Claude Sonnet 3.5 with GEPA-optimized prompts performs as well as Claude Opus 4 without GEPA on set of agentic tasks I designed 😲

This makes me think that when using an MoE LLM as a program in DSPy, it may be possible to prune a lot of experts. Has there been any work on mapping the ratio of experts in an LLM that are never used for task-specific generation? Say I am using an LLM only to translate cooking

This is somewhat true. There are many ways to have a full eval chain of your prompts though. And you can even abstract away from prompts entirely with DSPy et al

This is a good example of why having "prompt" & "context" terms add value. For using a chatbot this is 100% true. Prompting is vibes and anecdotes. When designing a context in an agent or app, the best are constantly running tests against evals. DSPy scales this further.

FYI the DSPy boyz have been saying this for 2 years guys

It's so very humbling to realize that on many occasions, the default prompt DSPy creates from its signatures and adapters can actually BEAT the initial prompt I might have written by hand (in other frameworks) based on my naive initial understanding of a problem or domain.

Chat, any AI wizards here looking for consulting gigs? Got a homie who’s looking for a AI-pilled contractor to run a bunch of experiments on autocomplete for the next month. DSPy experience a plus. Cracked team, sick startup.

Haha I love how, as a rite of passage for any language community, the DSPy community has decided to split on Adapters. Are they a fundamental first-class concept? Or are they just implicit low-level helpers? Don't think my opinion on this matters very much to the debate😆

I know I said not to use frameworks when building your Ai agents. But once you're done learning and understanding the challenges of making llm tasks reliable, look at DSPy/DSPy.rb Ignore compound modules if you like and just use it to create callable llm functions.

webllm running dspy in the browser. ax has come far i'm excited to build more, axflows, optimization, etc. models like gemma 3n make edge ai a very real thing.

GEPA optimizes LLMs without costly reinforcement learning venturebeat.com/ai/gepa-optimi…

The team at Chroma hit launch on Chroma Cloud a couple hours ago. In 10 min, I loaded my blog posts into a collection. (Including learning curve and account set up) In 12 lines, using DSPy, I had a tool-using agent that could search my blog content and write summaries.

mlflow has really nice integration with dspy I like how tracing for dspy modules works here

Have you heard the news? ⚡ The MLflow TypeScript SDK is here! Now you can bring #MLflow’s industry‑leading observability directly to your TypeScript and JavaScript applications. 🚀 The SDK enables: ✅ Automatic tracing of #LLM and AI API calls ✅ Manual instrumentation for

This reminds me we (the dspy community) should make a DSPy book. It could be titled: Guide to programming AI systems. Would you read it?

The DSPy community in Toronto is getting together tomorrow at New Stadium! Come by for a remote workshop with Maxime Rivest 🧙‍♂️🦙🐧, spacy and hang with local DSPy experts like Prashanth Rao and Robbie Pasquale. An evening of building elegant AI systems and good pizza.

Looking forward to heavily nerding out tomorrow evening in Toronto 🇨🇦discussing all things DSPy with fellow afficionados! Also looking forward to see Maxime Rivest 🧙‍♂️🦙🐧 present live - shame I don't get to meet him in person (maybe I gotta make the trek up to Ottawa some day to do

Interesting new paper with DSPy, for synthetic retrieval data generation. > replacing static prompt templates with dynamic, Chain-of-Thought (CoT) optimized prompts using the DSPy framework [reduces] the need for aggressive filtering while improving retrieval performance