The Rundown: Snap just unveiled its fifth-generation Spectacles, standalone AR glasses powered by the new Snap OS, integrating new AI capabilities and enhancing social interactions through augmented reality.

The details:

The new Spectacles feature a suite of cameras and sensors, enabling multi-modal AI, contextual understanding, and full freedom of movement while remaining lightweight and compact.

Snap OS, built specifically for Spectacles, allows users to interact with My AI, launch immersive Lenses, and extend Snapchat experiences into 3D.

Snap also launched AI-powered video and portrait generation tools, creating videos and images from a single photo to enhance creative expression.

Spectacles are available now to developers in the United States and require a $99/m subscription with a 12-month commitment.

Why it matters: While this is a major step in the right direction for boundless, cordless AR and AI devices — the 45-minute battery life and 46-degree field of view leave Spectacles with room for improvement. Regardless, Snap is pushing the boundaries of tech engineering, and it will be fascinating to see where the tech ends up in a few years.

Image source: 1X

The Rundown: 1X has developed its ‘World Model’, an AI-powered virtual simulator for robots that can predict complex object interactions and imagine multiple future scenarios from a single starting point, trained on thousands of hours of real-world robot data.

The details:

The model predicts complex interactions including rigid bodies, dropping effects, partial observability, and deformable object interactions.

It’s trained on thousands of hours of data from 1X’s EVE humanoid robots performing various tasks in homes and offices.

The model demonstrates action controllability, generating diverse outcomes for different commands from the same starting point, such as opening doors or grasping boxes.

1X is releasing over 100 hours of video data, pretrained models, and cash prizes to accelerate research in compression, sampling, and evaluation.

Why it matters: OpenAI-backed 1X is on a mission to create smarter, more capable robots to help people with everything from household chores to mobility issues. The new world model will also significantly enhance the capabilities of their recently unveiled NEO Beta humanoid and accelerate the path to truly useful home robots.

Image source: Midjourney

The Rundown: A new study from Hong Kong researchers revealed that large language models (LLMs) like GPT-4 possess a dynamic memory similar to human memory, challenging our understanding of AI capabilities and potentially bridging the gap between artificial and human intelligence.

The details:

The researchers defined memory as the ability to generate outputs based on specific inputs, rather than static information storage.

LLMs demonstrated near-perfect memorization of thousands of poems after limited exposure, outperforming humans.

The study proposes, however, that both human brains and LLMs operate on similar principles of “dynamic fitting”, or adjusting their responses based on input rather than retrieving fixed information.

Both humans and LLMs can only “remember” something after triggered by input, suggesting LLM reasoning might be limited by model size and training data quality rather than fundamental differences from human cognition.

Why it matters: If verified and LLMs do have memory capabilities similar to humans, it could change the way we understand artificial intelligence. Without fundamental cognitive differences between AI and LLMs, scaling AI capabilities may simply be an issue of improving hardware and expanding data resources.