Subvolts take AI learns science by doing real experiments, speeding up discovery. Oct 2, 2025 - Developers
- Learn about novel AI training methods that connect models to physical reality for scientific tasks.
- Business
- Understand how AI can accelerate R&D in fields like materials science and advanced manufacturing.
This interview explores a novel AI training method using physical experiments, potentially advancing scientific discovery.
Training an AI Scientist with Feedback from Reality, w- Liam Fedus & Ekin Dogus Cubuk (from a16z)
Audio streams from Erik Torenberg, Nathan Labenz. Your place is remembered on this device.
From Subvolts
Summary and analysis by Subvolts editorial, AI-assisted.
The Cognitive Revolution podcast episode "Training an AI Scientist with Feedback from Reality" features Liam Fedus and Ekin Dogus Cubuk from Periodic Labs. They discuss training frontier AI models using real-world physics lab experiments instead of digital rewards. This approach aims to accelerate scientific discovery, such as finding room-temperature superconductors, by integrating AI with automated materials synthesis and physical testing.
Imagine teaching an AI to be a scientist. Instead of just showing it books, this AI actually runs real-world physics experiments. This helps it learn faster and discover new things, like materials for better technology.
This episode offers insight into a novel AI training methodology that grounds models in physical reality. It's for researchers and developers exploring AI's application in scientific discovery and advanced manufacturing. The discussion on physically grounded reward functions and automated synthesis connects to frontier labs' efforts in building more capable and scientifically adept AI systems, moving beyond purely digital benchmarks.
- HostNathan Labenz
- GuestLiam FedusGeneral Partner, Periodic Labs
- GuestEkin Dogus CubukGeneral Partner, Periodic Labs
- Episode Title: Training an AI Scientist with Feedback from Reality, w- Liam Fedus & Ekin Dogus Cubuk (from a16z)
- Show: The Cognitive Revolution
- Length: 61 minutes
- Core Innovation: Physically grounded reward function integrating experiments directly with AI systems.
- Training Approach: Trains pre-trained language models on physics and chemistry using real-world lab experiments.
Today’s special crosspost features a16z General Partner Anjney Midha with Liam Fedus and Ekin Dogus Cubuk of Periodic Labs. They discuss experiments in training frontier AI models on real-world physics lab experiments—rather than digital reward functions—and how coupling language models with automated materials synthesis could accelerate the discovery of room-temperature superconductors and transform R&D in advanced manufacturing.
Shownotes brought to you by Notion AI Meeting Notes - try one month for free at: https://notion.com/lp/nathan
- Core innovation: Periodic Labs is creating a "physically grounded reward function" by integrating experiments directly with AI systems.
- Physical lab integration: Unlike traditional AI labs, Periodic includes actual physical experimentation facilities where AI can run tests in the real world.
- Novel approach to training: Periodic takes pre-trained language models and further trains them on physics and chemistry knowledge to improve performance in scientific domains.
Read the full transcript here: https://storage.aipodcast.ing/cache/combinedtranscript-20251002T100327544Z.html
Sponsors:
Linear:
Linear is the system for modern product development. Nearly every AI company you've heard of is using Linear to build products. Get 6 months of Linear Business for free at: https://linear.app/tcr
AGNTCY:
AGNTCY is dropping code, specs, and services.
Visit AGNTCY.org.
Visit Outshift Internet of Agents
Claude:
Claude is the AI collaborator that understands your entire workflow and thinks with you to tackle complex problems like coding and business strategy. Sign up and get 50% off your first 3 months of Claude Pro at https://claude.ai/tcr
Shopify:
Shopify powers millions of businesses worldwide, handling 10% of U.S. e-commerce. With hundreds of templates, AI tools for product descriptions, and seamless marketing campaign creation, it's like having a design studio and marketing team in one. Start your $1/month trial today at https://shopify.com/cognitive
PRODUCED BY:
- physically grounded reward function
- Liam Fedus & Ekin Dogus Cubuk
- AI training that uses real-world results from experiments, not just digital scores.
- automated materials synthesis
- Liam Fedus & Ekin Dogus Cubuk
- Using machines to automatically create new materials.
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
Same guests, other shows
Questions people ask
- What is Periodic Labs' core innovation?
- Periodic Labs creates a "physically grounded reward function" by integrating actual physical experiments directly with AI systems.
- How does Periodic Labs train its AI models?
- They take pre-trained language models and further train them on physics and chemistry knowledge, using real-world lab experiments for feedback.
Page generated Oct 19, 2026. Summaries are Subvolts' own; the audio, artwork and notes belong to Erik Torenberg, Nathan Labenz.