Twice a week, Data Points brings you the latest AI news, tools, models, and research in brief. In today’s edition, you’ll find:
Amazon gets into browser use with Nova Act
Claude goes back to school with program for higher ed
Google becomes a harder place for AI researchers to publish
Evaluating models’ ability to replicate cutting-edge AI research
But first:
OpenAI solicits feedback on a forthcoming model with open weights
CEO Sam Altman announced the company would be publishing a new open weight model with reasoning capabilities sometime in “the coming months.” OpenAI plans to host developer events in the U.S., Asia, and Europe to demonstrate the model, and published a web form soliciting developer ideas feedback. The new model would be OpenAI’s first open multipurpose language model since 2019’s GPT-2 and should give individual developers and large organizations more ability to customize their own versions and make them available for commercial or noncommercial use. (Sam Altman on X and OpenAI)
Runway’s new video generation model improves consistency of characters
Runway rolled out its new Gen-4 model to paid individual and enterprise customers. The new model uses a single reference image for characters or objects, plus text instructions for scenes, to generate scenes where the characters and objects do not morph into similar shapes, or transform into suddenly new styles, in the manner of many previous video models. These scenes can then be edited together to create coherent short videos. This greater narrative and dramatic continuity in generated videos could make them more useful for individual users and the entertainment industry. (Runway)
Amazon releases a research preview of its Nova Act model and SDK
Amazon AGI unveiled Nova Act, a new model designed for agentic computer use. Amazon says Nova Act outperforms Claude 3.7 Sonnet and OpenAI’s Computer Use Agent at interacting with text, icons, and UI elements on the web. Nova Act is available for developers for free in a research preview as part of a new website for its family of Nova models. (Amazon)
Claude for Education gives universities special chatbot access
Anthropic debuted a new program targeting students, instructors, and administrators in higher education. Claude for Education would give university users access to Anthropic’s chatbots, including a new Learning Mode that would try to guide students’ reasoning through problems rather than presenting answers. Launched with several high-profile university and educational technology partners, the program suggests different approaches to chatbot interaction may be more pedagogically useful, potentially smoothing AI adoption for teachers and students. (Anthropic)
Google DeepMind slows down publication of AI research
Current and former DeepMind researchers say that the company has introduced a tougher vetting process and more bureaucracy to make it harder to publish new AI and machine learning studies, especially if the results or methods might tip information to Google’s competitors. One new policy includes a six month embargo on research related to the company’s strategic AI priorities. This represents a significant shift for both the AI research community and for Google, whose public research has long been seen as essential in kickstarting the AI boom. (Financial Times)
Claude Sonnet 3.5-based agent tops new OpenAI ML research benchmark
OpenAI researchers released PaperBench, a new dataset and benchmark that tests large language models’ ability to recreate AI and machine learning papers from scratch, including understanding related research, creating and executing a codebase, and producing a version of the paper. Claude Sonnet 3.5 led all tested models by replicating 21 percent of selected papers, broken down into sub-components. OpenAI noted that none of the tested models outperform the human baseline. The benchmark could be useful in evaluating models’ overall performance in replicating technical research or at AI engineering, but public test materials like top conference articles often find themselves in models’ training data, contaminating future results. (OpenAI and GitHub)
Meme of the week
Still want to know more about what matters in AI right now?
Read this week’s issue of The Batch for in-depth analysis of news and research.
This week, Andrew Ng shared his approach to “lazy prompting”—a technique where you start with minimal extra input and refine only as needed.
“Contrary to standard prompting advice that you should give LLMs the context they need to succeed, I find it’s sometimes faster to be lazy and dash off a quick, imprecise prompt and see what happens. The key is whether you can quickly assess the output quality.”
