# Feb 8, 2025

## Daily News Highlights

### It Turns Out We Really Did Need RNNs
- **Accelerated convergence** in iterative reasoning frameworks achieves optimal rates of O(1/t²), indicating that error decreases **quadratically** with each iteration, especially in noise-free conditions; this is detailed in the [paper](https://arxiv.org/abs/2502.03787).

### GRPO fits in 8GB VRAM - DeepSeek R1's Zero's recipe
- **GRPO** now fits in **under 8GB of VRAM** for Qwen 1.5B, enabling significant resource savings with **Unsloth** and **LoRA/QLoRA** techniques, which reduce VRAM usage by **80%**.

### Why LLMs still have problems with OCR
- **LLMs struggle with OCR** due to their design prioritizing semantic understanding over precise character recognition, leading to significant errors in complex layouts and tables.

### Value-Based Deep RL Scales Predictably
- **Value-based off-policy RL methods** demonstrate predictable scaling behavior, challenging the notion of their erratic performance, as shown by the **Pareto frontier** that governs data and compute requirements based on the updates-to-data (UTD) ratio.

### How does DeepSeek work: An inside look
- **DeepSeek** is an open-source large language model (LLM) that utilizes a **Mixture of Experts (MoE)** architecture, activating only a subset of its **671 billion parameters** per task, enhancing speed and efficiency compared to traditional models like ChatGPT.

### Gold-medalist Performance in Solving Olympiad Geometry with AlphaGeometry2
- **AlphaGeometry2** has surpassed gold medalists in solving Olympiad geometry problems, achieving an impressive **84% solving rate** for geometry problems over the last 25 years, up from **54%** with its predecessor.

### Torchhd: A Python Library for Hyperdimensional Computing
- **Torchhd** is a Python library designed for **Hyperdimensional Computing (HDC)**, enabling fast, noise-robust learning through high-dimensional vectors, which diverges from traditional numeric computation methods.

### AI-Designed Proteins Take on Deadly Snake Venom
- **AI-designed proteins** offer a revolutionary alternative to traditional antivenoms, effectively neutralizing snake venom in lab tests with **80-100% survival rates** in mice, showcasing a potential breakthrough in snakebite treatment.