Hey, it’s Andreas.
Welcome back to Human in the Loop — your field guide to the latest in AI agents, emerging workflows, and how to stay ahead of what’s here today and what’s coming next.

This week:

• NVIDIA + OpenAI team up on a $100B commitment to deploy 10 Gigawatts of NVIDIA Systems

• Google’s surprings find that  “90% of developers now use AI tools”

• Deep Dive: This week, we’re looking past the models and the GPUs.
  Because the truth is: AI’s biggest bottleneck isn’t technical.
  It’s human.

Let’s dive in.

Weekly Field Notes

🧰 Industry Updates
New drops: Tools, frameworks & infra for AI agents

🌀 NVIDIA’s $100B OpenAI stake
→ Nvidia’s latest $100B commitment to OpenAI adds to a pattern of circular deals: investing in customers (OpenAI, CoreWeave et. al.) who in turn buy or lease Nvidia GPUs.

🌀 Meta unveils Code World Model
→ Trained on execution traces. Steps toward runtime-aware agents.

🌀 Meta launches Vibes
→ AI-first video feed inside Meta AI app. Agents as curators.

🌀 OpenAI launches ChatGPT Pulse
→ Pro users on mobile now get daily personalized update cards, with ChatGPT doing overnight research using chat history, memory, and optional Gmail/Calendar connections.

🌀 AWS Nova Act in VS Code
→ New VS Code, Kiro, and Cursor extension brings agent building, testing, and debugging directly into the IDE.

🌀 Microsoft launches AI Marketplace
→ Azure Marketplace + AppSource merge into one hub with 3,000+ AI apps and agents, integrated across Microsoft 365 and Azure.

🌀 Google DeepMind updates Frontier Safety Framework
→ New protocols for advanced AI risks. Safety governance maturing inside labs.

🌀 Cloudflare open-sources VibeSDK
→ One-click deployment of a full AI vibe coding platform: isolated sandboxes, project templates, caching, and GitHub/Cloudflare export built in.

🌀 GitHub Copilot Agent GA
→ Now live for all paid users: an autonomous dev agent that runs in the background via GitHub Actions, drafts PRs, and iterates from review comments.

🌀 Perplexity adds AI Email Assistant
→ Native for Gmail & Outlook.

🌀 Google Chrome adds DevTools MCP
→ Public preview of a Chrome DevTools MCP server lets AI agents debug live web pages with 26 built-in tools for tracing, network analysis, and simulating user actions. Works with Claude, Cursor, Copilot, and Gemini CLI.

🎓 Learning & Upskilling
Sharpen your edge - top free courses this week

📘 Google’s 5-Day AI Agents Intensive (Nov 10–14)
→ Free online course covering agent architectures, tools, MCP interoperability, memory, evaluation, and deployment. Daily assignments, live sessions, and a capstone project to build and showcase your own agent.

📘 OpenAI drops Prompt Packs
→ A free library of 300 job-specific prompts across 12+ roles (IT, sales, marketing, HR, product, execs, engineers, CS).

📘 Snowflake: Building & Evaluating Data Agents
→ Free 2-hour course on designing multi-agent workflows that plan, execute, and query data sources.

🌱 Mind Fuel
Strategic reads, enterprise POVs and research

🔹 Google on AI adoption
→ DORA report 2025 with an interesting finding: “90% of developers now use AI tools”.

🔹 Cognition on the future of programming
→ Scott Wu (CEO, Cognition) reflects on building Devin, acquiring Windsurf, and why he thinks AGI may already be here. Key points: fundamentals of CS still matter, AI-first workflows are reshaping software engineering, and reinforcement learning could unlock the next wave of coding agents.

🔹 OpenAI drops GDPval
→ New benchmark measuring models on real-world economic tasks across 44 occupations and 9 sectors (law, nursing, engineering, finance, journalism, more). Built with expert-vetted deliverables (briefs, blueprints, care plans, slides). Early results: Claude Opus 4.1 tops aesthetics, GPT-5 leads accuracy, and frontier models now complete tasks ~100x faster and cheaper than humans.

🔹 Google publishes 60-page guide on AI agents
→ Technical deep dive on building scalable agents with the Agent Development Kit (ADK): covers agent components, orchestration patterns, MCP/A2A interoperability, deployment on Vertex AI, and AgentOps for observability and safety.

🔹 WEF & Accenture on Responsible AI
→ Enterprise playbook for governance, transparency, accountability.

♾️ Thought Loop
What I've been thinking, building, circling this week

AI success isn’t just about smarter models.
It’s about smarter people.

AI adoption has never been higher — McKinsey says 78% of companies are now using or exploring it. And yet, a huge number of projects stall before production, as highlighted in the recent MIT study, which went viral.

The usual suspects get blamed:

→ Unclear ROI
→ Poor data quality
→ Integration headaches

But here’s the factor most leaders overlook: AI literacy.

What is AI literacy?

It’s not coding. It’s the ability to understand, trust, and challenge AI in context.

It means:
→ Employees know when and how to use AI safely in their daily work
→ Experts can link AI systems directly to real business needs and outcomes
→ Leaders understand what AI can (and can’t) do, so they make better decisions

Without it, organizations fall into the two biggest traps: blind trust or blind rejection.

A simple framework

I came across a brilliant breakdown that captures AI literacy in four levels (see below):

Level 1 — None: No clue how AI works. Still far too common.
Level 2 — Basic: Knows the concepts. Can follow, not lead.
Level 3 — Intermediate: Applies AI meaningfully in their work. The SME sweet spot.
Level 4 — Strong: Shapes AI strategy. Evaluates trade-offs. Connects models to business outcomes.

This is a practical starting point to:
→ Diagnose where your org sits today
→ Define role-specific training paths
→ Move groups up the ladder in a staged way

There’s no one-size-fits-all. Executives, engineers, and managers need different levels of literacy.

Why literacy isn’t enough

BUT AI literacy alone won’t deliver transformation. You also need the right skills on the right teams.

Enterprise AI requires a bench that spans:
→ AI & data engineers
→ Cloud & security experts
→ Domain leaders who know where AI adds value
→ Change leaders who drive adoption across the org
→ Product managers who translate use cases into outcomes

Most companies don’t have this balance. Even with strong AI hires, projects stall without cultural enablement.

The unlock: layered enablement

So what actually works? The unlock isn’t another one-off workshop — it’s layered enablement.

1. Broad literacy for EVERYONE — so people trust, adopt, and contribute

2. Role-specific training — so finance, HR, ops, and sales teams learn how AI augments their workflows

3. Deep expertise where it matters — so technical teams can ship scalable solutions

4. Change management and culture — so AI is seen as a tool, not a threat

AI transformation doesn’t fail too often because the tech isn’t ready. It fails because the people aren’t.
Think about it.

🔧 Tool Spotlight
A tool I'm testing and watching closely this week

This week Orchids caught my attention — mainly because they ranked #1 on UI Bench + Design Arena, ahead of Devin, Cursor, Bolt, Replit, and more. That’s what put them on my radar.

I’ve been doing a lot of vibe coding recently, and Orchids looks like another strong entrant:
→ Handles frontend, backend, auth, DB, and payments end-to-end
→ Generates prototypes, mockups, and production-ready apps in minutes
→ Pushes toward full-stack autonomy — but still benefits from human customization and product sense

The user experience is surprisingly clean, it’s essentially free to try, and I’ve had fun playing around with it over the past few days.

→ Try it here: Orchids.app (unlimited credits promo was live at launch — not sure if it’s still active).