TLDR AI 2024-08-30

Yes, that's not a typo. 100 million tokens in context for reasoning and programming agents. Magic Dev also announced a partnership to build two new supercomputers on Google Cloud. This comes based on the recent $320m fundraise to accelerate its product development.

California's controversial AI safety bill, SB-1047, has passed the State Assembly and is now one step away from Governor Gavin Newsom's desk. Newsom must decide by September 30 whether to sign it into law.

OpenAI and Anthropic have agreed to allow the US government early access to their major new AI models before public release to enhance safety evaluations as part of a memorandum with the US AI Safety Institute.

MTMamba++ is a new model architecture designed to enhance multi-task scene understanding. This approach uses a Mamba-based decoder featuring two core blocks, STM and CTM, to capture long-range dependencies and improve cross-task interactions.

Mambda models are hard to accelerate with speculative decoding because of the lack of a KV cache for backtracking. This paper, from some of the original authors, showcases some novel acceleration algorithms and distillation methods.

When serving many concurrent requests to a large language model, things often get bottlenecked because we don't know how long output generation will be, so we get head of line bottlenecks. However, you can learn to rank relative lengths between requests and serve the shortest ones first, which leads to a 6.5x higher throughput for multi batch generation.

RAW-Adapter is a new approach that adapts sRGB pre-trained models to handle camera RAW data effectively.

The Qwen team has released vision language models built on top of Qwen 2.

Great open dataset from Nous Research for training models to call functions.

Join us for the AI event of the summer in NYC! Enjoy great food, drinks, and networking while learning from the successful AI journeys of top companies. Hear about the tools used, lessons learned, and wins on their path to production.

🗓️ When:Wednesday, September 4

🕺 Who's attending: Engineers & experts from DataStax, PayPal, NVIDIA, Vercel, Unstructured & more

🥂 What's on: Panel discussions, tech talks, demos, networking & a Hack Night

🗽 Where: NYC | Pier Sixty at Chelsea Pier

Attendance is limited, so RSVP today to be considered!

This study has adapted the Segment Anything Model 2 (SAM 2), originally designed for video annotation, to streamline the annotation of 3D medical images.

Meta has released some statistics about Llama model usage. It found strong demand for its models being deployed in enterprise settings.

Anthropic has shifted to a cautious endorsement of California's AI regulation bill SB 1047 after key amendments. Changes include reduced penalties, clearer language, and increased reporting requirements. The bill aims to ensure safety in AI practices, gaining support from Anthropic yet facing opposition from OpenAI and Meta.

OpenAI announced that ChatGPT has surpassed 200 million weekly active users, doubling its user base from a year ago, and noted that 92% of Fortune 500 companies are using its products.

ByteDance has introduced Douyin Search, an app for searching short video content on China's version of TikTok, directly competing with Baidu's search dominance.

Flow is a Mac dictation app that lets you speak naturally and writes in your style in every application, 3x faster than a keyboard.