Dot Net Perls Blog

Yesterday I did some experimenting with Z-Image-Turbo in ComfyUI.

Vibe coding (using agentic AI to generate a complete program from prompts) can be kind of fun.

Recently I have been doing some spec-based Rust code generation.

Over the past few months I have been trying to understand how to use LLMs cost-effectively on OpenRouter.

I have a fair amount of Rust code that needs rewriting.

Yesterday I wrote 2 more specs for my program, and needed a LLM to generate the code.

I have recently been trying to replace some Rust functions with specifications written in Markdown.

Something about AI-based code generation seems lazy or sloppy.

So I wrote a 180-line Markdown file containing a detailed specification of a parsing function needed in my program.

I have spent some time implementing a simple MCP server that can do certain management tasks.

I spent some time using llama-cpp with Ternary Bonsai 2.

Ternary models represent each weight as one of 3 values (1, 0, and -1).

Today I wanted to test Gemini 3.8 Flash (a cloud model from Google).

One thing that annoys me on the Internet is when I start reading an article or blog post, thinking the content might be unique, and it turns out to be 100% AI-generated.

Recently a Rust library for displaying JPEG XL files was released, and Chrome and other browsers are planning to support JXL files on the web.

Today I downloaded and used Qwen 3.5 0.8B (Qwen3.5-0.8B-Q4_K_M.gguf) on my local system.

I have started to feel more doubts about the benefits of using agentic AI for coding (like in coding harnesses like OpenCode, Pi).

Kagi is an independent search engine without ads, tracking, or evil killer robots (it's hard to keep up with the things tech companies are doing these days).

I know that agentic harnesses are popular in software development these days.

I have a local MCP server that mostly just calls a handful of functions.

In open-source agentic coding harnesses, there seems to be 2 top choices: Pi and OpenCode.

Trying to find a decent solution for local AI coding has been difficult for me.

Recently I decided to create an MCP server in Rust.

There has been a subtle bug in my Rust program for years now.

Before I designed my own agentic AI system (an MCP server) I tried to keep an open mind about AI.

I have spent some time evaluating small language models (350M parameters) and they are significantly different from LLMs.

Recently my Rust program starting exhibiting a panic on a certain agentic tool call.

Large models with 2 billion or more parameters (such as Gemma 4 E4B) tend to be reliable in tool calling in agentic systems.

The AI bubble refers to the concept that the excitement about AI has led to excess investment.

I have had good results with small language models (such as 1.2B parameters) for interacting with my MCP server.

Yesterday I was trying to track memory usage of some programs I was running, and was using top in Ubuntu.

Last night I spent some time trying to make my MCP server work with a smaller local model, one that is just 1.2 billion parameters and 696 MB on disk.

Should tool calls be combined in a MCP (Model Context Protocol) server?

Yesterday I was thinking about my MCP server, and how it has 10 functions—if I could combine some of the tools, could I make it faster?

I have a small MCP server I use to deal with administrative tasks.

I have used both Pi and OpenCode in the past, and although I liked Pi I tended to just use OpenCode out of habit.

I was feeling a little left out over the course of this last week as most of the open-weights models released were too large for me to run locally.

I have been experimenting more with various chat bots available on the Internet.

Today I was using an online chat bot (a well-known one) and exploring changes to my Rust code.

I tested Muse Spark 1.1, a new close-to-frontier LLM from Meta Superintelligence Labs, on my Rust program.

Now that I have an agentic AI system set up (I wrote my own local MCP server) I can test different models on it.

I have mostly concluded my efforts in building (and optimizing) a local MCP server for llama-cpp.

An MCP server provides a list of tools, along with their parameters and descriptions, to an AI agent.

I have been trying to develop an agentic AI system with a local MCP server.

In my Rust program I have code that tries to normalize the arguments passed into it.

I am making some sort of customer service MCP server for my own use.

After using an LLM in llama-cpp to handle local tasks, I started wondering: why couldn't I write some code to handle the queries faster than the LLM inference?

I have been using ComfyUI on my RTX 3060 to generate desktop background images for a while.

Writing the descriptions in an MCP server is like writing prompts for the tools.

I have been experimenting with optimizing my local-LLM agentic AI workflow.