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.

Before we get into today’s edition, a quick thank you. We’re now 23,378 readers strong. And just a few days ago, Beehiiv (the platform I use) reached out to tell me that this newsletter is now one of the fastest-growing on the platform.

This week:

• NVIDIA + Intel team up: former rivals, building a new force in AI hardware.

• Notion, Gamma, and Zoom roll out mature updates with fresh agentic AI capabilities. Seems like slowly everybody’s joining the agent party.

• Deep Dive: How people are really using ChatGPT — and what it tells us about the future of human decision-making.

Let’s dive in.

Weekly Field Notes

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

🌀 Google makes Chrome an agentic browser with Gemini built-in
→ Chrome now embeds Gemini directly, turning the browser into an execution layer for agentic AI. I’ve said it many times before that the browser is the default interface for AI agents. Google has been on a strong run lately, shipping a steady stream of high-quality features.

🌀 Google unveils Agent Payments Protocol (AP2)
→ AP2 is designed for secure agent-to-agent payments. Early signs of a financial standard for agentic commerce.

🌀 NVIDIA + Intel team up on infrastructure and computing products
→ Joint push to build the next generation of AI compute. Two rivals cooperating is an interesting development.

🌀 Notion 3.0 adds autonomous agents
→ Docs now write, summarize, and restructure themselves via agents.

🌀 OpenAI updates Codex with GPT-5
→ The upgrade targets long autonomous coding sessions.

🌀 MCP Registry goes live
→ A catalog of agent tools and servers, standardizing discovery in the Model Context Protocol ecosystem (Tutorial to get started).

🌀 Zoom Companion 3.0 introduces agentic AI
→ The new update brings proactive scheduling, follow-up drafting, and even decision-tracking across conversations.

🌀 Gamma 3.0 adds an AI “co-pilot” for presentations
→ Strong focus on real-time deck creation, storyboarding, and automated polish.

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

📘 Google Gemini AI coding via CLI
→ Great tutorial on coding via Gemini CLI directly coming from Google.

📘 OpenAI launches “Grove program”
→ A new program for technical talent before the startup stage. Not quite an accelerator — instead, a 5-week in-person track at OpenAI’s SF HQ with mentoring, early access to unreleased tools, and a dense peer network.

📘 DeepLearning.AI – Build AI Apps with MCP Server: Working with Box Files
→ A fresh course teaching how to connect MCP servers with Box.

📘 IBM’s AI Skills Mission – Training 2M people in AI (free)
→ IBM is rolling out a massive upskilling push with free AI certifications. Here are some top courses (and more can also be found here).

🌱 Mind Fuel
Strategic reads, enterprise POVs and research

🔹 Google DeepMind warns of “Virtual Agent Economies”
→ Their caution: when agents trade, cooperate, and evolve incentives faster than humans can monitor, the system may slip out of human control.

🔹 Vercel on Vibe Coding
→ Their State of Vibe Coding 2025 gives some good insights on the future of coding.

🔹 Origin on AI Financial Advisors
→ Origin built the first SEC-regulated AI financial advisor with multi-agent orchestration, real-time market grounding, and compliance guardrails — outperforming both human CFPs and frontier LLMs on financial reasoning tasks.

🔹 Albania appoints world’s first AI minister
→ “Diella,” is basically the “world’s first AI-powered virtual minister”, which now oversees public procurement in Albania. Bold experiment — but raises questions about practicality.

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

OpenAI researchers, together with Harvard and Duke economists, published the largest study ever on ChatGPT usage. They analyzed 1.5M conversations across their 700M weekly users.

The growth of ChatGPT

ChatGPT went live on November 30, 2022, as a simple “research preview.”
Five days later, it had already crossed 1 million users.

By November 2023 — less than a year in — it hit 100 million weekly actives.
Since then, usage has doubled roughly every 7–8 months.

Fast forward to September 2025: 750 million weekly active users. Nearly 10% of the world’s adults — a pace of adoption we’ve never seen before.

The usage of ChatGPT

To analyze usage the study sampled 1.5M anonymized messages from consumer ChatGPT plans (Free, Plus, Pro) and ran them through automated classifiers that labeled each message by purpose. For demographics like education and occupation, they used a secure data clean room.

Here’s what they found:

1. Work vs. non-work: the surprise split

• In 2024, 53% of usage was non-work.

• In 2025, it’s 73%.
  → Productivity matters, but the consumer surplus of AI — tutoring, advice, everyday problem-solving — may be bigger than the workplace ROI everyone obsesses about.

2. What people actually do with ChatGPT

Nearly 80% of all conversations fall into just three buckets:

• Practical guidance (how-to, tutoring, ideation)

• Seeking information (like search, but personalized)

• Writing (emails, docs, editing — not coding!)

Computer programming? Only 4.2% of messages (this was a major surprise for me).
Companionship/therapy? Just 1.9%.
And the myth of “everyone using AI girlfriends” doesn’t show up here.

3. The new decision engine

The authors break the mental model down into three layers:

• Asking → guidance, advice, decision support (49%)

• Doing → task execution, outputs that plug into workflows (40%)

• Expressing → feelings, chit-chat (11%)

The killer stat: Asking is growing faster than Doing.
AI isn’t just a task robot — it’s becoming a decision-making co-pilot.

4. Who’s using it?

• Early gender gap closed — women now slightly outnumber men among active users.

• Almost half of usage comes from under-26s.

• Growth is fastest in lower-income countries.

• Educated professionals skew heavily toward work-related use, especially writing.

5. The real leverage of AI?

I see too many people still frame AI as pure automation. That’s a narrow lens. And the paper above shows the real leverage is so much broader:

• Decision support across knowledge work.

• Consumer value embedded in everyday life.

• Early signs of digital divide reversal, with faster uptake in lower-income markets.

What makes ChatGPT (and other LLMs) remarkable is not just capability, but accessibility. Anyone can use it — across geographies, genders, or skill levels (and currently 10% of the world’s population uses it weekly). You don’t need to be an AI engineer or data scientist to benefit. That’s real democratization: the ability to tap into AI without building your own model.

And it’s not just about “end-to-end task automation.” The pull is toward decision support systems — tools that sharpen questions, surface comparisons, and improve the quality of choices. That’s where adoption is currently massively accelerating. The disruption isn’t that AI writes your email — it’s that it changes how you decide what to write in the first place.
And what to do next.

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

A year ago IBM built Docling. And it’s slowly becoming one of the most successful open-source projects to come out of IBM — now part of the LF AI & Data Foundation and already embedded in thousands of AI pipelines (nearly at 40k GitHub stars).

So what is IBM Docling?
IBM Docling is an open-source toolkit that converts complex documents like PDFs, Word files, and slide decks into structured, machine-readable formats such as JSON and Markdown, making them ready for use with generative AI models and retrieval-augmented generation workflows.

It helps you to
→ Parses PDFs, Word, PowerPoint, Excel, HTML, images, audio, and more
→ Extracts tables, formulas, code, layouts, and metadata with precision
→ Runs fully local for sensitive or air-gapped environments
→ Exports into clean formats: Markdown, HTML, JSON, DocTags
→ Plug-and-play with LangChain, LlamaIndex, Haystack, Crew AI, and MCP servers

Instead of brittle OCR scripts or cloud-locked APIs, Docling gives enterprises a unified, extensible way to convert unstructured content into LLM-ready data. It’s fast, private, and designed for agentic AI workflows.

I’ve already recommended this to several companies - and the feedback has been nothing but excellent.

Try it now:
→ Explore Docling on GitHub