Google DeepMind has introduced Weather Lab—a new AI-based development designed to predict tropical cyclones. This service serves as a platform for testing experimental AI models capable of simulating storm development. Weather Lab relies on stochastic neural networks that can generate 50 possible storm scenarios over a 15-day period. This approach results in more accurate forecasts—during testing, the model shows positional deviations of only 140 km in 5-day predictions, outperforming most global solutions. It also provides better predictions of hurricane intensity and wind radius compared to regional physical models.
Another significant advancement is the new training methodology called Internal Coherence Maximization (ICM). Developed by experts from Anthropic, New York University, and George Washington University, this method enables models to learn independently without constant human intervention. The AI checks the logical consistency of its responses and their coherence with each other, improving overall output quality. In tests like TruthfulQA and GSM8K, the system produces results comparable to human-evaluated training, and in some cases, even surpasses it. For example, the model without specially planned training achieves up to 80% accuracy in determining the author’s gender—a higher rate than humans. In interactions with the Claude 3.5 Haiku chat model, ICM training helped win 60% of sparring matches against the version with human oversight. However, it’s important to note that the method works only with information the model already knows and struggles with long texts or new knowledge.
Regarding speed and efficiency: NVIDIA, in collaboration with Stability AI, optimized the generative model Stable Diffusion 3.5 using TensorRT. Thanks to this optimization, the model now runs faster and requires less VRAM. Previously, the Large version needed 18 GB of VRAM, but after optimization, that requirement was reduced to just 11 GB thanks to FP8 quantization technology. This makes the model accessible for a wider range of graphics cards, including RTX 40 series and Blackwell chips. TensorRT support transformed the model’s architecture and weights for powerful Tensor Cores, increasing rendering speed by 2.3 times for the large version and 1.7 times for the medium. Additionally, a lightweight SDK—eight times smaller—was developed, capable of dynamic on-the-fly compilation for Windows ML. Model versions are already available on Hugging Face, and a dedicated microservice called NIM will soon simplify integration into third-party applications.
In tool enhancement, Google has expanded Gemini AI’s capabilities within Workspace. The new version can automatically analyze PDF documents and Google Forms. When opening a PDF, the system creates brief summaries, suggests sentence writing, or poses questions—all supporting more than 20 languages. With Google Forms, AI begins to summarize responses to open-ended questions, highlighting key themes. This feature activates when there are three or more responses and will be available from June 26, initially only in English. Another useful feature is “Form Creation Assistance,” which generates templates based on descriptions or attached files (such as documents or spreadsheets). This will start rolling out to users from July 7.
Regarding international technology and sanctions circumvention, some Chinese engineers have found ways to operate abroad—such as in Malaysia. There, on leased data centers, they use 300 servers with prohibited export-to-China NVIDIA chips to upload their 80 terabytes of data for AI training. This is a response to strict US restrictions. Data transfer occurs via physical hard drive shipment rather than the internet, with shell companies established in Malaysia and equipment rerouted through third countries. However, increased monitoring by NVIDIA and stricter regulations in Southeast Asia may soon complicate such loopholes.