# Apr 26, 2025

## Daily

### Lossless LLM compression for efficient GPU inference via dynamic-length float
- **DFloat11** achieves a **30% reduction** in LLM size while ensuring outputs are **bit-for-bit identical** to the original, utilizing entropy coding for optimal compression.

### The Policy Puppetry Prompt: Novel bypass for major LLMs
- **HiddenLayer** introduces a **novel universal bypass** that enhances security for all major **LLMs**, effectively protecting against inference, bypass, extraction attacks, and model theft without complicating existing models or requiring access to raw data.

### Berkeley Humanoid Lite – open-source robot
- **Berkeley Humanoid Lite** is an **open-source** humanoid robot designed to be **accessible** and **customizable**, with a total hardware cost under **$5,000**, promoting community engagement in robotics.

### World Emulation via DNN
- The project demonstrates a **neural network** capable of generating a **playable world** from real-world video data, showcasing the unique ability of neural worlds to create environments from any video, not just game footage.

### Paper2Code: Automating Code Generation from Scientific Papers
- **PaperCoder** automates the transformation of machine learning papers into **functional code repositories** through a structured three-stage process: planning, analysis, and generation, utilizing specialized agents for collaboration.

### We compress any BF16 model to ~70% size during inference, while keeping the output LOSSLESS so that you can fit in more context or run larger models.
- **DF11 compresses BF16 models to ~70% size during inference** while maintaining **lossless output**, allowing for larger context or model sizes without sacrificing accuracy, as detailed in the [arXiv paper](https://arxiv.org/abs/2504.11651).

### LLMs can see and hear without any training
- **MILS enables LLMs to process visual and auditory data without prior training**, showcasing a significant advancement in cross-modal learning capabilities.

### Paper2Code: Automating Code Generation from Scientific Papers in Machine Learning
- **PaperCoder** automates the transformation of machine learning papers into functional code repositories, utilizing a **multi-agent LLM framework** that enhances reproducibility and accelerates research progress.

### Cross-Encoder Rediscovers a Semantic Variant of BM25
- **BERT-based cross-encoders** not only outperform BM25 but may also **reimplement it semantically**, revealing how MiniLM learns components akin to BM25 through mechanistic interpretability.

### CosAE: Learnable Fourier Series for Image Restoration
- **CosAE** (Cosine Autoencoder) innovatively combines **Fourier series** with a feed-forward neural network, enabling **extreme spatial compression** while preserving image detail during restoration.

### Intuition behind Load-Balancing Loss in the paper OUTRAGEOUSLY LARGE NEURAL NETWORKS: THE SPARSELY-GATED MIXTURE-OF-EXPERTS LAYER
- The **Load-Balancing Loss** in the paper "OUTRAGEOUSLY LARGE NEURAL NETWORKS" aims to ensure that all experts in a **sparsely-gated mixture-of-experts** model are utilized effectively, preventing any single expert from becoming overloaded while others remain underused.

### Accelerate PyTorch 2.7 on Intel® GPUs
- **PyTorch 2.7** enhances performance on **Intel GPUs**, achieving up to **3x speedup** in inference for models like Stable Diffusion, thanks to optimizations in **scaled dot-product attention** (SDPA) and the introduction of **torch.compile** on Windows.