Subvolts take Learn MiniMax's RL training secrets for open-weight models and real-world LLM debugging. Feb 22, 2026 - Developers
- Learn advanced RL training techniques and debugging strategies for open-weight LLMs.
- Business
- Understand how AI labs use feedback loops and agents to improve models and monitor the AI landscape.
This is a solid interview with a researcher from a notable AI lab discussing practical training methods for open-weight models.
Intelligence with Everyone: RL @ MiniMax, with Olive Song, from AIE NYC & Inference by Turing Post
Audio streams from Erik Torenberg, Nathan Labenz. Your place is remembered on this device.
From Subvolts
Summary and analysis by Subvolts editorial, AI-assisted.
Olive Song from MiniMax discusses training their open-weight M series models using reinforcement learning and product feedback loops. The episode covers interleaved thinking for agentic tasks, reward hacking, and using FP32 precision for RL training. Song also shares insights on debugging LLM failures and how MiniMax uses AI agents to monitor the AI landscape.
This episode explains how MiniMax trains its advanced AI models using learning from experience and user feedback. It covers how they make AI agents smarter for complex tasks and fix problems when AI makes mistakes, which helps build better AI for everyone.
This episode is for AI engineers and researchers interested in practical LLM training techniques. It delves into reinforcement learning strategies, a core area for frontier labs like OpenAI and Google Gemini, and addresses challenges like reward hacking. MiniMax's approach to open-weight models also aligns with Meta AI's strategy. The discussion on debugging and AI agents tracking the AI landscape offers valuable insights into the operational aspects of cutting-edge AI development.
- HostNathan Labenz
- GuestOlive SongMiniMax
- Guest: Olive Song from MiniMax
- Show: The Cognitive Revolution
- Length: 55 minutes
- Training Method: Reinforcement learning, tight product feedback loops, systematic environment perturbations
- Model Series: M series open-weight models
- RL Precision: Moved to FP32 precision
Olive Song from MiniMax shares how her team trains the M series frontier open-weight models using reinforcement learning, tight product feedback loops, and systematic environment perturbations. This crossover episode weaves together her AI Engineer Conference talk and an in-depth interview from the Inference podcast. Listeners will learn about interleaved thinking for long-horizon agentic tasks, fighting reward hacking, and why they moved RL training to FP32 precision. Olive also offers a candid look at debugging real-world LLM failures and how MiniMax uses AI agents to track the fast-moving AI landscape.
Nathan uses Granola to uncover blind spots in conversations and AI research. Try it at granola.ai/tcr with code TCR — and if you’re already using it, test his blind spot recipe here: https://bit.ly/granolablindspot
LINKS:
Conference Talk (AI Engineer, Dec 2025) – https://www.youtube.com/watch?v=lY1iFbDPRlw
Interview (Turing Post, Jan 2026) – https://www.youtube.com/watch?v=GkUMqWeHn40
Sponsors:
Claude:
Claude is the AI collaborator that understands your entire workflow, from drafting and research to coding and complex problem-solving. Start tackling bigger problems with Claude and unlock Claude Pro’s full capabilities at https://claude.ai/tcr
Tasklet:
Tasklet is an AI agent that automates your work 24/7; just describe what you want in plain English and it gets the job done. Try it for free and use code COGREV for 50% off your first month at https://tasklet.ai
CHAPTERS:
(00:00) About the Episode
(04:15) Minimax M2 presentation (Part 1)
(17:59) Sponsors: Claude | Tasklet
(21:22) Minimax M2 presentation (Part 2)
(21:26) Research life and culture
(26:27) Alignment, safety and feedback
(32:01) Long-horizon coding agents
(35:57) Open models and evaluation
(43:29) M2.2 and researcher goals
(48:16) Continual learning and AGI
(52:58) Closing musical summary
(55:49) Outro
PRODUCED BY:
SOCIAL LINKS:
Website: https://www.cognitiverevolution.ai
Twitter (Podcast): https://x.com/cogrev_podcast
Twitter (Nathan): https://x.com/labenz
LinkedIn: https://linkedin.com/in/nathanlabenz/
- reinforcement learning
- Olive Song
- A way AI learns by trying things and getting rewards or penalties.
- open-weight models
- Olive Song
- AI models whose internal workings are shared publicly, allowing others to build upon them.
- reward hacking
- Olive Song
- When an AI finds a shortcut to get a reward without actually doing the intended task correctly.
- FP32 precision
- Olive Song
- A standard way computers store numbers, affecting how accurately AI calculations are performed.
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
1 h 05
1 h 04
36 min
1 h 35Can AI Replace Teachers? Inside the $40M Company Using AI Tutors to Teach 200% Faster | #233
WithMackenzie Price, Joe Lamont
Questions people ask
- What are interleaved thinking and reward hacking?
- Interleaved thinking helps AI agents tackle long tasks. Reward hacking occurs when AI finds unintended ways to get rewards without truly solving the problem.
- Why did MiniMax move RL training to FP32 precision?
- The notes do not specify the exact reasons for moving RL training to FP32 precision, only that they did.
Page generated Oct 19, 2026. Summaries are Subvolts' own; the audio, artwork and notes belong to Erik Torenberg, Nathan Labenz.