AI

TypeSafe AI released Jev, a System One Model that returns typed, calibrated decisions instead of text, priced at $42 per billion input tokens.

New research shows SynthID-Text watermarking changes how LLMs handle harmful prompts and tool calls, a behavioral effect dubbed sampling drift.

Fulcra Dynamics launched universal agent-to-agent communication, letting any agent harness like Claude or ChatGPT connect across peers and groups.

Paper2Agent from Stanford converts research papers and codebases into MCP servers, with prebuilt AlphaGenome, Scanpy and TISSUE agents now available.

Mozilla's State of Open Source AI report finds frontier AI models buy a 4.4-month head start at five times the per-task cost.

A proposed UK settlement rule overhaul would double the wait for permanent residence, creating a retention problem for AI employers.

A Princeton researcher proposes Recurrent Looped Transformer (RLT), carrying decoder state across tokens with 96 logical blocks per token.

New Mexico's Supreme Court held lawyer Stephen Aarons in contempt for citing fake testimony from fabricated witnesses in a ChatGPT-generated brief.

AI agents can close the gap between supply chain detection and action, executing bounded decisions before exceptions get expensive.

Google ToolGrad flips tool-use data generation: verify the chain first, write the query second, hitting a 99.8% pass rate.

Microsoft's September patch release fixes a record 972 vulnerabilities, 112 rated critical, as AI-assisted bug hunting drives unprecedented patch volumes.

California man with bipolar disorder sues OpenAI, alleging ChatGPT delusions fueled religious mania and a suicide attempt.