# Oct 9, 2024

### Daily

- **Diff Transformer** enhances attention by focusing on relevant context while effectively canceling noise, utilizing a differential attention mechanism that computes scores as the difference between two softmax maps.

- The **L-Mul algorithm** approximates floating point multiplication using a single integer adder, achieving **higher precision** while significantly reducing computational resources compared to traditional methods. [Link to article](https://arxiv.org/abs/2410.00907)

- **David Baker** receives half of the **2024 Nobel Prize in Chemistry** for his groundbreaking work in **computational protein design**, while **Demis Hassabis** and **John M. Jumper** share the other half for their development of **AlphaFold2**, an AI model that predicts protein structures.

- The **2024 Nobel Prize in Chemistry** recognizes **Google DeepMind's AlphaFold** for its groundbreaking contributions to protein folding, highlighting the transformative impact of AI in scientific research.

- The **2024 Physics Nobel Prize** highlights significant issues of **plagiarism and incorrect attribution** in computer science, particularly regarding the **Hopfield network** and the **Boltzmann Machine**, which overlook foundational contributions by Shun-Ichi Amari and others.

- **FinetuneDB** enables rapid training of AI models with user data, achieving enhanced performance in **minutes** rather than weeks, while also reducing costs significantly.

- The app developed for **GPT-2** extracts interpretable **"circuits"**, revealing how information flows through the model and demonstrating the formation of features that identify grammatical patterns.

- **Diffusion models** are shown to be **evolutionary algorithms**, as they inherently perform selection, mutation, and reproductive isolation through a denoising process, leading to the development of the **Diffusion Evolution method** for refining solutions in parameter spaces.

- The **L-Mul algorithm** approximates floating-point multiplication using integer addition, achieving **higher precision** with significantly lower computational costs, potentially reducing energy consumption by **95%** for element-wise operations and **80%** for dot products.

- The **SWE-bench Multimodal** dataset introduces **617 new task instances** where AI agents tackle real bugs from **17 JavaScript GitHub repositories**, each accompanied by an image, enhancing the challenge for existing models.

- **NVIDIA's cuLitho platform** is revolutionizing semiconductor manufacturing at TSMC by enabling accelerated computational lithography, which significantly enhances chip production efficiency and performance.

- **TaRot** is a novel method for task adaptation in **Large Language Models (LLMs)**, utilizing learnable rotation matrices optimized via **Bayesian Optimization** to enhance performance on classification and generation tasks.

- **Quantum algorithm DQI shows promise** for outperforming classical heuristics in NP-hard optimization problems, specifically achieving a better approximation ratio for the Optimal Polynomial Intersection problem than any known classical algorithm, with a ratio of approximately **0.7179** compared to the classical best of **0.55**.

- A new **Python library** enables users to execute cloud functions directly from a **local Jupyter notebook**, allowing for seamless integration of local and remote computing environments.