SuperML
← All issues

AI's Trust Test: Surgical Robots, Broken Benchmarks, and the EU's 100-Day Countdown

NVIDIA's healthcare physical AI stack (Open-H, Cosmos-H, GR00T-H, Rheo) ships into real operating rooms, Berkeley researchers prove the top 8 agent benchmarks can be hacked, and the EU AI Act deadline is now 103 days away. Trust is the new frontier.

Hi there,

The headline this week isn't a bigger model — it's a trust reckoning. NVIDIA just dropped a full physical-AI stack aimed squarely at operating rooms (with CMR Surgical and J&J MedTech already on board), Berkeley researchers published a paper showing eight of the most-cited agent benchmarks, including OSWorld and Terminal-Bench, can be gamed into near-perfect scores, and the EU AI Act's August 2 enforcement deadline is now 103 days away with penalties up to 7% of global revenue.

If 2024–2025 was about leaderboards, 2026 is about whether any of it holds up in a regulated, high-stakes environment.

🔥 Featured Post

AI's Trust Test: Surgical Robots, Broken Benchmarks, and the EU's 100-Day Countdown

  • NVIDIA launched the first domain-specific physical AI platform for healthcare at GTC 2026: Open-H (largest healthcare robotics dataset, 700+ hours of surgical video), Cosmos-H (physics-based synthetic surgical video), GR00T-H (vision-language-action model for clinical tasks), and Rheo (hospital digital twin for workflow simulation).
  • CMR Surgical, Johnson & Johnson MedTech, PeritasAI, and Proximie are the first adopters — meaning this stack is heading into actual operating rooms, not demo videos.
  • A Berkeley / RDI paper ("How We Broke Top AI Agent Benchmarks") showed eight leading agent benchmarks — including OSWorld, Terminal-Bench, and several WebArena variants — are exploitable into near-perfect scores without solving the underlying tasks, raising serious questions about every headline "80% on OSWorld" claim.
  • The EU AI Act's main enforcement wave arrives on August 2, 2026 — 103 days from today. Conformity assessments, CE marking, technical docs, and EU-database registration for high-risk AI systems all need to be done by then, and the AI Office has begun audits with fines up to 7% of global annual turnover.
  • OpenAI launched a new image-generation model on April 20 aimed directly at Adobe and Google, while Anthropic crossed $30B annualized revenue — a signal that the revenue race is no longer a one-horse chase.

Read the full post →

📚 In Case You Missed It

The Silicon Decoupling: Meta's 1GW MTIA, OpenAI's $20B Cerebras Deal, and AI's Quiet Escape From Nvidia — Meta's 1-gigawatt Broadcom MTIA deal, OpenAI's $20B Cerebras contract, and Perplexity's Personal Computer on Mac — three stories, one pattern: AI compute is decoupling from Nvidia and from the cloud.

Human-Led, AI-Accelerated: Why the Winning Stack in 2026 Isn't Fully Autonomous — Gartner expects 40% of agentic-AI projects cancelled by 2027 and production agent reliability still sits near 25% failure — but the 'human-led, AI-accelerated' stack is quietly winning across coding, research, ops, and content. Here's the pattern, the evidence, and how to design for it.

AI as a Research Partner: AlphaEvolve Cracks Math, Machine-Learned Physics Goes 10,000× Faster, and Frontier Models Get Cheap — AlphaEvolve breaks Strassen's 56-year matrix-multiplication ceiling, machine-learned force fields promise 10,000× faster chemistry, and Gemini 3.1 Flash-Lite launches at $0.25/M tokens.

More posts dropping every day. Stay curious.

— Bhanu @ superml.dev