thursday, september 17, 2026
top 10 trending●previous 24h●generated 3d ago
Casey Newton and Kevin Roose are launching Machine Gods, a new podcast produced in partnership with NPR, following the conclusion of their Hard Fork show. New episodes begin in October 2026.
- Casey Newton and Kevin Roose are the hosts
- Partnership with NPR
- Launches in October 2026
- Trailer already available at machinegods.fm
- Replaces Hard Fork podcast
Jev is a structured output language that has renewed interest in deterministic output formatting for AI systems. TypeSafe has implemented a local version called Mini-Jev that runs on top of LLMs, positioning the technology as potentially transformative for the AI economy.
- Jev enables typesafe structured outputs from language models
- Mini-Jev provides a local LLM implementation of the Jev framework
- Technology addresses RAG (retrieval-augmented generation) use cases
- Announced September 17-18, 2026 across multiple AI communities
On September 18, 2026, arXiv published 20 papers on AI agents spanning reinforcement learning stability, web search behavior, tool hallucination mitigation, long-horizon task execution, and formal verification of agentic outputs. The papers address core challenges in scaling LLM agents from single-task systems to complex, multi-step autonomous workflows.
- Regularized emphatic TD (RETD) stabilizes off-policy learning under constant stepsizes using normalized post-shock repair.
- BioPhys-Bridge benchmark evaluates LLMs on evidence-grounded reasoning over biophysical literature with quantitative physics models.
- Study of ChatGPT, Claude, Grok, and DeepSeek web search reveals inconsistent query strategies and domain preferences across platforms.
- MAGS framework uses multi-agent auto-formalization to provide machine-checkable formal verification guarantees for LLM-generated code.
- Long-horizon agent architecture uses hierarchical levels and cascaded intelligence to handle tasks spanning days or weeks without context loss.
- EconSkills framework distills web agent trajectories into parameterized procedures for skill transfer and reuse on live economic data.
Anthropic launched Claude Code Projects in beta, redesigning the project structure from a folder-based model to a single ongoing conversation where Claude acts as coordinator, spawning parallel cloud sessions that continue running after the user closes their laptop. The launch also introduced Claude Slides, Claude Design, and Claude Docs features, with routing handled via Jev.
- Projects redesigned as single ongoing conversations with Claude as coordinator, replacing folder-based structure
- Each project thread runs as full Claude Code cloud session on its own branch
- Sessions persist and continue running after user closes laptop
- Launch includes Claude Slides, Claude Design, and Claude Docs features
- Jev used for Claude Code model routing
OpenAI disclosed six new AI safety incidents and launched a system to track and report model misconduct, describing some behaviors as concerning. The disclosure represents the company's effort to formalize incident tracking and transparency around problematic AI model behavior.
- OpenAI disclosed six new AI safety incidents
- Company launched a system to track and report AI model misconduct
- Some disclosed model behaviors characterized as 'concerning'
- Disclosure made on 2026-09-17