But here's what nobody's saying: Chinese AI just dominated the leaderboard.
7 of 8 top video models: China. Faster. Cheaper. Better. While everyone waited for Sora, they shipped the future.
RESEARCH
-
Chinese AI dominates video models: 7 of 8 top, faster, cheaper
By
–
-
Alibaba-backed AI (100M users) launches omni-model, memory attention, instant generation
By
–
Alibaba-backed. 100M users. Launched Jan 13.
— God of Prompt (@godofprompt) 17 janvier 2026
Three breakthroughs nobody else has:
Omni-model (everything speaks one language)
Memory attention (infinite worlds that remember)
1-4 step generation (instant response)
They unified what Google, OpenAI, Meta built separately. pic.twitter.com/irDCNCm78ZAlibaba-backed. 100M users. Launched Jan 13.
Three breakthroughs nobody else has: Omni-model (everything speaks one language)
Memory attention (infinite worlds that remember)
1-4 step generation (instant response) They unified what Google, OpenAI, Meta built separately. -

Intelligence bootstrapping begins with explosion of verified algorithms
By
–
Incredible. We are only a few weeks into this development. By the end of this year we’ll have a massive explosion in new *verified* algorithms and methods. These are the early stages of the intelligence bootstrapping itself.
-

The world model builder keeping America competitive in robotics AI
By
–
The world model builder, or the guy who is keeping America competitive with China in robotics AI. Most people don't know @bercankilic but you should. Almost every robotics company in San Francisco/Silicon Valley buys his data. data. He is largely seen as the best at
-

C2C outperforms traditional Text-to-Text on four benchmarks
By
–
8. C2C was rigorously tested on four major challenges: MMLU-Redux, OpenBookQA, ARC-Challenge, and C-Eval. It significantly outperformed the traditional "Text-to-Text" approach.
-

Efficient Cache Improves Model Performance Without Retraining
By
–
7. The paper demonstrated that improving a model's performance is possible by utilizing a more efficient cache. This enhancement does not require adding extra words or retraining, maintaining the same input length while enhancing understanding.
-

C2C introduces neural Fuser to connect model KV-Caches
By
–
5. Each model possesses a memory storage referred to as KV-Cache. C2C introduces a small neural "Fuser" that connects the memory of one model to another, facilitating information sharing and collaboration.
-

Architectural breakthrough: model pairs boost accuracy and speed
By
–
The results are insane: – 8.5-10.5% higher accuracy than individual models
– 3.0-5.0% better than text-based communication
– 2× speedup in latency
– Works across ANY model pair (different sizes, architectures, tokenizers) This isn't incremental. It's architectural. -

C2C is telepathy for AI, replacing slow telegram-like communication
By
–
Every LLM today communicates like humans sending telegrams. Model 1 generates text → Model 2 reads text → repeats. It's slow, expensive, and loses meaning in translation. C2C is telepathy for AI models.
-

Cache-to-Cache: LLMs talk via KV-Caches, 10% accuracy boost, 2x speed
By
–
Researchers just made LLMs talk to each other WITHOUT generating a single word. Cache-to-Cache (C2C) lets AI models talk directly through their KV-Caches, bypassing text entirely. 8.5-10.5% accuracy boost. 2× faster. Zero token waste. Here's the breakthrough (and why this