TLDR AI 2025-04-17

Baseten is now offering inference on NVIDIA B200 GPUs. B200s are ideal for workloads with aggressive throughput, latency, and cost requirements. By leveraging B200 GPUs, Baseten customers are achieving:

📈 5x higher throughput

📉 50%+ lower cost per token

📉 38% lower latency for the largest LLMs

With B200 GPUs, you can run models like DeepSeek, Llama, and Qwen to power code generation, search, reasoning agents, and more. Get access to Baseten's B200s and run performant applications at scale.

Run model inference on B200 GPUs with Baseten

Grok, the chatbot from xAI, now includes Grok Studio, a canvas-like tool to build documents and basic apps. It's now live for all users.

Candidate for best blog post of the year for folks interested in generative modeling. This post breaks down the history, intuition, and key innovations for learned latents.

OpenAI has released the new o3 and o4-mini models, which improve ChatGPT's tool use and enable smarter, faster reasoning with integrated web search, file analysis, and image generation.

NVIDIA's EquivDM is a framework that enhances video diffusion by using consistent noise for better motion tracking and 3D-consistent outputs with fewer sampling steps.

Prime Intellect has trained a 32B network, fully distributed, with reinforcement learning for reasoning. It has open sourced a substantial amount of its code and useful libraries.

M1 is a Mamba reasoning model that has been trained with extended test time computation. It doesn't fully match state-of-the-art models, but it shows strong performance, especially on long context and through-put.

103K examples of highly filtered and decontaminated math problems for reasoning model training.

Prima CPP is an extension of llama.cpp that tries to enable mmaping of memory for large models to enable them to run on low RAM environments.

Tile Language is a concise domain-specific language designed to streamline the development of high-performance GPU/CPU kernels (e.g., GEMM, Dequant GEMM, FlashAttention, and LinearAttention). By employing a Pythonic syntax with an underlying compiler infrastructure on top of TVM, it allows developers to focus on productivity without sacrificing the low-level optimizations necessary for state-of-the-art performance.

Hugging Face has expanded its HELMET benchmark to include more models and insights, helping researchers evaluate long-context LLMs like Phi-4 and Jamba 1.6.

Google's 2024 Ads Safety Report highlights how LLM upgrades blocked billions of bad ads, suspended 700K+ scam accounts, and reduced impersonation scams significantly.

Assort Health, a leading AI platform for managing patient calls, has announced new funding, bringing total capital to $26 million, to accelerate its mission of improving healthcare access. The company's technology has led to 8x revenue growth since late 2024 by reducing call hold times and increasing appointment accuracy, evidenced by high patient satisfaction. Backed by prominent investors, Assort Health integrates with EHR systems, achieving a 99% scheduling accuracy and over a 90% resolution rate.

Stability AI and AMD optimized several Stable Diffusion models for Radeon GPUs and Ryzen AI, improving speed and performance for AMD users.

OpenAI plans to acquire AI coding tool Windsurf for $3 billion to enhance its generative AI capabilities.

Liquid is a multimodal LLM that integrates visual comprehension and generation by tokenizing images into discrete codes.