Maximilian Schwarzmüller

Tailwind is facing HUGE financial problems.

The CSS library is more successful than ever => More downloads per month than ever before.
But the company behind it struggling and had to lay off 75% of its employees yesterday.

According to Adam Wathan, its founder, revenue reportedly down ~80%, website traffic down ~40%.

And that website traffic is the problem: Tailwind made money by selling Tailwind Plus & Catalyst to website visitors.

Now with AI assistants knowing all about Tailwind and generating code without the developer ever visiting the docs, that revenue source is disappearing.
Hence that business model of “give docs for free → convert on premium components” is eroding fast.

And, of course, this isn’t just a Tailwind story - it affects anyone who publishes high-quality free content or open-source tools. AI benefits from that ecosystem, but doesn’t necessarily sustain it.

If we want tools like Tailwind to survive, they need new business models - or support through their current ones.
I really hope they figure it out. Tailwind is an awesome library that changed the way a lot of us build UIs.

Udemy is getting acquired (by Coursera) & the Udemy brand will disappear.

I've created & shared my first Udemy course almost ten years ago.

And to this day, that "Angular - The Complete Guide" course (which - like many other courses - got updated many, many times by me) is one of my bestselling courses.

I can't tell what this acquisition by Coursera means for my future as a Udemy instructor. Time will tell.
I will definitely keep on teaching - on one platform or another.

But learning that a brand that was THE main part of my professional life for the last 10 years will go away is really very, very sad.

This is a different post. It's very personal and important to me.

10 years ago, I changed my life - without fully being aware of it back then. Because on October 19th 2015, I registered my first company - still in Berlin at the time (I eventually moved to Munich in 2017).

The photo attached to this post shows my original registration form (unfortunately only in German - sorry about that).

I started this company because I wasn't satisfied with "just doing my normal job". I wanted to try something, risk something.
But most of all, I wanted to write code & help others to learn how to code. Because as a 100% self-taught developer, I knew (and still know) the struggles of learning and mastering coding.

I started my journey to becoming a developer when I was 12 or 13 years old. Just by reading books and trying things out - we didn't even have an internet connection at that time in my home.
I've been coding since then - but always just as a hobby, (almost) never as a job - apart from a few holiday or side jobs. Instead, I studied business administration and started working in banks and consultancies.

But that changed in 2015 when I finally decided to follow my passion and work as a developer and teacher.

If you're able to read German (or use ChatGPT for translation), you can see that my original business registration form contains a pretty lengthy business purpose that mostly was about freelancing. Because I never would’ve dreamed that teaching could actually sustain me and provide a reliable income.

That's why I feel blessed that through lots of hard work, dedication, more support than I could've ever hoped for by my spouse, family and friends, and - of course - also luck, I was (and still am) able to work fulltime as a teacher and developer and make a very decent living off that.

So, looking back at the last 10 years, I'm incredibly thankful for those amazing years and how my life changed over that times pan.
And now with AI (and life in general) providing plenty of new challenges and uncertainties, I'm excited to see what the future holds.

I'm not going anywhere, I'll continue teaching, writing code and all the other things I enjoy, and I hope I'll be able to write a similar post after 20 years.

If you stuck along all the way to the end of this post, that's amazing! If you took one of my courses, watched some of my YouTube videos, read one of my articles or my book, or simply follow me here on LinkedIn, please know that it truly means the world to me!

Thank you so very much - you're all amazing!

Naturally, as a developer, what do I do when I need a new professional photo of myself?

Buy a new shirt?

Use an existing AI service that requires 10 photos of myself to train a custom model?

Or use ChatGPT and hope for the best?

Of course not! I build my own business headshots AI service.

It's good. It's fast. And it's affordable (no subscription).

Give it a try if you don't want to build your own one! :-)

I just launched a new course! 🚀 🎉

👉 AI Agents & Workflows - The Practical Guide

And, yes, I know: "AI Agents" is an extremely hyped up term. And most people actually just mean "stuff that uses AI" when they say "AI Agents".

But: You can build genuinely useful workflows & agents that utilize AI (LLMs) to solve common problems, provide useful automations and save you time.

And that's exactly why I created a course on that!

It'll give you the solid foundation and the key concepts you need to know to build your own AI-powered workflows and agents.

You'll learn how to programmatically (with Python) build workflows that combine steps and tools that use LLMs to generate content, make decisions, perform research and much more.

This course will walk you through the theory you need (e.g., "Workflows vs Agents", "Agent Memory: Short-term & Long-term", "Managing Control Flow") and provide plenty of concrete examples to ensure that you're able to build actually useful AI workflows & agents.

The US government staggering the rollout of GPT 5.6 is bad news - just as its decisions regarding Fable 5 by Anthropic were.

Obviously, it's bad for developers from the EU (like myself). But it'll be a huge problem for the US, too. It'll hand China the win in the AI race.

If the US government controls who may or may not use the frontier models by US companies, open weight models like GLM 5.2 will become drastically more important and relevant.

Revenue will become a problem for OpenAI and Anthropic as the customer base gets smaller. That will lead to slower training progress and, ultimately, make it much harder to stay ahead of Chinese models.

This would be a great chance for frontier EU model providers. Unfortunately, we have none.

Right now, it's extremely easy to feel behind as a developer.

Obviously, the big question is: "Is software engineering still a career in 1 or 2 years?"
And then there's so much AI-related stuff to learn - if you dive into the AI coding world: Different tools (GitHub Copilot, Cursor, Claude Code, ...), models, working with agents, subagents, skills, rule files, MCP, ...

The list is endless - and growing every day.
At least it can feel like that.

But it shouldn't. There's no reason to feel behind or lost.
AI coding is a field that's evolving so fast - you're not missing out on anything, because what's hot and important today is forgotten or built into tools tomorrow.

That does NOT mean that you should ignore it though.

I absolutely recommend learning how to work efficiently with AI. I'm convinced "agentic engineering" (or "AI-assisted coding" or however you want to call it) is the future.

Employers will not want developers who refuse to work with AI or only use it half-heartedly.
And, to be honest, you CAN be more productive with AI (when using it as an assistant and steering it - NOT with vibecoding).

And I guess that is my answer to the "is it still a career" question: Yes, absolutely. But it will change. It'll be more about designing systems, defining guidelines, breaking problems up into smaller pieces, defining patterns, rules and best practices - and of course also about reviewing code.
I don't think we'll write 0 code in the future but it will be far less for sure.

But there is no reason to feel behind.

So looks like Amazon is running a neat promotion for my "React Key Concepts" (second edition) book! 🎉

If you don't have it already, and if you're interested in a React deep dive (including RSCs, Next.js & more!), you find a link in the comments! :-)

In my bubble, the last week or so has been almost exclusively about OpenClaw (formerly Clawdbot).

It's an impressive example for how personal AI assistants may look like in the future. For me, it's also a great showcase for all the security nightmares we'll face in this future AND for the lack of actual use-cases I have for such an assistant right now.

In order to get most out of OpenClaw, you probably want to install it on your system (eg your MacBook).
But you should absolutely NOT do that!
It'll open your machine up to serious prompt injection attacks and other vulnerabilities.

OpenClaw is not just a kind of vibecoded piece of software that's riddled with security issues, but also a tool that suffers from two of the biggest flaws of LLMs: Unpredictability and vulnerability to prompt injection attacks (which can be orchestrated in various ways, including malicious emails, websites or supply chain attacks via ClawdHub).

A successful attack could take over your bot and therefore the system on which it's running.

Securing OpenClaw is possible (at least to some extent - eg, running it on a VPS, using sandboxing and requiring approvals) but every extra security layer vastly reduces it's utility and usefulness.

Personally, I also don't yet fully see the vision of solving all my digital problems via chat interfaces and AI.
I strongly believe in purpose-built user interfaces and services and don't think chatting (via text or voice) is a great way of getting stuff done.
But I may be wrong or just not creative enough. Time will tell.

OpenClaw certainly is a highly interesting case study and I'm sure OpenAI, Google and others are watching and will try to release something similar. Obviously, they don't have the benefit of "it's open source, security doesn't matter" though.

Claude Code is an everyday tool for many developers. And like all tools, using it efficiently, and knowing about its more advanced features, allows you to get more out of it.

That's exactly why I built a brand-new course about Claude Code!

In that course, I explore the base usage, configuration and more advanced features like building (and using!) custom subagents or agent skills. I use Claude Code to build a complete demo project and I share my personal learnings from heavily using Claude Code for the past couple of months.

By the end of this course, you'll be able to go beyond basic Claude Code usage and prompts, and leverage the more advanced features and concepts when needed!

Claude Code received a new desktop app!

If you worked with the Codex app or Cursor (its agent mode) before, you know what to expect:
- sidebar with projects and agent sessions on the left
- chat area in the middle
- code review / diff editor and preview window on the right

But it's actually pretty good I'd say. Feels pretty snappy, and has a nice design.
So naturally I sat down an updated my "Claude Code - The Practical Guide" course to include this new app, including features like "Dispatch" and more!

How time flies!

A little over a year ago I published the second edition of my "React Key Concepts" book!

It's the first edition in better. More content (up-to-date, of course), coverage of Next.js basics, a look behind the scenes of server components ... and so much more!

Now with AI writing so much of our code, having a solid foundation matters more than ever. You want to be able to steer your AI and understand the code it's writing.
Or you're vibecoding and living with the consequences.

If you're interested in the book, it's now on sale on Amazon. Might be a good deal :)

$165,000 in tokens is what the Zig => Rust rewrite of Bun cost.

The official "Bun ported to Rust" blog post is out and I read it for you. Here's my summary (hand-written by the way, just saying...):

1. This was NOT vibe coding
Jarred Sumner didn't just ask Claude to "rewrite Bun in Rust" and hope for the best => He built a fairly elaborate workflow around Claude Code (and with Claude Code):
- a porting guide
- lifetime analysis
- multiple implementation agents
- separate adversarial reviewer agents + fixer agents
- CI loops
He also had to tweak infrastructure (EC2 instance) to support running the complex test suite in a loop with Claude porting and fixing code.

2. Amazing scale
Bun had 535,496 lines of Zig code, excluding comments.
The rewrite happened over 11 days.
At peak, Claude (Fable) wrote about 1,300 lines of code per minute. There were 64 Claudes running through dynamic workflows. The final result included 6,502 non-merge commits and a landed diff of +1,009,272 lines.
The token usage was also … not tiny:
- 5.9 billion uncached input tokens
- 690 million output tokens
- 72 billion cached input token reads
=> around $165,000 at API pricing
Jarred estimates that doing this manually would've taken 3 engineers with full codebase context about a year.

3. Rust was chosen for stability, not because Zig is bad
The post is not about "Zig bad, Rust good".
Bun only exists because Zig made the original project possible. Jarred is very clear about that.
The problem was Bun's specific situation: JavaScriptCore, C/C++ libraries, garbage-collected JavaScript values, manually managed memory, Node.js compatibility, async behavior, and a huge runtime surface.
That combination led to bugs like use-after-free, double-free, memory leaks and more.

Rust gives Bun stronger guardrails through things like ownership, Drop, the borrow checker, and more explicit unsafe boundaries.

4. The AI reviewers were as important as the AI implementers
One of the most important parts: the code written by Claude was reviewed by other Claude instances (with fresh contexts).
The reviewers were adversarial => Their job was to assume the code was wrong and find the bug.
That caught real issues, including code that compiled cleanly but was still broken.

5. The rewrite still introduced regressions
Despite the complex setup and workflow, the rewrite introduced 19 known regressions.
All of them were fixed by now.
Most regressions came from semantic differences between Zig and Rust:
- Zig assert vs Rust debug_assert!
- release bounds checks
- Zig comptime formatting vs Rust runtime formatting

6. Bun got better after the rewrite
According to the post, Bun v1.4 fixes 128 bugs that reproduce in v1.3.14.
It also brings:
- lower memory usage
- smaller binaries
- roughly 2–5% better performance in benchmarks

Vibe coding 🤪
Agentic engineering 🤝

Today. Will you join me?

Yesterday, OpenAI announced "ChatGPT Agent" - a new capability (not available in the EU yet) that allows ChatGPT to perform actions with help of a virtual computer. Like browsing the web (including interacting with websites), work on spreadsheets etc.

Even though it's still limited and in early stages, it's an impressive piece of technology.
But it's not all gloomy...

The Good
In theory, ChatGPT Agent will allow you to outsource all kinds of tedious tasks.
OpenAI showed a demo where the agent would buy clothes and a gift for a wedding. But obviously, the idea is that you can do more meaningful, helpful things with it.

The Bad
Right now, it's still limited. It's slow, it may get stuck, it may not be able to do what you want it to do.
The results it produces may not have the quality you need. More manual refinement might be needed. Or you might need to redo it from scratch.
But that's not the worst thing about it...

The Ugly
I see at least two BIG PROBLEMS:

1) Security: Prompt injection becomes a HUGE issue if you outsource complex tasks to AI agents that have access to a broad variety of tools.
Bad actors may trick the agent into doing things you don't want it to do. Especially if you give the agent access to protected resources (e.g., bank account) it doesn't take a lot of fantasy to imagine what could happen.

2) The web as we know it may be dead:
Even if 1) can be tamed with proper controls, barriers etc, there's another big problem => We're more and more working with a web that's populated with content by machines for machines.
Already, lots of content gets consumed through ChatGPT, AI summaries etc.
With more capable AI agents that will likely accelerate.
The same is true for producing content. AI already produces lots of posts (not this one, by the way) and blog posts.
I'm not 100% sure if that's the future we need.

But it's coming anyways, and AI obviously has many advantages and helpful use-cases, too.
Still, many challenges await!

There are many good reasons to run web apps, workflows or AI agents on a VPS. Using a cost-efficient, isolated, always-on machine makes sense for many uses.

That's why I created a new "VPS Essentials - The Practical Guide" course where I bundled up my knowledge on how to properly create, configure and use a VPS. I dive into concrete examples (web-app, one-off workflow, recurring workflows, OpenClaw) and of course also cover SSH hardening - AND also other core topics like TLS encryption, reverse proxies and more.

The course is now live on Academind Pro and can, as always, be bought standalone OR as part of our subscription offering. Existing subscribers already got access!

Enjoy! I hope you'll get a lot out of it! :-)

Finally another livestream! Can't wait to start in a few hours!

Will you join me? 😃

Today I woke up to the news of Meta laying off 600 employees from their AI division. Only a few weeks after going on a hiring spree (with signing bonuses up to $100m) for AI top talent.

Is that the end of the AI hype?

Well, at first glance, it might look like the AI bubble is bursting. But the reality is the opposite: Meta is doubling down on AI.

The cuts are part of a radical restructuring that consolidates power under new Chief AI Officer Alexandr Wang and his elite “TBD Lab”, focused solely on building Meta’s next-generation superintelligence model.

Meta seems to primarily be laying off members of the "old" FAIR (Fundamental AI Research) team to focus on "top talent", remove friction and overlapping responsibilities, and focus on thew new TBD Lab.

So, no, it's not the end of the AI hype - instead Meta seems to be doubling down on their new hires and strategy.

My Look Back At 2025 (as a web developer)

In theory, this was supposed to be “the year of AI agents.”

But in reality, that only happened because everything that included AI was labeled an agent.
Built a program that reads PDF files and summarizes them via AI? Congratulations, you got an AI agent.

So ... yeah ... it wasn't really the year of AI agents I'd say.

BUT one area where AI & AI tooling DID improve was coding assistants. Cursor got better, Claude Code was released - writing code with AI assistance (not to be confused with vibe coding, which I'm not a fan of) is on another level than it was a year ago. And that trend will continue, I think.

Tools that help scaffold code, refactor, or explore APIs quietly became part of many dev workflows. And many companies are expecting their developers to leverage those AI assistants.

However, when you lean too heavily on AI without knowing what you’re doing, you risk trading expertise for “vibe coding” - a term that was coined by Andrej Karpathy in 2025.
But "going with the flow" and "not really caring about the code" might work for quick prototypes, experiments or one-off programs.
For real-world projects? You still need solid fundamentals, thoughtful architecture, and a critical mindset.
I think 2025 showed us that, as developers, we can get a lot out of AI by leveraging our skills and combining them with the breadth of the "knowledge" of those LLMs and the speed at which they can spit out code.

Besides AI (where I could share more - see the graphic attached to this post), we saw a web ecosystem which experienced various attacks on npm packages and vulnerabilities related to React.
Which is especially a problem since React.js belongs to the "default stack" AI loves to use: TypeScript, React, Next.js, Tailwind.

There's nothing wrong with that stack, but I'm concerned about a lack of innovation that may result from such a default stack. We're not seeing that yet, I think (and we do indeed have strong "newcomers" like better-auth in 2025) but it is something I worry about.

I think 2026 will be an exciting year. I'm hopeful that the dev job market will improve, that we can leverage AI to build useful stuff AND that we can work with exciting technologies & libraries, including "new" ones like, for example, TanStack Start (an amazing alternative to Next.js).

Back with another livestream today (in ~1.5 hours).

Upon popular request, we'll work on an actual (side) project I'm currently playing around with.

Going to add authentication with Better Auth and payments with Polar.

Only got 90 minutes, so let's see how far we get 💪

I'll continue building the AI captioning web app from last week - should be fun again!
You can also join if you missed last week, no worries. 😃