Screaming At My Screen by Timo Zimmermann

This week was one of the most random, unfocused weeks I spent working on LazerBunny and so many things moved forward in a meaningful way. The goal never was to get everything "done", but good enough to use it on a daily basis and validate design decisions and figure out potential adjustments that need to be made. I took the opportunity to have two agents running in parallel and have them prepare smaller chunks of code using Qwen 3.6 27b. Overall an okay experience. Code layout and design were…

It is rare for me to quote Mike Tyson, be he was onto something with this. Here is the thing with coding agents: you will not tell a coding agent to do something in one sentence and then something usable falls out. I mean… there is a chance - depending on your prompt. If you tell it to copy something that exists millions of times on GitHub. But anything even remotely more involved than that, things will be a bit more tricky, especially if you are dealing with an existing system and not one…

Building out a small fleet of coding and research agents for LazerBunny means I regularly have requests to LLMs running, which are hosted on various hardware in my local network. This is not a big deal most of the time, but things get a little bit more interesting when multiple agents compete for the limited local resources. I spent some time this week to look into potential solutions. There are some I ruled out: Buy more hardware. This has to wait till we hear "the big pop". Use online hosted…

I have worked on software for Apple devices on and off during my career. It even goes as far back as working on a desktop application, before things were called apps or the little glass slabs became our daily companion. Some of my highlights include having an app featured in the AppStore, working on an Apple Watch integration and having day one support (mostly, thanks review process) and some other fun stuff. I stopped working on mac-related apps in 2018 when Apple’s behavior towards develope...

This week was highly unstructured in regards to working on Lazerbunny, to the point that you could have rolled a dice to figure out which task from the todo list I will pick up at any given day. The two most important ones were progress on the avatar and adding context compression to the coding agent. It was clear from the beginning that I am rather wasteful with the context when using my coding agent. The theory worked very well for most tasks: Just do not bother. If I only give it a single ...

I am a very big fan of self-hosting services I use on a daily basis or that hold my data. Some services are mostly a backup, some are in day to day production use for my company, some are to toy around. Whenever one of the big providers messes up, has a data breach or is bought and shuts down people advocate for self-hosting similar services, and you might think I’d be one of the,. Sadly it is not that easy. I am not self-hosting because I do not trust Apple or Google to keep my data out of t...

One of the key components of any personal assistant is the ability to remember what you ask them to do and put it in the context of what happened before. So for this week I was mostly focused on building the memory layer of LazerBunny. A fun little exercise when trying to push search and retrieval with minimal resource usage. First of all I want a cookie point for not titling the blogpost "thnks fr th mmrs". Now, most people who spend any time with LLMs might have seen the term RAG - Retrieva...

Having an agent work while you do other things sounds like a neat idea. How well it works depends on a number of factors including the model you choose for the hallucination machine powering it, the prompts, tools and of course how patiently and you tell it exactly what to do. Having an agent write all the code has not worked for me, but I found some value in having one ready. So what better way to spend a Saturday watching LEC than writing one for LazerBunny? The primary use case I have for ...

I have a relatively simple setup to host services locally: One Caddy container acting as reverse proxy for all services. DNS is set up, containers talk via Docker networking and I run an ACME server for SSL certificates. The disadvantage of this is that each client needs to trust the root certificate. Easy on most desktops. Okay to do on mobile devices. The most painful thing I ever did on an Apple TV. But it is a bit inconvenient when we have visitors over, and on top of that my wife is real...

Do you know the feeling when you look at a system and how it behaves and think "this cannot be it"? This is how I felt when working on LazerBunny this week. Somehow, for some reason the MCP approach struck me as less favourable, and after some experimenting I completely dropped MCP from the existing timer service. I also had a bit more fun with voice cloning and slowly narrowing in on the final solution. The usual argument to avoid MCPs is that they fill up your context with 500 functions you...

I have seen more and more people talk about human.json lately and I think it is a pretty neat idea. From what I can tell it checks all the boxes I would expect from a protocol like this. The fact that it relies on browser extensions right now makes sense, but might become a limiting factor in future. Or the number of extensions needs to go up beyond the two easy ones and come to mobile as well. I am not sure this will be going anywhere beyond a few enthusiastic people, but you never know. Imp...

Back in 2024, I ordered the Worlds (basically their World Cup?) statue of League of Legends. I did not hear anything from Riot for over one and a half years. When I opened a support ticket, I got the very helpful response, “We did not cancel your order.” I checked twice if they snuck in the word "yet" somewhere. But just a few months later, my prepaid statue of Luxanna Crowngaurd finally arrived. I am not sure for whom these gloves are made, but surely not me. There is no way I can get my ...

When the RTX 5090 was released I was already very much not in favor of the whole DLSS marketing that went into it and how it gave a very unrealistic view of the performance compared to the previous generation. Companies obviously make numbers look as good as possible during release announcements. Thankfully most reviewers simply do their own benchmarking and we get numbers without imaginary frames. But like power connectors going up in flames Nvidia found a way to double down on stupid. First...