Starting from practical AI (artificial intelligence), Amazon Web Services has taken its own path in the fierce AI competition.
On December 19th, Amazon Web Services re: Invent recap was held in Shanghai. Dai Wen, General Manager of Solution Architecture for Amazon Web Services Greater China, gave a detailed introduction to a series of technology releases recently launched by Amazon at the re: Invent conference, covering three major areas: generative AI, data strategy, and cloud services, including the new self-developed Amazon Nova series and computing instances supported by the Amazon chip Trainium2.
It is reported that the re: Invent conference has a history of 13 years. This year, 60000 people gathered at the Las Vegas venue in the United States, and over 400000 people worldwide watched it online. Amazon Web Services also seized this opportunity and launched a full stack collaborative innovation covering infrastructure, models, and applications on site, and announced deepening partnerships with AI startups Anthropic and Apple.
No model can be applicable to all scenarios
Among a series of new products, Amazon's next-generation basic model, the Amazon Nova series, has attracted much attention. As a series model, Amazon Nova is not limited to a specific type, but meets the needs of different scenarios through a rich variety of types. This series includes six models, ranging from plain text model Micro, low-cost multimodal model Lite, high-performance multimodal model Pro, to Premier expected to be launched in the first quarter of 2025, as well as two more advanced models Canvas and Reel whose release dates have not yet been determined.
More importantly, the Amazon Nova series represents Amazon's understanding of AI development trends and user pain points. Amazon CEO Andy Jassy proposed the concept of "Practical AI" when releasing Amazon Nova, stating that at the current stage of development, no model can be applicable to all scenarios. Dai Wen also mentioned in his speech that Amazon Nova's positioning is precisely "practical AI", and Amazon is committed to "providing users with more cost-effective and practical choices, rather than just pursuing performance in rankings or public data testing".
From a practical perspective, Amazon Nova significantly reduces the cost of basic models. In their respective intelligent categories, Amazon Nova Micro, Lite, and Pro have application costs that are at least 75% lower than the best performing models in Amazon Bedrock, and are also the fastest models in their respective categories in Amazon Bedrock.
In an interview, Dai Wen said that Amazon Web Services focuses on users when releasing products and pricing: "Because we are a leading global cloud service provider with a sufficient number of users. Therefore, when building a product, in many cases, there are already some users who have corresponding needs, and our pricing is also aimed at this group of users
Dai Wen said that taking the Micro model in Amazon Nova as an example, as a model dedicated to text to text conversion, Amazon will consider the value-added points or service points of the model when pricing, and then make choices in various aspects of the model: "We have noticed that some customers do only need a pure text model, so in the construction of the model, we need to find ways to achieve the ultimate cost-effectiveness and lower latency
For similar considerations, Amazon has focused on enhancing the Amazon Nova Canvas model's ability to generate high-definition small photos and add watermarks to images. Dai Wen said, "We don't believe that Amazon Nova can solve all problems. Therefore, when Amazon Nova was born, the capabilities and price positioning of its different models were actually tailored to the needs and demands of actual users in different scenarios, and thus created models with different positioning
The software development logic centered around the "reverse working method"
In addition, Amazon has also demonstrated its larger development goals in the field of AI. Beyond the Amazon Nova series, Amazon has announced the release of a "speech to speech" model and an "Any to Any" multimodal model in Q2 2025, capable of processing text, images, audio, and video as input and output.
Dai Wen stated that since the launch of its first product on March 14, 2006, Amazon Web Services has developed into over 240 fully functional services, with the core software development logic of "reverse engineering", which means that "Amazon will only develop corresponding products when customers have needs, and will not work in isolation". At the same time, Amazon Web Services will analyze the root causes behind customer problems and carry out innovation, presenting a "matrix style, full stack style innovation form".
As a hot topic in the AI field this year, "Agent" is also a concept that Amazon Web Services attaches great importance to. At the re: Invent conference, Amazon's generative AI service, Amazon Bedrock, released a multi-agent collaboration feature where users can use natural language to describe their needs and quickly create agents through Amazon Bedrock to handle tasks such as orders, financial reports, and retention. Meanwhile, Amazon Web Services has launched three automated agents for the generative AI software assistant Amazon Q.
Dai Wen stated that with the rapid development of intelligent agent technology and the increasing variety of its types, it means that users must flexibly choose different models according to different application scenarios, "under the coordinated management of multiple models, accurately empower different intelligent agents". Amazon Web Services has realized that "enabling more efficient interaction and collaboration among multiple intelligent agents" is currently a very strong market demand, so it has launched targeted features or new functions in this area.
At this conference, another update from Amazon Bedrock also caught the attention of AI professionals. It is reported that this feature, called "automated reasoning check," can help solve the long-standing "illusion" problem that has plagued large models through mathematical verification, that is, the situation where the model output does not match the actual situation.
Specifically, users can define rules in the strategy and use this technology to detect and correct model illusions. Once the model shows signs of "illusion", it can quickly and accurately identify and promptly initiate the secondary processing flow. Dai Wen pointed out that this measure can be regarded as a key "weapon" to promote the output of the model towards production.