🤯Musk’s Most Powerful AI Project

Elon Musk's xAI begins training Grok LLM on world's most powerful AI cluster

Key Points 

• The cluster has 100,000 liquid-cooled H100 GPUs from Nvidia. 

• Musk mentioned that the goal is to create "the world’s most powerful AI by every metric" by December this year, and the Memphis Supercluster will give them a significant advantage in reaching this goal.

👨🏻‍💻News - Elon Musk announced on X today that xAI has begun training on the world's most powerful AI training cluster, the Memphis Supercluster in Tennessee. Reports say the supercluster is in the southwestern part of the city and is the biggest capital investment by a new-to-market company in the city's history.

In his post, Musk explained that the cluster has 100,000 liquid-cooled H100 GPUs from Nvidia, which are highly sought after by AI model providers, including his competitors at OpenAI. He also mentioned that the cluster uses a single RDMA fabric for efficient, low-latency data transfer between compute nodes without overloading the CPU.

🕵️What's it for though? xAI plans to train its own language models on the supercluster. Musk also mentioned that the goal is to train "the world’s most powerful AI by every metric" by December this year. He added that the Memphis Supercluster will give them a "significant advantage" in achieving this. 

Google’s new weather system combines AI with traditional physics for faster, cost-effective forecasts

Key Points 

• Google’s new open-source weather prediction model combines AI with traditional forecasting methods.

• This combination allows it to generate accurate forecasts more quickly and with less computational effort. 

• According to researchers, AI weather models, such as Google's GraphCast, use fewer lines of code than traditional models, making forecasting quicker and more efficient while retaining  benefits of conventional methods.

👨🏻‍🔬News - Researchers at Google have developed a new open-source weather prediction model called NeuralGCM, which blends machine learning with traditional methods to provide accurate forecasts at a lower cost. 

This model addresses a gap between two approaches: machine learning techniques that are quick but struggle with long-term forecasts, and general circulation models that are accurate but slow and expensive. 

🧐How so? Instead of choosing one over the other, NeuralGCM combines both approaches. Basically, the new model still relies on a traditional approach for understanding large-scale atmospheric changes. However, it depends on AI to handle smaller-scale details, like cloud formations and regional weather quirks (such as San Francisco’s fog), where the larger models tend to struggle. 

This combination allows NeuralGCM to generate accurate forecasts more quickly and with less computational effort, matching the quality of forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF), a partner in the research.

🌞What's the most important point here? According to the paper, AI weather models are much more compact. For example, Google’s GraphCast, once trained on 40 years of weather data, runs with fewer than 5,500 lines of code. In comparison, the National Oceanic and Atmospheric Administration (NOAA) model needs nearly 377,000 lines. This shows that AI can make certain parts of weather forecasting quicker and more efficient while still retaining the advantages of conventional systems.