OpenPipe

Today was the first time I used my fine-tuned model I used Llama 3.1 using @OpenPipeAI the results are encouraging. There are like 5 more to build so wish me luck 😂

@charles_irl @DecagonAI Don't get us started!

Read more about how @DecagonAI serves custom small models for interactive intelligence here: modal.com/blog/decagon-c…. Or ask the @OpenPipeAI team about small models. They'll talk your ear off!

Start in Python, then Rewrite It In Rust™ once you understand the problem and the requirements. Start with off-the-shelf large models, then migrate to custom small models once you've collected enough data and explored failure modes.

Coming soon!

We as a team are learning RL for the first time this weekend. We're totally new to this field, but @OpenPipeAI's docs really helped us kickstart! @wandb

.@corbtt did a great job on this pod, he's a cool guy, recommended youtube.com/watch?v=yYZBd2…

@Vtrivedy10 @corbtt @latentspacepod 🫡

some thoughts on why/how the standard paradigm for optimizing Task Specific Agents will be Harness Engineering + Rubric Based Task Specific RL. this write up codifies a lot of my thoughts on where harness engineering is going + inspiration from @corbtt on @latentspacepod Steps: 1. Obsessively hand/auto tune agent harness until you reach a baseline threshold of task performance. Goal: make sure the agent has roughly what it needs to succeed. 2. Do Task Specific RL to make the model better at operating in the harness, touching the model weights pushes us beyond what harness engineering alone can do. FAQs: 1. Why in this order? If your agent rarely succeeds on the Task, it can’t get enough of a reward, this makes RL difficult. Optimize the harness first —> “You can’t succeed if you don’t have the right tools” 2. Why Not Just Keep Optimizing your Harness, I thought Prompt Engineering is the way? Harness engineering in the latter stages is incredibly hard. It’s also combinatorially complex from the start. You have to jointly optimizes each component(System Prompt, Tools/Skills/MCP, Subagent definitions, Additional Context) but you have: - A Selection Problem: How do you intelligently select relevant tools from a sea of possibilities? Context is a precious resource and selecting too many tools is confusing and degrades perf. - Codependency: No component is optimized in isolation, it’s one big system (ex: changing the system prompt may change how/if a tool is called) 3. So how should I start? First painstakingly test everything in your harness. - try multiple models - hand tune system prompts, try GEPA - more/less tools, tool descriptions, compound tools - handing off tasks to subagents - preloading useful notes, docs, and instructions as references eventually when you hit a wall (on performance or human resources), you move on. You can also move on much earlier once you hit some performance threshold. 4. What does RL get you? Agents (Claude Code, Codex) from the labs are so reliable at using their tools (WebSearch, Multi-Edit, Grep) because they’re directly post-trained with them. We want that same for our tasks, we want to make the model more comfortable using its harness while training to increase Task performance. 5. How should I get started with Task Specific RL? A fantastic first place to start is RULER from @OpenPipeAI which relies on rubrics created by you + LLM as a judge across multiple generations. For your task, you probably already have a set of ideas on what’s good, codifying that in a rubric is all you need to get started working on writing up a walkthrough blog of this with code. really excited about building products that treat agent building as a harness optimization problem that you measure deeply + push further with RL

LIVE: Kyle Corbitt, Head of the OpenPipe team at CoreWeave, joins ThursdAI to talk about launching the first Serverless Reinforcement Learning capability. x.com/i/broadcasts/1…

The Custom SLMs era is upon us 🙌 - Nanochat by @karpathy - Thinker (PEFTaaS) by @thinkymachines - Tunix (Post-train in Jax) by @GoogleAI - Art (Agent RL) by @OpenPipeAI - Environments Hub by @PrimeIntellect - NeMo Microservices by @nvidia

@morgymcg @TrelisResearch @wandb @OpenPipeAI i am looking forward to @OpenPipeAI next week on @altryne podcast, see what's new and what's ahead. awesome! looking forward to it.

@MaziyarPanahi @TrelisResearch @wandb Serverless RL from the @OpenPipeAI crew (now also at Coreweave) with W&B inference is pretty sweet - new models being added soon too openpipe.ai/blog/serverles…

@l2k 🫡

We've started a great tradition at CoreWeave of shipping an integrated new product weeks after acquisition - congrats @OpenPipeAI on the serverless RL launch!

I'm in the unfortunate position to let you know that I've fallen for the RL-LLMs propaganda 100% with these results from openpipe I am now fully RL pilled and there is no turning back very sorry folks

@yacinelearning Welcome to the RL side. You'll like it here.

@yacinelearning One thing that surprised me pretty much is how well feedback from LLM is for automatic improvement. OpenPipe is one example with RL, but it also works well with prompt optimization (sometimes outperforming RL), see GEPA paper. arxiv.org/abs/2507.19457

Sept 8: @CoreWeave acquires @OpenPipeAI Oct 8: @OpenPipeAI ships Serverless RL

🚀 Big launch from @OpenPipeAI: We just launched Serverless RL — train agents faster and cheaper with zero infra headaches. Compared to running your own GPUs, Serverless RL is: - 40% cheaper - 28% faster wall‑clock - instantly deployed to prod via @wandb Inference

@wandb @CoreWeave the coreweave acquisitions is about to facilitate the ultimate end to end platform

@wandb @CoreWeave First of many launches! Stay tuned. 👀

RL X-mas came early. 🎄 For too long, building powerful AI agents with Reinforcement Learning has been blocked by GPU scarcity and complex infrastructure. That ends today. Introducing Serverless RL from wandb, powered by @CoreWeave! We're making RL accessible to all.