Hey, it’s Andreas.

Welcome to 2026 - and Happy New Year.

Pause for a moment and think back just one year.

In early 2025, we were still joking about ChatGPT failing to count the number of “r”s in “strawberry.” Reasoning models from Chinese frontier labs like DeepSeek had not yet reshaped the field. Open-source reasoning agents were niche. Claude did not have a dedicated coding agent. IBM’s Granite 3.0 had only just arrived. And the agent conversation itself was barely forming - MCP only started gaining traction in spring through some endorsement from Sam Altman

Fast forward to today, and the baseline has shifted completely. In AI, a single year now feels like a decade.

AI will not replace most professionals in 2026. But professionals who stop learning will be replaced by those who do not. I strongly recommend making it a habit this year to block dedicated time every week—at least a few non-negotiable hours—to learn, experiment, and apply AI hands-on.Continuous upskilling, hands-on experimentation, and real-world application are no longer optional—they are the cost of staying relevant and the easiest way to accelerate your career. That is why I will also keep a much stronger focus on practical learning this year. I’ m also co-authroing a practical, hands-on book on this topic, with an excellent group of people.

So let’s kick things off.

This week, I am starting with a curated set of nine free high-quality AI courses to help you level up your skills at the beginning of 2026. Practical, actionable, and focused on real workflows.

And we also cover:
→ Meta’s $2B acquisition of Manus AI
→ OpenAI’s rumored new device
→ How DeepSeek went viral with its latest research
→ And much more

Let’s dive in.

P.S.: If you have not seen it yet, I recently shared my AI predictions for 2026. We will revisit and pressure-test those ideas throughout the year.

Weekly Field Notes

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

🌀 Meta - acquisition of Manus AI for multi-agent coordination
→ Buys multi-agent coordination as platform IP to lock in the agent layer early.

🌀 SoftBank has completed its $41 billion investment in OpenAI
→ One of the largest private funding rounds in history, giving the Japanese conglomerate an 11% stake in the ChatGPT maker.

🌀 OpenAI - rumored AI Pen as a new ambient input device
→ Reports suggest OpenAI is exploring a lightweight, camera-integrated pen as a third core device. A strong push toward always-on, context-capturing AI interfaces beyond screens.

🌀 DeepSeek - mHC architecture for stable deep scaling
→ Stabilizes gradients in very deep models so you can scale depth without training collapse.

🌀 Z.ai - GLM-4.7 open-source-leaning coding model
→ Open-source-leaning coding model pushing agentic evals and tightening the gap to top proprietary coders.

🌀 QuestLabs - IQuest-Coder-V1 40B model for coding leaderboards
→ 40B coder claiming leaderboard wins, increasing pressure for reproducible evals and clean harnesses.

🌀 Alibaba Research - ALE ecosystem for scalable agentic LLM stacks
→ Open ecosystem for building agentic LLMs end-to-end, signaling stack standardization.

🌀 Netflix - real-time distributed graphs for personalization at scale
→ Graph infra at billions of nodes enables millisecond personalization at internet scale.

🌱 Mind Fuel
Strategic reads, enterprise POVs and research

🔹 IBM - AI year in review and signals shaping 2026
→ Good overview on four major trends from four different people on IBM.

🔹 Jack Clark - Anthropic co-founder on AI creating parallel worlds of work
→ Frontier agents now complete multi-hour coding tasks, with task length doubling every 4-7 months, rapidly widening the gap between those using advanced AI tools and those who are not.

🔹 a16z - why AI agents shift software from chat to action in 2026
→ a16z argues interfaces move beyond prompting toward execution, software becomes machine-legible, and agents turn from demos into deployable workers across voice, workflows, and core business systems.

🔹 Forbes - AI agents becoming core enterprise infrastructure by 2026
→ Forbes forecasts that 40% of enterprise applications will be agent-driven, signaling a shift from AI as an add-on to AI as a default execution layer inside business software.

🔹 Oxford - iterative deployment as a self-improvement lever
→ User-curated fine-tuning loops can drive step-change gains over repeated iterations.

🔹 Boris Cherny creator of Claude Code shows how he’s using it in practice
→ He runs multiple agents in parallel with minimal customization, reinforcing that modern coding agents are designed to work out of the box and scale through workflow habits, not complex setup.

🔹 Deloitte - Tech Trends 2026 and the shift to AI-native enterprises
→ Deloitte argues AI is no longer an add-on but the core operating logic of enterprise IT, with 2026 marking the transition from experimentation to AI-native systems by default.

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

If 2025 taught us anything, it’s this:

AI literacy is no longer a nice-to-have. It’s becoming the baseline for professional leverage!

If I could give you one piece of advice for 2026: make AI learning a habit. Block dedicated time every week - non-negotiable hours - to learn, experiment, and apply AI hands-on. Because the current pace is relentless. Some things you learned six months ago are already outdated.

Treat learning AI as a long game: learn, ship, unlearn, repeat. The real skill is building a fast feedback loop with the tools. Continuous upskilling, real experimentation, and shipping small outcomes are now the cost of staying relevant - and one of the simplest ways to accelerate your career.

So here are 9 free courses worth your time - clean, trusted, and practical.

1. DeepLearning.AI - AI for Everyone (Coursera)
   → Covers AI basics, what AI can and cannot do, how AI projects work, and how to think about adoption and impact.

2. Google - Introduction to Generative AI (Skills Boost) 
   → Covers core genAI concepts, LLM fundamentals, typical use cases, and responsible AI basics in bite-sized modules.

3. MIT OpenCourseWare - AI and Algorithms (SERC)
   → Covers social and ethical responsibilities, real-world harms, accountability, and how algorithmic systems shape outcomes.

4. Google - Machine Learning Crash Course
   → Covers supervised learning, common model patterns, practical exercises, and the core intuition behind training and evaluation.

5. Harvard (CS50) - Introduction to AI with Python
   → Covers search, optimization, machine learning concepts, and practical Python-based projects to build intuition and skill.

6. MIT - Introduction to Deep Learning (6.S191)
   → Covers neural networks, backprop, CNNs, sequence models, and modern deep learning workflows through lectures and labs.

7.  MIT - Foundation Models and Generative AI
   → Covers how foundation models work, training and adaptation concepts, genAI capabilities and limits, and how to apply them.

8. DeepLearning.AI - ChatGPT Prompt Engineering for Developers
   → Covers prompt patterns, iteration tactics, reliability techniques, and how to integrate LLMs into practical applications.

9. NVIDIA - Building RAG Agents with LLMs (DLI)
   → Covers RAG foundations, grounding with enterprise data, agent patterns, and practical implementation concepts for production-minded builds.

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

Claude is strong at general reasoning. But you also need repeatable execution. Claude Skills is Anthropic’s way to package procedural knowledge (and optional scripts) so Claude can run the same workflow consistently across Claude.ai, Claude Code, and the API. I became a heavy user over the holidays, and it’s very useful, though it feels like not many people use it or know about it.

Key traits:

• Less context rot because Claude loads only what’s needed

• Reusable workflow modules, not one-off prompts

• Consistent outputs across a team (format, brand, standards)

• Scales from personal productivity to automated pipelines

• Works across web, terminal, and agent builds

You can try it here: Claude.ai (settings - Capabilities - Skills - enable code execution and file creation).