The whole story of “Project Nile”, the core of Amazon’s generative AI strategy, has been revealed.Experienced salesperson-level interactive search capabilities

The new search feature was scheduled to be released in September, but has been “postponed,” one of the people said.

In-house testing is still ongoing, and while there is a possibility that it will be postponed again,It is expected to be released in the US starting in January 2024.” said another source.

An internal Amazon document obtained by Insider reveals the following:

“Project Nile is a confidential initiative to develop an interactive shopping agent for retail customers.”

Interactive search powered by AI is expected to improve the shopping experience for consumers and lead to increased sales, especially through mobile devices.

For this reason, the company’s top executives, including CEO Andy Jassy and Doug Herrington, head of the retail division (Worldwide Amazon Store CEO), are supporting this project, and it is a top priority within the company. It is said that it has become.

The success or failure of Project Nile will also have a major impact on consumers. That’s because Amazon search has become the default entry point for online shopping for many consumers.

According to a survey by Jungle Scout, which develops management and analysis tools for Amazon sellers, 61% of consumers in the United States use Amazon search first when shopping online.[Chart 1]

Fundamentally changing online shopping

Amazon employees first learned of Project Nile during a recent town hall meeting titled “Reinventing Amazon Shopping with Large-Scale Language Models.” It was a place for dialogue).

As mentioned at the beginning, Mr. Siros, wearing an “Artificial Intelligence Love” T-shirt, introduced Project Nile as a “super secret” project and said the following:

“AI is fundamentally changing online shopping, using the power of large-scale language models to understand shoppers’ questions and requests and provide answers.”

The new generative AI layer will provide shopping experiences such as:

If you type “What kind of coffee maker do you think I should buy?” into the search bar, multiple options will appear, including drip type, pod type, and espresso type, saving you the trouble of checking each product introduction page.

You can then ask, “What’s the best pod product?” and it will show you the top choices based on a variety of metrics, including the number of positive reviews.

Shoppers can then compare products based on specifications and features such as brew time and cup size, or ask for more information about a specific coffee maker, such as the Keurig.

Additionally, based on individual order history and “wish lists,” AI will provide a summary of customer reviews for the product and suggest personalized product recommendations.

Just like the experienced staff at a brick-and-mortar store.

Sirosh likened the new AI capabilities to sales staff at a brick-and-mortar store, who know each shopper’s individual preferences.

“Before e-commerce, the sales staff in the store acted as a search engine. They knew everything about the product.

I was able to figure out what customers wanted just by observing them when they came to the store. This is because similar customers have visited the store in the past and behaved in the same way. If someone you have previously served a customer with returns to your store, they may even remember their preferences.

In-store sales staff were able to integrate all kinds of information and provide appropriate support to shoppers through natural interactions.

I wonder if the same thing can be achieved with Amazon’s huge e-commerce site. Isn’t that the mission we should accomplish in the future?”

According to Siros,The conversational AI layer will currently be implemented in addition to the existing search bar..In addition, the new search functionPossibility of using multiple large-scale language modelsIt is said that there are also.

Expecting an increase in sales from mobile devices

The new search feature appears to be rolling out first to Amazon’s mobile app.

Although mobile searches account for nearly 80% of all searches on Amazon’s e-commerce site, the conversion rate (the percentage that results in a product purchase) is lower than that of desktop searches.That’s the reason.

“If we can provide a great (interactive search) experience, highly professional answers, and increase conversion rates through mobile apps, we have the potential to make a significant contribution to our overall revenue,” Sirosh said.

Amazon’s active moves in the field of generative AI are not limited to Project Nile, which we introduced above.

At a new product launch event held on September 20th, the company announced that its smart assistant, Alexa, will be equipped with an Alexa Chat feature that utilizes generated AI. Alexa’s responses are personalized based on the user’s profile, and its understanding of context helps keep conversations on topic.

In addition, as reported in an article dated July 14, Amazon Web Services (AWS), the cloud division, has established a new organization that specializes in supporting customer companies in the use of generative AI technology.

Prior to that, in April, it released Bedrock, a service that provides multiple platform models, including Amazon Titan, to help customers build generative AI-based apps. .

In addition, although this is different from a direct generative AI implementation project, Anthropic, the developer of Claude, the company’s biggest rival, is praised for having more natural conversations and higher creativity than OpenAI’s ChatGPT. It is still fresh in my memory that they announced on September 25 that they would invest up to $4 billion (approximately 590 billion yen).

Additionally, in response to these moves, CEO Jassy will appoint Rohit Prasad, senior vice president and head scientist for Alexa, to report to him this summer to develop Amazon’s “most ambitious” large-scale language model at Amazon. He has been appointed as the person in charge of a “newly established team that will play a central role” toward the future.

Project Nile is undoubtedly one of the most important projects for Amazon’s retail division. “Project Nile is our ‘absolute top priority’ and we are urgently increasing our staffing,” Sirosz told his team.

Utilizing ChatGPT and humans for AI training

The Project Nile concept was one of many ideas included in an internal document Amazon compiled in early 2023 called Generative AI – ChatGPT Impact and Opportunity Analysis.

According to people familiar with the situation, in addition to proprietary data obtained from its vast product lineup, Project Nile also collects information about users’ actions on the site (such as searches and clicks), their shopping cart saves and purchase history, etc. Amazon seems to think it can develop competitive AI models because it can use customer review information as training data.

Amazon is also hiring human trainers to check the output (answers) of AI to improve the quality of search results. The AI ​​trainer evaluates the quality of the answers and compares them to those of competing AIs like ChatGPT, Microsoft’s Bing, and Google’s Bard.

According to one official, “We are also considering integrating with Alexa to enable not only text searches but also voice searches.”.

The biggest concern within the company is “the accuracy of search results,” the person said.

Conversational AI, such as ChatGPT, can generate false information that is not based on fact, known as “hallucination.”

By using content correction tools, human reviewers, and refusing to answer sensitive questions about sensitive products like healthcare products (which are subject to many legal restrictions), Amazon We hope to alleviate this problem.

The same person added that Amazon is also considering the option of seeking third-party cooperation.

“We are considering leveraging ChatGPT to provide answers to more sensitive questions, and may partner with other conversational AI services like YouChat.”

Search engine drawbacks and security limitations

At the aforementioned town hall meeting, Sirosz addressed the shortcomings of Amazon’s search engine.Current search engines are unable to leverage users’ past search data to understand context (i.e., search intent) and reflect that in search results.That’s what it means.

Increasing competition means Amazon needs to move quickly, but security and compliance issues around privacy make that difficult, Sirosz said.

“Unlike OpenAI (which makes conversational AI the mainstay of its business), unfortunately our actions will be restricted for data security reasons (because we have a huge amount of personal information).

Compared to our competitors, we have so much to lose (if we run into privacy-related security issues).”

Still, Siros is confident in the role of generative AI in dramatically changing the Amazon search and shopping experience.

“Generative AI is the new normal. Everything in the software world (not just our company) will undergo fundamental changes.”