Amazon Nova: a series of AIs that understand even sarcasm

Amazon Nova is a family of generative AI models designed to be multimodal, that is, capable of processing text, images, and video. These models include specific versions such as Micro, Lite, Pro and Premier, optimized for different needs and with a focus on accessibility and cost efficiency. In addition, Nova integrates advanced tools such as Canvas for text-based image creation and Reel, which makes it easy to produce cinema-quality videos.

One of the highlights of Nova is its “distillation” process, which allows large models to be tuned to specific cases, reducing costs without sacrificing performance. Matt Garman, CEO of AWS, emphasized at the Re:Invent event the capacity for applications such as language translation and emotion or sarcasm detection, which represents a significant advance in contextual understanding.

The initiative is part of Amazon’s effort to democratize artificial intelligence through AWS and Bedrock, its AI system, offering accessible tools even for small businesses, while driving broader adoption of these technologies.

Vishal Sharma, vice president of Generative Artificial Intelligence, pointed out that in learning and speech models, there are non-verbal elements such as laughter, pauses or hesitations, which the model must interpret to express naturalness.

“As for text, the models’ ability to understand sarcasm depends on the pre-training and data used. Although it is not perfect yet, it is constantly improving. Additionally, our current models can analyze images, videos or text and answer questions like “What does this mean?” or “What’s the mood?” Our short-term goal is to develop “any-to-any” models that combine text, video, audio and other formats, reflecting human multimodality,” said Sharma.

In communication, sarcasm often depends on tone of voice, body language or context so that the recipient understands the intention. In written texts, sarcasm can be more difficult to detect, although it is often accompanied by signs such as exaggeration, obvious irony or, more recently, the use of quotes or emojis. However, interpreting it is difficult even for humans.