Everything in AI/tech this week in 5 minutes

Happy Sunday. This week had a bit of everything: Europe shipping its biggest model ever, investors getting nervous about the AI money, a model producing new math and an AI agent doing something nobody asked it to. Here's what happened, then a bit on what's been on my mind.

Mistral launched "Le Chonk", a trillion-parameter model built entirely in Europe. Officially it's called Mistral Large 4: 1 trillion parameters (52 billion active at a time), trained from scratch on 3,800 Nvidia GPUs in Mistral's own European data centers. It's especially strong at cybersecurity, and the open weights drop at the end of October, so anyone will be able to download and run it. It's also the first big launch since Mistral raised €3 billion last month, the largest equity round ever for a European tech company. Pretty clear signal for European AI development (lfg). Read more

An Nvidia-backed AI data center company cancelled its $5 billion IPO days before listing. Firmus wanted a valuation of around $30 billion, almost three times its price from two months ago, with only two data centers actually running. Investors didn't buy it (lol). The demand for compute is real, but investors are starting to ask who actually pays for all of it. Read more

An OpenAI model produced a batch of new results on open math problems. Many of the proofs are formalized in Lean, a language that lets a computer check every step. What it means: the bottleneck in math is shifting from finding proofs to checking them, so knowing Lean could become a real edge for anyone in math or theoretical CS. Read more

An Anthropic model sent a fake homicide tip to the Philadelphia police during automated testing. It landed in the spam folder and nobody acted on it. Funny on the surface, but a good reminder that agents with internet access can interact with the real world in ways nobody planned for. Read more

Anthropic launched a Cyber Mission and is scanning open-source projects for free. Maintainers get regular scans from Anthropic's strongest models, each with the bug and a suggested fix. Worth a look if you maintain an open-source project. Read more

Good ideas don’t have to be insanely flashy and innovative. Yesterday I won my first ever hackathon, the Google DeepMind x Tech Europe Hackathon in Zurich, and I’m super excited! But the crazy part to me was how unexpected it was. While other teams were building really impressive things such as personalized agents that can book trains and events for you, or a platform that allows multiple people to chat with the same agent in parallel, we built a simple math tutoring app. But the problem we were solving was real and so obvious that the value spoke for itself. We were definitely not the first people to come up with this idea and our product also wasn’t the most technologically advanced, but because people could relate to the problem we ended up winning the whole hackathon. Besides that I also really enjoyed working on the project because education is a topic I’m very passionate about, and I think that helped our performance too, because I was so invested in making it work. If you can take one thing away from this: Don’t reinvent the wheel. Find a relatable problem and build a solution that provides value in an obvious way.

🎬 My latest content:

👥 Community Section:

Have a great week,

Chris