Good afternoon. It’s Friday, December 27th.

Did you know: On this day in 1982, Time Magazine named the personal computer "Machine of the Year."

In today’s email:

• Unitree Robot Dog

• AI Soccer Referee

• Simulated Agents and the Rise of a Hive Mind

• 5 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

Unitree B2-W robot dog shows off human carrying, flips, and off-road skills

Unitree Technology’s B2-W robot dog is proving its mettle—literally. A year post-mass production, this machine now flips, slides, and scales steep slopes with ease, all while carrying a 40-kilogram load and jumping from 2.8 meters in height. It’s not just a flashy demo; the robot dog’s real value lies in its potential for industrial applications, from hazardous inspections to fire rescues.

This blend of agility, power, and endurance is quickly turning the B2-W into a tool capable of handling the toughest physical challenges. In other words, the future of robotics just got a whole lot more mobile and functional. Read more.

New AI system recognizes soccer fouls, evaluates severity, and provides commentary on key plays

Researchers at Shanghai Jiao Tong University and Alibaba unveiled MatchVision, an AI system trained on SoccerReplay-1988—the largest dataset of its kind, featuring 2,000 matches and over 3,300 hours of European league footage. MatchVision identifies 24 game events, including fouls, and evaluates their severity with up to 84% accuracy.

This AI system not only tracks on-field actions but also generates human-like commentary by analysing context and technical play. Outperforming existing models, MatchVision could streamline match analysis, produce automated highlights, and assist referees. The dataset and model are set to debut on GitHub, promising broader access to advanced sports analytics. Read more.

Nvidia’s Jim Fan Projects Simulated Agents and the Rise of a Hive Mind

Nvidia's Jim Fan projects that the majority of embodied agents will first take shape in virtual environments, before being zero-shot deployed in the real world. These agents, linked by a "hive mind," will exchange latent embeddings to synchronize actions in multi-agent tasks. A glimpse into this future is the City of Tokyo’s 3D digital twin, offering a high-res simulation of the entire city for download—just one example of how physical spaces are increasingly migrating into the digital realm.

In this landscape, robots will train not in isolation but within a vast “iron fleet” running on real-time graphics engines, creating trillions of training tokens at scale. Nvidia’s Santa Clara headquarters was designed and rendered in Omniverse, its GPU-accelerated platform, before any physical construction began. Read more.

• World’s only private F-16 to soar with AI, jammers, advanced missiles

• 2024 Showed It Really Is Possible to Rein in AI

• OpenAI's Sébastien Bubeck Predicts AI Models Will Soon Solve Major Problems in ‘AGI Time’

• Microsoft and OpenAI have a financial definition of AGI: Report

• DeepSeek-V3 Sets a New Benchmark in Open-Source AI Performance

• Anthropic says most tasks don't require complex AI agents

• Google's Veo Physics Test Sets New Benchmark

• Altman Taps Users for OpenAI's 2025 Roadmap—Sora Upgrade First on Deck

• ChatGPT and Sora went down for the second time this month

• Nvidia’s Christmas Present: GB300 & B300 – Reasoning Inference, Amazon, Memory, Supply Chain

• ChatGPT Search can be tricked into misleading users, new research reveals

• Watch: China unveils ultra-acrobatic humanoid robot with 320° waist spin power

📄 Token-Budget-Aware LLM Reasoning

📄 3DGraphLLM: Combining Semantic Graphs and Large Language Models for 3D Scene Understanding ​​

📄 Fourier Position Embedding: Enhancing Attention's Periodic Extension for Length Generalization

📄 YuLan-Mini: An Open Data-efficient Language Model

📄 A Silver Bullet or a Compromise for Full Attention? A Comprehensive Study of Gist Token-based Context Compression

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!