Anthropic recently published a fascinating paper with an unusual title: “On the Biology of a Large Language Model.” But what does biology have to do with AI? As it turns out, quite a lot. The researchers are essentially performing a digital dissection of Claude’s “brain” to understand how it thinks—and finding surprising parallels to biological systems along the way. In this post, I’ll break down this complex research into digestible parts, explaining how scientists are starting to understand what’s happening inside these AI systems when they answer our questions or write poetry.

I’ve been using GitButler for about a week now, and it’s already transformed how I interact with Git. As someone who regularly juggles multiple feature branches and context switches between tasks, I was looking for a tool that could simplify my Git workflow. GitButler caught my attention with its promise of virtual branches and simplified collaboration features. My interest was further piqued after watching presentations by Scott Chacon, one of GitButler’s creators and also a co-founder of GitHub.

The Brazilian National Institute of Industrial Property (INPI) is embarking on a significant transformation of its patent services. A recent presentation detailed plans to replace the current web search system with a comprehensive, unified platform designed to streamline the entire patent process. This initiative promises to enhance efficiency, reduce user errors, and improve the overall experience for patent applicants and stakeholders. Current Challenges in the Brazilian Patent System Currently, the INPI’s patent system relies on several disconnected systems.

We’ve all been there: headphones on, music pumping, fingers flying across the keyboard, lost in the “flow” with your favorite AI agent. This, my friends, is vibe coding. It’s when you’re in the zone, seemingly effortlessly producing code. It’s the idea that you can create great software simply by immersing yourself in the feeling of coding, trusting your intuition and riding the wave of inspiration. But lately, I’ve realized that for me, at least, the vibe is…off.

So, I was watching this video the other day, and it got me thinking… you know, about how everything’s changing with AI. It seems like the old ways of doing things in software are just not cutting it anymore. Remember how companies used to try and, like, trap you in their ecosystem? Make it impossible to leave? That’s software lock-in, and honestly, I think AI is making that a thing of the past.

This article details a troubleshooting process for a Proxmox cluster authentication failure that prevented users from logging into the web interface. The root cause was identified as a missing /etc/pve/access.cfg file, crucial for user authentication, likely due to corosync communication issues. The problem was resolved by restarting the pve-cluster service, forcing quorum, manually creating a minimal access.cfg file, and restarting the pveproxy service. Preventative measures include regular cluster health checks, automated configuration file backups, ensuring proper shutdown procedures, maintaining network stability, monitoring storage health, applying Proxmox updates, and careful permissions management.

This is a quick fix to a problem I experienced with Open WebUI, which was that custom models were not showing in the models list. There was nothing in the documentation or in the issues pointing to a solution. I decided to open a new issue 4 days ago and today the user @dcolley gave the missing information and it worked. For some reason, custom models are not working when you use Direct Connections.

A really simple script for those that, for some odd reason, decided to run the Open WebUI server from inside a LXC. The problem I really like the interface in Proxmox and i prefer it than dealing with Portainer for some services that are more accessed. This is the case for Open WebUI, which I use daily. The problem I faced was the update process of the service, which I knew little about.

I needed an easy and cheap option to integrate my vault with artificial intelligence. The easiest option could be to pay for subscriptions, which would be a high cost for me, since I am would not use the service that often to justify. Another option would be to pay for a Cloud GPU service to run big queries that would be unfeasible to run locally. This option seemed too cumbersome for my case.

The problem This week I have been searching for solutions to use domains with my applications running in containers in my MacBook Pro server. Yes, I have converted my MacBook Pro into a server, at least for now. The first solution was to use a Pi-Hole as a local DNS and it works fine, but the problem is that I have to specify ports to get into the applications and it gets annonying.