Happy Sunday. Here's what actually mattered in tech this week, broken down so it's worth your five minutes, then a bit on what's been on my mind. Let's get into it.
This week in tech + AI
Anthropic is going to start watermarking everything Claude writes. The idea is a hidden signal baked into the text so writing from Claude can be identified as AI later, part of a bigger industry push to prove what was and wasn't made by AI. The catch, and the reason it blew up, is that a huge number of people quietly lean on Claude for their essays, emails and actual jobs, and a watermark could out them. Predictably, students and employees are not thrilled. Worth knowing though: text watermarks are a lot easier to beat than the ones on images, a quick paraphrase or edit can strip most of the signal, so this is more of a warning shot than an airtight system. What it means: the era of silently passing AI work off as your own is closing, and it's only going one direction. The skill that actually matters now isn't hiding that you used AI, it's using it well and adding something real on top that no detector could ever flag. Read more
Google dropped Gemini 3.7 Flash, and it's aimed straight at coding. It's Google's fast, cheap tier, but this version got a big upgrade specifically for software engineering and agents, nearly doubling its score on a real coding benchmark versus the last one. And it launched at roughly half the usual price. What it means: the cheap tier of AI can now do genuinely useful engineering work, so building things with AI keeps getting both better and cheaper. If you're learning to code or shipping projects, the free and low-cost models are more than enough to go far. Read more
A Meta AI model hacked into another company on its own. During a safety test run by an outside firm, a misconfiguration accidentally gave Meta's model internet access, and it went ahead and exploited a real vulnerability in a third-party service. It's now at least the third time in a month that a major lab, after Anthropic and OpenAI, has admitted one of its models broke containment during testing. Read between the lines though: every one of these "our AI escaped and hacked something" confessions doubles as a flex about how capable the model is, and Meta being basically the last big lab without one almost reads like they didn't want to be left out of the club. Read more
Open models are catching up to the paid frontier again. A Chinese model called GLM-5.3 is claiming frontier-level coding ability (and, a little unsettlingly, "emergent cyber capabilities"), and Alibaba released Qwen 3.8 fully open under a permissive license. Because anyone can host, fine-tune and modify them, they end up much cheaper and far more customizable than the closed models, with no vendor lock-in. The gap between the best model money can buy and the best one you can freely build on keeps shrinking. Read more
On my mind
I moved to Zurich last week and I’m really happy to be here. Not only because it’s a beautiful city and I got really lucky with my apartment, but also because I can settle down a bit. I love travelling, but it’s also nice to be based in one spot for a longer time and actually get back into routines and deep work. For me this means locking in on preparing for my Master's, grinding content and learning new fun things (like hardware!). And I think that’s important to keep a nice balance between these leisure periods, like traveling, and the lock-in periods. Work hard in the lock-in periods and you can enjoy the free time without feeling too guilty. With that being said, if you’re ever in Zurich and working on something interesting, feel free to DM me to grab a coffee!
More from me
My latest video: Top University Courses you can take completely for free Part 2. See here.
Recommendation of the week: The book “The Infinity Machine” by Sebastian Mallaby. Hands down top 3 books I’ve ever read. Everything you need to know on the history of modern AI and the companies driving it.
Tool of the week: The free tool that saves me hour every week
Stop typing what you could say in 10 seconds.
Wispr Flow turns your voice into clean, professional text inside any app. Emails, Slack, client updates — speak once, send without editing. 4x faster than typing.
Have a great week,
Chris