scripod.com
The next big breakthrough will be AIs learning on the job

Overview

Shownote

Highlights

Transcript

Chapters

Pins

The next big breakthrough will be AIs learning on the job

Dwarkesh Podcast

Jun 26
Dwarkesh Podcast

Dwarkesh Podcast

Jun 26
OverviewShownoteHighlightsTranscriptChaptersPins

Shownote

Read it here [https://www.dwarkesh.com/p/the-next-paradigm]. Thanks to Mercury for sponsoring this essay. Mercury [https://mercury.com/] 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 [http://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 Get full access to Dwarkesh Podcast at www.dwarkesh.com/subscribe [https://www.dwarkesh.com/subscribe?utm_medium=podcast&utm_campaign=CTA_4]

Highlights

This podcast explores the key research bets and challenges in developing artificial general intelligence (AGI), focusing on how AI labs are approaching training and learning. The discussion moves beyond simple scaling to examine the nuances of verifiability, grindability, and continual learning in AI systems.
00:00
Scaling can overcome current deficits like data inefficiency
02:12
A domain must be grindable for scalable AI training.
06:10
Deployment reveals valuable real-world learning opportunities
14:13
RL preserves existing knowledge better than supervised learning
15:22
Dreaming could become a fourth scaling axis
17:23
AI shifts from pre-deployment training to learning from deployment

Chapters

The big research bet the labs are making
00:00
Grindability is just as important as verifiability
02:12
Will RLVR alone generalize?
06:10
Getting the learning back to the weights
08:41
Dreaming
15:22
What 2027 looks like
17:23

Transcript

Dwarkesh Patel: So here's the big research bet that all the labs are making. They think that if we train AIs to accomplish millions of verifiable tasks across thousands of diverse RL environments, then we will have basically built AGI. Because this kind of...