Hey, it’s Andreas.
Welcome back to Human in the Loop. We’re heading into our last full working week of the year, so the news flow is likely to start quieting down.

This week we cover:
→ OpenAI releases GPT-5.2
→ IBM launches four new learning paths on agentic AI
→ Microsoft drops the largest study on conversational AI
→ And a deep dive into why a trillion dollars of AI hasn’t moved the world yet.

Slightly different format today; let me know what you think.

Let’s dive in.

Weekly Field Notes

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

🌀 OpenAI releases GPT-5.2 → adding stronger reasoning, safety and multimodal depth.

🌀 Disney announces a $1B equity investment and licensing deal with OpenAI → enabling video and image generation with 200+ Disney, Pixar, Marvel and Star Wars characters across Sora and ChatGPT Images.

🌀 Google releases a wave of audio AI updates → positioning Gemini as the backbone for real-time speech across consumer and developer products.

🌀 Google releases AlphaEvolve → an autonomous agent that optimizes algorithms end-to-end, and it’s already being used internally “to solve Google’s hardest engineering problems”.

🌀 OpenAI, Anthropic + Block donate AGENTS.md + MCP to Linux Foundation → massive move towards formalizing open agent standards.

🌀 Cursor releases a visual editor for the Cursor Browser → enabling drag-and-drop UI edits, direct component inspection, and prompt-based changes

🌀 Shopify launches SimGym and Agentic Storefronts → enabling AI-driven customer simulations.

🌀 Stripe announces its Agentic Commerce Suite → making online stores agent-ready.

🌀 LangChain introduces Polly → an agent engineer for debugging and improving agents.

🌀 Mistral AI releases DevStral2 and the Vibe CLI → enabling project-aware coding agents.

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

📘 IBM releases four new learning paths covering 15 hands-on tutorials on agentic AI → expanding its practical enablement for building and orchestrating AI agents.

📘 Qdrant launches a short course on Multi-Vector Image Retrieval → hands-on training with ColBERT, ColPali and MUVERA.

📘 Daily Dose of Data Science releases a free 75-page MCP cookbook → celebrating one year of Model Context Protocol with hands-on projects.

📘 Anthropic on AI Safety Research (Fellows Program) → opened applications for the May and July 2026 cohorts.

🌱 Mind Fuel
Strategic reads, enterprise POVs and research

🔹 Google DeepMind’s Shane Legg says there are no fundamental barriers to continual learning and visual reasoning in AI → argues this could enable minimum AGI-level systems within the next two years.

🔹 Microsoft publishes the Copilot Usage Report 2025 → Largest study on conversational AI ever, revealing real-world agent behavior at scale.

🔹 Stanford releases its latest Foundation Model Transparency Index → with IBM leading the 2025 ranking at 95/100 across 100 transparency indicators.

🔹 Yohei Nakajima publishes a synthesis of NeurIPS 2025 research on self-improving AI agents → mapping emerging techniques for agents that reflect, self-train, and modify their own behavior over time.

🔹 Anthropic publishes a community Q&A with Anthropic philosopher Amanda Askell → sharing insights on how Claude’s character is shaped and reflections on the future.

🔹 Bloomberg reports on Meta rethinking its AI strategy → as the company shifts from open-source leadership toward building its first major closed, monetizable model.

🔹 OpenAI publishes State of Enterprise AI 2025 → highlighting widening ROI gaps across companies.

🔹 Perplexity releases its 2025 AI Agent Adoption Report → tracking global agent usage trends.

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

Just one month ago, none of this existed:

• Opus 4.5

• NanoBanana Pro

• GPT 5.2

• Gemini 3.0

• Grok 4.1

We are no longer living through a standard innovation cycle. This isn’t incremental progress - it’s a collapse of the time between research and reality.

Never before in human history have core cognitive systems multiplied and commoditized at this velocity. We used to measure the jump from research lab to “real world” in decades.

Then years.
Then quarters.
Now, we measure it in weeks.

We are living through an extraordinary chapter of human history that future generations will study as a tectonic shift, yet it feels strangely ordinary. This is how the most profound transitions happen: they don’t announce themselves with fanfare; they disguise themselves as a slightly more efficient Monday.

The consequence isn’t just better chatbots. It is a permanent state of acceleration where capability outruns our institutions, our regulations, and our biological capacity to adapt. Calling this a civilizational inflection point might feel hyperbolic - but pretending it is "business as usual" is a delusion.

The Trillion-Dollar Disconnect

And yet, here is the friction point.

It has been three full years since ChatGPT changed the world. Since then, we’ve entered a weekly rhythm of increasingly powerful models - releases that would have been considered breakthroughs not long ago, now arriving as routine updates. We have seen a trillion dollars in capital deployed and a massive migration of the world’s brightest minds into the AI sector.

BUT beyond the synthetic benchmarks and the hype cycles, we have yet to see a documented, macro-scale impact on global GDP or labor productivity. We are seeing a “capability explosion” without an “economic explosion.”

So where does the disconnect actually originate?

Is there a lag between tool adoption and institutional reorganization?
Are we overestimating the value of intelligence when it isn’t paired with physical-world transformation?
Are the structures of our world simply not capable of absorbing the speed we’ve created?
Or is this the familiar pattern of general-purpose technologies - powerful at first, but slow to register in GDP and productivity metrics?

🧩  Sector Brief
A report worth reading this week

Zendesk just released its CX Trends 2026 report, based on a global survey of 11,000 people across 22 countries, asking a single question: how will AI reshape customer experience by 2026?

 The research points to a decisive shift: high-performing CX will be anchored in AI-first operations. Here are the five key trends that define what high-performing CX will look like in the years ahead.

1. Memory-rich AI drives real personalization
Eighty-five percent of leaders say contextual memory will be essential for truly personalized journeys.

2. AI-powered self-service becomes the new standard
Instant, accurate resolution is non-negotiable. Eighty-five percent fear customers will switch brands over unresolved issues.

3. Multimodal support turns channels into one continuous conversation
Seventy-six percent of customers want to move between text, images, and video without restarting.

4. Prompt-driven analytics democratize decision-making
Eighty-one percent believe giving every employee the ability to ask AI questions will transform decisions.

5. Transparency becomes mandatory
Demand for explainable AI is up 63 percent, making clarity a core trust driver.

You get the full report here: „CX Trends 2026 report".

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

I’ve been working the last two weeks heavily with Nano Banana Pro (Google) and it clearly stands out. If you haven’t tried it yet, you should - it’s a significant step up from Nano Banana and well worth a look.

There’s no doubt this is by far the strongest image model for infographics, diagrams, and structured visuals. Layout control - including branded designs - iterative refinement, and text-in-image quality are all clearly ahead of the pack.

How to access it

• Gemini - Free, iterative, but you get a small watermark

• NotebookLM - Free, but you can’t iterate (less practical)

• Google AI Studio - Full control, no watermark, paid (~$0.13–0.24 per image)

If you do visual thinking or AI explainers, this one is worth your time.