Amazon’s AI Evolution: From Chatbots to Autonomous Shopping Agents

To review this article again, go to My Profile, and then check out your saved stories.

Amazon Envisions AI Assistants Dedicated to Your Shopping Needs

While Amazon may not possess ChatGPT, its future plans involve the creation of sophisticated artificial intelligence technologies, with a focus on AI assistants determined to aid in your purchasing activities.

The online retail firm has begun integrating ChatGPT-inspired artificial intelligence across its digital platforms, revealing today that it has introduced, among various upgrades, AI-crafted guides for shopping across numerous product areas. The firm's leaders have mentioned that their development team is delving into more advanced AI features, such as self-governing AI shopping assistants that suggest products to consumers or automatically place products in their shopping basket.

"Trishul Chilimbi, a Vice President and distinguished scientist at Amazon focusing on integrating the company's primary AI into its offerings, mentioned, 'It's in our plans. We are in the process of developing and testing it, and once we believe it meets our standards, we will launch it in the most appropriate format.'"

Chilimbi believes the initial move towards AI assistants will probably involve chatbots that automatically suggest items by understanding your preferences and activities, along with keeping up with wider market movements. He admits that ensuring these interactions are not perceived as invasive is vital. "If it becomes irritating and ineffective, people will ignore it," he states. "However, if it manages to offer unexpected and captivating recommendations, then its usage will increase."

In February 2024, Amazon introduced a new chatbot feature named Rufus to its service. Rufus is designed to respond to various inquiries regarding Amazon's extensive product catalog. This chatbot operates on a custom-built large language model, similar to the technology behind ChatGPT, which is also referred to as Rufus.

The Rufus language model has been developed by Amazon, initially learning from a broad spectrum of text found on the internet, such as the content from openly accessible websites. It is then specialized for commercial purposes through exposure to a selection of Amazon's own exclusive data. According to Chilimbi, Amazon's language model is built on "hundreds of billions of parameters." (Parameters can be seen as a basic indicator of a model's abilities; for context, the biggest language model from Meta that is available to the public consists of 405 billion parameters.) He acknowledged that Amazon is working on a more advanced model but refrained from giving specifics about its size or the new functionalities it is expected to bring.

Amazon, like other technology enterprises, is shifting its focus from mere conversation towards exploring the capabilities of what are termed as agents. These agents, powered by Large Language Models (LLMs), are designed to perform meaningful tasks for users, whether that involves generating code in real-time, entering text, or controlling the movement of a computer's cursor. In the future, these AI agents could potentially navigate through multiple websites to resolve a parking fine, or manage a computer to complete a tax filing. However, the challenge lies in enabling LLM-based applications to execute these complex activities consistently, as they demand a higher degree of accuracy and dependability than straightforward inquiries.

"Every leading business is currently engaged with AI agents," states Ruslan Salakhutdinov, a computer science expert at Carnegie Mellon University focusing on AI agents. He finds the technology thrilling as it offers the potential to streamline myriad daily mundane activities. "In terms of ecommerce, if agents are able to secure the most favorable results for me, that would be incredible," he remarks.

Salakhutdinov and his team at Carnegie Mellon University created a simulated online shopping site within a testing environment known as Visual Web Arena, aimed at evaluating AI agents. The primary hurdles involve improving the agents' ability to understand visual data and teaching them to navigate through numerous potential choices to identify the appropriate option, a task that might necessitate enhanced cognitive skills.

However, according to Salakhutdinov, possessing extensive knowledge regarding the methods consumers employ for essential and routine activities such as purchasing could play a key role in maintaining their engagement. "Data will play a significant role," he asserts.

Shipping Made Simple

Amazon's representatives are primarily focused on assisting customers in discovering and purchasing their needs or desires. For instance, a Rufus agent might identify when a new installment of a book series a customer is following is released and could then suggest it, automatically place it in their shopping cart, or even purchase it on their behalf, explains Rajiv Mehta, a vice president at Amazon who specializes in conversational AI for shopping. "It might propose, 'We've already bought one for you. We can dispatch it today, so it arrives at your doorstep tomorrow morning. How does that sound?'" Mehta mentions. Furthermore, he indicates that Amazon is exploring ways to integrate advertising into its recommendation system.

Chilimbi and Mehta suggest that one day, an agent could end up making numerous purchases on behalf of a customer who simply states, "I need supplies for a camping trip, get me whatever is necessary." They propose a somewhat far-fetched yet conceivable situation where agents autonomously determine a customer's needs and proceed to purchase and deliver items directly to them. "Perhaps you could allocate a spending limit," Chilimbi mentions, smiling.

Today, Amazon unveiled its latest innovation at the Reinvent conference in Nashville: AI-driven shopping guides. Initially launching on the company's mobile site and app in the US, these guides mark a step forward towards the goal of creating a highly intelligent shopping assistant. Leveraging the power of the Rufus LLM, these guides can quickly produce the kind of detailed comparisons and insights that would normally require hours of web browsing to compile. Brett Canfield, a senior product manager focusing on personalization at Amazon, explains, "Navigating a new product category can be quite daunting, requiring a lot of time to understand the basics, the available features, and the different options."

Canfield presented WIRED with buying guides for TVs and earbuds that highlighted crucial specs, defined essential terms, and, naturally, suggested which items to purchase. The foundational language model taps into a broad database of product details, consumer inquiries, critiques, and feedback, along with shopping patterns of users. "Such an endeavor is truly feasible only through the use of generative AI," Canfield remarks.

The latest shopping guides showcase the capabilities of generative AI in ecommerce by focusing on product categories that are usually considered too specialized to receive such attention. An example provided is "The definitive hedge trimmers."

Manual Resources

Moreover, the manuals illustrate how transformative AI poses a risk to overturn the financial dynamics of search and purchasing activities, while extensively utilizing content from traditional publishers.

Search results created by artificial intelligence now frequently offer evaluations and viewpoints on products, steering visitors away from platforms such as WIRED, which generate revenue through crafting buying guides, critiques, and various write-ups. Interestingly, the information used to generate these AI-driven outcomes is initially sourced from these very websites.

Canfield refrained from disclosing any specifics about the extra training materials utilized in developing the fresh AI shopping guide tool. (In August of this year, Condé Nast, the parent organization of WIRED, formed an alliance with OpenAI, the creators of ChatGPT.)

Worries of such kind are not expected to dampen enthusiasm for artificial intelligence at Amazon or elsewhere in the ecommerce sector. The application of machine learning in ecommerce, for purposes like analytics, search functions, and suggesting products, is already a common practice. As Large Language Models (LLMs) pave the way for novel applications, a study by analysts predicts that the AI market in ecommerce is set to expand from a valuation of $6.6 billion in 2023 to $22.6 billion by the year 2032.

"LLM agents have revolutionized customer service," states Mark Chrystal, the CEO of Profitmind, a firm that employs artificial intelligence to offer analytics to retailers.

According to Chrystal, major corporations such as Amazon stand to gain significantly from advancements in generative AI, owing to their vast data reserves. He suggests this will result in more sophisticated AI systems that enhance customer service and foster innovations in products and delivery. However, he cautions that this trend essentially means those with abundant data will further prosper, while those lacking in data resources will fall behind.

Amazon reports that its Rufus LLM exhibits distinct capabilities particularly beneficial for ecommerce operations. Chilimbi shared an example where an Amazon executive requested the LLM to identify top Batman graphic novels, and was unexpectedly presented with a selection that included the iconic, yet unrelated, dystopian novel Watchmen. Upon inquiry about its unconventional choice, the Rufus model explained that the narrative and character depth found in Frank Miller's acclaimed 1980s Batman series, The Dark Knight Returns, shares thematic similarities with Alan Moore's Watchmen. "Sometimes it amazes you with its insight," Chilimbi observed.

Amazon's Rufus LLM diverges from the typical training regimen of most large language models (LLMs) by undergoing a unique form of fine-tuning. Instead of just enhancing its ability to conduct fluid conversations and steer clear of unsuitable remarks, this additional training sharpens Rufus's skills as a shopping assistant. Chilimbi explains that this process involves analyzing various indicators, such as a user's interaction with a product recommendation, including clicking on it, adding it to their cart, and ultimately making a purchase.

Chilimbi highlights that Amazon has created a unique shopping standard to evaluate Rufus and enhance its intelligence. Unlike a typical Large Language Model (LLM) which might be gauged on its capacity to tackle general trivia, or address mathematical or scientific inquiries, Amazon's standard assesses the model's proficiency in assisting customers to more effortlessly locate their desired items.

Amazon aims to enhance the intelligence of its artificial intelligence, with the goal of empowering its autonomous, retail-focused AI entities.

"Salakhutdinov from CMU expresses that we haven't fully arrived at that point yet, indicating his hesitation to entrust his credit card to an AI agent at this time. He points out that certain decisions are irreversible, emphasizing, 'Once you've made the purchase, it's done.'"

You May Also Enjoy …

Delivered to your email: A selection of the most fascinating and peculiar tales from the vault of WIRED

Elon Musk poses a threat to national security

Interview: Meredith Whittaker Aims to Disprove Capitalism

What's the solution for a challenge like Polestar?

Occasion: Don’t miss out on The Major Interview happening on the 3rd of December in San

Additional Content from WIRED

Insights and Tutorials

© 2024 Condé Nast. All rights are reserved. Purchases made through our site may result in WIRED receiving a commission as part of our Affiliate Partnerships with retail partners. Content from this site cannot be copied, shared, distributed, or used in any form without explicit written consent from Condé Nast. Choices regarding advertisements apply.

Choose a global website