AWS Conference Highlights: Next-Gen Neural Networks, Chips, and Partnerships

At the annual AWS re:Invent conference, Amazon unveiled the new generation of AI models, Amazon Nova, designed for a broad range of tasks. These neural networks can process text, images, and video.

? Huge #AI news from #reInvent this morning. Meet Amazon Nova, our new generation of AI foundation models with state-of-the-art intelligence across a wide range of tasks and industry-leading price performance.

Our four Amazon Nova understanding models deliver exceptional… https://t.co/zwCBOQezQQ

— Amazon (@amazon) December 4, 2024

“[…] Our new Amazon Nova models are designed to address the needs of internal and external teams, providing engaging intelligence and content generation, as well as significant advancements in latency, cost-effectiveness, personalization, information provisioning, and agent capabilities,” commented Rohit Prasad, Senior Vice President of Amazon’s AI division.

Amazon Nova features six neural networks:

• Micro — a text model offering minimal latency at low cost;
• Lite — an affordable multimodal neural network that quickly processes images, video, and text;
• Pro — a high-performance multimodal tool, offering the “best combination” of accuracy, speed, and cost for a wide range of tasks;
• Premier — the most capable neural network for complex tasks and reasoning;
• Canvas — a powerful model for image generation;
• Reel — creates videos.

The latter can generate a short clip from an image.

Micro has a context window of 128,000 tokens, while Lite and Pro have 300,000 tokens.

The models are integrated into Amazon Bedrock—a service offering high-performance AI applications from leading companies through a single API.

In 2025, the company plans to introduce two additional models: one for speech processing and a universal multimodal tool.

In late November, media reported on Amazon’s new generative AI and additional $4 billion investments in Anthropic.

Amazon Chips for Anthropic and Apple

Amazon plans to cluster hundreds of thousands of Trainium2 semiconductors to provide Anthropic with the capability to train large language models. This solution will increase the startup’s computing power fivefold, the company announced at AWS re:Invent.

The firm began offering chips on December 3. The new cluster, named Project Rainier, contains over 100,000 microprocessors, noted Gadi Hutt, who works in this area.

Apple also announced the use of Amazon’s AI chips for its services, such as search, and is evaluating the potential of the new processor for training Apple Intelligence models.

“We have established strong relationships, and the infrastructure is reliable and capable of serving our customers worldwide,” stated Benoit Dupin, Apple’s Senior Director of Machine Learning and Artificial Intelligence.

According to him, the Cupertino-based company has used AWS for over a decade for various services like Siri, Apple Maps, and Apple Music. Amazon chips have led to a 40% increase in efficiency.

Dupin suggested that Apple might use Trainium2 for training Apple Intelligence, expecting a 50% increase in performance.

Data Center Upgrades

In addition to new chips, Amazon announced a series of solutions to enhance data center efficiency. These include an improved cooling system, the use of diesel fuel for backup generators, and server rack layout optimization.

Liquid cooling is provided for the most powerful chips, as fans alone cannot reduce the temperature sufficiently.

The company plans to spend $75 billion on modernization.

Combating Hallucinations

At the AWS re:Invent conference, the company’s cloud division announced the launch of a tool to combat artificial intelligence hallucinations—Automated Reasoning. The service verifies neural network responses by analyzing other information sources.

The tool is similar to a correction feature from Microsoft, introduced in September. Google offers a similar tool within its Vertex AI platform.

Automated Reasoning uses “logically precise” and “verifiable reasoning” to reach conclusions.

Back in November, Binance announced the integration of AWS generative AI into its products.