Dhaval Adjodah

1. Key Themes

Real RSI Means Swapping the Optimizer for the Optimizee

The founders argue "RSI" has become an overloaded buzzword. Their litmus test is to ask who is improving and who is being improved. If a human uses AI tools to build better AI, that is "autocatalysis," a pattern that has accompanied nearly every technological epoch. True RSI puts the model itself on the receiving end of the improvement.

Owen He: "I think the important thing when you are differentiating them is to ask the question: what is the optimizer and what is the optimizee? When you're self-improving, who is improving?" [00:03:25]

Owen He: "Now, for the first time, we could swap the optimizer as the optimizee, so the the brain of the agents, which is LM, the model, can be the thing that you're optimizing and this lead to a compound, this reinvestment of cognitive improvement into itself." [00:04:46]

A Closed Loop Changes the Math: Constant or Accelerating Marginal Gains

Dhaval Adjodah frames capability growth as performance divided by cost. In prior technologies the numerator was fixed, so progress plateaued. If the improver is the thing being improved, the numerator correlates with capability, and the dynamics change qualitatively.

Dhaval Adjodah: "If you have a closed loop, you still have to pour in resources, but the marginal gains ideally get either constant or larger over time, and that's the fundamental difference." [00:05:45]

Dhaval Adjodah: "The simplest case would be that the numerator is also correlated with capability... the math cancels out, and you basically get constant improvement over time... And then there's one more case, which is now that the numerator could be correlated with a higher power of capability, and then you get exponential takeoff." [00:07:01]

Closed Loop ≠ Takeoff: It's an Empirical Question

Martin Casado pushes back on the popular leap from "circle with an arrow" to "foom." The founders agree that outcomes depend on an exponent nobody can derive analytically, and that the field lacks even a learning theory of deep learning.

Dhaval Adjodah: "It's almost not useful to speculate on what we think could happen because this is really an empirical problem. And the real way to know what's happening is to test things, and testing things is really, really hard." [00:08:11]

Dhaval Adjodah: "We don't even have a learning theory of deep learning. We don't know fundamentally why it works... I think the closest things we have are scaling laws, and these are empirical laws." [00:08:59]

The Denominator Is the Honest Test: Account for All Resources

The key discipline is to refuse to credit "RSI" for gains bought by pouring in more capital, compute, or human effort. Casado notes that model release cadence is speeding up, but dividing by capital raised changes the picture.

Martin Casado: "If you look at like the model release time, that it is reducing... But if you divide by the actual amount of capital raised by these companies, that's no longer the case, right?" [00:09:59]

Dhaval Adjodah: "If you had get improvement and they just come at more and more resources being poured in, it's not RSI. You have to be able to back your claim of how much improvement you're getting." [00:10:48]

The team includes human experimentation time in the denominator and runs an internal "bleederboard" tracking experiment cost per person.

Optimize the Whole Stack Jointly, Because Components Interact

Owen He argues that RSI efforts focused on only one layer (harness-only or weights-only) are incomplete, since layers affect each other and any layer left un-optimized becomes the bottleneck.

Owen He: "Anything that you don't optimize will later become your bottleneck." [00:09:49]

Owen He: "Weights and harnesses are not isolated components. They have impact on each other." [00:17:18]

Dhaval Adjodah: "You can think of RSI at many different levels of the stack, right? You can think of prompts, harness, weights, you can go all the way to, basically, chips." [00:00:16]

Humans Are Agents Inside the System, Not Outside It

Contrary to the "replace the researchers" narrative, makermaker treats humans as one kind of agent in a multi-agent system that is recursively improved.

Owen He: "In this multi-agent system, human researchers and engineers are actually special case of these agents, and it's this hybrid system of multiple agents working together that's being recursively improved." [00:09:34]

Dhaval Adjodah: "If you have good researchers and good engineers, that they can bring to the table that's going to be very hard to replace by agents." [00:22:27]

Inference Optimization Is the Ideal Proving Ground

makermaker picked inference serving because it is close to AI itself, improves both numerator and denominator, and is economically urgent as everyone wants to serve open-weight models efficiently.

Dhaval Adjodah: "There's not that many problems that are super close to AI and gives you improvement on both the numerator and denominator. And one of those is the inference problem." [00:18:28]

Dhaval Adjodah: "Everybody in the world right now wants to serve their own model, and everybody wants to use the latest, for example, open-weight model, Kimi k3 or whatever, just coming out, and they want to be able to serve it as efficiently as possible." [00:19:31]

Intelligence Is a Vector, and Amdahl's Law Governs Takeoff by Domain

Owen He argues that the slowest component caps overall speedup, so fast-evaluation domains like code and math will accelerate far faster than domains gated by atoms, such as drug discovery.

Owen He: "Intelligence is not a single scalar. It's more like a multi-dimension vector, and there will be domains and fields where this RSI will give you a takeoff, and there will be dimensions where you might plateau." [00:28:18]

Owen He: "For domains where you still need to move atoms that interact with physical world that's really slow, you're still going to be bottlenecked by that... those fields probably won't be accelerated as fast as, you know, coding or mathematics, where you have very fast and cheap evaluations." [00:29:11]

Measurement Rigor Is the Hardest Part

Variance, GPU-level noise, and reward hacking make it extremely hard to separate real gains from artifacts.

Dhaval Adjodah: "The technical answer is, you have to be really, really good at measuring things." [00:26:00]

Dhaval Adjodah: "So far, we have not found a replacement for what we call quality time. Literally, running hundreds of experiments, looking at it, developing an intuition, trying to figure out what is wrong, what is not wrong." [00:26:43]

2. Contrarian Perspectives

Most "RSI" Is Not RSI, and Most RSI Papers Don't Survive Scale

Dhaval Adjodah argues that much of the published and announced RSI work fails replication or reduces to noise. This challenges the current hype cycle, where "we use AI to create AI" is treated as RSI.

Dhaval Adjodah: "A lot of the papers that claim to do RSI, when we test it in our internal infra, we find that they don't work. And a lot of it is because they're cool ideas that humans try to introspect and say, 'Oh, you know, this is what I think should happen.'" [00:14:09]

Dhaval Adjodah: "Every few months, you know, there's like an announcement of something RSI-related, we look at it... we try to replicate it, and we realize, 'Oh, it's noise, you know, or it's cherry-picking.'" [00:25:42]

Simple Closed Loops Beat Clever, Human-Inspired Architectures

Complex, introspection-inspired mechanisms make good papers but don't scale. The simple thing that closes the loop wins, which runs against the instinct to build elaborate cognitive architectures.

Dhaval Adjodah: "They're complicated, they sound cool, they get you a good paper, but then when we test it at scale, you know, on benchmarks, on real economic problems, they don't scale. And what we found is that the simpler thing that closes the loop is the thing that scales." [00:14:27]

The Same Hardware Is Getting Dramatically More Capable Through Algorithms Alone

Owen He and Casado point to the nanoGPT speedrun: GPT-2, once weeks on many GPUs, now trains on a single GPU for roughly $100, with much of the recent progress coming from agents rather than humans. This reframes "AI progress is just money and hardware."

Owen He: "It's basically asking you to retrain GPT-2, which, when it just came out, is a very expensive model to train and it takes weeks, I think, on multiple GPUs. And right now people can train it on a single GPU within hours." [00:12:11]

Martin Casado: "It's $100, I just saw it." [00:12:59]

Owen He: "Same GPU, same data." [00:13:02]

Owen He (on the source of gains): "Agents, a lot of them are auto research results." [00:13:07]

The Next Frontier May Hide in Boring Corners, Not Fashionable Problems

Rather than aiming RSI only at trendy hard problems, the founders suspect the largest unlocks lie in unglamorous workflows. They frame it with a company joke grounded in an Einstein/GPS analogy.

Dhaval Adjodah: "The complexity of the world is such that the real challenges probably exist in spaces that are not being applied to, right? And so, that's the joke, that there is some workflow somewhere in a garden hose company in Wisconsin that if we solve it, we get, you know, foom." [00:31:20]

Capability Per Unit Resource May Keep Rising, Bounded Only by the Physics of Computation

Casado doubts that output can rise while energy input falls without violating information theory. Adjodah responds that it's happening empirically and will asymptote only at physical limits.

Dhaval Adjodah: "It's not just possible, I think it's happening." [00:23:31]

Dhaval Adjodah: "You know, how many how many particles are there? 10 to the 80. Like we know speed of light, you know... You multiply those two, and you basically get an upper bound of computation." [00:24:10]

3. Companies Identified

makermaker.ai

Startup founded by Dhaval Adjodah, Owen He, and Grace, building true recursive self-improving AI systems with an initial focus on inference-serving efficiency. Mentioned as the subject of the episode, with a rigorous, math-first, measurement-obsessed culture and claimed early evidence of improving performance and efficiency together.

Dhaval Adjodah: "We are trying very hard to pick areas of the AI space that are as close as possible to AI itself. Again, we try to be recursive." [00:18:14]

Dhaval Adjodah: "We have been so thorough, we have not released things until we're super sure it's correct." [00:00:43]

DeepMind

Google's AI research lab. Mentioned as Owen He's prior employer, where he worked as a research scientist after a PhD in continual learning and meta-learning.

Owen He: "For my PhD, I worked on continual learning and meta-learning, then worked for a few years at DeepMind as a research scientist." [00:02:49]

MIT Media Lab

Research lab at MIT. Mentioned as the setting for Dhaval Adjodah's PhD in reinforcement learning.

Dhaval Adjodah: "I did PhD in reinforcement learning at the MIT Media Lab." [00:02:17]

nanoGPT benchmark (GPT-2 retraining speedrun)

A community benchmark for retraining GPT-2 as fast and cheaply as possible. Mentioned as evidence that algorithmic and kernel improvements, many found by agents, are cutting cost on fixed hardware.

Owen He: "People can train it on a single GPU within hours, and that number has been shrinking constantly." [00:12:28]

Kimi (Moonshot's open-weight model line, referenced as "Kimi k3")

Open-weight model family. Mentioned as the type of fast-moving open-weight model everyone wants to serve efficiently, which creates economic demand for RSI-driven inference optimization.

Dhaval Adjodah: "Everybody wants to use the latest, for example, open-weight model, Kimi k3 or whatever, just coming out, and they want to be able to serve it as efficiently as possible." [00:19:40]

OpenAI (GPT-2 / GPT-5)

Mentioned only as reference points. GPT-2 is the benchmark whose training cost has collapsed, and GPT-5 is the frontier model for which Casado says similar cost declines aren't being observed.

Martin Casado: "What you want to see is like you can do that with GPT-5, which we're not seeing, right?" [00:13:14]

NVIDIA (B200 GPUs)

Mentioned in the context of hardware-level noise. Two nominally identical B200 GPUs can behave slightly differently, and this variance propagates to high-level experimental results.

Dhaval Adjodah: "This GPU and this other GPU, they're both, you know, B200, but they have slightly different behaviors like on the physics, and you can see some of that being replicated at the high level." [00:27:47]

4. People Identified

Co-founder of makermaker.ai; physicist-trained, with a PhD in reinforcement learning from the MIT Media Lab. Mentioned as the lead articulator of the closed-loop math framework and the measurement-rigor culture.

Dhaval Adjodah: "We always start with a very grounded mathematical framework that basically lays out the map of the whole space." [00:14:37]

Owen He

Co-founder of makermaker.ai; math background, PhD work on continual learning and meta-learning, former DeepMind research scientist. Mentioned for the optimizer/optimizee framing and the Amdahl's-law view of takeoff by domain.

Owen He: "So, the digital part we could optimize and accelerate, you could get like 1,000 times faster, a million times faster, but if it's only 20% of your overall time, you're only going to get 20% speed up, right?" [00:28:59]

Grace (makermaker.ai co-founder)

Third co-founder of makermaker, named only in passing.

Dhaval Adjodah: "I'm very excited to be building a cool startup, makermaker, with Owen and Grace." [00:02:17]

Martin Casado

a16z General Partner and the host. Networking background, which he brings to the control-systems framing (oscillation, convergence, non-linear dynamics). Mentioned as the interlocutor pressure-testing the claims, including the capital-divided-by-release-cadence point.

Martin Casado: "When I have dealt with, you know, non-linear control systems, there's a bunch of degenerate behavior you have to deal with. It's not just convergence and divergence, you have things like oscillations." [00:27:18]

Richard Feynman

Physicist. Cited by Owen He for the distinction between "stamp collections" and physics, used to argue that RSI research needs a unified framework.

Owen He: "I think Richard Feynman said once that there are two kinds of sciences. There's stamp collections, and then there are physics." [00:16:21]

Albert Einstein

Physicist. Invoked in the "garden hose company in Wisconsin" analogy: without relativity, GPS engineers would have needed an unexplained empirical fudge factor.

Dhaval Adjodah: "If Einstein hadn't discovered relativity, he would not have these laws that tell you how speed and, you know, contraction of time works." [00:30:28]

The "Viceroy of Variance" (unnamed makermaker engineer)

An internal engineer whose explicit job is to attack the team's own results and find noise or errors. Mentioned as an exemplar of adversarial self-verification.

Dhaval Adjodah: "We named him the Viceroy of Variance. His job is to go and like destroy the results we have and find what's wrong." [00:26:00]

5. Operating Insights

Institutionalize a "Red Team" Role for Your Own Results

Makermaker assigns a dedicated engineer to try to break internal results, reinforcing the role with team rituals like a Slack emoji. In a domain full of noise, cherry-picking, and hype, making disproof someone's job protects against shipping false positives and builds credibility when you do announce.

Dhaval Adjodah: "His job is to go and like destroy the results we have and find what's wrong. And we even have little like Slack emojis that we have." [00:26:25]

Make Cost Visible and Slightly Embarrassing: The "Bleederboard"

Instead of only ranking experiment performance on leaderboards, makermaker publicly tracks the spend of every experimenter. This builds a cost-aware culture where the denominator is as salient as the numerator.

Dhaval Adjodah: "We have something called a bleederboard, where we keep track of the cost of experiments of everybody who's running experiments... any day, can look at like who's high up on the bleederboard." [00:11:52]

Compress Complexity Into a Few Abstractions Before Building

Rather than building 200 bespoke systems and hoping a Rube Goldberg machine works, the team invests early in first-principles math to find the ~12 abstractions that generalize. This lets them recognize that many papers are subclasses of one framework and avoids getting lost in the search space.

Dhaval Adjodah: "Maybe you realize, you know, they're actually just 12 abstractions that generalize to these 200 things." [00:15:26]

Let the System Choose What to Optimize, and Keep Humans Hands-Off

The team deliberately lets the agent decide which components and which LLMs to use, even letting the master prompt itself be improved, to avoid human-imposed bottlenecks.

Owen He: "To be honest, today it's not really us who are deciding what to optimize, it's our agent." [00:20:47]

Spend "Quality Time" Staring at Experiments Before Trusting Any Metric

No tooling has yet replaced hundreds of hands-on runs to build intuition about variance versus real signal versus subtle reward hacking, a practical lesson for anyone running agent-heavy R&D.

Dhaval Adjodah: "What is this weird thing that actually shows that there's, you know, some kind of reward hacking like two levels down? The agent that built the agent that built the agent, did hack something." [00:27:02]

6. Overlooked Insights

"Which LLM Is Used for Which Task" Can Itself Be Delegated to the System

Buried in a brief exchange, Dhaval says model selection is not always chosen by the founders: "We let the models choose." This implies a meta-layer of routing and orchestration being optimized recursively. This is a potential moat, because the routing and optimization policy (the "improver of the improver") accumulates proprietary empirical knowledge that cannot be copied from a model checkpoint.

Dhaval Adjodah: "The choice of which LLM to use for which thing is also not always up to us. We let the models choose." [00:20:40]

The Most Valuable RSI Domain May Be Where Evaluation Is Cheap, and Nobody Is Pricing That In

Owen He's Amdahl's-law aside implies something investable: takeoff will be highly uneven, concentrating in domains with fast, cheap, automated evals (code, math, inference optimization) and lagging sharply in wet-lab, clinical, and physical domains. Investors who assume AI speeds everything uniformly may overestimate drug discovery or robotics timelines and underestimate the compounding in software-native verticals, while the "slowest component" (often the physical or human eval loop) becomes the true scarce asset.

Owen He: "Every new idea have to go through a month or multiple months evaluation. So, those fields probably won't be accelerated as fast as, you know, coding or mathematics, where you have very fast and cheap evaluations." [00:29:18]

Industry-Wide Compute Efficiency per Unit Output Is a Guarded Secret and a Leading Indicator

Dhaval mentions in passing that "how much compute you can get out of this" is "one of the most closely guarded secrets in the industry." Combined with the point that cheaper rollouts mean more trajectory sampling and therefore smarter models, denominator improvements are quietly feeding numerator improvements at frontier labs. Cost-efficiency of training and serving may be a better signal of true recursive progress than headline benchmark scores.

Dhaval Adjodah: "One of the most closely guarded secrets in the industry is exactly this thing, which is how much compute you can get out of this, because, in some sense, even if all you were improving was your denominator, you could get faster training because now you can get more rollouts." [00:13:21]