Category: ai
Grok Bot for Home Assistant with their MCP
A week ago I started using Grok Bot for some hobby projects and found out that there’s a Home Assistant MCP from the Open Source community. And wow: It’s great! Read moreSo, Ministral tried to generate SQLAfter connecting it to Grok Bot I asked it a few simple questions: What lights do I have? And…
Watermarking AI output helps AI Companies
I fully understand why AI companies comply with the EU to watermark AI-generated content. This makes it a lot easier for them to train on human-made content, since it’s watermarked. It reminds me of this study, where if you put images of the health effects of smoking on cig. packages, people actually smoke more. Of…
Code was carved in stone, it now is clay but will become water.
The title says it all. Before the era of agentic-programming software had to be designed so it could be not touched for a very long time. It was carved in stone. With the speed you can build, update and maintain features nowadays (and not even talking about parallel agents working on things) things have started…
Claude’s voice mode is awesome!
Since the release of GPT’s proper voice mode, where it sighs and such, I’ve been using voice agents for ideating and learning more. Recently, Anthropic updated Claude’s own voice mode and I think it’s powered by ElevenLabs their voice models (just an assumption based on how it sounds) and I have to say: I’m impressed.…
MiniMax M2.1 still beats Kimi K2.5
Even if Kimi is better on BenchMarks, MiniMax seems to be a lot better in backend code in the wild. Our codebase is well over 500k lines of code and MiniMax seems to navigate it a lot better than Kimi. And you can clearly see that on the statistics of Kilo, that people like MiniMax’s…
Backend Git Bisect with an Agent
Today I committed a few “wip” commits and during one of those commits I broke some code and caused tests to fail. I got curious and wanted to see if an LLM could git bisect and the answer was: yes, very fast! The LLM (Gemini 3 Flash in this case) quickly started the bisect and…
Qwen’s response to Grok Code 1 Fast?
Grok Code 1 Fast has been a smash hit since it’s launch, mainly because it was a free model with a peak of 240B tokens on one day: This is the only thing I can think about when I read the new model’s information of Qwen3 Coder Next: that it’s a competitor to Grok’s coding…
Kilo’s new CLI is here!
A week ago, I was setting up the Kilo CLI on one of the laptops of a colleague and we saw something surprisingly happen: the new version of Kilo’s CLI was loaded (1.0.3 at that time, where 0.26.1 was the @latest tag). So, I decided to set it up on my laptop and I immediately…
My favorite model: Gemini 3 Flash
It’s a fairly affordable model but it has a lot of power. It’s capable of generating wonderful UI like this in one shot: The model is in the top of intelligence, the top of speed and it’s in the low area of the token pricing which makes it a model you can easily debug and…
The migration to OpenRouter
A few months ago we opted to use OpenRouter at the company I work for. And moving to it was one of the best decisions we’ve made for our AI implementation. The reason? Unmatched stability. OpenRouter handles the load balancing across providers so we can just focus on building features that add value for the…
Kimi K2.5 is out!
And I’m absolutely loving it’s behaviour! Currently using it in Kilo and it’s a good model. It’s not that fast yet, but that’s mainly a provider issue as you can see on OpenRouter: Most of the providers haven’t allocated enough resources or haven’t optimised it enough, that’s an assumption but what I usually see is…
You don’t need the best model
When people work with AI, they always opt for “the best model” hen people work with AI, they always opt for “the best model” without having thought about it.This is bad since it: Yes, the best models are “the best.” Which is wonderful, but using those models is not necessarily a requirement.For less complex tasks…
