news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

New Intelligence Report

Editor: Editorial Department

【New Wisdom Introduction】The entrepreneurial husband-and-wife shop hired AI to do the design, and the monitors became the No. 1 brand in the industry with a huge sales volume. Behind this, it is not only the confidence given by the overseas business of 300 million users, but also because a team focused on doing three things well.

The recent news about self-driving taxis has sparked a heated discussion across the Internet - has AI taken over human jobs?

The answer is no.

You know, in the real world, most small and micro businesses are actually facing the dilemma of talent shortage. In this case, AI not only does not replace anyone, but instead provides a steady stream of talent for small and micro businesses.

As the industry that was the first to be exposed to AI and has been most transformed, cross-border e-commerce has now become a huge market with a scale of 10 trillion yuan.

So, when it collides with today's big model, in what aspects will it be reconstructed?

When I tried to build my own overseas business, it was an eye-opener.

One person opened a store and sold it to the United States.

Step 1: List your products

First of all, there is a problem at the step of putting the product on the shelves.

How can KV designs for different styles of clothes be eye-catching? What kind of models should be hired? How to translate introduction pages for different countries? The combination of these options will inevitably lead to the best solution.

The huge expenditure behind it is an unbearable burden for ordinary businesses like us.

Take the cost of e-commerce photography for example: In China, it costs at least 200 yuan to make a set of product photos. If you want to hire a European or American model to shoot, it will cost at least several thousand yuan. After the shooting, you have to ask the artist to retouch the photos, and you have to spend an average of 8,000-15,000 yuan per month.

This is not all. If you want to sell a product to the global market, you have to log on to different platforms and also need localized expressions of language and materials.

After completing a set of work processes, not only is it time-consuming, but people are also exhausted.

But these problems are small cases for AI——

The following pictures are all generated by the picture tool pic-copilot produced by Alibaba International: https://www.piccopilot.com/zh

Batch Cutout

Before putting clothes on the shelves, we need to have the main picture of the product.

Faced with a large number of women's clothing, how can you produce main pictures of multiple pieces of clothing in batches without any effort?

With the power of AI, 20 photos can be cut out at a time.

Whether the background is the floor, the desktop, or a dazzling carpet, AI can accurately remove the background with one click.

With these floor plans, we can not only apply other backgrounds as homepage displays, but also let virtual models use them to present different effects.

Virtual Fitting

Next, just select a piece of clothing, then choose the desired model and pose, and AI can quickly generate the upper body effect.

It can be seen that the rendered clothes are highly consistent with the actual ones in terms of texture and style.

Model skin change

Of course, for those outfits that have already been photographed, you only need to "start with a picture" and leave the rest to AI.

If there are areas that need fine-tuning, you can also use the tool to manually select the area.

Next, choose an AI model that matches your potential audience.

And after determining the hair color and background.

You can get the dressing effects of models of different genders, ages, and skin colors with just one click.

It can be seen that only the model itself changes, and the selected clothes and bags will not be affected in any way.

Even the sunglasses that were not clearly marked were perfectly preserved by AI.

In this way, we only need to sit in front of the computer and click the mouse to put all the products in our hands on the shelves, which greatly saves the high cost of model shooting and post-production.

According to statistics, with the help of AI virtual try-on, not only the sales volume of products can be doubled, but consumers' preference for products has also increased by 100%.

Title suggestions

How to use product titles and descriptions to attract readers' purchasing desire is also a big deal.

AI trained with massive e-commerce data can easily generate product titles.

All we have to do is post all the information related to the product, and AI will automatically eliminate useless content and quickly generate high-traffic titles based on category characteristics and user purchasing decisions.

Step 2: Place an ad

Next, it’s time to place ads.

An outstanding advertising image has mysteries in design layout, creative copy, and background image, which will have a real impact on the effectiveness of advertising.

In the past, these ads were basically produced by designers manually, which not only took a long time and was costly, but the effects of the materials were not necessarily ideal.

On the other hand, our operational space for the placement of activation advertisements is also very limited.

Because most activation ads are delivered in the form of a dynamic "product pool", and a product pool contains tens of millions of products, it is impossible to optimize the materials by manual delivery alone.

With AI, everything is very different.

Generate ads with one click

AI, supported by large models, can quickly generate large quantities of advertising materials in batches, solving the bottleneck of insufficient material production.

In terms of text, a large model for generating multilingual creative copy is trained based on massive product data, which can generate more personalized and attractive selling points based on the product information, country of sale, and target population.

In terms of pictures, AI learns and abstracts the designer's design paradigm based on massive e-commerce advertising materials across the entire network.

This includes the placement of a product and text, the alignment relationship, product background, and design element style matching rules.

Once the input design content is given, we can let AI adapt the typesetting and layout and synthesize creative pictures in real time.

As a women's clothing store, it's not too much to sell shoes (manual dog head)

Precision delivery

After the advertisement design is completed, AI will also score its own generation effect and deliver high-scoring pictures in a targeted manner.

Finally, through historical delivery data, AI can also summarize experience, for example, products with text selling points perform better in advertising channels.

For those product pictures that do not have high-quality selling points, AI can be used to extract selling points.

Based on massive amounts of advertising effectiveness data, Alibaba International has specially trained a multimodal creative scoring model.

It can not only estimate the delivery effect, but also dynamically adjust the layout according to actual conditions to align the optimal delivery effect.

In the above process, there are multiple AI capabilities such as multimodal recognition + multilingual text generation + AI image processing + delivery effect RL. As a result, the advertising production cost was reduced by 3% and the advertising ROI was increased by 5%.

Step 3: Store Operation

In terms of store operations, AI can help merchants better understand their customers by analyzing their preferences.

Respond to customer reviews

At the same time, it can provide customers with personalized responses, so that every customer comment can get accurate and thoughtful feedback.

In response to this user's comment "cheaply made", the AI ​​will reply: "We know that our cheap prices may not always meet your expectations. But we have always been committed to providing high-quality products at affordable prices."

Real-time chat translation

No matter what language the buyer uses, we can rely on AI to translate the conversation smoothly.

Even if there are spelling errors, AI can automatically correct them

Intelligent customer service

For customers, intelligent customer service is a natural fit for AI usage scenarios.

During the conversation, by calling product APIs and logistics APIs, AI can answer most of the user's questions about products and logistics.

Step 4: After-sales service

The last and most difficult part is after-sales service.

Faced with consumers' requests for unconditional returns and refunds, the cost would be very high if the goods were shipped back from overseas every time; but if the goods were not shipped back, the damage to the goods would be very serious.

In this scenario, AI can be used to negotiate partial refunds and no returns, thereby reducing the return and refund ratio and recovering losses.

Specifically, the model will first learn how to handle various historical manual judgment cases during training.

Then, with the help of multimodal technology, we analyze data such as user messages, refund vouchers (the degree of damage to the goods), transaction snapshots, logistics routes, etc., to understand the cause of the dispute and calculate the amount of partial refund.

Finally, we provide consumers with a more satisfactory "partial refund without return" solution.

In this way, it is possible to improve refund efficiency, reduce refund losses, and optimize customer satisfaction (estimated to reach 3%).

In addition to returns and refunds, there is another more troublesome dispute - Chargeback.

After encountering a malicious payment rejection, it takes merchants an average of 20 minutes to complete the appeal materials, which are then reviewed and supplemented by the platform staff.

In contrast, Chargeback Agent uses the multimodal capabilities of the model to understand relevant information such as orders, fulfillment, logistics, goods, and reasons for defense, and then automatically collects and assembles evidence to ultimately generate detailed defense materials with one click.

It is understood that this AI Agent can help cross-border merchants recover tens of millions of yuan in losses in a year.

In short, from product listing to marketing, store operations, and after-sales service, AI has been fully integrated into the entire chain of cross-border e-commerce.

AI e-commerce enters its second year

People say that 2023 is also the first year of AI e-commerce. After more than a year of development, AI e-commerce has long entered the mature application stage from the hype stage.

AI e-commerce has become one of the core application scenarios in the era of big models and is also the best testing ground.

The reason why this field is becoming increasingly hot is, firstly, that the scale of global e-commerce is already large, and secondly, with the increasing number of AI e-commerce applications, its implementation path is gradually becoming clear.

A report from Goldman Sachs stated that global e-commerce sales reached US$3.6 trillion in 2023, and is expected to grow by 8% year-on-year in 24 years and reach US$5 trillion in 28 years.

It can be seen that the AI ​​e-commerce track itself is a place worthy of deep exploration of innovation and the release of huge commercial value.

From e-commerce operations to supply chain and consumer end, the power of AI is penetrating into every link.

On the operational side, e-commerce platform giants such as Amazon and Alibaba International have launched AI tools for merchants.

For example, Amazon’s AI Listing feature helps sellers write attractive copy more easily and efficiently.

From the supply side, the rise of AI has accelerated the process of innovation for some companies. For example, jewelry companies have to go through a cycle of at least several months from design, mold making, model shooting, product testing, and new product launch.

Now, the workflow for companies to launch new products has been reshaped by AI, from design, to AI-generated renderings, eye-catching copy, product testing, to big data analysis, and even AI can create a more efficient matching mechanism.

Merchants have already skillfully used these AI tools to generate copy, handle translations, design product images, etc., greatly reducing a lot of repetitive and tedious work.

While reducing costs and increasing efficiency, it also better caters to consumers' purchasing psychology.

Another Deloitte report confirms that 26% of marketers surveyed are already using GenAI to generate marketing content, and 45% plan to use this technology by the end of 2024.

Not only that, the impact of AI on consumers is also revolutionary.

Amazon released an AI review integration function that extracts past buyers' reviews and summarizes them in a paragraph and pins them to the top. The AI ​​fit function allows users to try on clothes online.

In February this year, Amazon’s “e-commerce version of ChatGPT” – Rufus was released, which can help buyers make decisions in the form of questions and answers.

Obviously, domestic and foreign e-commerce platforms have ignited the fire of the big model revolution.

E-commerce is an industry that is bound to be changed by AI, and this is only the first step.

300 million user business, supported by AI+e-commerce

As an e-commerce giant, Alibaba International's performance is also very impressive.

The reason why it was able to find a landing point quickly is that Alibaba International has a natural overseas business with 300 million users.

Therefore, in model training, the team has a good prior background and professional and diverse data.

What Alibaba International needs to do is to provide everyone with a shared AI infrastructure and then solve the AI ​​needs of all these businesses in e-commerce.

What’s interesting is that if you want to do this, you need to make a product-level abstraction of “what AI can do” in e-commerce.

We need to imagine, what difficulties will a factory in Shenzhen that makes 3C products, a trader in Yiwu that deals in department stores, or an e-commerce company that has been doing stall business in Guangzhou on Taobao and Pinduoduo for many years encounter when they want to export?

How do you know what overseas consumers want to buy when you don’t speak a foreign language?

The needs of East Asian and European and American customers are different. How to meet them?

For most cross-border entrepreneurs, the key is to explore opportunities in niche scenarios.

The Alibaba International team has found the way by embracing AI.

In the early days, the Alibaba International team quickly and loosely integrated AI capabilities into existing business scenarios and product systems based on more than 40 scenarios.

As the process gradually deepened, the team reached a large-scale stage and developed a large number of standardized services.

At this point, the business differences between different countries and cultural backgrounds become apparent.

For example, the product expression, model presentation, and product information compliance requirements of Asian and Middle Eastern customers are very different from those of European and American customers.

On the other hand, it is also necessary to use platforms such as PAAS on large-scale products, combined with a more unified model approach, to avoid fragmentation and obtain better inference costs and greater economies of scale.

In addition, the team also needs to provide more on-demand customization capabilities to the front-line business team based on the original products.

In the past, everyone viewed the same content product, but now AI can generate content infinitely and in large quantities, and can generate content that is tailored to each person.

Does AI search simply replace traditional machine learning models with a large model?

It is not just that. In the view of the Alibaba International team, this may involve the entire system and the entire data design model, as well as the expression and presentation of products at the bottom of the algorithm.

At present, human language expression has reached a sufficiently high level of abstraction. The languages ​​of different countries and nationalities can already well express all abstract logic and the logic of specific physical objects. However, in terms of vision, it is still an open field.

Focus on only three things

After last year's "100 Model Wars", it is clear who the world's leading basic model company is.

Founded in April 2023, Alibaba International AI Business has already grown into a team of over 100 people, but its self-positioning is: We are not a team that trains basic models.

They emphasized that they would focus on three things:

First, it is multilingual.

The team will use multilingual enhancement to improve the multilingual capabilities of large models and use them more efficiently in multilingual tasks.

Take translation for example. It has always been undertaken by small models, and there is no clear answer as to how to translate large models.

Furthermore, in order to control costs, many business tools need to control costs. Alibaba International hopes to do a good job in multilingual translation with lower costs and better results, so it lets LLM do multilingual translation.

At the conference, they unveiled for the first time the multi-language enhanced large model MarcoPolo and more technical details behind it.

It is trained based on massive high-quality multilingual data, of which 2.5T tokens are expected for minority languages. It can support 30+ minority languages ​​and has models with different parameter sizes such as 8B/57B/72B.

Second, it is multimodal.

In the e-commerce arena, having a large multi-language model is often not enough, as this involves a rich and complex multimodal scenario.

Suppose a customer returns a piece of clothing because it is the wrong color and takes a photo. AI needs to use its “eyes” to identify whether the clothing is dark blue or black.

From the merchant’s perspective, they hope that after uploading a product picture, AI can optimize and complete it.

In addition, when it comes to what are the same items and what are similar items, AI recommendations and identifications all require innovation in underlying technologies.

Alibaba International's multimodal large model MarcoPolo-VL is based on the industry's original structured embedding alignment model (SEA) training and tuning, and can provide 7B/14B models. Moreover, with the same parameter effect, it surpasses the known open source models.

Finally, on the engineering and architecture side, build the entire model service PAAS.

AIGC has different business platforms. For merchants, the most convenient way is to complete the entire operation process on one platform.

Therefore, the birth of a universal underlying platform is inevitable.

Alibaba PAAS service can greatly reduce the cost of model inference services and can support tens of millions of AI service calls per day with hundreds of inference cards.

The following is their layout on PAAS. Based on the underlying Alibaba Cloud and other infrastructure, they have built an end-to-end complete technology chain for training, reasoning, and application. Different application scenarios can be efficiently iterated on unified AI application engines, model workbenches, application builders and other products, thereby providing a set of shared application facilities for each business to make good use of AI.

After more than a year of hard work and experimentation, Alibaba International has empowered 500,000 small and medium-sized businesses in 40+ application scenarios, and information on more than 100 million items has been optimized.

Why is the number 40 so important?

Zhang Kaifu said that when the scenario can generate actual value with AI, what has happened in the past six months is that the use of AI has begun to increase significantly.

Data from the past six months show that on average, merchants’ calls for AI double every two months.

Now, the platform's average daily call volume has exceeded 50 million times, and its size doubles every two months.

More than 100 million products have been completely rewritten through AI.

Through translation, model change, blacklisting, selling point generation and other steps that are all done with AI, the team is very confident that when the product is presented to American consumers, they will understand it, want to buy it and be willing to place an order!

Moreover, in Zhang Kaifu's view, the smaller and medium-sized businesses are, the more they can benefit from the application of AI.

So, what about using AI?

The following are some real business cases.

One merchant said that before AI was introduced, there was only a picture of the product taken with a mobile phone. Although it is expensive to shoot real-life images, they look ordinary when put on the platform.

With AI, you can quickly generate picture backgrounds.

The lighting and shadows of the generated images are not only comparable to those in real photography, but can also be applied to a large number of marketing image templates. These templates are designed by overseas designers based on local styles. Not only does this reduce costs, but it also increases click-through rates.

For example, Zeuslap, a top monitor merchant on the AliExpress platform, began using Alibaba AI's image generation function in November 2023.

Whether it is the store decoration banner, product scene picture, business detail picture, YouTube and TikTok cover picture, etc., they are all completed by AI.

In the past, merchants needed to spend a lot of time looking for materials and designs, and they had to do PS by themselves. Now, as long as you select a template and upload the product, you can easily get the product image and marketing image, which saves a lot of time.

Today, with the help of Alibaba International AI, Zeuslap has grown from a small, husband-and-wife business to the leading brand in the platform industry.

The dramatic changes that AI will bring to e-commerce overseas are still to come.

References: