NVIDIA Blackwell Sweeps MLPerf Training 6.0

Good morning 👋 NVIDIA's Blackwell swept every category at MLPerf Training 6.0, including brand-new MoE benchmarks built around DeepSeek-V3 671B. If you're deciding which infrastructure to bet on for your next training run, the gap just got clearer.

In today's issue:

• 🔭 NVIDIA Blackwell sweeps MLPerf Training 6.0
• 🧠 GLM-5.2, Strands + LeRobot, Bedrock Guardrails
• 🔬 Red-teaming Anthropic's newest models, agent-written test debt, prompt injection in templated LLMs
• 🏛️ EU's Digital Decade: progress, but structural gaps to 2030
• 🛠️ Persistent agent memory how-to + 5 trending tools

🔭 THE ONE THING

🏆 NVIDIA Blackwell Sweeps MLPerf Training 6.0

MLCommons released Training 6.0 results yesterday, and NVIDIA's Blackwell platform took first across all seven benchmarks. The headline number: GB300 NVL72 systems trained 1.6x faster than GB200 NVL72 at equivalent scale. On DeepSeek-V3 671B (one of two new MoE workloads added this round), NVIDIA scaled to 8,192 GPUs, the largest Blackwell submission in MLPerf Training history. (NVIDIA blog, NVIDIA Technical Blog)

The real editorial is the benchmark addition itself. MLPerf has historically tested dense LLM and vision workloads; MoE at production scale is a different problem. Blackwell performing cleanly on 671B sparse architecture (the topology powering many of today's most efficient models) tells you the hardware advantage isn't locked to a single paradigm. For founders sizing compute contracts and teams planning pre-training runs: GB300 is worth pricing in if you were planning on GB200.

🧠 MODELS & RELEASES

• 🤖 GLM-5.2 ships on Hugging Face with explicit positioning for long-horizon tasks. Open weights, framed for multi-step agentic workloads rather than single-turn benchmarks. HuggingFace
• 🦾 Strands Agents + LeRobot (Amazon + Hugging Face) connects the HF Hub model ecosystem to physical robot hardware. Upload a policy to the Hub, deploy it to your robot via Strands. Closes the loop between open-source research and hardware-in-the-loop deployment. HuggingFace
• 🛡️ Amazon Bedrock Guardrails adds InvokeGuardrailChecks, an API for applying individual safety checks at any point in an agentic pipeline without creating full guardrail resources. Built for mid-flow auditing in multi-agent systems. AWS

🔬 RESEARCH HIGHLIGHTS

• Red-teaming Anthropic Fable 5 and Opus 4.8: independent researchers tested both models against 7,826 harmful intents across 10 harm categories using the HackAgent automated jailbreak framework. The paper maps which attack families succeed and at what rates. Useful baseline if you're building on either model and need to understand the residual attack surface. arXiv
• All Smoke, No Alarm: an audit of 932,000+ AI-agent-authored pull requests across 116,000 GitHub repos found that test files routinely lack oracle signals. Assertions, expected values, anything that would actually catch a failure. Agents write the test skeleton; the verification logic often isn't there. Worth checking your own codegen pipeline. arXiv
• Structural Role Injection in Handlebars-Templated LLM Prompts: Handlebars' triple-brace {{{x}}} syntax skips HTML escaping and is the default in Microsoft Semantic Kernel. The paper shows this creates a structural injection vector where malicious input can redefine roles within the prompt. If you're using templated prompts, default to double-brace and audit any {{{ usage in your codebase. arXiv

🏛️ POLICY & REGULATION

• European Commission published the 2026 State of the Digital Decade report, the fourth annual audit of EU progress toward 2030 digital targets. Infrastructure has advanced; structural gaps remain in AI adoption, digital skills, and cross-border data flows. The Commission frames these as investment gaps, not regulatory ones, suggesting the next lever is capital mobilization over new rules.

🛠️ TOOLS & HOW-TO

Give your AI agent memory that survives restarts with MemPalace

1. pip install mempalace and initialize a store with your agent's name as the namespace.

2. After each conversation turn, persist key facts: memory.write(fact, importance=0.8).

3. At session start, retrieve relevant context: ctx = memory.read(query=task_description, limit=10).

4. Inject the retrieved context into your system prompt before the first user message.

5. Tune the importance threshold to control how aggressively the store prunes low-signal memories over time.

Prompt: You are a helpful assistant. Relevant context from prior sessions: {memory_context}. Use this to maintain continuity — don't re-ask questions already answered, and reference prior decisions when relevant.

5 trending right now

• 🪨 caveman: Claude Code skill that cuts 65% of token spend by having the model respond in compressed, minimal-syntax output. GitHub
• 🏛️ MemPalace: highest-benchmarked open-source AI memory system, free and self-hostable. GitHub
• 🕶️ NVIDIA XR AI: public-beta SDK for building multimodal AI agents that run on AR glasses and XR devices. NVIDIA
• 🖥️ hermes-desktop: desktop companion for the Hermes Agent workflow, brings agent interactions off the terminal. GitHub
• ⚡ SageMaker container caching: AWS feature cutting end-to-end latency up to 2x on generative AI models during scale-out events, no code changes required. AWS

🔗 QUICK LINKS

• OpenAI released Deployment Simulation, a pre-release evaluation method using real conversation data to predict model behavior before deployment, a step toward closing the lab-to-production gap in safety evaluation.
• Google DeepMind is partnering with the UK government on an AI-powered planning prototype for faster housing approval decisions, the first DeepMind product aimed at planning bureaucracy.
• HPE + NVIDIA expanded the HPE AI Factory at HPE Discover Las Vegas to include NVIDIA Vera CPU and NVLink Fusion support, targeting enterprises moving agentic AI from proof of concept to production.

See you tomorrow.

Pradeep Perugu