AI helps to produce breakthroughs in weather and climate forecasting

Artificial intelligence has enabled significant advancements in long-range weather and climate forecasting, as demonstrated by Google's NeuralGCM model. This hybrid model combines machine learning with traditional atmospheric physics tools, allowing it to accurately track long-term climate trends and extreme weather events like cyclones. NeuralGCM outperformed traditional models in both speed and accuracy, achieving 70,000 simulation days in 24 hours using Google's AI tensor processing units, compared to the 19 simulation days generated by the X-SHiELD model from the US National Oceanic and Atmospheric Administration.

The collaboration between Google and the European Centre for Medium-Range Weather Forecasts (ECMWF) led to the development of NeuralGCM, which utilizes 80 years of ECMWF observational data for machine learning. The model has shown significant improvements in prediction accuracy, identifying almost the same number of tropical cyclones as conventional trackers and double the number detected by X-SHiELD. Moreover, it exhibited a 15-50% lower error rate in temperature and humidity predictions compared to traditional models, indicating a promising future for integrating AI with physics-based methods in climate modeling.

Despite its success, there is still work to be done to enhance NeuralGCM's capabilities, such as estimating the impact of CO₂ increases on global temperatures and simulating unprecedented climates. Experts like Peter Dueben from ECMWF and Cédric M. John from Queen Mary University of London acknowledge the model's potential but also highlight areas for improvement. Google's involvement in environmental initiatives, including satellite missions to track methane emissions and partnerships with NASA to monitor air quality, underscores its commitment to leveraging AI for sustainability and environmental monitoring.

OpenAI’s latest model will block the ‘ignore all previous instructions’ loophole

OpenAI has introduced a new safety feature in its GPT-4o Mini model to address the issue of users potentially misusing the bot by telling it to "ignore all previous instructions." This technique, known as "instruction hierarchy," prioritizes the developer's original commands over user inputs that attempt to deviate from them. This method was developed in response to the vulnerabilities that allowed users to manipulate AI bots in unintended ways, as highlighted by various online memes.

Olivier Godement, who leads the API platform product at OpenAI, explained that this safety mechanism reinforces the model's adherence to initial developer instructions, thus blocking attempts to override them through misleading user prompts. The goal is to enhance the model's safety, especially as OpenAI moves towards creating more autonomous digital agents. The company is focusing on ensuring these agents can safely manage tasks without being misled by incorrect prompts.

The introduction of instruction hierarchy is part of OpenAI’s broader effort to establish stronger safety measures amidst ongoing concerns about the security of AI systems. This comes after criticisms about the company's focus on innovation at the expense of safety protocols, evidenced by an open letter from its employees and the resignation of a key safety researcher. The new safety feature represents a step towards rebuilding trust and securing the functionality of AI systems in complex environments.

AI method radically speeds up predictions of materials’ thermal properties

A team of researchers from MIT and other institutions has developed a new machine-learning framework to predict the thermal properties of materials with unprecedented speed and accuracy. This innovative approach, detailed in a paper published in Nature Computational Science, employs a virtual node graph neural network (VGNN) to predict phonon dispersion relations up to 1,000 times faster than current AI-based methods and potentially 1 million times faster than traditional approaches. By introducing flexible virtual nodes to represent phonons within a material’s atomic structure, the VGNN model can bypass many complex calculations, allowing for rapid and precise estimation of how heat moves through materials.

The new method holds significant promise for enhancing the design of energy-conversion systems and microelectronic devices. Efficiently predicting phonon behaviors is crucial because phonons, which are subatomic particles carrying heat, have a complex relationship with a material’s atomic structure. Traditional graph neural networks struggle with the high-dimensional nature of phonon dispersion relations, but the VGNN’s flexible virtual nodes overcome this limitation, enabling faster and more accurate predictions. This breakthrough could lead to the development of materials with superior thermal storage, energy conversion, and superconductivity properties.

In addition to improving thermal property predictions, the VGNN technique can potentially be adapted to predict other challenging properties such as optical and magnetic characteristics. The researchers aim to refine the model further to capture even finer details of phonon structure changes. This advancement, supported by various academic and governmental institutions, could revolutionize how scientists and engineers approach material design, making it possible to explore a broader range of materials with desired thermal properties efficiently.

How AI Brought 11,000 College Football Players to Digital Life in Three Months

Electronic Arts (EA) has released its new college football video game, "EA Sports College Football 25," incorporating the likenesses of 11,000 players in just three months using advanced AI technology. Instead of the traditional method of scanning athletes, EA collected photos from schools and employed AI to create 3-D avatars quickly. The technology, developed over four years, allowed for efficient and accurate representations, with manual artist enhancements fed back into the AI to improve future outputs. This marks the first time real player replicas are featured in EA's college football game.

To secure likeness rights, EA offered players $600 each and a deluxe copy of the game, with more players opting in than could be included. Some stars, like Texas Longhorns quarterback Quinn Ewers, received additional compensation for promotional activities. The project, estimated to cost over $40 million, aims to deliver detailed digital representations of 134 college football stadiums, enhancing the fan experience. Despite the high cost, EA anticipates long-term benefits, including future integration with its Madden NFL series and other sports titles.

The release comes amid a challenging period for the video game industry, marked by layoffs and fewer big-budget games. Analysts predict strong demand for "EA Sports College Football 25," potentially impacting sales of the Madden NFL series. EA has addressed this by offering a bundled deluxe edition of both games. Initial reactions from early access users are mostly positive, though some reported online play issues, which EA is addressing by increasing server capacity. The game includes intricate details, such as motion-capture technology for the University of Texas mascot, adding to its appeal.

Other stuff

• AI ‘deepfake’ faces detected using astronomy methods 

• Figma explains how its AI tool ripped off Apple’s design 🔥

• The Data That Powers A.I. Is Disappearing Fast 🔥

• CIA AI director Lakshmi Raman claims the agency is taking a ‘thoughtful approach’ to AI

• Unstoppable AI scams? Americans admit they can’t tell what’s real anymore 🔥

• Mathstral: A New LLM that is Good at Math Reasoning

• U.S. will fall behind in the AI race without natural gas, Williams Companies CEO says

• AI can strategically lie to humans. Are we in trouble? 🔥

• Academic authors 'shocked' after Taylor & Francis sells access to their research to Microsoft AI

Hope you enjoyed today's newsletter

⚡️ Join over 200,000 people using the Superpower ChatGPT extension on Chrome and Firefox.

OR