The Effortless Podcast

Recursive Self-Improvement: AI Models, Agents and Memory - The Effortless Podcast - Episode 24

Episode Notes

In this episode of The Effortless Podcast, Dheeraj Pandey and Amit Prakash return after the summer break to unpack recursive self-improvement (RSI) across the three legs of the AI stack: the model, the agent harness and memory. They start with the news cycle, from small, fast classifier models like Jev to 10,000 parallel agents working on Navier-Stokes, and what it means for routing, discovery and how agents communicate.

The conversation moves to where humans still sit in the loop as models help build the next generation of models, what Bing and Google AdSense taught about measurement and fast feedback loops, and why self-improving memory and agents need metrics that act as leading indicators. They close on the difference between generalized intelligence and institutional intelligence, the limits of ideas that never meet physics, and why bits ultimately have to meet atoms.

Key Topics & Timestamps

• 00:00 - Intro
• 00:30 - Back after summer and the recent news cycle
• 02:29 - Jev and the case for small, specialized models
• 04:44 - From BERT to Jev: what actually changed
• 06:36 - Routing, OpenRouter and the policy engine
• 09:20 - Discovery and routing: MCP, A2A and agents as IP addresses
• 10:25 - How 10,000 parallel agents learned to communicate
• 11:47 - ALOHA, Ethernet and lessons from networking
• 12:44 - Distributed consensus: RISC vs CISC as a model for agents
• 15:21 - Separation of concerns and the three legs of the AI stack
• 16:58 - What recursive self-improvement means
• 20:21 - Models helping build models: GFS1 to GFS2
• 23:52 - Shrinking the 90% perspiration
• 24:59 - Lessons from Bing and AdSense: measurement and fast loops
• 27:46 - GFS vs Bigtable: when structure pays off
• 29:32 - Self-improving memory and junk CRM data
• 34:22 - Self-improving agents: code, prompts and judgment
• 35:07 - Metrics, evals and observability
• 36:28 - From lagging to leading indicators
• 39:58 - Generalized vs institutional intelligence
• 42:07 - Who owns institutional intelligence: training data and privacy
• 45:36 - Can models generate novel ideas?
• 47:50 - Math vs physics: where ideas meet limited resources
• 49:41 - Toxicity in drug discovery and in CI/CD
• 53:39 - Taste, context and simulated populations
• 55:16 - An agent that interviews you for story memories
• 57:54 - "Only two jobs left" and the limits of physical AI
• 1:00:51 - Quantum simulation and Earth as a giant computer

Hosts:

Amit Prakash – CEO and Founder at AmpUp, former engineer at Google AdSense and Microsoft Bing, with extensive expertise in distributed systems and machine learning

Dheeraj Pandey – Co-founder and CEO at DevRev, former Co-founder & CEO of Nutanix. A tech visionary with a deep interest in AI, systems, and the future of work.

Follow the Hosts:

Amit Prakash

LinkedIn – https://www.linkedin.com/in/amit-prakash-50719a2/

Twitter/X – https://x.com/amitp42

Dheeraj Pandey

LinkedIn – https://www.linkedin.com/in/dpandey/

Twitter/X – https://x.com/dheeraj

Share Your Thoughts

Have questions, comments, or ideas for future episodes?

Email us at EffortlessPodcastHQ@gmail.com

Don't forget to Like, Comment, and Subscribe for more conversations at the intersection of AI, systems, and product design.