# News Highlights

### The NO FAKES Act has changed, and it's worse
- The **NO FAKES Act** has evolved from a targeted approach against generative AI misinformation to a sweeping legislation that imposes **broad censorship** and **overbroad filters**, threatening both free speech and innovation in digital spaces.

### The FPGA turns 40
- The **FPGA** celebrates its **40th anniversary**, evolving from the **Xilinx XC2064** with **64 configurable logic blocks** to modern devices boasting **8.9 million system logic cells**, showcasing a remarkable leap in capability and complexity.

### MCP is eating the world
- **MCP (Model Context Protocol) is gaining traction due to its simplicity and timely execution**, enabling the creation of agents and workflows atop LLMs, unlike previous complex attempts that faltered under integration challenges.

### The Bitter Lesson is coming for Tokenization
- **Tokenization's fragility** in LLMs is a significant bottleneck, with the potential to be replaced by more efficient methods like the **Byte Latent Transformer (BLT)**, which aims to leverage byte-level modeling for improved performance and scalability.

### Gemini Robotics On-Device brings AI to local robotic devices
- **Gemini Robotics On-Device** introduces a powerful **on-device VLA model** that enhances robotic dexterity and task adaptation, operating independently of data networks for improved latency and robustness in various environments.

### Minimal Boolean Formulas
- **The minimum number of AND or OR operators** required to express any Boolean function of five variables is **33**, a finding computed by Russ Cox and Alex Healy, which had not been previously established.

### The best Hacker on HackerOne is now an AI
- **XBOW** has achieved a historic milestone as the first autonomous penetration tester to reach the **top spot** on the US HackerOne leaderboard, demonstrating its capability to discover vulnerabilities in real-world environments.

### LongWriter-Zero: Mastering Ultra-Long Text Generation via Reinforcement Learning
- **LongWriter-Zero** introduces a novel **reinforcement learning (RL)** approach for ultra-long text generation, eliminating the need for costly synthetic data and achieving superior quality in outputs.

### Research Scientist roles in CV/DL at top companies
- **Research Scientist roles** at top tech companies increasingly prioritize **DSA skills**, with algorithm interviews becoming a significant hurdle despite strong publication records in CV/DL fields.

### 7B RL model outperforms larger models
- The **7B RL model** outperforms larger models like DeepSeek-R1 by providing **stronger distillation** and **cold-starting** through optimized step-by-step explanations, enhancing the learning process for its students.

### QuEra Quantum System Leverages Neutral Atoms
- **QuEra's quantum system utilizes neutral atoms** held by laser beams, allowing for room-temperature operation and compact design, which enhances scalability and reduces cooling requirements compared to other modalities.

### Issues with ClsToken, AvgPool for transformer embedding models
- **ClsToken, AvgPool, and MaxPool** are often chosen for Transformer embeddings based on empirical performance rather than mathematical motivation, leading to potential inefficiencies in summarizing embeddings.

### Effects of quantization on long-context tasks
- **4-bit quantization** can lead to substantial performance drops on long-context tasks, with losses reaching **up to 59%** compared to full-precision models, particularly when using languages other than English.

### Dynamic modulation of sampling parameters in LLMs
- **Dynamic modulation of sampling parameters** in LLMs could enhance creativity and precision by allowing models to learn when to adjust **temperature**, **top-p**, and **top-k** during token generation, rather than relying on static values.

### CommVQ: Commutative Vector Quantization for KV Cache Compression
- **CommVQ** introduces a novel **Commutative Vector Quantization** method that compresses the key-value (KV) cache for long-context LLMs, achieving an **87.5% reduction** in FP16 cache size with **2-bit quantization** while maintaining high accuracy.