🎯 Top 3 Things to Know
1. Google's Pentagon Gemini deal goes "any lawful purpose" — ~600 Googlers sign an open letter against it. GenAI.mil now runs Gemini 3.1 Pro on classified data for "any lawful government purpose" — the exact language Anthropic refused and got blacklisted for. The internal backlash is now public; Fortune's read is this won't be a Project Maven repeat, since tech-sector layoffs have eroded employee leverage. Anthropic remains the only major lab that walked away from the procurement standard everyone else signed. Fortune · 9to5Google
2. IBM: 76% of large orgs now have a Chief AI Officer, up from 26% a year ago. IBM Institute for Business Value's annual CEO study (2,000 leaders, 33 countries) shows CAIO penetration tripled in twelve months. By 2030, CEOs expect 48% of operational decisions where consistency and guardrails can be codified to be made by AI without human review (up from 25% today). 29% of employees expect to need reskilling for a different role between 2026 and 2028; 53% need upskilling for their current role. IBM Newsroom
3. Grafana open-sources o11y-bench: an agent benchmark that runs on a real Grafana stack, not a static dataset. 63 tasks across PromQL, LogQL, TraceQL, incident investigations, and dashboard editing — graded on what the agent changes in the system, not what it claims to do. Across 29 model variants, Claude Opus 4.7 (reasoning off) leads on consistency, Qwen 3.6 Plus is the top open-source. The methodology — measure system mutation, not transcript — is more interesting than the leaderboard. Grafana Labs · GitHub
🚀 Frontier Models & Features
Quiet day — no new frontier-model releases in the last 24 hours. Cost compression keeps playing out from late-April releases (DeepSeek V4, Mistral Medium 3.5, Gemini 3.1 Flash-Lite) rather than fresh launches.
🔬 Research Worth Reading
"Shorthand for Thought: Compressing LLM Reasoning via Entropy-Guided Supertokens" (arXiv:2604.26355). Entropy-guided merge of reasoning tokens into "supertokens" — direct attack on CoT verbosity, framed as cost optimization. Verdict: read abstract; full paper if you're optimizing CoT cost in production.
"LLM Reasoning Is Latent, Not the Chain of Thought" (arXiv:2604.15726). Position paper: CoT surface text is decoration; reasoning lives in latent trajectories. If true, mechanistic interventions on hidden states beat prompt engineering — and eval rubrics that grade the trace are measuring the wrong thing. Verdict: read full paper — this reframes a load-bearing assumption.
"Rethinking Model Efficiency: Multi-Agent Inference with Large Models" (arXiv:2604.04929). Empirical case that one big model with fewer output tokens often dominates many small models with long outputs on cost-adjusted benchmarks. Counter-narrative to the "swarm of cheap agents" trend. Verdict: read full paper if you're scoping a multi-agent harness.
"ThinkJEPA: Empowering Latent World Models with Large Vision-Language Reasoning" (arXiv:2603.22281). Dual-branch JEPA: dense frame branch for fine motion, VLM "thinker" branch for long-horizon semantic guidance. Beats VLM-only and JEPA-only baselines on hand-manipulation prediction. Verdict: read abstract; full paper if tracking JEPA-family work.
🏢 Enterprise in the Wild
Microsoft Foundry ships per-agent, per-model token telemetry as standard. Eval results, traces, latency, token usage, quality metrics all publish to Azure Monitor and are KQL-queryable within minutes. Per-agent cost attribution down to the token — roughly where agent FinOps lives in production now. Microsoft Tech Community
Japan launches national push for AI-driven wet-lab automation. Researchers deploying AI-powered robots to run repetitive biology experiments. Notable as AI-for-science framed as national policy, not a vendor pitch. The Japan Times
🛠️ Tooling & Ecosystem
Grafana o11y-bench — see Top 3. Lift for your own harness: grade on the side effects the agent caused, not on its narrated transcript. Portable to any framework with persistent state. Most agent evals score what the agent says it did; this one scores what the world looks like afterward. Blog
⭐ ANTHROPIC — Official Blender MCP connector ships. Anthropic joined the Blender Development Fund as a patron and released an official Blender connector — 3D scene analysis and batch script edits over MCP. Pattern continues: creative-tool ecosystems shipping MCP integration before native AI features. Anthropic release notes
⚖️ Policy & Regulation
- EU AI Act Omnibus trilogue postponed; next round May 13. Parliament, Council, and Commission failed to agree on April 28 over how the Act overlaps with sectoral regulations. August 2 GPAI enforcement is still legally binding — but the gap between "deadline binding" and "Code of Practice ready" widens with each delayed trilogue. IAPP · Holland & Knight
📌 Watch List
- Cost-aware / token-economy agents. "Shorthand for Thought" attacks the same handoff overhead Recursive MAS measured last week — different abstraction, same target. CoT compression is now an active research lane (arXiv:2604.26355).
- Latent-vs-surface reasoning. A position paper arguing CoT text is decoration, not computation, would change which interventions work. Watch for empirical follow-ups (arXiv:2604.15726).
- Agent eval methodology. o11y-bench grades system mutations, not transcripts — second eval in two weeks to score side effects over self-reports (Grafana).
- World models / JEPA. ThinkJEPA's dual-branch design suggests JEPA is moving from "interesting alternative" to "shippable" faster than expected (arXiv:2603.22281).
- Frontier-lab government procurement. Anthropic's "all lawful purposes" refusal is now the definitional axis: every other major lab signed; only Anthropic held. Google's 600-employee letter is the first internal pushback to surface (Fortune).