A week in the world of AI: America's technological edge. Amazon plans a giant computing network, Meta tests new glasses and Google helps predict storms

Artificial intelligence is increasingly influencing various areas of technology - and the United States is playing a major role. From data centres to smart glasses to weather forecasts, US tech giants are showing their strength. Amazon is investing tens of billions in powerful computing complexes, Meta is testing new glasses with a view from the user's eyes, and Google is helping to better predict hurricanes. On top of this, US researchers are exploring how AI actually thinks - and have found that models often "explain" their answers differently than we would expect. Here's a round-up of the highlights.

Amazon is building a supercomputer network for Anthropic

Amazon introduced Project Rainierwhich will connect huge data centres into a single powerful network, called the ultracluster. It will mainly be used for training large AI models that require huge computing power. The move is intended to strengthen Amazon's position in the AI field and enable faster development of advanced systems for robotics, cloud services or the Alexa voice assistant, for example.

• Amazon plans to build up to 30 data centres, the first to be established in Indiana and other US states.
• It will use its own powerful chips Trainium 2 and 3that are more economical than conventional alternatives.
• The centres will be connected by their own network technology Elastic Fabric Adapter.
• The key partner of the project is the company Anthropicin which Amazon has invested $8 billion.
• If the cooperation fails, Amazon will use this infrastructure for its cloud services AWS.

Meta introduces the second generation of Aria smart glasses

Meta revealed a new version of its Aria Gen 2 smart glassesthat allow you to collect data from a human perspective in real time. They are primarily intended for augmented reality research and the development of robots that could learn better by mimicking human behaviour. With built-in cameras, microphones and sensors, the glasses capture everything a human sees, hears and does - providing valuable training material for future AI.

• The glasses weigh only 75 g and one charge lasts up to 8 hours.
• Contained at 5 cameras, two of which are for 3D space sensing, and 7 microphones for audio, including the user's voice.
• The sensors monitor, for example eye, hand, heart rate and environmental movements.
• The glasses can capture the surrounding space in 3D and track precise hand movements.
• The data can be used for training robots or developing smart assistants.

AI helps predict hurricanes more accurately

US National Hurricane Center (NHC) cooperates with Google to develop a new AI system that helps predict the development of tropical storms with greater accuracy. Using machine learning and extensive historical data, the AI model can estimate where and how hard a hurricane will hit - even more than a week in advance. This can help protect lives and allow authorities to better prepare for emergencies.

• New model can track a hurricane up to 15 days in advance.
• It uses a special AI network that processes data from previous years (1979-2022).
• Compared to previous models, the up to 140 km more accurate with a five-day forecast.
• The model better estimates wind speed and storm track than older systems.
• More accurate predictions can save lives and reduce damage in vulnerable areas.

The models' chain of thought often does not explain their decisions

Research by Anthropic showed that language models such as Claude 3.7 Sonnet or DeepSeek-R1 sometimes make up "explanations" for answers that don't actually match how they arrived at the correct answer. Even if the model is influenced by the wrong clue, it often fails to mention it in the subsequent "explanation". This suggests that while the models appear to be reasoning logically, their actual decision-making process remains hidden.

• Scientists have been feeding the models misleading hintsthat led them to the wrong answer.
• Even when models have been influenced by the hint, they often did not mention in their explanation (chain of thought).
• Claude only mentioned the clue in 25 % cases, DeepSeek v 39 %.
• This means that the "chain of thought" of a model is not always a reliable indicator, why he chose that answer.
• Research shows that AI models cannot yet be fully trusted to explain their decisions.

The Batch - DeepLearning.Ai by Andrew Ng / gnews.cz - GH