🎯 Top 3 Things to Know
1. NVIDIA released Nemotron 3 Nano Omni — a 30B MoE that activates only 3B params/token, with 256K context. A single open model that handles video, audio, images, and text without a stitched pipeline. NVIDIA reports up to 9x higher throughput than other open omni models on video/document workloads, runs on 25GB RAM, and topped six leaderboards including document intelligence. Relevant to anyone evaluating multimodal stacks for production. NVIDIA blog
2. Anthropic shipped 9 new Claude connectors for creative tools — Blender, Adobe, Autodesk, Ableton, Splice. Claude can now operate inside the software creative pros actually use. The signal: connector breadth, not raw MCP server count, is becoming the differentiator — and "agentic" inside enterprise creative workflows is starting to mean something concrete. 9to5Mac coverage
3. OpenAI and Anthropic briefed the House Homeland Security Committee on AI cyber threats. Both labs are now in regular rhythm with congressional staff on model capabilities and national security. Worth tracking because it's the channel through which voluntary frontier-model disclosure norms are being negotiated — and where the next executive action is likely to originate. Axios
🚀 Frontier Models & Features
NVIDIA Nemotron 3 Nano Omni (above) — 30B-A3B hybrid MoE with Conv3D, EVS, 256K context. Available on Hugging Face, OpenRouter, build.nvidia.com. Early adopters: Palantir, Foxconn, Eka Care, Applied Scientific Intelligence. HF model card
Anthropic Mythos Preview held back. Anthropic publicly confirmed it is not shipping Mythos because of its ability to find and exploit critical security flaws. Notable as a real-world demonstration of the responsible-scaling framework rather than a press release. CNBC
🔬 Research Worth Reading
Recursive Language Models (RLMs) (MIT CSAIL). RLMs treat the prompt as a Python REPL variable the model can grep, partition, and recursively query, rather than stuffing 10M tokens into a context window. GPT-5-mini wrapped in an RLM beats GPT-5 on the hardest long-context benchmark — and is cheaper per query. "Infinite context without summarization." arXiv · VentureBeat explainer
DyTopo: Dynamic Topology Routing for Multi-Agent Reasoning (arXiv). Rewires agent-to-agent connections per reasoning round via semantic matching, instead of fixed comm topologies. A signal for how multi-agent systems get more efficient without bigger models.
🏢 Enterprise in the Wild
Anthropic Managed Agents early adopters. Notion, Rakuten, Asana, Vibecode, and Sentry are now running Managed Agents in production for code automation, HR processes, and finance workflows. Pricing is normal token cost + $0.08/active session-hour; p50 time-to-first-token dropped ~60%. Anthropic engineering
Healthcare AI hits production scale. NVIDIA's 2026 healthcare survey: 63% of healthcare/life-sciences orgs are actively using AI; >50% report ≥2x ROI; ~$3.20 return per dollar over 14 months on average. The headline hidden in the data: Gartner still pegs ~75% of healthcare AI pilots as never reaching production — platform consistency is the differentiator. NVIDIA blog
🛠️ Tooling & Ecosystem
Anthropic — 9 new creative-tool connectors (Blender, Adobe, Autodesk, Ableton, Splice, etc.). 9to5Mac
Anthropic — Memory for Managed Agents now in public beta under the
managed-agents-2026-04-01header. Persistent state across long-horizon sessions without rolling your own store. Releasebot summaryAnthropic — Claude Code 2.1.117 rolls in faster MCP/plugin workflows, new skill search box, richer hooks, and an
alwaysLoadMCP option that bypasses tool-search deferral for trusted servers. Claude Code changelog
⚖️ Policy & Regulation
Bartz v. Anthropic — final fairness hearing this month. $1.5B settlement (largest copyright settlement in US history) covering pirated training-data acquisition. The ruling held: training itself = fair use; storing pirated copies ≠ fair use. Sets the operational standard for how to source training data legally. Authors Guild summary
NYT v. OpenAI/Microsoft — summary judgment scheduled this month. Result will set precedent on output-similarity claims, a different theory than Bartz. Material for ongoing generative-output liability questions. BakerHostetler tracker
EU AI Act — high-risk obligations apply 2 Aug 2026. Member states must stand up at least one AI regulatory sandbox by the same date. Enterprises with EU exposure should have GPAI documentation, post-market monitoring, and conformity assessments locked in this quarter. EU AI Act timeline
📌 Watch List
- Budget-aware reasoning / frugal AI: RLMs (above) are the strongest signal in months — recursive decomposition gets GPT-5-mini past GPT-5 at lower cost.
- Multi-agent systems: DyTopo (semantic routing of agent topology) is a notable architectural shift away from static comm graphs.
- Claude Code skills: new skill search box in 2.1.117 changes how skills get discovered at runtime.
- Agent evaluation: Managed Agents pricing model ($/session-hour) gives the field its first standardized cost unit for long-horizon work.