Hey, it’s Andreas.
Last week was full of drama in the AI world.

What started as controversy around the Pentagon deal between OpenAI and Anthropic quickly turned into a full-blown public fight. Anthropic CEO Dario Amodei reportedly accused OpenAI of spreading "straight up lies" in an internal memo, Anthropic sued the Pentagon, ChatGPT uninstalls surged, Claude downloads climbed, and even OpenAI employees appeared to side with Anthropic.

This story is clearly not over, but it deserves close attention. Because this is no longer just about what AI can do. It is about who gets to decide what it should not do.

But instead of getting stuck in the drama, let’s focus on what happened over the last few days.

In today’s issue:

• Microsoft launches Copilot Cowork inside Microsoft 365, powered by Anthropic’s Claude

• OpenAI releases ChatGPT 5.4 with a 1M of context window

• Anthropic drops new insightful research on labor market and AI exposure

• And much more.

Let’s get into it.

Weekly Field Notes

🧰 Industry Updates

🌀 Google releases NotebookLM video overviews → New cinematic video overviews available now in NotebookLM. Helps you to transform uploaded files into polished video explainers rather than simple narrated slides.

🌀 OpenAI drops GPT-5.4 → GPT-5.4 is the newest model, which finally brings up to 1M tokens of context and “improved tool use for longer, more autonomous workflows”. 

🌀 OpenAI works on a GitHub alternative → Reports say OpenAI is building an internal alternative to GitHub

🌀 Microsoft introduces Copilot Cowork → Microsoft just launched Copilot Cowork, a new Microsoft 365 capability that uses Anthropic’s Claude system to run tasks in the background across apps. It arrives with the $99 enterprise bundle.

🌀 IBM released Granite 4 1B Speech → Model topped the Open ASR leaderboard while staying small, fast, and built for enterprise use.

🌀 Google releases Workspace CLI → Google launched an open-source CLI for Workspace with 40+ built-in agent skills. It is designed to plug more easily into agent-based platforms.

🌀 Alibaba on turbulence inside Qwen → Alibaba lost Qwen tech lead Junyang Lin just days after its Qwen 3.5 small model launch.

🌀 SambaNova releases new chip for inference-first AI infrastructure → SambaNova introduced the SN50 chip, built for better inference speed, latency, and cost across chained model calls.

🎓 Learning & Upskilling

📘 OpenClaw course for beginners → Good course that walks you through OpenClaw from setup to real-world use cases. Good entry point if you want to get started without fumbling through the basics.

📘 DeepLearning.AI on building and training an LLM with JAX → New course with Google that teaches you to build a 20M-parameter GPT-style model in JAX, from training and checkpointing to inference. Solid pick if you want a hands-on look at how modern LLM pipelines actually work.

📘 Hugging Face cheat sheet for beginners → Beginner-friendly cheat sheet that covers the basics of Hugging Face and points you to the key resources for getting started with open-source AI.

📘 MCP vs Skills → One of the most important distinctions in agent architecture right now: Skills tell the model how to work, while MCP tells it what it can access. In short, Skills shape reasoning, MCP shapes connectivity - and strong agents need both.

🌱 Perspectives & Research

🔹 Thariq Shihipar on building Claude Code → Good read from Anthropic’s Claude Code Lead on how Claude Code evolved, including why the AskUserQuestion tool became necessary.

🔹 Polsia Podcast after hitting $1M ARR → First podcast with Polsia, a company that builds agents to autonomously run your business, right after they crossed $1M ARR, up from just $50k on February 1.

🔹 Arc Institute and Stanford on Evo 2 → An open-source AI trained on 9 trillion DNA base pairs across all domains of life. It can identify complex features like regulatory regions and splice sites.

🔹 IBM Research released TerraStackAI → IBM Research introduced TerraStackAI, an open-source platform for preparing satellite data, fine-tuning geospatial AI models, and deploying them into production.

🔹 Amplifying - What Claude Code defaults reveal? → When Claude Code is asked to build, it keeps always reaching for the same stack: Stripe, shadcn, and Vercel.

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

Anthropic just published new labor market research this week. And it’s one of the clearest signals yet that the labor market is being re-priced in real time.

And the uncomfortable part is this:

The jobs we told people were safe because they required degrees, credentials, and years of training may be far more exposed than the jobs rooted in the physical world.

Computer programmers, customer service reps, data entry workers, medical record specialists, and financial analysts rank among the most exposed roles in the dataset. In several of these occupations, Claude can already perform a surprisingly large share of the underlying tasks.

That alone is significant.

But I do not think the most interesting part of this data is which jobs are exposed.

I think the most important part is the gap between capability and actual usage.

AI can theoretically perform a large share of tasks across knowledge professions already. Yet most organizations are still using only a small fraction of that capacity today.

Because technological disruption rarely happens the moment something becomes possible.
It happens when companies redesign workflows around it.

Right now, in most companies, AI is still being layered on top of existing processes.
It helps write faster.
Research faster.
Analyze faster.

But the structure of the work often stays the same.

But the real inflection point comes when the process itself is rebuilt around AI.

That is when roles change shape.
That is when headcount models change.
That is when the economics of teams start to move.

So I do not think the core story is that most professions will fully disappear.

It is that leverage per worker will rise sharply.

One person will be able to produce more.
Some teams will get smaller.
And the traditional entry-level pipeline may narrow as companies need fewer junior people to do the work AI can already absorb.

That is how labor markets change quietly at first.

Not always through mass layoffs.
Often through slower hiring, fewer junior openings, and a gradual redesign of how work gets done. 

Very interesting to see is also that some of the safer jobs are still bartenders, dishwashers, lifeguards, and other roles anchored in physical reality. Not because they are low value. Because they are harder to automate end to end.

Physical work still has a moat.

Now here is the part I do not see enough people talking about:
This may be an American report based on American workers and an American model.

But the implications are global.

India has millions of IT services workers concentrated in roles that sit close to the exposure zone.

The Philippines built a major share of its economic model around BPO and knowledge work services.

Many emerging markets scaled by becoming the world’s back office.

AI changes that equation fast.

If knowledge work can be automated, augmented, or delivered with dramatically fewer people, then this is not just a workforce story.
It is a geopolitical and economic development story.

Some countries will adapt faster than others.

And the countries that move fastest on AI adoption, workflow redesign, and education reform will not just gain a productivity edge.
They may reshape the balance of economic power.

That is why I do not think the real takeaway is "degrees are dead."

Degrees are not dead.

But they are no longer enough.

The big question is no longer whether AI will affect knowledge work.

It will.

The real question is whether your country, your company, and your career are on the right side of the adoption curve.

Full report: https://lnkd.in/eSYscHrB