Subvolts take AI models may soon learn continuously, improving on the job like humans. Jun 26, 2026 - Developers
- Understand cutting-edge research directions in AI model training and continuous learning.
- Business
- Consider how AI systems that learn on the job could transform operations and service delivery.
This episode explores a significant potential shift in AI development, focusing on continuous learning and its implications.
The next big breakthrough will be AIs learning on the job
Audio streams from Dwarkesh Patel. Your place is remembered on this device.
From Subvolts
Summary and analysis by Subvolts editorial, AI-assisted.
Dwarkesh Patel's podcast episode "The next big breakthrough will be AIs learning on the job" explores how AI labs are betting on continuous learning. The discussion covers the importance of "grindability" alongside verifiability in AI development, the potential for Reinforcement Learning from Human Feedback (RLHF) to generalize, and how to integrate new learning back into AI models. It also touches on AI "dreaming" and future AI landscapes.
Imagine an AI that keeps learning and getting better as it works, not just from pre-set lessons. This episode discusses how AI developers are trying to make this happen, which could lead to smarter and more helpful AI tools for everyone.
This episode is for AI researchers and enthusiasts interested in the future of model training. It delves into the technical challenges and potential breakthroughs in making AI systems learn continuously, much like humans do on the job. This directly relates to the ongoing efforts by frontier labs like Google Gemini and OpenAI to develop more capable and adaptable AI, moving beyond static training datasets towards dynamic, real-time learning.
- HostDwarkesh Patel
- Episode Title: The next big breakthrough will be AIs learning on the job
- Show: Dwarkesh Podcast
- Length: 20 minutes
- Timestamp: What 2027 looks like: 00:17:23
Read it here.
Thanks to Mercury for sponsoring this essay.
Mercury has automated basically my entire bill pay process for my business. I just give contractors a dedicated email address, and when they send an invoice, Mercury automatically creates a draft payment for me to review. I no longer have to hunt through my inbox for invoices or deal with messy spreadsheets to track my bills. Mercury handles it all. Learn more at mercury.com
Timestamps:
(00:00:00) – The big research bet the labs are making
(00:02:12) – Grindability is just as important as verifiability
(00:06:10) – Will RLVR alone generalize?
(00:08:41) – Getting the learning back to the weights
(00:15:22) – Dreaming
(00:17:23) – What 2027 looks like
This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.dwarkesh.com
- RLVR
- Dwarkesh Patel
- Reinforcement Learning from Human Feedback (RLHF), a method where AI learns from human preferences.
1 h 371 h 201 h 37AI researchers debate how close we are to recursive self-improvement
WithJohn Schulman, Beren Millidge, Charlie O’Neill
2 h 21Questions people ask
- What is the main research bet labs are making?
- The main research bet is on AIs learning on the job, implying continuous learning and adaptation.
- What is grindability in AI?
- Grindability is presented as a crucial aspect of AI development, alongside verifiability, suggesting the ability to handle extensive training or processing.
Page generated Oct 19, 2026. Summaries are Subvolts' own; the audio, artwork and notes belong to Dwarkesh Patel.