Weekly Tinkering by Junod.dev for Jan 21

Models, models, everywhere. The pace of new model announcements continues to accelerate, and expand. Models have their own origin stories, and even variations based on run location.

Choosing the right model for tasks is going to become instinct to AI practitioners, and could cost those who don’t know. After a morning of coding with Claude and Cline I sent a few million tokens. The whole session cost me around 5 bucks.

Claude’s token price is 3$/M. Deepseek is V3 is $0.14/M. As we see below, though, cost is not the only consideration. I use Openrouter to easily swap models, and even run locations.

Things to read/watch

Kevin Xu from Interconnected recapped a portion of a recent podcast appearance from Zuck about the danger of Chinese models. He then tested them in various runtimes, and they responded differently. Interesting indeed, a Chinese model running out of bounds both doesn’t phone home, nor censor?
- Was Zuck Right about Chinese AI Models?
Cole Medin is a constant source of great information. I started down the Pydantic.ai path because of a recent video of his. He also does a lot of N8N agent work. I hope he shows us how to run Pydantic in N8N soon, I started down the path but it looks perilous.
- Pydantic AI + DeepSeek V3 - The BEST AI Agent Combo

Things to try

Agents working with web data is a huge piece of the puzzle. Web search is moving along, most major providers have it, or you can mix in services like Brave or Exa for search. Agentic access is still in formation mode, Stagehand looks to solve that. I need to play with this more, very promising.
- Stagehand - An AI web browsing framework focused on simplicity and extensibility.
ComfyUI is a workflow system for AI image generation. It provides flexibility and reusability in AI image generation tasks. My favorite AI generation tool Fai has workflows built right in. I’ve been working through the article below to learn how to use them.
- Comfy UI Beginner’s Guide

Things I’m tinkering with

I continue to practice with generating code. I started a new project this past week, one that had been on my mind for a couple years. I started with a paragraph in a .md file.

I then opened Claude Desktop and said something like “Let’s write some requirements in c:\folder”. After semantically directing the production of requirements, user stories, and some technical direction I had an idea.

I then said, Claude, generate me prompts that I will feed to cline to execute on our plans. Claude started making prompts.

I then loaded Cline and told Cline to read the requirements and help me get started, and did it ever. That’s also when I saw Cline’s new memory bank in action.

AI assited coding is a skill, one you need to hone. Semantic access to code is going to be a super power, for me at least.

Cline Memory Bank

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.