The ecommerce business is now sprinkling ChatGPT-like AI over its site and apps—today announcing, among other additions, AI-generated buying guides for hundreds of different goods categories. Executives at the company claim its engineers are also exploring more ambitious AI services, including automatic AI purchasing agents that recommend goods to a client or even put items to their vehicle.
“It’s on our blueprint. We’re working on it, prototyping it, and when we think it ’s good enough, we’ll release it in whatever form makes sense, ” says Trishul Chilimbi, a VP and distinguished scientist at Amazon who works on applying the company ’s core AI to its products and services.
Chilimbi says the first step toward AI officials will likely be bots that actively advocate products based on what they know of your habits and interests, as well as a knowledge of broader trends. He acknowledges that making this feel nonintrusive will be crucial. “If it ’s no good and annoying, then you’ll tune it out, ” he says. “ But if it comes up with surprising things that are interesting, you’ll use it more. ”
Amazon added a chatbot called Rufus to its platform in February 2024 that can answer a wide range of questions about Amazon’s many products. The bot uses a bespoke large language model—the kind of algorithm that powers ChatGPT—also known as Rufus.
The Rufus LLM is trained on vast amounts of internet text—like those on publicly available websites—and is then fine-tuned into a commerce-focused model by being fed a carefully curated diet of Amazon’s proprietary data. Chilimbi says Amazon’s LLM has “hundreds of billions of parameters. ” ( Parameters are a rough measure of capability; for comparison, Meta’s largest publicly available LLM has 405 billion. ) He confirmed that Amazon is training a larger model, but declined to say how large it is or what capabilities Amazon hopes it will unlock.
Like many tech companies, Amazon is looking beyond chat and turning its attention toward the potential of so-called agents, which use LLMs but attempt to carry out useful tasks on users ’ behalf either by writing code on-the-fly, inputing text, or moving a computer’s cursor. Future AI agents might, for instance, navigate various websites to sort out a parking ticket, or they might operate a PC to file a tax return. Getting LLM-powered programs to do this reliably is elusive, however, because such tasks are vastly more complex than simple queries and require a new level of precision and reliability.
“Every major company is now doing [AI] agents, ” says Ruslan Salakhutdinov, a computer scientist at Carnegie Mellon University who is working on AI agents. The technology is exciting, he says, because it promises to automate countless routine tasks that people do every day: “On the ecommerce side, if agents can find the best possible outcome for me, that’s amazing. ”
Salakhutdinov and colleagues at CMU developed a dummy ecommerce website as part of a platform called Visual Web Arena for testing AI agents. Key challenges include enabling agents to better make sense of visual information and training them to explore vast arrays of possible options while zooming in on the correct one—something that may require more advanced reasoning abilities.
But Salakhutdinov says that having a wealth of information about how users go about common and important tasks like shopping might be a crucial ingredient for getting them to stay on track. “Data is going to be very important, ” he says.
Ship It
Amazon’s agents are, of course, likely to be more focused on helping customers find and buy whatever they need or want. A Rufus agent might notice when the next book in a series someone is reading becomes available and then automatically recommend it, add it to your cart, or even buy it for you, says Rajiv Mehta, a vice president at Amazon who works on conversational AI shopping. “It could say, ‘We have one bought for you. We can ship it today, and it will arrive tomorrow morning at your door. Would you like that? ’” Mehta says. He adds that Amazon is thinking about how advertising can be incorporated into its model’s recommendation.
Chilimbi and Mehta say that eventually, an agent might go on a shopping spree when a customer says, “I’m going on a camping trip, buy me everything I need. ” An extreme, though not impossible, scenario would involve agents that decide for themselves when a customer needs something, and then buy and ship it to their door. “You could maybe give it a budget, ” Chilimbi says with a grin.
Amazon’s new AI-generated shopping guides, announced at its Reinvent conference in Nashville today and initially available on the company ’s US mobile website and app, are a small step toward the ultimate vision of a superintelligent shopping assistant. The Rufus LLM is used to autogenerate the sort of information and insights that could take someone hours of online research to gather. “If you ever try to shop in a category you’re not familiar with, it can be pretty time-consuming to understand the lay of the land, the different features available, and the different selections, ” says Brett Canfield, a senior product manager on the personalization team at Amazon.
Canfield showed WIRED shopping guides for televisions and earbuds that noted important technical features, explanations of key terminology, and, of course, recommendations on which products to buy. The underlying LLM has access to the vast corpus of product information, customer questions, reviews, and feedback, and users ’ buying habits. “This is really only possible with generative AI, ” Canfield says.
The new shopping guides highlight generative AI’s potential in ecommerce, creating guides for product categories too niche to normally get the treatment. “The definitive hedge trimmers, ” for instance.
Guide Supplies
The guides also, however, show how generative AI threatens to upend the economics of search and shopping while borrowing liberally from conventional publishers.
AI-generated search results often now provide product comparisons and opinions. This diverts traffic from outlets, like WIRED, that make money by producing shopping guides, reviews, and other articles, even though the AI results are produced using data scraped from such websites in the first place.
Canfield declines to say what additional training data was used to build the new AI shopping guide feature. ( WIRED’s parent company, Condé Nast, entered into a partnership with OpenAI, the company behind ChatGPT, in August of this year. )
“LLM agents are a customer service game changer, ” says Mark Chrystal, CEO of Profitmind, a company that uses AI to provide retailers with analytics.
Chrystal says that big players like Amazon might benefit most from the rise of generative AI because they have so much data to feed to their models. This should “lead to increasingly capable AI systems that not only improve customer service but also lead to product and delivery innovations, ” he says, although he notes that “in essence, the data-rich will continue to get richer and the data-poor will get poorer. ”
Amazon says its Rufus LLM already demonstrates some unique abilities that are especially useful for ecommerce. Chilimbi recounts an incident involving an executive at Amazon who asked the LLM to recommend the best Batman graphic novels, and was surprised when it came back with a list that included the non-Batman dystopian classic Watchmen”}” href=”https://en.wikipedia.org/wiki/Watchmen” rel=”nofollow noopener” target=”_blank”>Watchmen. When asked why it chose the book, the Rufus model stated that the themes and characters in Frank Miller’s popular 1980s Batman series The Dark Knight Returns carry a similar resonance to those in Alan Moore’s Watchmen”}” href=”https://en.wikipedia.org/wiki/Watchmen” rel=”nofollow noopener” target=”_blank”>Watchmen. “Occasionally you say, ‘Oh wow, how does it do this? ’” Chilimbi says.
Amazon’s Rufus LLM is n’t only fed a different diet from most LLMs; it also gets a different kind of fine-tuning. The additional training that normally helps chatbots engage in coherent conversation and avoid saying inappropriate things is used by Amazon to train its model to be a better “shopping concierge. ” “There are multiple signals ” fed to the model as fine-tuning, Chilimbi says, including whether someone clicks on a recommendation, adds it to their cart, and eventually buys it.
Chilimbi adds that Amazon has developed its own shopping benchmark for testing Rufus and helping it get smarter. But while a conventional LLM might be tested on its ability to answer general knowledge questions or solve math or science problems, Amazon’s benchmark tests the model’s ability to help a customer find what they are looking for more easily.
Amazon hopes that raising the shopping IQ of its AI might eventually enable its independent, shopping-centric AI agents.
“We aren’t quite there, ” says Salakhutdinov of CMU, who notes that he would n’t be comfortable giving an AI agent his credit card just yet. “There are some actions you can’t really reverse from, ” he says. “You know, like you already bought it. ”
