Weekly Tinkering by Junod.dev for Feb 4th

My daily workflows are becoming centered around AI tools. At the same time, I spend too much time daily mucking through AI data to find the things that matter to me, the things that push my workflows forward. Headline fatigue is hard.

In the past two weeks the flood of AI news looks more like a water main break than a firehose. International developments, domestic releases, hastily written laws and bans, it’s a lot. I saw a post on linkedin on how Deepseek has now infiltrated enterprise networks and is stealing data (it has not), be weary of what you read.

We’re going to stay focused on productivity and becoming bionic.

Things to read/watch

Sam Witteveen is one of my favorite AI youtubers. Smolagents has been on my list of things to research, and I finally watched his video, and it’s worth a watch. Smolagents are HuggingFace second go at agents, and this one is code centric. That means to thinks, and creates code context. Sam does a great job showing that off.
- smolagents - HuggingFace’s NEW Agent Framework
You get a Sam two pack this week. On my list is to dive deeper into both Smol and Pydantic.ai agents. I’d like to see how far I can get locally with a DeepSeek distillation, and of course with openrouter, the main way I inference these days. Sam’s intro to Pydantic as a research agent helped me a lot.
- [Pydantic AI + DeepSeek V3 - The BEST AI Agent Combo](PydanticAI - Building a Research Agent)

Things to try

Goose is a new open source (Rust) agentic toolset from Block, Jack’s company. Goose works with multiple LLMs, and acts as an AI assistant in a folder. Goose also supports MCPs, so my most used MCP exa is getting a goose workout. Goose has a CLI for Linux/WSL, or a full app for Mac. They work identically. One of my first uses was Valentines day research, and it killed it.
- Goose
I’m on a mission to use MCPs everywhere, and I have found another client that supports them out of the box. Hyperchat, a Chinese based open source LLM client, supports MCPs. I have tested it a bit and so far it’s not bad, and will help me experiment with MCPs outside of the Claude app or Cline.
- Hyperchat

Things I’m tinkering with

I am heavily utilizing the “work with me in this folder” pattern. I am also honing my skills on asking AI to create the things that will help AI create the things. Meaning, my first swings at AI productivity asked AI to do a thing, now I ask AI to think about it, write it down, then move to do a thing, which is far more productive so far.

I’m also realizing AI can help me create any kind of thing. Any list, any document, it’s not just code. AI assistant tooling line Cline and Goose are tuned for code, but just as happy to create a list of ingredients in a csv.

I’ve spent the most time tinkering with Goose in the past week, and it also has the notion of gooserule like cline. One thing I’m starting to do is each time I start a new “thing”, ask the ai to make the {thing}rules specific to what we’re doing, then start gathering data.

Practice is the best way to become bionic.

I hope you enjoyed this week’s post. Please let me know if you have any feedback, and thank you for making it to the end.