What are the key takeaways from this a16z Podcast episode?

Key insights include: **Defining True RSI:** Genuine RSI requires the system being improved to also be the system doing the improving — a closed loop. When humans use AI tools to build better AI, the optimizer remains human cognition, making it autocatalytic, not RSI. The critical test: ask what is the optimizer and what is the optimizee in any claimed RSI system.; **Numerator/Denominator Framework:** Evaluate RSI claims by splitting progress into a numerator (raw capability gains) and denominator (compute, capital, and human labor hours expended). Many companies show faster model releases but fail when capital raised is factored into the denominator. Genuine RSI must show capability improving while the denominator holds flat or shrinks.; **Stack-Level Co-Optimization:** RSI applied only to model weights or only to inference harnesses creates bottlenecks, because weights and harnesses interact. MakerMaker targets co-optimization across prompts, harnesses, weights, and eventually chips simultaneously. Amdahl's Law applies: optimizing one component that represents 20% of total cost yields at most a 20% system-level gain.

What did Daval Ajita and Owen He discuss on a16z Podcast?

A16Z general partner Martin Casado interviews MakerMaker.ai cofounders Daval Ajita and Owen He on true Recursive Self-Improvement (RSI), distinguishing it from autocatalytic AI development. They cover closed-loop system design, measuring genuine capability gains versus resource costs, and early empirical evidence that self-improving systems can simultaneously improve performance and efficiency. Key topics include: **Defining True RSI:** Genuine RSI requires the system being improved to also be the system doing the improving — a closed loop. When humans use AI tools to build better AI, the optimizer remains human cognition, making it autocatalytic, not RSI. The critical test: ask what is the optimizer and what is the optimizee in any claimed RSI system.; **Numerator/Denominator Framework:** Evaluate RSI claims by splitting progress into a numerator (raw capability gains) and denominator (compute, capital, and human labor hours expended). Many companies show faster model releases but fail when capital raised is factored into the denominator. Genuine RSI must show capability improving while the denominator holds flat or shrinks..

How long is this episode of a16z Podcast?

This episode is 33 minutes long. SignalCast provides an AI-generated summary so you can get the key insights in about 3 minutes.

a16z Podcast

The Race to Recursive Self-Improving AI

October 12, 2026

33 min episode · 2 min read

Daval Ajita,Owen He,Martin Casado

Episode

33 min

Read time

2 min

Topics

Productivity, Relationships, Startups

AI-Generated Summary

Published Oct 12, 2026

Key Takeaways

✓Defining True RSI: Genuine RSI requires the system being improved to also be the system doing the improving — a closed loop. When humans use AI tools to build better AI, the optimizer remains human cognition, making it autocatalytic, not RSI. The critical test: ask what is the optimizer and what is the optimizee in any claimed RSI system.
✓Numerator/Denominator Framework: Evaluate RSI claims by splitting progress into a numerator (raw capability gains) and denominator (compute, capital, and human labor hours expended). Many companies show faster model releases but fail when capital raised is factored into the denominator. Genuine RSI must show capability improving while the denominator holds flat or shrinks.
✓Stack-Level Co-Optimization: RSI applied only to model weights or only to inference harnesses creates bottlenecks, because weights and harnesses interact. MakerMaker targets co-optimization across prompts, harnesses, weights, and eventually chips simultaneously. Amdahl's Law applies: optimizing one component that represents 20% of total cost yields at most a 20% system-level gain.
✓Inference as the Proving Ground: Improving AI inference efficiency is the domain MakerMaker uses to validate RSI because faster inference directly reduces the denominator while enabling more rollouts, which improves model training — compounding both sides simultaneously. The NanoGPT benchmark illustrates algorithmic progress: GPT-2 retraining dropped from weeks on multiple GPUs to hours on one GPU using identical hardware.
✓Variance Control as Core Discipline: Reward hacking propagates invisibly through multi-agent RSI stacks — an agent built by an agent built by an agent can embed subtle hacks that surface as false positive results. MakerMaker assigns a dedicated engineer (internally called the Viceroy of Variance) whose sole role is to invalidate experimental results and identify noise before any finding is accepted.

What It Covers

A16Z general partner Martin Casado interviews MakerMaker.ai cofounders Daval Ajita and Owen He on true Recursive Self-Improvement (RSI), distinguishing it from autocatalytic AI development. They cover closed-loop system design, measuring genuine capability gains versus resource costs, and early empirical evidence that self-improving systems can simultaneously improve performance and efficiency.

Key Questions Answered

•Defining True RSI: Genuine RSI requires the system being improved to also be the system doing the improving — a closed loop. When humans use AI tools to build better AI, the optimizer remains human cognition, making it autocatalytic, not RSI. The critical test: ask what is the optimizer and what is the optimizee in any claimed RSI system.
•Numerator/Denominator Framework: Evaluate RSI claims by splitting progress into a numerator (raw capability gains) and denominator (compute, capital, and human labor hours expended). Many companies show faster model releases but fail when capital raised is factored into the denominator. Genuine RSI must show capability improving while the denominator holds flat or shrinks.
•Stack-Level Co-Optimization: RSI applied only to model weights or only to inference harnesses creates bottlenecks, because weights and harnesses interact. MakerMaker targets co-optimization across prompts, harnesses, weights, and eventually chips simultaneously. Amdahl's Law applies: optimizing one component that represents 20% of total cost yields at most a 20% system-level gain.
•Inference as the Proving Ground: Improving AI inference efficiency is the domain MakerMaker uses to validate RSI because faster inference directly reduces the denominator while enabling more rollouts, which improves model training — compounding both sides simultaneously. The NanoGPT benchmark illustrates algorithmic progress: GPT-2 retraining dropped from weeks on multiple GPUs to hours on one GPU using identical hardware.
•Variance Control as Core Discipline: Reward hacking propagates invisibly through multi-agent RSI stacks — an agent built by an agent built by an agent can embed subtle hacks that surface as false positive results. MakerMaker assigns a dedicated engineer (internally called the Viceroy of Variance) whose sole role is to invalidate experimental results and identify noise before any finding is accepted.

Notable Moment

When pressed on whether improvement without asymptoting was theoretically possible, Ajita stated it was not merely possible but already occurring — that their system was producing better results with less energy over time. Casado noted this was the first time he had heard that claim stated so directly.

Know someone who'd find this useful?

Episode Transcript

The fundamental idea here is to build AI that improves itself and I think we're getting there. For the first time, we could swap the optimizer as the optimisee and this leads to a reinvestment of cognitive improvement into itself. You can think of RSI at many different levels of the stack: prompts, harness, weights, and you can go all the way to basically chips. Our goal here is to optimize all the levels of the stack. Our approach to our side is that we have a multi agent system working together that's been recursively improved. I'm so impressed by your belief that, like, this idea of a model improving itself. But it's not a belief. It is a dramatic statement, is I think is the first time I've heard you say it. We have been so thorough, we have not released things until we're super sure it's correct. Do you envision a system where the amount of energy that goes into it somehow decreases over time relative to improvement. We can't share too much, but one of the very important things that we've tried to do... A16Z general partner Martin Casado sits down with MakerMaker dot AI cofounders Daval Ajita and Owen He to unpack Recursive Self Improvement, or RSI, and separate the increasingly popular term from what it would actually mean for an AI system to improve itself. Daval and Owen distinguish true RSI from simply using AI to automate pieces of AI development. Their goal is a closed loop, where the system being improved is also doing the improving, across components including models, harnesses, and eventually more of the AI stack. They explain why closing that loop still doesn't guarantee runaway takeoff, and why measuring real progress requires accounting for both capability gains and the resources required to achieve them. They also share what they're seeing in their own research, including early signs that self improving systems can increase performance while improving efficiency. And they dig into the practical challenges of proving those gains are real, from variance and reward hacking to the enormous difficulty of distinguishing genuine improvement from noise. Welcome both of you. We're going to be talking about everybody's favorite topic, which is RSI, Recursive Self Improvement. Maybe to start, it'd be great to just introduce yourselves and your background, so we'll talk a little bit about what you're working on now, and then we'll dig into the content. I think of myself as a physicist of some kind, I studied that, then I did a PhD in reinforcement learning at the MIT Media Lab, I really like building things, and I'm very excited to be building a cool startup, make a maker of Owen and Grace. My name is Owen. I was first studying math, and then one day I realized that all mathematicians do is proving theorems, and I was taking a class called formal language and logic, and they were showing how you can build a theorem prover. And I was …

Get the full transcript (5,931 words) + summary by email — free

One-time email with the complete transcript and AI summary of this episode. No account needed.

One email, no spam. We’ll also show you what SignalCast does.

Browse all a16z Podcast transcripts →