# Jun 20, 2024

- **Safe Superintelligence Inc. (SSI)** aims to develop **the world's first safe superintelligence**, addressing both safety and capabilities through **revolutionary engineering**.

- **Claude 3.5 Sonnet**, the latest release in the Claude 3.5 model family by Anthropic, **outperforms its predecessors and competitors** in intelligence, speed, and cost, offering **free access** on Claude.ai and its iOS app, with **cost-effective rates** for higher usage tiers.

- **Ilya Sutskever**, a prominent figure in AI, has co-founded **Safe Superintelligence Inc.** with a focus on developing **Artificial Superintelligence (ASI)** without the distraction of product cycles.

- **Umbra Space has released nearly 1,000 satellite images** of ships using its Synthetic Aperture Radar (SAR) satellites, capable of capturing high-resolution images through clouds and at night, with the dataset totaling almost 7.5 TB.

- **Unique3D** is an official implementation designed for **generating high-quality and efficient 3D meshes from a single image**, with a focus on high fidelity and diverse textured meshes that can be produced in **30 seconds** from single-view wild images. [Paper](https://arxiv.org/abs/2405.20343) | [Project page](https://wukailu.github.io/Unique3D/) | [Huggingface Demo](https://huggingface.co/spaces/Wuvin/Unique3D) | [Gradio Demo](https://u45213-bcf9-ef67553e.westx.seetacloud.com:8443/) | [Online Demo](https://www.aiuni.ai/)

- **Character.AI has achieved a significant milestone in AI inference efficiency**, serving over **20,000 queries per second**, which is about **20% of Google Search's volume**, highlighting the scale and efficiency of their LLMs. [Statista, 2024](https://www.statista.com/statistics/195140/new-user-generated-content-uploaded-by-users-per-minute/?ref=research.character.ai)

- **ControlNet** is leveraged to animate the **Game of Life**, utilizing a unique approach that preserves the grid's structure through each iteration by employing a "control image."

- **Instruct-MusicGen** introduces a **novel text-to-music editing approach** by finetuning a pretrained MusicGen model, enabling music modification through text queries without the need for training specific models from scratch.

- The quest for a **cost-effective setup** to run 400B models locally explores alternatives to the prohibitively expensive 8 x H100 GPUs, suggesting options like **AMD EPYC CPUs** and the anticipated **Mac Studio M4 Ultra**.

- The **PixelProse dataset** enriches existing datasets like **CC12M, RedCaps, and CommonPool** with **over 16M pairs of image and dense caption** using **Gemini-1.0 Pro Vision** technology, aiming to enhance machine learning projects with more detailed visual descriptions.

- **NVIDIA AI Workbench** is a **free tool** designed to simplify the development, testing, and prototyping of AI applications, enabling developers to easily experiment with and customize models, including those based on retrieval-augmented generation (RAG).

- The **Data Is Better Together initiative**, a collaboration between **Hugging Face and Argilla**, has successfully engaged the open-source ML community in creating impactful datasets, highlighted by the **prompt ranking project** and the **Multilingual Prompt Evaluation Project (MPEP)**.

- Microsoft Research introduces a **new theoretical framework** that synergizes habitual and goal-directed behaviors using **variational Bayesian methods**, challenging the traditional view of these behaviors as distinct and separate.

- PyTorch has integrated **semi-structured (2:4) sparsity**, demonstrating a **10% inference speedup** and a **1.3x speedup** in model training, notably on NVIDIA A100 GPUs, with minimal impact on accuracy for tasks like image segmentation and DINOv2 ViT-L training. [Segment-anything](https://github.com/pytorch/ao/tree/main/torchao/sparsity#segment-anything) | [Benchmarking](https://github.com/pytorch/ao/tree/main/torchao/sparsity/training#benchmarking)