On 20 July 2023, Chen Haidong, director of Solution Architect,Alibaba Cloud Intelligence, participated in the keynote of this Midsummer Open Source Night.The topic shared was that open source technologies to protect the ownership of personal data.Here's what he said.

2023年7月20日，阿里云智能解决方案架构总监，陈海栋，参加本次仲夏开源之夜的主题演讲。分享的主题是：保护个人数据所有权的开源技术。以下是他的演讲内容。

Haidong (David) Chen focuses on Ecommerce Industry Solution and Compliance, Cloud Native, and Data and Artificial Intelligence solution. As Daraz's former infrastructure director, he migrated Daraz from AWS to Alibaba Cloud, and leveraged the PAI platform to extend gitops to MLops.

陈海东专注于电子商务行业解决方案与合规、云原生、数据和人工智能解决方案。作为 Daraz 前基础架构总监，他将 Daraz 从 AWS 迁移到阿里巴巴云，并利用 PAI 平台将 gitops 扩展到 MLops。

演讲全文

Hello everybody，so this is Haidong from Alibaba Cloud.Welcome and let's get started.

In the case of open source,everybody wants to  just stop talking and show me the demo.There will be lots of codes today,so feel free if you are very interested and just keep full.

▲ Haidong (David) Chen

So first,let's have a brief over what is generous of AI.Previously we are having that discriminate AI,which means we can do sentimental analysis,we can understand if you're happy or not,we can do the object detection,like Yolo,to understand in this picture this is a human or this is dog,but right now we are being able to generate contents like emails,pictures,compose music.

In a data world,you need a data engineer to get a data,you need data analyze what is in the data and you need data scientists to understand the conversion rate to understand what we can do,to boost yourselves to help promote your products.But right now in AIGC,we can directly chat with your database,so you no longer need that long chain of data team,you can just talk directly with your data,there will be a demo just to understand this and after sales event this is quite when you hear,so many scam costs technology we are using for the outselves events,so we engaged customers with 24-7 and also with the sentimental analysis,so you could be able to feel the emotion of that.

▲ Image source: PPT presentation

Finally,search and recommendation.If you're a local user of Lazada and you screw up this there will be a Lazada chat called Lazi Chat,and you can just directly talk with it,for example,if you know what to buy,you can just go to search,but if you don't know what to buy and tomorrow is your mother's birthday,you are wondering what's my mom  be like,what is her favorite birthday present,so that's something the search bar or the recommendation service condition.Traditional doesn't give you,but ChatGPT or our lazi Chat can give you as recommendations of a female who would be love and that's something like,this would give us the generative AI.So much for the talking.

▲ Image source: PPT presentation

Let's look directly at the demos.For the text to image, there will be art design,entertainment,brand awareness,architecture design and user interface,these are the demos of what we can generate.And this is the demo of how we generate.

so,you can see here,you put the prompt here,and you will specify what is the pixel size and how many you want,you can just generate one batch of four and pick whatever you like.These are the models can be selected and be customizable.For example, if you love a Laura,a web 3 level or a cute cartoon icon,you can use the one demo and one model,but if you are more interested in like real life,you can just change to another demo and another model.I think you can see it will take up sometime and to generate all images.So close up a portrait how you specify the picture.Is it a close up?Or is it a long range portal?Or is it far away?and then Asia Princess and finally the setup of the picture.Is it detailed?Or is it in a cinematric and detailed on face?So in this way, you can specify how you categorize these pictures. This close up is usually used in a scene of the cartoon image, so something how you can have a prompt library and something we are embedded in your prompt making.

▲ Image source: PPT presentation

This is an e-commerce platform,for example,if you are in FMCG product company,you have millions of product every year and you will be thinking that I have a new product and I need a new set of marketing people to get my marketing materials ready,however,if you are using this who can bring you a new way and efficiency.This is your previous old product,like poster,and this is your new product,and then we can merge them together using your existing to design assets and your new product's picture to generate a new existing marketing assets and your new product photo.So in this way,you can see all the design can be changed and all the service the detects on the picture can be changed.It's an integrated marketing portal,so you no longer need some so much, like marketing human capital,and you can do the high level design and high level of efficiency.

▲ Image source: PPT presentation

How we achieve this in the bottom layer we have the PAI-Algorithm Platform?If you have your own model or if you have your own data science team,you wanna train your own model,you can use the bottom layer,we provide the platform and email ops,like the whole tool chain,what if you want just API service and web portal service,just as what we saw just now and what we can provide.We can provide API service to wear a web portal service to give all the client,to give the previous day,the work of data scientists,to the work of data engineer,to directly the marketing people the stakeholders,so they can directly use the AIGC content to boost their marketing events.

There are also some contents that we showcased here.As you can see,in the first picture,the girl is holding the candle in her right hand,and now we are shifting to her left hand,this is what AIGC can do with this model,and with this model we have already the girl wearing the dress.And now we change the dress into pants in this model.It's a changing of style and closes,for example,Lora is a model which is quite popular in open source community,previously is the kind of dress,later on,we change to the kind of dress and here is the ControlNet.ControlNet can control the outlook of the picture to make sure the outlook doesn't change,but detailed in terms of style and layout can change.

▲ Image source: PPT presentation

So in a summary,what we can do and we have in advantage,comprehensive features of API and Web UI,so doesn't matter if you're just a user,anyone can use and cost an efficiency.

Now let's see.What we just mentioned in the retail in industry,we wanna know how we can close up the loop of data scientist,data engineer and data end list,what if you know this morning CEO say,"I want to understand this year this monster's sales data",and nobody can provide all the data,teams say"I need 1 month or one week to provide",but in this case you no longer need to take a e-commerce recommendation to use case example.For example,in e-commerce company in Singapore region,we identified as a type of clothes which is selling very great,this clothes style may be a dress.However,we want to see if this type of dress can be sell also good in Malaysia region,and this is something we want to recommend to the major uses.But how can we verify our idea?How can we get the data?What kind of dress is very popular here?

Previously we need the data engineers and right now no longer.As you can see here we hope.Hook up directly to our postgrass database and then ask directly,"show me the top-3 districts with the top sales records"and this is the SQL automatically generated.After a generated,we can just run directly to the database.So,you can see the results come out Asia,Europe and Africa region,it has the sales data and this is exactly what we want to showcase with you.The language model not only can chat with you,but also can generate code generate sequel,and this sequel can directly plug in back to the database.So in this case,the language model not only needs to be able to chat or connect with you,but also it needs to know what we have,what data we have,what are the data tables,what are the fields we have,and that's something we do.We integrated everything together to our Alibaba car,but also it's available open source,you can also do the connection from the open source to your database.And as you can see here,the sequel generation from a group buy,from select,from all complicated sequel operators,it can support all,and we want to do more,emphasize on this.

As you can see here,we can do group by all the limit where select,so mostly we cover all of them,and this is the sequel which is generated,and this is a result which is a run a run result,if you have doubts on a sequel it just go change directly,and then the results will change as well.

▲ Image source: PPT presentation

So not only we can chat with our database,but also we can chat with the documents.

This is also a demo of ChatwithPDF，we can just go upload your PDF directly and then it will pass the content of the PDF.In this case,we are not uploading this data to the language model,you can have your own database,so you don't have the risk of data leak or data privacy.This is a document describing what is ECS of valuable cloud and you can just ask,not only can ask give you the answer but also can give you the reference to make sure it's an accurate answer.There's a risk of AI illusion which means it will give you random answers or answer,which doesn't make sense he made up,so this reference help you to check with what is the original document say and is it the correct way of answering this.

It's also as you can see here. We emphasized here on the quote reference case and it's also available on get up.So feel free to copy and paste to get pulled.

▲ Image source: PPT presentation

We want to introduce our large language model on bio information.In this model, what we do is we can input the bio code of atcgaucg the protein series and we used the language model of sequential language model to generate and to predict the street structure of the protein.In this way, it's very similar to the work of AlphaFold.

This is how we generalize from LUCA to our bell technology,the bell diverse world.In this bell diverse world we are all these proteins generated,derived from the LUCA,and that's something which was very fascinating and that's something which we are working on.After we know the protein's sequential data,then we can predict the structure through this structure of the protein.These 3D structure of protein finally will define the outlooked appearance of,if you have a green eye,a blue eye or a red eye,if you have a big nose or small nose,that's something that's very fascinating and also it's very much used in the drug industry,so for the new drug invention or will be very much boosted by via the 3D structure prediction.

▲ Image source: PPT presentation

This is our official guitar,and this is the way we make this happen.First we have the data of sequential data and structure data,structure is output,sequential data is input,and then we do cross validation,and then we figure out the highlights,and then this is our am modeling we have by LSTM,and we have all this large language model embedded,finally we can predict the structure.After we have the structure, we can know the features of it to identify the potential risk of the new protein or the new drug,as for the successful use case.

▲ Image source: PPT presentation

As you can see,for Lazada Lazi Chat,it's already online and everybody can try it with your phone.The Lazi Chat is available with the newest version of Lazada,and then the RAS is also available.For example,as you can see here,in this case of the RAS,what do I need for my suitcase for the upcoming visit of Dhaka.Dhaka is a city of a Bangladesh,in that area it's mostly sunny,that's why you need a hat and that's why you need also some normal travel cases.This is some kind of domain knowledge which we don't have,and this is also the value of the e-commerce platform which can give you as a domain expert.How we deploy?This is the key question.

▲ Image source: PPT presentation

After we have described so many stuff,if I'm lazi,if I don't want to go through all the effort of deploying this,then you can just go to Olibaba Cloud,go to the Computer Nest,and click official,just one click and then you can have the URL everything deployed.After you access the URL what you will see.If you use stable diffusion,like text image model,you can just plug in prompt care the picture.If you are using the ChatwwithPDF model,you can also just go into the endpoint and directly start using.So it's very simple just 3 steps:press step,computer nest;second step,wait for the domain to be ready and click in;and final step,just having fun with it.

Partners

About KCC Singapore

KCC Singapore, founded on July 20, 2023, is the first step in the open source community's global strategy. Our mission is to empower developers to embrace and contribute to open source. Through partnerships with universities, tech companies, and government departments, we aim to promote open source adoption in Singapore's digital economy. Working closely with local open source communities and forging global connections, we amplify the voice of Chinese open source. Together, we empower open source for a brighter digital future.

▲ From left to right: Alfred Wu, Shu Min, Wang jianmin

点击“阅读原文”获取PPT

编辑丨李楠

Night Of AI｜KCC@新加坡 X 早早聊喊你来参加 AI 之夜啦～

仲夏开源之夜演讲｜ FOSSG 联合发起人 max:解读人工智能和开源技术的最新热潮

KCC Singapore 助力首个出海的 AI 顶会 | GAIR 2023 议程公布