# Jun 25, 2024

## AMD's Giant MI300X
- AMD's **Radeon Instinct MI300X** significantly outperforms NVIDIA's H100 in various benchmarks, marking a pivotal shift in the GPU compute market dominated by NVIDIA due to its CUDA ecosystem and superior hardware.

## Researchers run high-performing LLM on the energy needed to power a lightbulb
- UC Santa Cruz researchers have developed a **large language model** that operates on **just 13 watts** of power by eliminating matrix multiplication, achieving **energy efficiency** over 50 times greater than typical hardware. [Read more](https://arxiv.org/abs/2406.02528)

## EGAIR – European Guild for Artificial Intelligence Regulation
- The **European Guild for Artificial Intelligence Regulation (EGAIR)** is rallying artists, creatives, and publishers across Europe to **protect intellectual property and personal data** from exploitation by AI companies, emphasizing the need for informed consent and the establishment of a "training right."

## Sohu – first specialized chip (ASIC) for transformer models
- **Sohu**, the **fastest AI chip** to date, achieves **over 500,000 tokens per second** with Llama 70B, setting a new benchmark for AI product development.

## Hackers 'jailbreak' powerful AI models in global effort to highlight flaws
- **Hackers globally are 'jailbreaking' AI models** like Meta's Llama 3 and OpenAI's GPT-4o to expose vulnerabilities, with some even advising on illegal activities, showcasing the models' potential for misuse.

## [R] [CVPR 2024] AV-RIR: Audio-Visual Room Impulse Response Estimation
- The **CVPR 2024** presentation introduces **AV-RIR**, a novel method for **Audio-Visual Room Impulse Response Estimation**, enhancing audio quality in diverse environments.

## AI discovers new rare-earth-free magnet at 200 times the speed of man
- **AI identified a rare-earth-free permanent magnet named MagNex**, analyzing over **100 million** compositions, focusing on cost, supply chain security, performance, and environmental impact.

## [R] M3-AUDIODEC: Multi-channel multi-speaker multi-spatial audio codec
- **M3-AUDIODEC** introduces a **neural spatial audio codec** capable of efficiently compressing multi-channel speech in scenarios involving single or multiple speakers, while preserving the spatial location of each speaker. [Paper](https://arxiv.org/pdf/2309.07416.pdf) | [Code](https://github.com/anton-jeran/MULTI-AUDIODEC)

## [R] Grade Score: Quantifying LLM Performance in Option Selection
- The **Grade Score** metric, introduced to assess **Large Language Models (LLMs)**, evaluates **consistency and fairness** by measuring **order bias** through Entropy and **choice stability** via Mode Frequency, offering a new lens on LLM reliability and impartiality.

## [N] ESM3: Simulating 500 million years of evolution with a language model
- **ESM3**, a **language model** by EvolutionaryScale, simulates **500 million years of evolution**, generating **functional proteins** far from known ones, showcasing its ability to reason over protein sequence, structure, and function.

## [R] [IEEE VR 2024] Listen2Scene: Interactive material-aware binaural sound propagation for 3D scenes
- **Listen2Scene** introduces **interactive material-aware binaural sound propagation** for 3D scenes, enhancing virtual reality experiences by simulating realistic audio based on materials present in the environment.

## Confidence Regulation Neurons in Language Models
- **Entropy neurons** in large language models (LLMs) **scale down logits** by operating in an unembedding null space, impacting the residual stream norm with minimal direct effect on the logits, a mechanism critical for managing uncertainty in next-token predictions.

## EvolutionaryScale Debuts With ESM3 Generative AI Model for Protein Design
- **EvolutionaryScale's ESM3 model**, leveraging **NVIDIA H100 GPUs**, introduces a **revolution in protein design** by enabling detailed analysis of protein sequences, structures, and functions, aiming to accelerate discoveries in fields like cancer treatment and environmental sustainability.

## 🤗XLSCOUT Unveils ParaEmbed 2.0: a Powerful Embedding Model Tailored for Patents and IP with Expert Support from Hugging Face
- **XLSCOUT's ParaEmbed 2.0**, developed in collaboration with **Hugging Face**, enhances patent analysis by leveraging **AI** to understand complex patent documents with improved accuracy.

## [R] I finetuned GPT-2 and BERT 135,000 times to see if it's fair to pretrain on unlabeled text from the test set
- **Pretraining on unlabeled text from the test set** shows no significant evaluation bias across 25 text classification tasks, suggesting it's a fair practice for enhancing model performance without compromising fairness.