# Recent AI Developments

- **Deep-TEMPEST** tackles the challenge of **eavesdropping on HDMI** by using deep learning to interpret electromagnetic waves, a significant leap from the unclear results of analog systems.

- **Meta's SAM 2 model** extends the capabilities of its predecessor by introducing a unified approach to segment objects in both images and videos, using a variety of inputs like clicks, boxes, or masks.

- **Activision has released the Caldera map from _Call of Duty: Warzone_ as an open-source data set**, aiming to foster innovation and learning in AI and game development through academic and research communities.

- **Meta's FAIR** has **launched SAM 2**, an **enhanced model** capable of segmenting objects not only in **images but now in video** as well, maintaining its **open-access** ethos.

- The **Ampere AmpereOne Aurora** is a **512-core AI CPU** designed for cloud-native AI computing, marking a significant advancement in processor technology.

- **Gemma Scope** is a new suite of tools designed to enhance the interpretability of language models, particularly focusing on **Gemma 2's** inner mechanisms through hundreds of open sparse autoencoders.

- **Amanda Bertsch** discusses the **"Unlimiformer"** architecture, aiming for **unlimited context length** in NLP, at the **Oxen.ai Paper Club** this Friday. [Oxen Community Call](https://oxen.ai/community)

- **ThinK** introduces a **query-dependent KV cache pruning method** that targets the channel dimension's redundancy, effectively reducing memory costs without sacrificing model accuracy.

- **Neurosymbolic AI** combines **symbolic reasoning** and **deep learning** to enhance reasoning over **knowledge graphs**, which represent complex, multi-relational data.

- **JaColBERTv2.5** significantly advances **Japanese neural information retrieval** by optimizing multi-vector retrievers for **constrained resources**, outperforming multilingual models in capturing linguistic nuances.

- **SAM2 outperforms popular object tracking frameworks** like bytetrack, botsort, and deepsort in tests, without the need for fine-tuning or parameter adjustments.

- The **Annotation Vocabulary** introduces a **transformer-readable language** for proteins, focusing on **biochemically relevant properties** without relying on amino acid sequences, thereby creating a new dimension for protein embeddings. [Read the paper](https://www.biorxiv.org/content/10.1101/2024.07.30.605924v1)

- **CLEFT** introduces an **efficient language-image contrastive learning method**, leveraging large pre-trained models and fine-tuning prompts to bridge the gap between complex clinical data and simple class labels.

- **Torchchat** is a new library by PyTorch that enables **efficient running of Llama 3, 3.1, and other large language models (LLMs)** on various platforms including laptops, desktops, and mobile devices, expanding upon previous work with native PyTorch 2.0 and CUDA for enhanced performance across more environments.

- **OmniBal** significantly **reduces the computational load imbalance** in **vision-language instruct-tuning models** by addressing data distribution and model architecture heterogeneity.