Chat with Squirrel Ai Learning | ASU+GSV Summit 2024

Summary

このスクリプトは、 scoro AIという会社が提供するAIアダプティブ学習システムについて紹介しています。scoro AIは、数学や科学など、子供たちに幅広い科目でアダプティブ学習を提供しています。彼らのプラットフォーム「L」は、学習内容の構築とハードウェアの提供を含む大きなアダプティブモデルです。その主な製品であるスマート学習タブレットは、ディストリビューターを通じて市場に販売されており、学生が自己学習を通じて学ぶためのシステムです。また、scoro AIは、中国の2,000以上のディストリビューターと60,000以上の公的な学校を支援し、2,400万人以上の学生がそのシステムを使用しているとされています。さらに、 scoro AIは、AI技術の進歩を最大限に活用し、教育の難しさに対処する取り組みを行っています。

Takeaways

• 🌟 スコロAIは、子供たちに適応型学習を提供し、数学、科学などの科目があります。
• 📚 scoro AIは、AI適応学習を9年前に開始し、今年で10周年を迎えます。
• 💡 Lという新しいプラットフォームは、大規模な適応モデルです。コンテンツビルディングとハードウェアを提供しています。
• 📈 scoro AIは、中国の2,000人のナショナルディストリビューターと60,000以上の公立学校を支援し、2,400万人以上の学生がシステムを使用しています。
• 🏆 scoro AIは、Deoの研究によりトップ50の急速成長テック企業の一つであり、MITテクノロジーレビューによって最も賢い企業の50社にランクされました。
• 📝 scoro AIは、AI適応学習と大規模モデルに関する研究で50以上の論文を発表し、さらに50を書く予定です。
• 🔍 scoro AIは、カリフォルニア大学、スタンフォード研究インstitute、中国のアカデミーと共同研究ラボを設立しています。
• 🧠 AIエージェントは、オープンAIのシステムと適応システムの間で動作し、評価、推奨、予測などの機能を提供します。
• 📊 scoro AIは、学生の学習行動データをリアルタイムで収集し、学習目標のナノレベルの分解を使用して、学生の学習ギャップを特定します。
• 🌐 scoro AIは、中国とアメリカの両方のデータセンターを使用し、データの安全性とプライバシーを重視しています。
• 🏢 scoro AIは、中国にサービスセンターを持ち、米国市場に進出し、数学のコンテンツを今年、来年は科学などの他の科目のコンテンツを構築予定です。
• 📱 スマート学習タブレットとデジタルパッドを使用して、学生は手書きの数学解答を入力する必要がなく、システムはすべての学習セッションを内部で処理します。
• 👩‍🏫 scoro AIの物理センターでは、デバイスの販売と学生の自己学習の監視、データ分析が行われ、オンラインスーパーバイズも提供しています。

Q & A

スクーロAIのサービスは何を提供していますか？

スクーロAIは、子供たちに合わせた学習を提供しています。数学、科学などの科目があり、さらに他の科目にも拡大していく予定です。また、AIアダプティブ学習を用いて、学生が自分に合った学習パスを見つけることができます。

スクーロAIのプラットフォーム「L」とは何ですか？

「L」はスクーロAIの新しいプラットフォームで、大きなアダプティブモデルをシステムに取り入れています。コンテンツのビルディングが行われ、さらにハードウェアも提供しています。

スクーロAIのスマート学習タブレットはどのような機能を備えていますか？

スマート学習タブレットは、AIアダプティブ学習モデルを搭載し、学習コンテンツを提供します。また、ペンやスピーカーなどの周辺機器も用意されており、学習をサポートする機能が集約されています。

スクーロAIのビジネスモデルはどのようなものか説明してください。

スクーロAIは、中国には2,000人のナショナルディストリビューターがあり、60,000以上の公立学校がアダプティブ学習システムを使用しています。また、スマート学習タブレットは過去1年間に200,000台以上の販売実績があります。

スクーロAIはどのようにして学生の学習効果を向上させていますか？

スクーロAIは、AIアダプティブ学習を用いて学生が自分に合った学習パスを見つけることができます。また、教師の役割は、学習内容との直接的な対話よりも、監督とデータ分析に変わっています。

スクーロAIの学習システムはどのようにして学生の学習を個別化していますか？

スクーロAIは、学生がテストを行うことで学習状態を把握し、強みや既に習得している学習目標、未習得のギャップを特定します。その後、個々の学生に合わせた学習コンテンツが提供され、学習セッションが進むごとに学習目標が動的に変化します。

スクーロAIの学習データはどのように活用されていますか？

スクーロAIは、学生の学習データを収集し、学習行動や歴史を記録、リアルタイムでの分類と予測を行い、学習内容の最適化に活用しています。また、学習データは学生の感情的な表情や学習プロセスを理解するのに役立ちます。

スクーロAIの学習システムはどのようにして学生の間違いを特定し、フィードバックを提供していますか？

学生が解答を提供する際には、デジタルパッドを使用して解答を記入します。システムはこれらの解答を分析し、学生が問題を正しく答えられない理由に基づいてフィードバックを提供します。また、問題解決の方法を示すビデオも提供されます。

スクーロAIはどのようにして学生の学習を監督していますか？

スクーロAIは、学習センターでの自習を監督するとともに、オンラインでの監督も行っています。学習セッション後、スーパーバイザーは学生の進捗や学習状況をダッシュボードから確認し、必要に応じてサポートを提供します。

スクーロAIはどのようにして学生の学習効果を予測していますか？

スクーロAIは、学生が行ったテストの結果や学習行動をもとに、学習効果を予測します。また、学習データアナリストは、データに基づいて学生が今後どれだけの学習時間を必要とするかを予測し、フィードバックを提供します。

スクーロAIの学習システムはどのようにして学生の学習目標を細分化していますか？

スクーロAIは学習目標をナノレベルまで細分化し、学生が習得していない学習目標を特定します。これにより、学生は小さな学習目標に焦点を当て、より効率的に学習することができます。

スクーロAIはどのようにして学生が異なる科目間のつながりを学ぶことができますか？

スクーロAIは、異なる科目間のつながりを学ぶために、学習目標の間の関係性を理解し、クロス科目の知識を組み合わせる知識グラフを用います。これにより、学生は異なる科目間のつながりを理解し、総合的な学習を深めることができます。

Outlines

00:00

😀 公司紹介とAI学習プラットフォームの概要

scoro AIは、数学、科学などの科目で子供たちに応じる学習を提供する企業です。新しいプラットフォーム「L」は、大規模なアダプティブモデルです。製品はスマート学習タブレットで、ディストリビュータを通じて市場に販売されています。AIアダプティブ学習は、学生が自己学習の道を見つけるのに役立ちます。

05:02

📈 ビジネスモデルと実績

scoro AIは中国で2,000人のディストリビュータと協力し、60,000以上の公立学校を支援しています。スマート学習タブレットは、昨年に200,000台以上の販売を記録しました。また、技術的な進歩により、MIT技術レビューによって「最も賢い企業」のランキングにランクされました。

10:03

🧠 AIアダプティブエンジンの3つのレイヤー

scoro AIのAIアダプティブエンジンは3つのレイヤーから構成されており、目標、コンテンツマップ、学習経路、そして間違いの理解に関するオントロジーが含まれています。さらに、学習記録スター、リアルタイムインベントリコレクターが学生の学習行動を記録し、次の学習コンテンツを予測します。

15:06

📚 ナノレベルの学習目標とコンテンツ

scoro AIは、学習目標を極めて細かく分割することで、学生が非常に小さな学習目標を認識し、効果的に学習ギャップを埋めることができます。これにより、学生は短時間で効果的に学習することができます。

20:08

🌐 学習デバイスと機能

scoro AIはスマート学習タブレットをはじめとする学習デバイスを提供しています。これらのデバイスには、アダプティブ学習と大規模モデルシステムが組み込まれており、さらに他のテクノロジーやアプリケーションも搭載されています。これは安全で健康的な学習環境を提供し、親が子供の学習状況を監視することができます。

25:10

🤖 AIエージェントの構造とデータの活用

scoro AIは、オープンAIを使用してAIエージェントを構築しており、学生とのインタラクションをサポートしています。また、学生から収集されたデータを使って、学習体験を改善し、製品を時間とともに向上させています。データの安全性とプライバシーについても重視しており、国ごとのルールに従ってデータ分析を行います。

30:10

📝 数学学習の相互作用と物理センターの役割

scoro AIの数学学習では、デジタルパッドを使用して解答を記入します。これにより、解答を再入力する必要がなくなります。また、学習センターでは、デバイスの販売や学生の自己学習の監視、データ分析が行われます。オンラインでの監視も行い、学習の進捗や問題点をオンライン会議で話し合うことができます。

35:12

🏢 学習センターの機能とオンラインスーパーバイジョン

学習センターは、デバイスの販売と学生の自己学習のサポートを提供しています。また、オンラインスーパーバイジョンを通じて、学習の進捗や問題点をリモートでサポートすることができます。これは、学生が自宅で学習する際にも適用され、効果的な学習サポートが提供されています。

Keywords

💡scoro AI

scoro AIは、子供たちのためのアダプティブ学習を提供する企業です。数学、科学などの科目をカバーし、さらに多くの科目に進出する予定です。この企業は、AIを活用して学習を支援し、子供たちが自分に合った学習パスを見つけられるようにしています。

💡アダプティブ学習

アダプティブ学習とは、学習者の個人差や進度に応じて学習内容を調整する学習方法です。scoro AIでは、AIを用いて学習者の学習ギャップを特定し、個々の学習者に合わせたコンテンツと練習問題を提供しています。

💡Lエンジン

Lエンジンは、scoro AIが開発した学習プラットフォームの名称であり、大規模なモデルをシステムに組み込んでいます。このエンジンは、コンテンツのビルドやハードウェアの提供など、 scoro AIの学習システムの基盤をなしています。

💡スマート学習タブレット

スマート学習タブレットは、scoro AIの製品であり、学習者自身が学習を進めるためのデバイスです。ディストリビュータを通じて販売され、学習者が持って家で学習することができます。

💡自己学習

scoro AIのシステムでは、人間の教師がコンテンツと直接交渉する必要はなく、学習者は自己学習が可能です。教師の役割は、学習者の進歩を監視し、データ分析を行い、学習結果を支援することに変わっています。

💡MCM（モデル思考力と方法論）

MCMは、scoro AIが導入した学習目標の1つで、学習者が知識だけでなく思考力やスキルを養うことを目指しています。異なる科目からMCMの要素を取り入れ、学習者に対して将来のキャリアに役立つスキルを育成するようにしています。

💡ナノレベルの学習目標

scoro AIでは、学習目標を極めて詳細な「ナノレベル」に分解しています。これにより、学習者の学習ギャップをより正確に特定し、効果的に学習を進めることができます。たとえば、分数の加減算は大きな学習目標であり、 scoro AIではこれをさらに細分化して、より小さな学習目標に分割しています。

💡AIエージェント

AIエージェントとは、scoro AIの学習システムと大規模モデルシステムの間に位置するエージェントで、評価、推奨、予測などの機能を果たします。これにより、学習者が正確で有益な情報を得られるようにサポートします。

💡データプライバシー

scoro AIでは、特に子供たちのデータプライバシーについて重視しており、中国とアメリカの異なる規則に従ってデータの安全性を確保しています。学習者の名前や個人情報は一切使用せず、プライバシーを守ります。

💡デジタルパッド

デジタルパッドは、 scoro AIの学習系统中に使用されるデバイスで、学習者が解答を紙に書く代わりにデジタルパッド上で記入できます。これにより、解答がシステムに入力されるまでの手間が省け、学習プロセスが効率化されます。

💡オンラインスーパーバイズ

オンラインスーパーバイズは、scoro AIが提供するリモートサポートサービスです。学習者が自宅で学習しても、スーパーバイザーはシステムからデータを取得し、学習者の進歩状況を監視し、必要に応じてオンラインミーティングを通じてサポートを提供します。

Highlights

scoro AI is an adaptive learning company focusing on math, science, and expanding into more subjects.

The company's new platform, L, is a large adaptive model that serves as the engine for their content building and hardware provision.

scoro AI's smart learning tablet is a key product, sold to distributors who then sell to customers.

The L system facilitates student self-directed learning, reducing the need for human teachers to interact directly with learning content.

Teachers' roles have shifted to supervising and data analysis, aiding in student learning outcomes.

scoro AI has addressed challenges in traditional education by personalizing learning paths through AI adaptive learning.

The company has a presence in 60,000 public schools and has served over 24 million students.

scoro AI has sold over 200,000 smart learning tablets in the past year.

The company has been recognized by Deloitte as one of the top 50 fast-growing tech companies.

scoro AI has published over 50 papers in AI and educational conferences and is working on another 50.

The company has established joint labs with institutions like Stanford Research Institute and China's Academy.

scoro AI's adaptive engine has three layers: goals, learning records, and real-time data collection, with a focus on multimodal behavior analysis.

The company uses a large adaptive model that includes a multimodal education large language model and knowledge graphs for different subjects.

scoro AI has developed AI agents to provide evaluation, recommendation, and prediction services, ensuring accurate information for students.

The company's engine breaks down learning objectives into nano-level detail for precise identification of learning gaps.

scoro AI's system collects real-time student data, including emotional and behavioral responses, to tailor learning content.

The company has physical centers in China that sell devices, supervise learning, and provide data analysis for students.

scoro AI is entering the US market, developing content for math this year and planning to expand to science and other subjects next year.

The company uses open AI APIs for student interactions and has an AI agent structure that interfaces between the open LLM and their adaptive system.

Transcripts

[Music]

so yeah good morning and I hope you all

enjoy the sunshine here in San Diego and

my name is Pearl and I'm with scoro AI

so welcome to the session uh scoro AI

has been producing uh adaptive learning

for for kids we have subjects in math

science and we're going to Explore More

into more subjects so um I'll have

Jolene here making uh introduction

session to the

company hi everyone um it's a very

freestyle today's session so you know

it's called chat with squir AI so anyone

who is interested with like to have a

chat so the first half of the session uh

I will do some introduction about the

company the product the technology and

business model that we are doing and

then the second half if anyone has any

questions free feel free to ask so it's

really really

freestyle um so as per mentioned uh

scroll AI learning is the name of the

company so basically we use AI adaptive

learning from 9 years ago and then we

adapt

uh large model into the

system so the com um currently our new

platform is called L which is large

adaptive model so that's the engine we

have and then on top of the engine we

have the contents building and then

provide Hardware so the hardware uh we

will show later there's another slide

showing it's kind of ecosystem of all

different uh types of the hardware the

main one will be the smart learning

tablet and then with the Smart Learning

tablet is our product and then we sell

to the Distributors and then

Distributors will sell to the customers

so basically we are not in school

program so we are additional Learning

System for the

market and more important is the whole L

system is student self-directed learning

which means no no human teacher needs to

be teaching or interacting with the

learning contents itself but the roles

of the teachers have been changed in our

whole system which is supervising and

doing data analyzing for the students

and then to help the students um you

know the learning

outcome so I think um the first pages I

I can actually just skip most of the

people in this conference everyone knows

the difficulties at current um

traditional education like um the same

speed in the classroom but the students

actually having different

personalized uh learning needs and also

in some areas it's the same situation in

China us and anywhere um around the

world that we don't have ai sorry we

don't have Superior teachers and well

educated teachers everywhere maybe New

York maybe San Diego maybe you know La

maybe be Shanghai but not some rural

areas and also other countries

and and another thing is most of the

schools somewhere um they teach

students sorry they teach students of

the knowledge only but the more

important is the student need to be

trained with ability mod of thinking you

know their skills not the knowledge only

so with the current Cent difficulties in

traditional education in the classroom

even though with online learning we call

them traditional learning

already it it can't really customize

each different students learning needs

so with AI adapt learning that's why we

started nine years ago and this year is

actually our 10th year so our goal was

to change these current difficulties to

help the students to find the

personalized learning path when we say

personalized I know everyone here is

talking about personalized it's really

personalized so four students in the

classroom and each of you will be

learning your own learning path maybe

you are in the same grade your levels

may be similar but you know from the AI

adaptive and a large Modo system we can

find out each different VI have

different learning gaps and then the

system will recommend to you all the

different learning contents and

practices and problems and not at the

same level even though you might be in

the same grade so that's what AI

adaptive learning can bring to the

students so uh this is a page about our

business model and currently we have um

2,000 National distributors in China and

we've helped over 60,000 Public Schools

um you know using the Adaptive Learning

System and accumulated we have had 24

million students used our system and for

the intelligent devices which is the

Smart Learning tablet we've got over um

200 uh 200,000 pieces already sold in

the past year and we were one of the uh

top 50 uh fast growing at tech company

uh that was the research by Deo and also

we were one of the 50 smartest company

uh ranked by MIT technology review that

was few years ago and also uh one of the

owner we had

was awarded by UNESCO uh we were one of

the AI Innovation companies in

2020 yes 2020

yeah so this picture shows uh we've been

giving speech

and around the world's International

conferences and most important we've got

many papers uh published in AI academic

conferences and AI educational

conferences and now we have over 50

papers and we are writing another 50 so

next year we're going to have over 100

papers published it's all about AI

adaptive with large model uh research

and also we had uh a few joint

um Labs so in

200 um 19 I remember so we had joint lab

with car Mali University and in 2015 we

started working with uh Stanford

Research Institute and also one of the

um China's Academy joint lab and this is

something we were you know really proud

of in from

2 um I think 2019 we started uh we we

found a scroll AI Award with triple AI

and Euro AI so that award was

announced um

to the scientist who can bring out the

technology and the products that will

benefit for our Humanity so that award

was $1 100 million dollars a year and

also we were one of the Stanford NBA um

case study so um let me introduce a

little bit about our engine so this is

the engine about adaptive not adaptive

with uh large model so we've got three

layers

um so this is the first layer so on the

first layer of the Adaptive engine we

have the goals and the goals will be

dynamically changed for the students and

then we have the content map learning

map and mistake reasoning onology so the

content maps are learning objective

learning object itive videos problems

questions and

solutions um all sorts of learning

materials that is digitalized and very

Nano level for the learning objective

and for the learning map is each

different students have their learning

pathway so they have their different

Knowledge Graph and of course for the

learning objective we have the knowledge

graph for all the subjects and for the

mistake reasoning

ontology we because each when the kids

learn they always make mistakes but the

back um you know the background reason

of why they make mistakes are totally

different for example um the answer a is

correct and B C D they are all wrong

answers so when I answer B and you

answer B so the reasons will be

different because our knowledge graph

our knowledge state are different so the

reasons are

different and for the second ler we have

the learning record star which is to

recording um all the learning behaviors

and records and all the histories for

the students learning and also doing the

real time uh and classification and

doing the prediction of the next step

learning um

contents and for the third layer is the

real time invent collector so here we

are collecting not only um

right or wrong or learning speed and

also the multimodal behavior that is

generated from the real human uh the

real students so that is kind of

interaction receiving the data and

transfer to the second

layer and here is the large adaptive

model so after um like one and a half

years ago we started to do res R&D in

the large model so when we say large

model it's not large language model

because we have multimodal Behavior

Analysis so it's not only the

language and here is the framework of

our l so for the first layer we can see

is the application layer and then the

second layer will be the model layer and

the third layer will be the data

layer so for the uh for the application

layers we have the adaptive learning uh

system which I mentioned early and also

you know the first one is basically the

Adaptive learning layers um I mentioned

the three layers and for the model layer

we have this is the main one for our

large adaptive model so we have the

multimodal education L uh large language

model and also the knowledge graph for

the learning objective for different

subjects and also when there different

subjects such as science math they have

cross subject knowledge that components

that can be related to each other so not

only within that subject it could be

cross subjects for the knowledge graph

and also the rack and also the AI agents

so for the AI agents we have different

agents between the Adaptive system and

the large model system because the

agents can can do kind of

evaluation uh recommendation

prediction and that is very important

otherwise everyone knows that from large

language models there's always some

confusions and wrong answers and maybe

for entertainment for adults that's fine

but for students learning the results

the information that the system give to

the students need to be accurate that is

very important and for the data layer

some of the part is from the Adaptive

system which I mentioned uh the learning

map the content map and also the

learning outcomes from the students

learning uh across the whole system and

there's another one we called Nano level

learning objective I think uh there will

there will be a page about the Nano

level and there's thousands of um we

have had over one uh sorry uh 100

million uh learning Behavior data in our

system and here is the engine for the a

agents so we have different agents which

I mentioned early um yeah you can take a

picture and then I will go to next

page so this is the autology of the

learning map learning goals and content

map which is the first layer so this is

the one I just mentioned the Nano level

breakdown of the learning

objective um we were actually at that

year we were the first company who did

really really Nan level for example if

the

normal numbers of the learning objective

from uh subject like common call in us

and common call in China is like 300 and

we other companies would uh have breaken

them down from 300 to 3,000 and what we

have done is from 3,000 to 30,000 so

it's goes to really really Nano level so

why do we go so Nano level because

it's very precise then when we when the

system recognize the students learning

Gap it will be really really tiny

learning objective then the system will

do root tracing to find out that

learning Gap then the students can only

um you know learn that Nano level of

their learning objective and once the

students mastered and then they can come

back so instead of a bigger learning

objective we use Nano level so that will

save lots of time to the students

because sometimes we find this big

learning objective the student cannot

understand cannot Master cannot answer

the question it's not because the whole

learning objective because this learning

objective can be break down into say 50

Nano level it could be 20 out of the 50

of the learning objectives the student

didn't master in the previous years

which means the students only need to

learn those 20 Nano level L then the

student can come back it will save lots

of time for the students so for example

addition and subtra subtraction

subtraction of fractions this is a

bigger uh Big L O and then we can break

down you know to second layer and then

to the third layer so uh from three

layers to nine layers that's the average

layers we've gone it really depends on

which learning objectives

so this is just one example and this is

a system we collect the students

learning data at real Time That's the

third layer of the Adaptive

engine and as I mentioned early we do

have uh learning goals that is set up

for the students once the students has

started learning and doing some testing

in the system we can have the goals and

the goals will dynamically be changed

along with the students learning

progress and also the system can do

prediction for the for the students for

example once I learned 10 hours in the

system and then from the the data

dashboard the data analyst which is the

human teacher can do a prediction based

on the data saying hey Jolene you you're

going to need another 50 hours and then

you can catch up for your current level

Uh current grade in your classroom so

that's prediction we can see from the

screen and then the prediction might be

changed because it really depends on how

long I'm going to use and how effective

I can learn so that will be prediction

for the students the students will kind

of have a plan how long I need to learn

to manage my uh learning and also doing

the recommendation of uh each different

students oh this is one example I want

to show so you can see the graph the

first line is the great sorry the first

line is grade 10 and the bottom one is

grade seven for example I'm a grade 10

students there are some learning

objective that's in yellow dot that I

didn't master so I can't really do some

you know test or answer some questions

or I can't really manage some of my

assignments so why can't do it it's

probably because some of the learning

objective from grade eight or grade

seven I didn't Master at that time what

I need to do is I need to learn those

knowledge Gap so the system will

automatically find out the learning

objective the gaps from grade seven or

eight all students have different gaps

and then I need to go back to those gaps

and learn and master and then I can just

automatically come back back to the

current grade 10 learning

level and there's another example of

three different students we can see the

uh the blue learning objectives are the

ones the students have already mastered

which is 80% mastered and the pink ones

are not mastered which is the gaps and

another students similar level same

grade but the gaps are different the

pink ones are different and the third

one the pink ones totally different

which means we need to have older

students even though in the same

classroom they should be learned

differently so our system is designed

for students not for teachers to teach

it's for students to learn so they will

be using the system learn

differently and this is another uh

element I mentioned at the very

beginning traditional Learning Classroom

actually especially in China and maybe

some Asian countries um the students

learn knowledge more more than some

skills from the classroom so we have the

system called MCM which is model

thinking capacity and methodology so

what we do is we divided those learning

objective from different careers for

different MCM learning objective or we

call MCM elements I won't go really

details and those MCM elements can be

from language subject can be from math

can be from physics can be from biology

it can be from any subject so we need to

do the analysis of which subject can

create those mcms and those MCM anims

will have connections with the careers

so that's kind of connection for the

students you learn the knowledge and

also at the same time you learn the

knowledge you can train your mcms and

those M MCM elements will be uh useful

for your future

career and this is um the the product uh

we have so in the middle is the Smart

Learning devices uh Smart Learning

tablet and we also have the pen and uh

the speaker and so on and sorry this is

the price I just want to show uh what we

have different types of the Smart

Learning

devices and those are the fun functions

for the learning uh tablet so we have

the Adaptive with large model system in

the in the learning tablet but we also

have other additional technology and

applications that is useful and even

games for the students it's a close

ecosystem in the learning tablet so it's

very safe and healthy for the students

to use and it's also very good for the

parents to kind of have monitor but

without saying too much but is kind of

monitoring understand your your young

kids what they have gone uh which uh

application they have used and also

another thing is they cannot go outside

of the other uh websites that is not as

restricted and this is the service

center uh we have in China and how how

it looks like we have the uh counters

that to sell the products and also we

have the self-directed Learning

Classroom for the student students to

learn so um this year we are entering US

market and we are designing the contents

for math this year and then next year

for Science and for other subjects so

what we need is we are building the

contents first and then build uh and in

integrate those contents onto our large

adaptive model engine and then the

products will be ready and then we are

going to expand our um centers

later on after the products uh is

pronounced is announced yeah okay thank

you thank you very much and maybe per

you want to come here and answer some

questions if any

yeah yeah I have like a million

questions sorry hi can you hear me okay

can you hear me okay so my name is Dylan

and I'm coming from the Atlanta area and

I just wanted to say first and for

foremost that this is a very very

impressive Enterprise I am blown away by

the amount of time it took to build this

the second thing I'd like to say is you

probably felt so lucky with changes in

the AI landscape because you were in the

perfect position to take advantage of

this type of technology so wow how

fortunate I have many questions for you

and I'll just go with down on the stack

and you can just take this microphone

away from me whenever you're ready okay

so uh this is a question for either of

you um that is what is the uh Primary

llm in your stack is it a custom

fine-tune llm that you use for

interaction with the

student that's part of it do you want to

answer or okay yeah it's it's part of it

so

um yeah I just turned to the l l m page

so it's very important that interaction

with the students so when we talk about

interactions there are different

interactions so one interaction is when

the students is using the Smart Learning

tablet to learn it will generate many

Behavior many data that is not talking

to the system interaction is the

students learning by themselves

interaction that's one interaction and

another interaction is in the system we

have a we call it a AI virt to um

assistant but that assistant is not

to teach the system is to answer the

questions for example um I learned this

I've learned this learning objective I

will say hey um assistant I do have a

question why I got it wrong right and

the assistant will tell you oh maybe

because your prerequisite L was not Mar

so you have to continue learning uh

don't give up you're going to be good

it's more like

emotional encouragement assistant with

the student so that's another

interaction the student really talk to

the system so for the conversation that

little bit of conversation what is the

are you using an open source llm for

providing that communication or are you

at Liberty to talk about it are you

using it's not open source we use open

AI of of course yeah okay got it so

you're using like API calls to chat gbt

4 or yes and but between that there will

be an AI

agent okay between our system to open AI

yeah so I have a question about AI

agents so how are you actually

architecting the AI agent structure are

you using something like autogen or crew

AI or do you have your own system for

creating these llm based agents

we don't have our own system for lrm

because we are not you know that kind of

company and that will requires lots of

funds money to to to uh to train so we

just uh we what we build is our AI

agents so that agents is kind of

isolated from um open AI system but it's

in between the open llm and with our

adaptive system so we do evaluation

classification and and many other you

know functions so that's what we do so

we don't have our own

llm okay yeah um so my next question has

to do with uh how you use data from

users to improve experience and improve

the product over time you are collecting

a lot of Behavioral data you're also

collecting a lot of data over where

students are making mistakes again and

again so how are you using the data that

you ingest to improve the product and

experience over

time yeah that's pretty deep uh

questions from the tag side from the AI

side um so we have let me show you this

um here where is

it the Adaptive that is here yes it's

more like the second layer and the third

layer here so because

we Sorry

here yeah so we collect the data on the

second

layer all the data comes into the system

and the algorithm will need to do

calculations so when we when we say

calculations for

example um we

capture the students emotional facial uh

expression that's one

data

if for example two students right and we

all answer this question and correctly

within three minutes for example but my

facial recommendation was kind of very

happy and concentrate and when per was

doing that she was like so Wonder around

you know that's different data the

system collect to analyze our different

learning behavior and also the system

will need to based on the prerequisite

previous learning uh history on and

learning behaviors and also our

different knowledge states of

understanding the whole chapter of the

learning objective and there's

relationships between each different

learning objective and our learning

history so that could be one part and

also for example for math when we answer

some questions there so the students

need to write down the solutions why you

know the step by steps and we have a pad

so they can write on the uh pad and then

they will be captured as a picture so

even though we both answer that question

correctly but the learning the solution

passway is different between us the

system will understand differently so

all sorts of um f Fe back will be taking

into con account concern right to to

analyze different students knowledge

stats and then for the next step of the

learning we going to have different mhm

yeah learning content to be you know

recommended to us yeah yeah I think I

can have some ad on here first of all we

have two separate servers for the data

for the Chinese students and here in US

students

we're concerned about the data safety

and we follow all the rules for

different countries and also when we

analyz data uh especially for kids and

under age we two countries we have

different rules right here and for China

or but there are some common common

characteristics for both countries of

course we will uh eliminate and covered

all the students name

and their uh special about their privacy

so that's the top priority we always

care in our company we always prioritize

what's the safe also we are also aware

the um we we emphasize on the education

side even though L they have lots of

Advantage as we hear all over across

this conference for this year we are

aware how should we adapt the modern

technology in the use of Education we

want we don't yes we want be ahead of in

the technology to use the most modern or

to use the most State ofth art to help

our students but we always want to

discuss with our peer companies with

peer um what we need to concern first

about our cats because it's about um uh

uh young kids not adults so we

are I think we want to contribute our

effort and we want to share our

technology with all the companies and um

to talk about to discuss to open

discussion with how can we help the kids

while we we are very concerned and keep

everything under control

right like quotation mark under control

for for Education sure well thank you

for that answer Pearl um so I've used a

system that's analogous to this so it's

not your system but I used a system

that's analogous to this that uses the

Adaptive approach plus Ai and for math

learning and I worked with it for about

six weeks and one thing that I noticed

was a shortcoming of the system was that

I had to constantly take my handwritten

math Solutions and type them in to the

system and if I didn't show my work or

show my stages of solution the system

couldn't help me with that you say that

you have a multimodal system I'd just

like to learn more about how the actual

user interacts in a math lesson with

your product and service okay

um we we actually don't call our systems

uh contents as math lessons because when

we call them lesson is more like um

classroom atmosphere so in the in our

system there's no lesson uh it's

like let me put in this way so the

students log into the system first what

they need to do is kind of testing the

testing won't be so long it could be

like 10 to 30 questions it's really

various for different students even it's

the same sub uh same chapter so after

after those little um tests the system

will understand my knowledge State and

what are the strengths what I have

already mastered and what are my gaps

and then I might need to be going back

to the previous uh learning objective to

master and then move step by step so if

for maass subject um after the test it

will be some learning contents started

to recommend to me so we call that

session is learning session the learning

session is I need to learn the

previous uh gaps which I have so that

will be pop up like um the system will

pop up some like three minute three

minutes around videos for each n level

of the learning objectives that I didn't

master in the previous levels and then

after this learning Maybe 20 minutes and

then the next St will be

practices practice of uh answering some

problems questions doing exercises and

then during that time there will be some

calculations or Solutions uh that I need

to write down so we have another pad

which is digital pad the students will

just write on the pad not on piece of

paper because if on piece of paper they

will need to be retyped into the system

so all the answers when students does

when students do is just type on the

keyboard and then the answer but when

they do the solutions there will be a

pad and writing on the pad and then will

be saved and captured so that data is

calculated into the system so there's no

extra like paper things that will need

to be retyped in the system and after

the practices the next session will be

the system will evaluation of the

students

learning uh outcome like the result the

result is not the score result but if

the student has mastered the learning

sessions all the learning objectives by

doing the

exercises maybe yes maybe not maybe

partial I don't know and then the next

one will be if the students has

successfully uh mastered all the

learning objective and the student can

be moved to next level of the learning

objectives learning otherwise it could

be kind of uh relearned but when they

relearn uh it will be different contents

even though it's for the same learning

objective but the contents of the

learning will be different and if the

students when they do practices there's

some wrong answers and there will be

popup of the problems solution

recommendation and Analysis and it's

kind of an the teaching uh for the

students how to Ender because the reason

the student cannot answer the problems

correctly could be understanding of the

problems it could be some other things

and this you know that video will show

some um analysis for the students answer

and tell them how to improve you know

this is kind of so everything is done

within the Smart Learning devices it

won't be any outside of like paper and

re retype so it's really student

centered um system it's either the

learning tablet or SAS system which is

the learning account uh based on just

Internet it's all done with some devices

it won't be any um external inputs thank

you additional external inputs um so I

have one last question and then I'll

hand the mic off three minutes yeah last

question is um could you go into detail

about the role of the physical centers

okay

I'll try my best three minutes so in the

physical centers two functions one

function is um the distri the owner of

the centers will sell devices yeah in

China or if it's not Hardware it could

be just a learning account and another

function is to supervising and doing

data analyzing for the students who are

in the centers doing their self-directed

study for example we are in one

classroom from different schools and

different sub subjects and different age

we can just sit there in our own desk

and learning by our own and the teacher

or the supervisor is somewhere else and

monitoring us and there's a dashboard

for the supervisor and the supervisor

can monitor each different students

progress and learning if there's any

like alert like joling is frustrating or

they can the supervisor can also monitor

our facial recognition right what

happens if any student has any concerns

or problems or need support the

supervisor will you know come in and

talk to me and maybe drag me to another

classroom and have some uh uh

encouragement or if are all okay and

then after this like one hour uh I

finish I need to go home or I finish

this session of the learning the

supervisor will have a talk with me as

Jen this is the report we've got

generated from the system I can see

strength weaknesses learning speed blah

blah blah there's so many details about

the report and then maybe doing some

encouragement uh so next step what you

need to do is learning uh you you know

the root tracing system will recognize

you have some other gaps or maybe you

need to move forward you know that kind

of discussion and then I just go home

maybe that's for the centers or if the

students decide to learn at home that's

fine and everything is done at home but

there's no supervising at home the

parents cannot monit supervising at all

right but what the supervisor can do is

uh you are learning from home but I can

also see the report and all the data and

learning behavior from the system I can

call you and we call it online

supervising it's that after you study I

can call you and say let's have a zoom

meeting I found out some questions and

good things about your learning and then

you know you can doing the same things

through the zoom meeting so we call it

online supervision sorry 30 seconds left

yeah I think our session is almost there

but we have another session in the

afternoon so yeah more more than welcome

to come back and drop by we we also want

to talk to you like we like to ask for

advice and suggestion from you as well

all right thank you everyone thank you

[Music]