Lakshya A Agrawal
@LakshyAAAgrawal
Followers
3K
Following
4K
Media
64
Statuses
2K
AI PhD @ UC Berkeley | GEPA Creator (https://t.co/EdPqvzj7k4) | Created https://t.co/YxPZsXZJeS | Past: AI4Code Research Fellow @MSFTResearch | Hobbyist Saxophonist
Berkeley, CA
Joined December 2013
What's stopping you from trying GEPA right now? P.S.: Please go try GEPA!🥹 https://t.co/uS9xm8nNCz
Both DSPy and (especially) GEPA are currently severely under hyped in the AI context engineering world
12
8
128
@dhrtha @LakshyAAAgrawal @DSPyOSS Yeah there's a good tutorial of image prompt iteration in the dspy docs, just run GEPA on it: Image Generation Prompt iteration - DSPy
0
1
3
This is a cool idea! Will put it on my list. Stay tuned for more to come in the repo early next week ⬇️
GEPA can work with any reflection_lm, which rewrites the current prompt and uses Pareto-based selection prompt to try new data points. I expect one can get further gains simply by running Prompt-MII+GEPA! Anybody would like to try this out? https://t.co/o7rQW5Tsbv
1
1
8
It took a good amount of effort to get all this working, but it was really a fun, learning project. I will hopefully write up a longer blog with proper explanations. And before anyone asks, yes 98% of this was vibe coded using @cursor_ai and @claudeai . /🧵
0
1
3
Special thanks to @zaph0id for all the support with this and @IndhicAI . When we get the time next, we will be working on a ColBERT model for Sanskrit / Kannada - built from the ground up. Happy to collaborate with others if interested. https://t.co/WkEAMGYxrH +
huggingface.co
2
3
7
Now we needed to load this directly in the browser for embedding inference. So we use https://t.co/KU3AjBAODc But you cant load .safetensors in the browser, so you need to convert this to onnx and quantize it to fp8 so that the file size downloaded in the browser is smaller. +
huggingface.co
1
1
2
Finally, the newly fine tuned model showed good improvements against the original embedding gemma https://t.co/mziTkXhT5n +
1
1
3
And Fine tuned the Embedding Gemma model - took around 2-2.5 hours. Lots of learning here in the Fine Tuning Process - needs a full blog post. https://t.co/sXX7QCvAub +
1
1
3
“GEPA is a text evaluation engine” Continue to be impressed. Wonder if it can be connected to the real world and write better copy than humans :)
@harshad_geek @AsfiShaheen In this context, GEPA works as a prompt optimizer, so the end result is a prompt (or multiple prompts for a multi-agent system, one for each component). However, one aspect that does not get highlighted enough is that GEPA is a text evolution engine: Given a target metric, GEPA
0
1
6
GEPA can work with any reflection_lm, which rewrites the current prompt and uses Pareto-based selection prompt to try new data points. I expect one can get further gains simply by running Prompt-MII+GEPA! Anybody would like to try this out? https://t.co/o7rQW5Tsbv
I'm excited to see all the interest in our new prompt optimization method, Prompt-MII! Some people are asking how it works when we train larger models. Anyone want to lend us 32-64 H100s for a bit? 😅
2
0
41
Sanskrit NLP is really close to my heart, and I am excited to see GEPA+@DSPyOSS being used to optimize a query generation pipeline to differentiate between pair of docs, generating 50k samples for Gemma embed fine-tuning. Superb work @dhrtha @zaph0id! https://t.co/z82cmMrfcZ
🕉️ A multi-agentic RAG to answer questions regarding the RgVeda - works completely in your browser. No server needed. #RigVedaHack @indiainpixels
https://t.co/jJz9HMbZnL 🧵
2
5
44
Always fun learning and working on this. As @karpathy says building and learning in depth opens up a world of infinite ideas and possibilities.
Special thanks to @zaph0id for all the support with this and @IndhicAI . When we get the time next, we will be working on a ColBERT model for Sanskrit / Kannada - built from the ground up. Happy to collaborate with others if interested. https://t.co/WkEAMGYxrH +
0
1
3
Please check out this thread by @dhrtha on submission to #RigVedaHack Endless ideations and possibilities! Ping @dhrtha and me if you are interested!
🕉️ A multi-agentic RAG to answer questions regarding the RgVeda - works completely in your browser. No server needed. #RigVedaHack @indiainpixels
https://t.co/jJz9HMbZnL 🧵
0
1
2
We do the AI persona training with @DSPyOSS btw. Really fun use case for GEPA!
0
0
2
GEPA+@DSPyOSS can optimize AI to generate human-like responses! Checkout this release by @hammer_mt! https://t.co/16OKQkIyBy
It turns out the secret ingredient was people. We did interviews with 100 real people (so far...) and trained AI personas to respond like them until a superhuman LLM-judge couldn't tell them apart. Now our responses in Ask Rally don't suffer from any of the issues that purely
4
3
42
The GEPA paper is a really good example of AI engineering. The basic gist is you can algorithmically iterate on the system prompts of an agent if you have some way to quantify 'good', and it can actually be better than tuning weights / doing RL.
2
2
14
🧵 (8/8) But how can my LLM handle long contexts? Not part of DSPy, but @a1zhang proposes Recursive Language Models (RLMs)! Instead of YOU deciding how to chunk/retrieve/decompose, the LM recursively calls other LMs to figure it out. It treats your context as a variable in a
2
1
4