TLDR AI 2024-04-01

xAI announced its next model, with 128k context length and improved reasoning capabilities. It excels at retrieval and programming.

Microsoft and OpenAI are reportedly planning a joint data center project that could reach $100 billion in cost, culminating in the launch of a massive AI supercomputer named “Stargate” by 2028.

China is now producing almost half of the world’s top AI researchers, surpassing the US, with 18% coming from US undergrad institutions. Despite pioneering AI breakthroughs, the US relies heavily on Chinese-born researchers, with Chinese talent making up 38% of top US-based AI professionals. The trend of Chinese researchers staying in China rather than moving to the US could impact global AI leadership dynamics.

Data, evals, and compute are essential for strong performing AI. This is especially true in enterprises. Evals may be one key moat that allows organizations to improve their AI products.

Researchers have introduced a new approach to understanding outdoor environments, overcoming obstacles like the varying conditions and lack of data that have previously limited advancements.

This paper introduces a control framework that combines AI and predictive models to facilitate smooth and safe lane changes in dense traffic, emphasizing cooperation with nearby drivers.

Zero shot voice cloning and speech generation capabilities in a powerful 700m parameter model.

EvoEval is a new benchmark suite that tests the coding abilities of Large Language Models more rigorously than ever before.

In natural conversation, people sometimes interrupt or talk over one another. This can be key to quickly coming to a consensus. This AI assistant predicts tokens while the person is talking and if it predicts enough in a row it will interrupt.

On April 2, hang out with Airbyte's Bindi Pankhudi and DataStax's Carter Rabasa as they dive into building a GenAI RAG app that can query the xkcd webcomic canon. This fun and interactive livestream will be chock full of code, covering the nitty gritty of:

▪️ Ingesting data at scale from multiple sources

▪️ Vectorizing that data using LLMs

▪️ Storing the information in a vector database

▪️ Using RAG to build a smart search engine

Space will be limited, so make sure to register today!

Anthropic's Claude 3 Opus has surpassed OpenAI's GPT-4 for the first time on Chatbot Arena. Chatbot Arena is a leaderboard run by the Large Model Systems Organization, a research organization dedicated to open models. Its site allows visitors to rate outputs from various models, enabling it to calculate the best models in aggregate. While Claude's rise is notable, GPT-4 is now over a year old.

Autonomous racing is advancing AI and machine learning in high-stress conditions, with competitions like the Indy Autonomous Challenge accelerating innovation in vehicle safety. Researchers and students use platforms like F1tenth for algorithm development, pushing the limits of autonomous vehicle capabilities on actual racetracks. These high-speed challenges contribute to a better understanding of machine perception, decision-making, and control systems for real-world traffic applications.

Qwen MoE is equivalent in performance to a strong 7B model with 1⁄3 of the activated parameters.

The embodiment hypothesis argues true intelligence necessitates physical interaction, prompting advancements in AI through simulations and real-world testing, despite challenges like the "sim-to-real gap," leading to the cautious deployment of AI robots in industries.

Microsoft's Copilot AI will soon run locally on PCs, requiring future AI PCs to have built-in neural processing units capable of over 40 TOPS.

Harness the power of AI and incorporate it directly into your workflows.