a16z

Dhaval Adjodah and Owen He on why real self-improvement is a closed loop

Dhaval Adjodah & Owen He· Co-founders at MakerMaker.AI
·~32 min·English·a16z
AgentsReasoningTrainingInference
TL;DR

Two MakerMaker.AI founders draw a hard line between real recursive self-improvement and using AI to help build AI, argue takeoff is an empirical question rather than a belief, and explain why they attack it through inference while treating humans as one more kind of agent.

01Core Mental Model

The Optimizer Becomes the Optimized

<strong>Real recursive self-improvement is not "AI helps build AI" but the loop closing, where the thing being improved is also the thing doing the improving.</strong>

a pragmatic thing would be just from a control theory perspective. If you have a gap in the system, you can't have a closed loop and autocatalytic has no closed loop, right?

— Dhaval Adjodah and Owen He, a16z
Key Insight
Framing RSI as a control-theory closed loop is a quiet filter that keeps only systems where the model itself is the optimizer and rejects every pipeline whose improvements still originate outside it. That one test is what lets the founders call most of the field's 'self-improving' work autocatalysis instead.
02The Math

Divide Performance by Cost

<strong>They model progress as a rate, performance over cost, so a fixed top line plateaus while a top line that scales with capability compounds into takeoff.</strong>

the first derivative of capability with respect to time is the thing that everybody is trying to improve

— Dhaval Adjodah and Owen He, a16z
Key Insight
Putting cost in the denominator reframes the whole race: a lab that posts bigger numbers only by raising and burning more money is improving the numerator while quietly inflating the denominator, so its real rate of improvement may be flat. The equation is as much an accounting discipline as a physics model.
03Epistemology

Not a Belief, a Measurement

<strong>They refuse to predict takeoff because the loop has no closed-form answer: you can write the differential equation but still have to measure its exponent empirically.</strong>

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.

— Dhaval Adjodah and Owen He, a16z
Key Insight
Their insistence that "we don't even have a learning theory of deep learning" cuts against both the doomers and the dismissers: nobody can prove takeoff or rule it out from an armchair, so confident predictions in either direction are a tell that someone skipped the measurement.
04Discipline

The Viceroy of Variance

<strong>Their hardest work is measuring honestly, so a dedicated engineer is paid to destroy their own results and find what is wrong before anyone celebrates.</strong>

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

— Dhaval Adjodah and Owen He, a16z
Key Insight
Naming an internal adversary (the viceroy of variance) and a cost ledger (the bleederboard) turns skepticism into an org chart. It is a bet that in a field drowning in cherry-picked RSI claims, the moat is not a secret algorithm but the discipline to tell signal from noise.
05Approach

Optimize Every Layer or It Becomes the Bottleneck

<strong>They optimize the whole stack, prompts through harness, weights, and eventually chips, because whatever layer you leave alone becomes the system's next bottleneck.</strong>

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.

— Dhaval Adjodah and Owen He, a16z
Key Insight
The demand for one framework over 200 ad-hoc systems is really a bet about scaling: local tricks that win a benchmark often fail to compose, so collapsing them into a handful of abstractions is what lets the gains survive contact with real, large-scale problems.
06Design Decision

Humans Are a Kind of Agent

<strong>The whole company is the multi-agent system: human researchers and engineers are modeled as one more kind of agent, so the people are recursively improved alongside the software.</strong>

human researchers and engineers are actually special case of these agents and it's this hybrid system of multiple agents working together that's been recursively improved.

— Dhaval Adjodah and Owen He, a16z
Key Insight
Keeping humans inside the loop is a hedge against their own thesis: if some gains genuinely require human judgment that agents cannot yet supply, a design that treats people as agents keeps improving even where full automation stalls — the opposite of the replace-humans framing most labs use.
07Key Bet

Why They Started With Inference

<strong>They attacked inference first because it sits closest to AI itself and is the rare target that lifts both performance and cost at the same time.</strong>

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

— Dhaval Adjodah and Owen He, a16z
Key Insight
Choosing inference is the numerator-over-denominator thesis applied to themselves: it is one of the few targets that is both recursive (closer to AI itself) and economically urgent right now, since everyone wants to serve the latest open-weight model cheaply. Even their master prompt, they say, has been improving itself.
08Where It Stops

The Slowest Component Wins

<strong>Intelligence is a vector, not a scalar, so RSI takes off where feedback is fast and cheap and crawls wherever the world still forces you to move atoms.</strong>

let's say drug discovery you still need to do wet lab experiments and then clinical trials and that take months then every new idea have to go through a months or multiple months uh evaluation

— Dhaval Adjodah and Owen He, a16z
Key Insight
The real brake on takeoff is latency, not intelligence: a digital optimizer can run a million times faster and still wait months on a wet-lab assay or a clinical trial, so the frontier advances domain by domain — coding and math first, anything that moves atoms far behind — rather than all at once. Feedback time is the binding constraint long before any physical ceiling is.