The release announces Jev 1.14.0 and a set of experimental "danger zone" features aimed at creative users who want to push boundaries rather than immediately deploy to production. The team frames this as a balance between reliable, battle-tested models for production builders and more adventurous capabilities for "dreamers," exposed via API header flags. Jev 1.14.0 is available under jev-preview with modest overall improvements (notably better multilingual support), jev-latest will be updated in a week, and jev-1.13.0 remains available with a planned deprecation timeline (60 days warning for 1.13.0, 30 days for 1.14.0). Users can opt in to share queries to help diagnose jaggedness (data will not be used to train models), and the team will publish internal intelligence measurements for open-weight models with explicit caveats about bias and evolving methodology.
Concrete experimental features include composable images (multimodality enabled by inserting image assets via a JSON interface - state, instructions, criteria - with the warning that multimodal performance lags pure text) and a new "box" type that returns a single bounding box for spatial outputs, available in jev-1.14.0. Longer context lengths can be unlocked with a danger-zone header despite decreased reliability on long contexts. The roadmap emphasizes incremental, stability-focused releases and features intended to be polished into production functionality over time.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
LFM2.5-24B-A2B is a 24-billion-parameter model designed for fast, on-device reasoning and agent tasks, offering high speed and competitive performance. It features an expanded vocabulary, longer context window, and extensive training, making it suitable for private, real-time applications. (liquid.ai)
An AI-powered interviewer interacts with an AI interviewee, leading to unpredictable and chaotic exchanges. The event highlights potential challenges and behaviors when AI systems communicate with each other directly. (theregister.com)
Anthropic's revenue increased to $16.8 billion in the third quarter, and the company claims to be profitable. The company is known for developing AI models and safety-focused AI technology. (ft.com)
Open source text-to-CAD tools in Elixir integrate build123d, OpenSCAD, and smlibpy for 3D printing applications. The repository has 17 stars and includes code for converting text to CAD models using AI and scripting tools. (github.com)
Google Research's Deep Review tool provides detailed AI feedback on complex documents to enhance understanding and analysis. The initiative is part of a broader effort to advance AI research across multiple disciplines and collaborate globally. (research.google)
Behavioral harnesses significantly improve the reliability of cost-optimized AI models, achieving consistent results comparable to larger models at lower costs. Despite increased token usage and runtime, harnesses can make smaller models more dependable and economically viable. (martinfowler.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.