AI Engineer
← All organizations

Model fine-tuning and agent reinforcement learning

OpenPipe

OpenPipe develops tools for customizing language models and training AI agents. Its original managed platform helped application developers capture production prompts and responses, turn them into training data, and deploy task-specific models through OpenAI-compatible endpoints. That workflow supported applications such as data extraction and classification. Its commercial model-training and inference functionality now resides within Weights & Biases.

Founded in 2023 by Kyle Corbitt and David Corbitt, OpenPipe extended its work from fine-tuning models to reinforcement learning for multi-step agents. Its Agent Reinforcement Trainer (ART) integrates GRPO training into Python applications, separating application workflows from GPU-based training and inference through a client/server architecture. Developers assign rewards to agent trajectories, and the training loop updates LoRA adapters before running further trials. Its RULER tooling supplies automatic reward generation, helping developers construct the feedback used in reinforcement learning.

CoreWeave completed its acquisition of OpenPipe in 2025, and OpenPipe’s team remains there working on model reliability. Following migration of core functionality, including model distillation, its legacy platform retirement ended model training and inference on OpenPipe.ai effective July 30, 2026. Customers were offered automated migration of existing models to W&B Inference; open-source ART remains supported rather than being retired with the hosted service.

openpipe.ai ↗

3 talks

Newest first

1 speaker at AIE

Affiliations reflect their AIE appearances, not necessarily current employment.

Kyle CorbittCo-founder and former chief executive of OpenPipe

Start here

  1. The GenAI Maturity Curve (or: You Probably Don’t Need Fine-Tuning)

    Start here to understand the progression from general-purpose prompting to task-specific models, including multi-stage inference and fine-tuned Llama 3 examples.

    Kyle CorbittAI Engineer World's Fair 2024

  2. How we scaled 500m AI agents in production with 2 engineers

    Learn how an initially effective GPT-4 workflow encountered production costs and limits in prompting, caching, and scaling at Method Financial.

    Mustafa Ali · Kyle CorbittAI Engineer Summit 2025

  3. How to Train Your Agent: Building Reliable Agents with RL

    Examine Corbitt’s reported comparison of a trained Qwen 2.5 14B email agent with prompted frontier-model baselines, including training effort and inference economics.

    Kyle CorbittAI Engineer World's Fair 2025

Messages from the stage

Specialization needs production evidence

Corbitt connects fine-tuning decisions to real production prompts, representative training data, and evaluation. Quality, cost, latency, and consistency provide the criteria for deciding whether a specialized model addresses a demonstrated need.

Financial outputs need validation

The joint discussion with Mustafa Ali examines fine-tuned open-source models for processing unstructured liability data. It pairs model error comparisons with financial-output validation and the tradeoffs between cost, accuracy, and latency.

Rewards shape agent reliability

The ART·E case study examines verifiable rewards and reward-hacking failure modes alongside reported improvements in agent accuracy. Reward design is a central technical concern in training the email assistant.

Affiliations reflect each recorded session, not necessarily current employment.

Company sources · checked 2026-08-28