying it's a reference system

00:15:46.800 --> 00:15:48.900
So it can quickly learn

00:15:50.800 --> 00:15:51.600
Let me repeat one point

00:15:51.600 --> 00:15:53.700
That is, you think it won't

00:15:53.700 --> 00:15:57.100
Large model companies won't train everything into their models

00:15:57.100 --> 00:16:02.100
Instead, an external reference is needed for that

00:16:02.100 --> 00:16:02.700
Right

00:16:02.700 --> 00:16:06.200
I think the reason is that doing it this way with large models is neither economical nor efficient

00:16:06.200 --> 00:16:08.700
That is, using a general-purpose model now

00:16:08.700 --> 00:16:12.000
It's impossible to solve all standing committee-level tasks

00:16:12.000 --> 00:16:13.200
This is a major judgment of mine

00:16:13.200 --> 00:16:14.700
Because there's recent research called

00:16:14.700 --> 00:16:16.400
It's called Agent Last Exam

00:16:16.400 --> 00:16:17.700
Actually we've also done research on this

00:16:17.700 --> 00:16:19.700
It actually compared

00:16:19.700 --> 00:16:24.000
Maybe 150 to 200 specific real-world job types

00:16:24.000 --> 00:16:26.500
All of which are relatively complex

00:16:26.500 --> 00:16:27.800
And involve operating multiple pieces of software

00:16:27.800 --> 00:16:29.200
And require multiple steps

00:16:29.200 --> 00:16:32.900
And even for frontier SOTA models on these

00:16:32.900 --> 00:16:34.600
One shot

00:16:34.600 --> 00:16:38.000
The end-to-end success rate is only about 2%

00:16:38.000 --> 00:16:40.400
So on this front, I think

00:16:40.400 --> 00:16:44.300
It's really hard to achieve a one-size-fits-all solution within three to five years

00:16:44.300 --> 00:16:47.400
We believe that each domain should actually build up its own accumulation

00:16:50.900 --> 00:16:51.600
Let me add a bit more

00:16:51.600 --> 00:16:53.400
Because I'm genuinely quite interested in this

00:16:53.400 --> 00:16:55.800
Because what we can see now is that even it

00:16:55.800 --> 00:16:56.900
As I just said, right

00:16:56.900 --> 00:16:58.800
Even when large models answer questions

00:16:58.800 --> 00:17:01.500
Starting from a certain year, they no longer rely solely on

00:17:01.500 --> 00:17:03.600
Looking from within their own model's training data

00:17:03.600 --> 00:17:07.200
But also need to look at external internet searches

00:17:07.200 --> 00:17:09.800
This will actually continue going forward

00:17:09.800 --> 00:17:11.700
So external new learning is always useful

00:17:11.700 --> 00:17:16.200
So now this is being extended to the behavioral layer of agents

00:17:16.200 --> 00:17:19.600
From the agent's perspective

00:17:19.600 --> 00:17:21.400
Right, because this actually echoes a bit

00:17:21.400 --> 00:17:22.900
With the direction you're working on

00:17:22.900 --> 00:17:25.400
That is, you're actually providing a person's social relationships

00:17:25.400 --> 00:17:29.900
This social aspect, this asset

00:17:29.900 --> 00:17:31.400
As an API

00:17:31.400 --> 00:17:34.800
Or as an agent, a network that can be deployed

00:17:34.800 --> 00:17:38.000
Right, and I actually take the knowledge and experience in everyone's mind

00:17:38.000 --> 00:17:39.200
And turn it into this kind of network as well

00:17:39.200 --> 00:17:40.400
So that people can deploy it

00:17:40.400 --> 00:17:42.100
Right, so this logic

00:17:42.100 --> 00:17:43.500
I think has some commonalities

00:17:43.500 --> 00:17:45.400
Because as long as people exist

00:17:45.400 --> 00:17:46.200
Their relationships

00:17:46.200 --> 00:17:50.200
Their experience, professional judgment, and intuition

00:17:50.200 --> 00:17:54.400
Are actually very hard for AI to fully exhaust and learn

00:17:54.400 --> 00:17:55.900
So at this point an idea occurred to me

00:17:55.900 --> 00:18:02.500
Which is this professional database

00:18:02.500 --> 00:18:04.400
A professional behavioral-layer database

00:18:04.400 --> 00:18:05.500
It's no longer the knowledge layer

00:18:05.500 --> 00:18:06.900
The knowledge layer has already been solved by others

00:18:06.900 --> 00:18:09.400
It's more like how to operate a computer

00:18:09.400 --> 00:18:11.800
That kind of layer

00:18:11.800 --> 00:18:14.000
Is it because the data in this layer

00:18:14.000 --> 00:18:15.600
Produces a lot of dynamic changes

00:18:15.600 --> 00:18:17.900
So you need it to be real-time

00:18:17.900 --> 00:18:20.300
He said it's just because it's so common

00:18:21.700 --> 00:18:24.400
This thing isn't worth a large model internalizing and training on its own

00:18:26.100 --> 00:18:27.200
It's probably both

00:18:27.200 --> 00:18:29.700
Indeed it's too common and too scattered

00:18:29.700 --> 00:18:31.800
So you can't learn everything

00:18:31.800 --> 00:18:34.100
And second

00:18:34.100 --> 00:18:39.300
It's actually also a behavior that requires a lot of reasoning

00:18:39.300 --> 00:18:41.400
Because when people operate different software

00:18:41.400 --> 00:18:43.500
Different keyboard and mouse operations

00:18:43.500 --> 00:18:44.400
They actually think it through

00:18:44.400 --> 00:18:46.900
At each step they go through a thought process

00:18:46.900 --> 00:18:51.400
Thinking about which pieces of information to examine

00:18:51.400 --> 00:18:53.300
And then take into account, right

00:18:53.300 --> 00:18:54.500
What factors were considered

00:18:54.500 --> 00:18:56.600
And who they might have communicated with, about what content

00:18:56.600 --> 00:18:58.800
And only then finally completed this case

00:18:58.800 --> 00:19:00.800
To submit or share, right

00:19:00.800 --> 00:19:04.200
And in this process there's actually a lot of complexity

00:19:04.200 --> 00:19:07.500
This is essentially a multi-stamp thinking chain

00:19:07.500 --> 00:19:08.800
There's a chain of thought embedded in it

00:19:08.800 --> 00:19:13.900
This kind of thing actually isn't well captured or recorded right now

00:19:13.900 --> 00:19:17.100
So what we hope is that, beyond the log of everyone's operations

00:19:17.100 --> 00:19:21.800
there's also this layer that can attach on top of the raw data

00:19:21.800 --> 00:19:23.800
so that ultimately it can be handed off to AI, saying

00:19:23.800 --> 00:19:25.300
I've encountered a similar task

00:19:25.300 --> 00:19:27.900
can we learn to think about it the way a human would

00:19:27.900 --> 00:19:29.300
OK

00:19:29.300 --> 00:19:31.100
So since you graduated from Yale

00:19:31.100 --> 00:19:33.700
you came to the US as an undergrad

00:19:33.700 --> 00:19:34.100
Right

00:19:34.100 --> 00:19:35.600
I actually came to the US in high school

00:19:35.600 --> 00:19:36.900
Came to the US in high school

00:19:36.900 --> 00:19:39.900
Then, given our age group

00:19:39.900 --> 00:19:43.700
you must have a lot of classmates who are already at large model companies

00:19:43.700 --> 00:19:44.100
Yes

00:19:44.100 --> 00:19:48.400
In terms of our network as founders

00:19:48.400 --> 00:19:51.900
because we get in touch with a lot of researchers

00:19:51.900 --> 00:19:55.600
who move between different frontier labs

00:19:55.600 --> 00:19:55.700
Right

00:19:55.700 --> 00:19:59.200
And we have an advisor who works on external machine learning at scale AI

00:19:59.200 --> 00:20:02.700
so his network, and many of his students, are also at frontier labs

00:20:02.700 --> 00:20:02.900
Right

00:20:04.300 --> 00:20:07.600
So based on your past entrepreneurial experience

00:20:07.600 --> 00:20:09.900
you have the ability to handle enterprise clients

00:20:09.900 --> 00:20:12.500
as well as personal relationships

00:20:12.500 --> 00:20:17.200
so you're able to reach large model companies

00:20:17.200 --> 00:20:19.300
Going back to the monetization need mentioned earlier

00:20:19.300 --> 00:20:20.100
Right

00:20:20.600 --> 00:20:22.500
And we've actually communicated with them directly

00:20:22.500 --> 00:20:24.300
so this need is indeed real

00:20:24.300 --> 00:20:24.500
Right

00:20:24.500 --> 00:20:25.700
This is a very real need

00:20:25.700 --> 00:20:29.700
Because recently I heard that companies like Salesforce

00:20:29.700 --> 00:20:31.500
spent two hundred thousand dollars

00:20:31.500 --> 00:20:35.400
to buy twenty long-thread operation trajectory datasets

00:20:35.400 --> 00:20:38.500
meaning using their own software to complete a task

00:20:38.500 --> 00:20:43.100
and one such high-quality trajectory is worth about ten thousand dollars

00:20:43.100 --> 00:20:43.600
Right

00:20:43.600 --> 00:20:48.800
Then all these other companies will need this kind of operational trajectory data to train their own AI

00:20:48.800 --> 00:20:54.400
So we think this is the next major opportunity and bottleneck

00:20:54.400 --> 00:20:59.000
That is to say, if AI really wants to achieve the so-called AGI goal of Long Horizon tasks

00:20:59.000 --> 00:21:01.800
It still needs a large amount of human experience

00:21:01.800 --> 00:21:05.300
To demonstrate and essentially teach the AI

00:21:05.300 --> 00:21:07.000
So this demand is very real

00:21:07.000 --> 00:21:07.500
And then

00:21:07.500 --> 00:21:10.800
You'll also get access to the next step, which is finding people to do the work

00:21:10.800 --> 00:21:11.300
Right, right, right

00:21:11.300 --> 00:21:11.600
Yes

00:21:11.600 --> 00:21:11.900
Yes

00:21:11.900 --> 00:21:12.100
Right

00:21:12.100 --> 00:21:17.300
Right now we're focused on building an early prototype of this kind of product

00:21:17.300 --> 00:21:19.400
And then getting it launched as quickly as possible

00:21:19.400 --> 00:21:22.300
And then selling the first batch of contracts

00:21:23.700 --> 00:21:25.100
I feel like it's actually a pretty good fit

00:21:25.100 --> 00:21:28.500
Because I previously worked on online recruiting in the virtual world too

00:21:28.500 --> 00:21:32.500
This actually sounds quite similar

00:21:32.500 --> 00:21:32.800
Right

00:21:32.800 --> 00:21:33.100
Yes

00:21:33.100 --> 00:21:33.400
Yes

00:21:33.400 --> 00:21:33.600
Right

00:21:33.600 --> 00:21:37.400
It's really in line with my previous interests and background

00:21:37.400 --> 00:21:41.400
And we also hope to have a more productized, standardized solution

00:21:41.400 --> 00:21:44.300
That can source and export at scale

00:21:44.300 --> 00:21:46.000
And then give them corresponding incentives and rewards

00:21:46.000 --> 00:21:47.900
And then let them onboard

00:21:47.900 --> 00:21:50.300
And then we can be more dynamic about it

00:21:50.300 --> 00:21:52.400
Depending on different needs

00:21:52.400 --> 00:21:54.100
Sourcing different kinds of expertise

00:21:55.300 --> 00:21:55.600
Right

00:21:55.600 --> 00:21:56.800
This is very similar

00:21:56.800 --> 00:21:59.600
Because what you did before was basically dealing with top-tier major clients

00:21:59.600 --> 00:22:01.700
Top-tier major clients have recruiting budgets

00:22:01.700 --> 00:22:02.300
So what about now

00:22:02.300 --> 00:22:04.000
These model companies also have recruiting budgets

00:22:04.000 --> 00:22:05.200
So they're not short on money

00:22:05.200 --> 00:22:05.800
Right, right, right

00:22:05.800 --> 00:22:09.900
Because each model company basically has a budget of one billion

00:22:09.900 --> 00:22:11.700
For data procurement

00:22:11.700 --> 00:22:11.900
Right

00:22:11.900 --> 00:22:13.700
And this part is actually very certain

00:22:13.700 --> 00:22:15.000
Also a very large market

00:22:15.000 --> 00:22:15.500
They lack people

00:22:15.500 --> 00:22:18.000
And then that recruiting event, on the other side

00:22:18.000 --> 00:22:19.100
It's people coming to find jobs

00:22:19.100 --> 00:22:20.000
Right, right, right, right, right

00:22:20.000 --> 00:22:22.600
So, actually, dating is similar too

00:22:22.600 --> 00:22:24.000
It's also a two-sided market

00:22:24.000 --> 00:22:24.200
Right

00:22:24.200 --> 00:22:25.600
So I think a marketplace

00:22:25.600 --> 00:22:29.900
Is still part of my architecture and approach as a product manager entrepreneur

00:22:29.900 --> 00:22:30.300
It's like

00:22:30.300 --> 00:22:30.900
Oh no, oh no

00:22:30.900 --> 00:22:36.300
Isn't this the common dream of product managers with a China industry background

00:22:36.300 --> 00:22:38.900
Would anyone think we're outdated

00:22:38.900 --> 00:22:39.600
I don't think so

00:22:39.600 --> 00:22:42.700
In the AI era there's still wisdom in this kind of marketplace

00:22:42.700 --> 00:22:44.100
And once a marketplace gets started

00:22:44.100 --> 00:22:45.600
It still has a flywheel effect

00:22:45.600 --> 00:22:48.000
So it will have relatively fast growth

00:22:49.000 --> 00:22:53.100
My feeling is that the people around us all quite like this network

00:22:53.100 --> 00:22:53.500
and things like that

00:22:53.500 --> 00:22:56.300
But some people think this is a leftover from the mobile era

00:22:57.700 --> 00:22:59.100
Right

00:22:59.100 --> 00:23:00.800
How do you see this, are we old-fashioned

00:23:00.800 --> 00:23:04.600
Or is it actually the next generation of product-type opportunity that should emerge

00:23:06.100 --> 00:23:07.200
I think there is an opportunity

00:23:07.200 --> 00:23:09.900
Otherwise the two of us wouldn't be sitting here discussing this topic

00:23:09.900 --> 00:23:12.000
Doing similar things, they all have a network, right

00:23:12.000 --> 00:23:13.400
Directions with potential

00:23:13.400 --> 00:23:20.600
But I don't think every field or direction is suitable to be solved with network effects

00:23:20.600 --> 00:23:20.800
Right

00:23:20.800 --> 00:23:23.100
So for example, some vertical AI in the hammer category

00:23:23.100 --> 00:23:25.100
It may not have a network effect at all

00:23:25.100 --> 00:23:28.000
It might just need to be forward deployed and implemented one by one

00:23:29.000 --> 00:23:30.800
Mainly it's another point of view

00:23:30.800 --> 00:23:33.100
As soon as you hear this, it sounds very mobile-internet-era and outdated

00:23:33.100 --> 00:23:33.400
Right

00:23:33.400 --> 00:23:35.600
Your mind has been fixed by the mobile internet era

00:23:35.600 --> 00:23:37.400
This is also a very common view

00:23:37.400 --> 00:23:38.100
How do you refute it

00:23:39.900 --> 00:23:42.600
I think we still need to start from first principles

00:23:42.700 --> 00:23:44.600
That is, to really solve this problem

00:23:44.600 --> 00:23:47.300
What kind of architecture and approach is best to solve it

00:23:47.300 --> 00:23:47.400
Right

00:23:47.400 --> 00:23:50.600
Then when we look at this kind of data bottleneck now

00:23:50.600 --> 00:23:55.400
It's actually a fairly scattered long tail supply

00:23:55.400 --> 00:23:56.900
This is actually quite suitable to be solved with this architecture

00:24:01.700 --> 00:24:03.400
So will you stay in the US going forward

00:24:03.400 --> 00:24:03.600
Right

00:24:03.600 --> 00:24:05.900
I'm now relocating to be based here in San Francisco

00:24:07.800 --> 00:24:08.300
Okay

00:24:08.300 --> 00:24:09.300
Anything else you'd like to talk about

00:24:09.300 --> 00:24:09.500
Okay

00:24:09.500 --> 00:24:11.100
That's about it for now

00:24:11.100 --> 00:24:11.400
Thank you
