I launched Timeboat Adventures, an experimental narrative game where you rescue historical figures and merge them into a single meta-entity that rewrites the 20th century. Powered by @DSPyOSS and Gemini-2.5-Flash from @googleaidevs. Try it here https://t.co/4OeKHeThph #AI #LLM

Useful DSPy tip: You can return feedback in the reward function for dspy.Refine. This gives you more control over your generations. I wish this information were included in the DSPy documentation. It’s one of my most-used modules when I don’t have time to optimize programs.

I’d love to hear your thoughts! I’m planning to write more tutorials and share write-ups about my experiments. What topics would you be most interested in reading about?

My first tutorial on creative task optimization with DSPy and GEPA is online! Not everything went as expected, but it’s a solid start toward better results. https://t.co/3UaaUCCcCB

and it sucks that substack doesn't have good syntax highlighting for code snippets

Caught a fever and sore throat earlier this week, so the tutorial got delayed. Feeling much better now and back to work - wrapping it up soon!

It feels like a heuristic, but if you tell an LLM in the system prompt that the text was generated by another LLM, it tends to evaluate more strictly and accurately. no more “you did a good job, for a human being”.

It’s kinda sad that there’s no easy way to access hidden states in DSPy. The dependency on LiteLLM is quite limiting when working with local LLMs. I might implement a custom handler for vLLM to test CLUE. it seems like an awesome alternative to the LLM-as-a-judge approach.

Finished my first tutorial on improving the quality of smaller LLMs for creative tasks using a teacher + GEPA. I'll publish it on Monday. What style of signature do you prefer? I usually go with classes, but for this tutorial I decided to use inline

Personally, I'd still prefer a plain UML diagram without any artifacts and piss filter. But anyway What's your approach to creating diagrams?

Here is the final result. I spent more time writing this thread than creating the diagram

Pro tip: Paste your MermaidJS code along with the image to Nano Banana This helps it understand the structure better and slightly improves output quality

3. Go to Google's AI Studio and give your diagram to Nano Banana Use a prompt like: "Generate a diagram based on the schema's content and structure, but with a cool sci-fi aesthetic"

2. Go to https://t.co/RuFkOCnx8s and render your diagram Take a screenshot or export as PNG

So here is the better way 1. Ask ChatGPT to generate a MermaidJS diagram based on your tutorial/concept

Here is the problem: Current image generation models simply can't create coherent technical diagrams. They're just disfigured abominations with no logic and sense.

I hate AI-generated diagrams in blogs They're confusing, ugly, unreadable When I see one, I want to close tutorial immediately I get it, sometimes we don't have enough time So, here's a mini-tutorial on how to generate much better diagrams with Nano Banana in 5 minutes🧵

wow I've reached 500 followers

I GAVE MY AGENTS TOOLS! in case of ASCII art generation, agents now can render text to png. It feels like that helps them make better analyses and improves quality. it's interesting that gemini-2.5-flash-lite uses tools more often than 2.5-pro. Pro prefer just thinking.

I want to build a cool visualization project: a TV-debate show with different agent/persona/LLMs streaming 24/7. Short on time/resources - would love help from anyone who would like to support the project. Please DM me!

Anyone interested in implementation of a multi-agent debate system with DSPy? rn I'm updating my optimization scripts for GEPA. What I love about DSPy is how it easy to optimize the whole multi-agent system at once. no need to tune each agent separately. on Github soon