AI News 22 Jan 2025

A major US "AI Manhattan project" called Project Stargate, led by OpenAI and Softbank, with support from Oracle, MGX, Arm, Microsoft, and NVIDIA, was announced. This project involves a $500 billion investment in AI infrastructure, a sum that is about 1.7% of the US GDP. The investment is planned over four years. This project is likened to the Apollo Program in terms of its ambition and scale. The project has sparked discussion about a potential AI arms race.

Mistral AI is planning an IPO and expanding into the Asia-Pacific market.

Noam Shazeer announced a second Gemini 2.0 Flash Thinking with improvements on the 2.0 Flash and a 1 million long context window available for use. The model also has a 64K output token window. Additionally, AI Studio gained a code interpreter.

DeepSeek AI has made a significant impact with the release of its R1 model, which is noted for performing on par with or even exceeding OpenAI's o1 model, particularly in web-enabled tasks. The R1 model is open-source and MIT licensed, enabling free use and commercialization. It has demonstrated strong performance in coding and reasoning tasks, and it is being integrated across various platforms like Cursor, Codeium, and Aider. There have been discussions regarding the model's token length limitations and censorship, and the model has been used for math tutoring, praised for its step-by-step solutions. Also, a new DeepSeek V2 model was released with reduced operational costs and performance enhancements.

Liquid AI's LFM-7B model is described as the best-performing model in its size class, with a non-transformer architecture for low memory usage and is optimized for multiple languages.

A new paper on Mind Evolution, an evolutionary search strategy, has shown over 98% success on planning tasks with Gemini 1.5 Pro.