AI News

1. Arcee AI Bets Half Its Venture Capital on Trinity-Large-Thinking, a 400B Open Reasoning Model

Arcee AI. The 30-person startup invested $20M of its total funding on a single 33-day training run using 2,048 NVIDIA B300 GPUs to produce Trinity-Large-Thinking, a 400-billion-parameter open-source reasoning model (Apache 2.0) built for agentic tasks. The MoE model activates only 13B parameters per token and scores first place on Tau2-Airline and second on PinchBench while costing 96% less than Claude Opus 4.6 on inference. Source

2. Researchers Define What Counts as a World Model

OpenWorldLib. An international team from Peking University, Kuaishou Technology, NUS, and Tsinghua published OpenWorldLib, a unified framework establishing that world models must satisfy three core requirements: perceiving environmental input, interacting with surroundings, and maintaining persistent memory. The framework explicitly excludes text-to-video models like Sora, which lack real-world feedback loops, aligning with Yann LeCun's position on their fundamental limitations. Source

3. Agent Skills Excel in Benchmarks but Collapse Under Realistic Conditions

UC Santa Barbara, MIT CSAIL, and MIT-IBM Watson AI Lab. Researchers tested 34,198 real-world agent skills and found that performance degrades sharply under realistic conditions, with Claude Opus 4.6's pass rate dropping from 55.4% with pre-loaded curated skills to 38.4% when independently selecting skills versus a 35.4% no-skill baseline. Weaker models performed worse with skills than without, and the best skill retrieval methods achieved only 65.5% recall. Source

4. Nebraska Passes Conversational AI Safety Act Requiring Chatbot Disclosure

Nebraska Legislature. Nebraska passed LB 525, combining a Conversational AI Safety Act with the Agricultural Data Privacy Act, requiring operators of conversational AI services to disclose when users are interacting with AI rather than a human. The bill adds safeguards for youth accounts and mandates appropriate responses to sensitive topics like mental health and self-harm, and now awaits the governor's signature. Source

5. Maryland Passes AI Predatory Pricing Act Targeting Dynamic Pricing

Maryland Legislature. Maryland passed HB 895, the Protection From Predatory Pricing Act, prohibiting food retailers and third-party delivery services from using AI-driven dynamic pricing or consumers' protected-class data to set prices. Violations constitute unfair or deceptive trade practices subject to enforcement and penalties. Source

6. Wave of State AI Bills Advance Across California, Connecticut, Oklahoma, and Hawaii

State Legislatures. Beyond Nebraska and Maryland, multiple states advanced significant AI legislation during the week. California moved SB 867 (chatbots in toys), SB 947 (employment AI), and three healthcare AI bills through committees, Connecticut's SB 5 passed the Joint Judiciary Committee, Oklahoma's SB 1521 cleared the House committee, and Hawaii's HB 1782 and SB 3001 were placed on notice for floor votes. Source

7. Tesla FSD Supervised Receives First EU Type Approval in the Netherlands

Tesla. The Dutch vehicle authority RDW granted Tesla the first European type approval for Full Self-Driving Supervised, with rollout to Dutch owners beginning April 11. The certification followed 1.6 million kilometers of European road testing and 4,500 closed-track tests under UN Regulation 171, with Germany, France, and Italy expected to issue national recognitions within 4-8 weeks. Source