Скачать или смотреть MoR: Google DeepMind's New Architecture | Faster, Smarter, and the Future of AI Beyond Transformers

In 2017, Google Brain gave the world Transformers, an architecture that made groundbreaking AI models like Chat GPT, Claude, Gemini, and even Sora possible. Now, in 2025, Google DeepMind introduces a revolutionary new architecture that might just be their successor: Mixture of Recursions, or MoR.

Discover why MoR is being hailed as leaner, faster, and possibly the future of AI. This isn't magic; it's brilliant design that rethinks how LLMs reason. Instead of processing every word through the same 100+ layers, MoR's smart router acts like a traffic cop, letting simple words exit early while complex ones go for a second or third loop, ensuring computation is only used where needed.
The benefits are astounding:
• Two times faster inference.
• 50% smaller memory needs.
• Half the training compute.
• Better performance than much bigger vanilla models.

MoR doesn't hold memory like Transformers; only active tokens get cached, leading to massive RAM savings. In rigorous tests, an 118 million MoR model outperforms a 315 million Transformer on few-shot accuracy, and at 1.7 billion scale, MoR excels with just one-third the weights. It's not just more efficient; it's smarter, allowing the model to decide in real-time how much computation each word actually needs.

Is this the end of Transformers? Not yet, but we are entering a new era. Just as Transformers replaced recurrent neural networks, MoR could become the next frontier for AI. Google seems to be writing the book on modern AI once again with this fast, memory-smart, recursively thinking architecture that might lead us into the next era of intelligence.