Ad
Short

William Fedus, who served as Vice President of Post-Training at OpenAI, has announced his departure from the company to pursue AI applications in scientific research. In an internal memo to colleagues, Fedus explained his plans to focus specifically on developing AI systems for physics applications. Post-training involves optimizing pre-trained AI models through additional training methods. Recent advances in this field include using reinforcement learning to enhance models' capabilities in mathematics and coding.

Fedus joins a growing list of senior executives who have left OpenAI in 2024, including the company's Head of Technology Mira Murati and Chief Scientist Ilya Sutskever, who departed to launch their own ventures. However, Fedus emphasizes that his departure remains amicable - OpenAI plans to invest in his new startup, viewing advances in scientific AI as an important pathway toward achieving artificial superintelligence (ASI).

Short

Cognition AI has secured hundreds of millions in new funding at a nearly $4 billion valuation, with venture capital firm 8VC leading the round. The investment doubles the startup's previous valuation, according to Bloomberg, which cited people familiar with the matter. The funding round attracted several prominent investors, including Founders Fund, Khosla Ventures, Elad Gil and Conviction Partners. The company's flagship product, Devin, which it markets as "the world's first AI software engineer," competes in an increasingly crowded AI coding landscape where competitive advantages remain unclear. Most existing coding tools function as AI-enhanced workspaces, powered by large language models from established labs such as OpenAI and Anthropic.

Short

Elon Musk's AI company xAI has acquired video startup Hotshot, which has developed three video generation models over the past two years: Hotshot-XL, Hotshot Act One and Hotshot. The startup says it's "excited to continue scaling these efforts on the world's largest cluster, Colossus, as part of xAI." The company's own service has been shut down. Existing users will be able to download their videos through the end of March.

 

Ad
Ad
Short

Mistral AI has released an updated version of its Small 3 model, now available as Small 3.1 under the Apache 2.0 license. The latest iteration brings enhanced text processing abilities, multimodal understanding, and an expanded context window that can handle up to 128,000 tokens. According to Mistral AI, Small 3.1 delivers better performance than similar models like Google Gemma 3 and GPT-4o-mini, achieving inference speeds of 150 tokens per second. The model runs on consumer hardware like a single RTX 4090 graphics card or a Mac with 32 GB of RAM. You can download Mistral Small 3.1 via Huggingface as a base and instruct version, or use it through the Mistral AI API and on Google Cloud Vertex AI. The company plans to expand availability to NVIDIA NIM and Microsoft Azure AI Foundry platforms in the coming weeks.

Image: Mistral AI
Ad
Ad
Short

Google has integrated native video understanding into its Gemini models, enabling users to analyze YouTube content through Google AI Studio. Simply enter a YouTube video link into your prompt. The system then transcribes the audio and analyzes the video frames at one-second intervals. You can, for example, reference specific timestamps and extract summaries, translations, or visual descriptions. Currently in preview, the feature permits processing up to 8 hours of video per day, with limitations of one public video per request. Gemini Pro processes videos up to two hours in length, while Gemini Flash handles videos up to one hour. The update follows the implementation of native image generation in Gemini.

Video: via Logan Kilpatrick

Ad
Ad
Google News