Subvolts take Learn advanced AI fine-tuning techniques to boost open-source model performance and cut costs. May 1, 2026 - Developers
- Learn advanced RL techniques like GRPO and LLM-as-judge to optimize open-source models for performance and cost.
- Business
- Understand how fine-tuning methods can improve AI model efficiency, reduce operational costs, and enhance product capabilities.
This episode features a practical discussion on advanced AI training techniques with a founder from an AI tooling company, offering insights into model optimization.
The RL Fine-Tuning Playbook: CoreWeave's Kyle Corbitt on GRPO, Rubrics, Environments, Reward Hacking
Audio streams from Erik Torenberg, Nathan Labenz. Your place is remembered on this device.
From Subvolts
Summary and analysis by Subvolts editorial, AI-assisted.
Kyle Corbitt, founder of OpenPipe, joins Nathan Labenz on The Cognitive Revolution to discuss reinforcement learning (RL) and custom fine-tuning for AI models. The episode covers how RL differs from supervised fine-tuning, the importance of GRPO and LLM-as-judge techniques, and their impact on open-source model performance, latency, and cost. They also explore reward hacking, evaluation design, LoRA adapters, and distillation methods used by Chinese labs.
This episode explains how to train AI models using methods like reinforcement learning, which is like teaching by rewarding good behavior. This helps make AI models smarter, faster, and cheaper to run, benefiting anyone who uses or builds AI.
This episode offers a deep dive into advanced AI training techniques, particularly reinforcement learning and its practical applications for improving open-source models. It is for AI researchers, developers, and engineers looking to optimize model performance, reduce costs, and understand cutting-edge fine-tuning methods. The discussion on GRPO and LLM-as-judge connects directly to ongoing efforts by frontier labs like Anthropic and Google to enhance model capabilities and safety through sophisticated training paradigms.
- HostNathan Labenz
- GuestKyle Corbittfounder, OpenPipe
- Guest: Kyle Corbitt, founder of OpenPipe
- Show: The Cognitive Revolution
- Length: 107 minutes
- Techniques discussed: Reinforcement learning, GRPO, LLM-as-judge, LoRA adapters, distillation
- Benefits for open-source models: Improved performance, latency, and cost
Kyle Corbitt, founder of OpenPipe, breaks down reinforcement learning and custom fine-tuning for modern AI models. He explains how RL differs from supervised fine-tuning, why GRPO and LLM-as-judge post-training matter, and how these techniques can improve performance, latency, and cost on open source models. The conversation also covers reward hacking, evaluation design, LoRA adapters, and how Chinese labs are using distillation to fast-follow frontier models.
Sponsors:
Sequence:
Sequence handles the full revenue workflow for complex pricing, from quoting and metering to invoicing, revenue recognition, and collections. Book a public demo at https://sequencehq.com and use code Cognizant in the source field to save 20% off year one
AvePoint:
AvePoint is building the control layer for AI agents so you can securely govern, audit, and recover every action at scale. Design trusted agentic outcomes from day one at https://avpt.co/tcr
VCX:
VCX, by Fundrise, is the public ticker for private tech, giving everyday investors access to high-growth private companies in AI, space, defense tech, and more. Learn how to invest at https://getvcx.com
Claude:
Claude by Anthropic is an AI collaborator that understands your workflow and helps you tackle research, writing, coding, and organization with deep context. Get started with Claude and explore Claude Pro at https://claude.ai/tcr
- Reinforcement Learning (RL)
- Kyle Corbitt
- A way to train AI by rewarding it for good actions and penalizing it for bad ones, like teaching a pet.
- GRPO
- Kyle Corbitt
- A specific method for reinforcement learning used to fine-tune AI models after initial training.
- LLM-as-judge
- Kyle Corbitt
- Using a large language model to evaluate and score the outputs of another AI model.
- LoRA adapters
- Kyle Corbitt
- A technique to efficiently fine-tune large AI models by only training a small number of extra parameters.
1 h 27AI:AM: Where's the Steering Wheel? My Agent Set the Password, Hidden Valence, Proofs Find Linux Bugs
WithCameron Berg, Everett Hildenbrandt, Grigore Rosu, Phil Maher
2 h 192 h 021 h 28AI:AM: A Level We Shouldn't Pass? Notes from The Curve + Tokens vs. Salaries & Is SaaS Cooked?
WithThomas Sohmers, swyx, Evan Miyazono, Edward Hu
51 minLoris Degioanni: Why AI Is Breaking Cybersecurity, and What Comes Next
WithLoris Degioanni
38 minWhy AI needs a new kind of supercomputer network - Episode 18
WithMark Handley, Greg Steinbrecher
Questions people ask
- What is GRPO and why is it important?
- GRPO is a reinforcement learning technique discussed for its role in post-training AI models, aiming to improve their performance and efficiency.
- How can reinforcement learning help open-source models?
- Reinforcement learning, along with techniques like LLM-as-judge, can enhance the performance, reduce latency, and lower the cost of running open-source AI models.
Page generated Oct 19, 2026. Summaries are Subvolts' own; the audio, artwork and notes belong to Erik Torenberg, Nathan Labenz.