DSPy is the highest-bandwidth language to talk to computers in a just-precise-enough way. First 36 lines below: 1. Take arbitrarily long content: `chunks`. 2. Build a global Table of Contents. 3. Distribute chunks into sections to be written. 4. Recursively write each section.
Working on some major updates to DSPyground. Launching soon 🔜
Creative writing with GEPA + @DSPyOSS (this time, to generate dialogues in Naruto's style!) by @ML_deep, covering Japanese audiences too! zenn.dev/cybernetics/ar…
Intrinsic Labs finds that GEPA+@DSPyOSS can help cut OCR error rates by upto 38% (91.5%->94.8%) on @getomni_ai OCR Benchmark, and also lifts model performance across different model classes (Gemini 2.5 Pro, 2.5 Flash, 2.0 Flash). Check out their full report in the thread!
An emerging pattern I am seeing for GEPA+@DSPyOSS deployment: 1. Init & Deploy: Use a small, high-quality initial dataset (e.g., labeled examples with explanations). Run GEPA + DSPy to optimize the program/agent and deploy. 2. Monitor the deployed system, collecting user…
Yes, this is a standard pattern that can be applied with GEPA. You could also have it run with batching directly from users feedback. @AsfiShaheen has a superb hybrid method, combining human labelling, but without introducing latency in gepa optimization itself:
To use GEPA (best optimizer in DSPy) you need two things 1. Hand labeled data set that’s accurate 2. Text explanation for WHY it’s accurate Here’s a worked example applied to tagging financial statements. This approach can save you both time and money. 🧵
I'm sorry I had specific work things in my talk that I can't share. but I used reference code from the great talk that @hammer_mt did! Check out his talk here youtube.com/watch?v=gstt7E…
youtube.com
YouTube
Mike Taylor: The Evaluator–Optimizer Pattern in DSPy with GEPA
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. open.substack.com/pub/meandnotes…
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
Amazing article by @mike_pavlukhin explaining how to tune a small model like Gemma-3-1B to write good quality creative content (stories) using GEPA with @DSPyOSS. Many interesting patterns and insights for creative task optimization!
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. open.substack.com/pub/meandnotes…
DSPy and its optimizer give it more of a machine learning feel, while the framework's philosophy is to treat prompts as a program with inputs and outputs.
And not just that, DSPy provides a simple abstraction for the whole LM stack (prompt, agent orchestration, tools, LM weights) all accessible from a convenient and simple interface, that works with all state of the art optimization techniques! For example, we (@dilarafsoylu,…
Now I understand @DSPyOSS from this video, and I will attempt to break it down. Basically if you build software that uses LLMs, it helps you manage, test and optimize backend prompts. When a new LLM comes out, you can swap it in and get the new best prompts for your existing…
Folks like the ones involved in @DSPyOSS and GEPA are why I believe in open-source above all else. They make humanity strictly better off. Planning on another deep dive thread soon. In the meantime, check out my previous posts highlighting their fantastic work!
Love this philosophy so much. Makes me always want to support and help nurture what we have 🫶🏽
yeah we know, and that's why we're not one this is foundational open source and open science, the linux, git, python, or pytorch of AI software systems
If @DSPyOSS was a startup right now it would have massive valuation. Just saying.
I absolutely love programming with @DSPyOSS . I genuinely feel happy while figuring out things. It is my go to. And I have said this before and I will keep saying it : DSPy is going to become the default of every AI stack within 3-4 months. Throw GEPA into the mix and you are…
the openai event highlighted the shift to building with ai. our models are good enough we just have to use them more. it also confirmed that dspy was the right bet
with all this talk around agent builders this is how simple it is to have a multi agent program in ax all driven by dspy signatures and optimizable.
My new DSPy slogan: “Come for the prompt optimization, stay for the maintainability and DX.”
United States 趨勢
- 1. Jets 67K posts
- 2. Aaron Glenn 2,720 posts
- 3. Justin Fields 4,627 posts
- 4. Garrett Wilson 1,601 posts
- 5. HAPPY BIRTHDAY JIMIN 57K posts
- 6. #JetUp 1,434 posts
- 7. Peart 1,570 posts
- 8. #BroncosCountry 2,228 posts
- 9. Tyrod N/A
- 10. #DENvsNYJ 1,573 posts
- 11. #OurMuseJimin 93.1K posts
- 12. Sherwood 1,162 posts
- 13. Kurt Warner N/A
- 14. Good Sunday 65.9K posts
- 15. Bam Knight N/A
- 16. Hail Mary 2,352 posts
- 17. #30YearsofLove 86.9K posts
- 18. Rich Eisen N/A
- 19. Brownlee N/A
- 20. Troy Franklin N/A
Something went wrong.
Something went wrong.