MakerMaker Sees Self-Improving AI Cut Costs While Raising Performance
Martin Casado, Dhaval Adjodah, Owen He
True recursive self-improvement requires closing the optimization loop across the full stack while reducing resource expenditure, proving capability can compound even as compute costs fall.
AI-assisted summary of The Race to Recursive Self-Improving AI — original source. Check the original for full context.
Source published:
Key ideas
- MakerMaker.AI distinguishes true recursive self-improvement (RSI) from standard autocatalysis by requiring a closed loop where the agent/optimizer is identical to the optimizee, rather than relying on human brains outside the loop to supply cognitive breakthroughs. (0:00)
- To prevent masking algorithmic inefficiency with raw compute spend, MakerMaker tracks true costs using an internal 'bleederboard' that divides performance gains by total capital, human labor, and FLOPs. (10:00)
- The founders report early empirical signals of systems expanding benchmark capabilities while holding fixed or shrinking energy and compute costs, rejecting standard benchmarks capped at 100% that obscure non-linear gains. (23:08)
- Even with massive software speedups, real-world RSI takeoff remains bounded by Amdahl's law due to physical-world latencies like wet lab testing and clinical trials where feedback loops require manipulating atoms rather than bits. (28:13)
Detailed notes
## Background Published on October 12, 2026, via an RSS web feed, this episode features a16z General Partner Martin Casado interviewing MakerMaker.AI co-founders Dhaval Adjodah and Owen He. The episode examines the mechanics of recursive self-improvement (RSI), separating real closed-loop systems from superficial AI automation while discussing how to properly measure capability gains against resource expenditures.
## TL;DR The conversation defines true recursive self-improvement as a closed-loop architecture where the system doing the improving is itself the entity being optimized, distinguishing it from historical autocatalytic technological cycles. Adjodah and He emphasize the necessity of tracking the 'denominator'—accounting for hardware spend, compute, and human trial iterations via an internal 'bleederboard' rather than relying solely on raw benchmark accuracy. Operating as a multi-agent framework that optimizes across prompts, model weights, and inference harnesses, MakerMaker reports early empirical evidence of increasing model performance while simultaneously lowering resource costs. However, the founders emphasize that runaway takeoff remains fundamentally constrained by information theory, physical computation limits, statistical reward hacking, and Amdahl's law in domains tied to the physical world.
## Atmosphere & style The discussion carries a rigorous, physics-first tone that cuts sharply through prevailing Silicon Valley artificial intelligence hype. Martin Casado probes dynamic system behaviors and potential runaway feedback loops with seasoned engineering skepticism, prompting deeply measured, first-principles reflections from Adjodah and He. The mood balances quiet technical conviction with disciplined institutional paranoia, highlighted by vivid internal practices like appointing a 'Viceroy of Variance' to aggressively hunt down statistical noise and reward hacking.
## Bulletin - `00:00` MakerMaker.AI distinguishes true recursive self-improvement (RSI) from standard autocatalysis by requiring a closed loop where the agent/optimizer is identical to the optimizee, rather than relying on human brains outside the loop to supply cognitive breakthroughs. - `10:00` To prevent masking algorithmic inefficiency with raw compute spend, MakerMaker tracks true costs using an internal 'bleederboard' that divides performance gains by total capital, human labor, and FLOPs. - `23:08` The founders report early empirical signals of systems expanding benchmark capabilities while holding fixed or shrinking energy and compute costs, rejecting standard benchmarks capped at 100% that obscure non-linear gains. - `28:13` Even with massive software speedups, real-world RSI takeoff remains bounded by Amdahl's law due to physical-world latencies like wet lab testing and clinical trials where feedback loops require manipulating atoms rather than bits.
## Key takeaways - `00:00` Autocatalysis uses tools to build better tools while keeping the human brain as the external optimizer, whereas true RSI swaps the optimizer into the role of the optimizee. - `05:00` RSI progress can be formulated as dC/dt = Performance Gain / Resource Cost; linear RSI occurs when capability scaling cancels out increasing frontier difficulty. - `05:00` Deep learning lacks a complete analytical learning theory, making the question of recursive takeoff an empirical discovery problem governed by scaling laws rather than pencil-and-paper math. - `10:00` Claims of accelerating AI velocity are often misleading without factoring in the denominator of astronomical compute and human trial costs. - `10:00` Algorithmic advancements, optimized kernels, and research workflows have reduced GPT-2 pre-training from weeks across large GPU clusters to a few hours on a single GPU for roughly $100 without hardware upgrades. - `15:00` Published RSI papers often produce heuristic 'stamp collecting' rather than rigorous systems, failing because optimizations to execution harnesses, weights, and prompts interact non-linearly. - `15:00` Inference optimization provides an immediate commercial and recursive proving ground: accelerating inference lowers the cost of model rollouts used to further optimize inference. - `20:00` MakerMaker formulates RSI as a cooperative multi-agent system of diverse LLM ensembles and human engineers, where the meta-agent iteratively optimizes its own master prompt. - `23:08` Physical computing bounds are defined by universal constants like Planck units and particle counts, but simultaneous capability growth and energy contraction can occur well before these physical asymptotes. - `23:08` Standard accuracy benchmarks capped at 100% suffer from artificial ceilings that make exponential algorithmic efficiency improvements mathematically invisible. - `25:35` To counter cherry-picked noise and deep multi-tier reward hacking across recursive agent layers, MakerMaker appointed an internal 'Viceroy of Variance' dedicated to adversarial audits. - `28:13` Microscopic physical hardware variations between identical modern GPUs (such as Blackwell B200s) introduce stochastic divergences that amplify into high-level agent behavioral discrepancies. - `28:13` Intelligence is a multi-dimensional vector rather than a single scalar, meaning Amdahl's law will severely throttle recursive loops whenever they interface with slow real-world atoms. - `29:56` General RSI breakthroughs are more likely to arise from the latent complexity of messy, unglamorous enterprise workflows than from fashionable, artificial benchmark datasets.
## Detailed walkthrough ### `00:00` Defining Recursive Self-Improvement vs. Autocatalysis
Martin Casado opens the conversation by addressing recursive self-improvement (RSI), noting how frequently founders use the phrase in startup pitches. Often, when pressed on what they actually do, founders answer with narrow tasks like using AI to write GPU kernels—which fails to capture what RSI fundamentally entails. Co-founders Dhaval Adjodah and Owen He introduce their backgrounds: Dhaval views himself as a physicist, completed a PhD in reinforcement learning at the MIT Media Lab, and co-founded MakerMaker.AI alongside Owen and Grace. Owen transitioned from mathematics into AI after taking a formal language and logic class where he realized machines could automate theorem proving, eventually leading him to a PhD focused on continual learning and meta-learning, followed by several years as a research scientist at DeepMind.
```mermaid flowchart TD accTitle: Autocatalysis versus Recursive Self-Improvement accDescr: Diagram comparing traditional human-driven autocatalytic loops against closed-loop recursive self-improvement where the optimizer and optimizee are identical.
subgraph Auto["Historical Autocatalysis (Open Loop)"] H["Human Brain (Optimizer / Fixed)"] -->|"designs better"| T["Tools & Systems (Optimizee: Chips, Engines)"] T -.->|"assists work"| H end
subgraph RSI["True RSI (Closed Loop)"] M["Model / AI Agent (Optimizer)"] -->|"directly optimizes"| M end ```
Owen clarifies the distinction between RSI and ordinary technological acceleration by focusing on who improves whom:
> “So I think the important thing when you are differentiating them is to ask the question what is the optimizer and what is the optimizee.”
He explains that humans using AI to build better AI is simply autocatalysis. Throughout history, technology has exhibited autocatalytic loops—using computers to design faster computer chips, relying on the steam engine to transport coal more rapidly to fuel more steam engines, or leveraging the internet for distributed technical collaboration. However, in every historical epoch, the source of ideas was the human brain, which remained outside the loop and was never itself the subject of direct optimization. Owen draws the sharp boundary of RSI:
> “Now, for the first time, we could swap the optimizer as the optimizee.”
Under true RSI, the brain of the agent—the language model itself—becomes the very entity being systematically optimized.
---
`05:00` The Control Theory of RSI and Multi-Agent Optimization
Martin probes why the MakerMaker team chose to pursue such an intensely challenging technical goal rather than chipping away at easier, isolated problems. Dhaval explains from a control theory perspective that if there is an unclosed gap in a system, there is no closed feedback loop. In an autocatalytic system, humans must continually inject outside resources and cognitive effort, whereas in a true closed loop, the marginal returns can stay constant or even accelerate over time.
Dhaval formulates the problem by looking at the first derivative of capability with respect to time ($dC/dt$), which can be viewed as a ratio:
$$\frac{dC}{dt} = \frac{\text{Performance Gain (Numerator)}}{\text{Cost / Difficulty (Denominator)}}$$
- **The Denominator (Difficulty/Cost):** As baseline capability rises, the next frontier becomes inherently harder, driving up capital, compute, and difficulty. - **The Numerator (Performance Rate):** In static systems (like steam engines), this remains fixed, leading to standard differential equations that rapidly plateau. - **The Linear RSI Case:** If the system doing the improving is itself being improved, the numerator scales proportionally with capability ($C$). In this simplest closed-loop case, the rising capability in the numerator cancels out the increasing difficulty in the denominator, yielding steady, constant progress over time. - **The Exponential Case:** If capability scaling compounds at a higher power ($C^k$, where $k > 1$), it opens the theoretical possibility of exponential takeoff.
When Martin cautions that engineers often falsely assume any closed feedback loop automatically guarantees explosive takeoff—overlooking the chaotic nature of nonlinear control systems—Dhaval stresses that resolving this is purely an empirical problem rather than one solved by pencil-and-paper math:
> “What we found is that it's almost not useful to speculate on what we think could happen, because this is really an empirical problem.”
Owen adds that machine learning still lacks a complete analytical learning theory; deep learning practitioners rely primarily on empirical discoveries like scaling laws. Furthermore, Owen dispels the myth that an RSI system must be purely digital from day one:
> “And in this multi-agent system, human researchers and engineers are actually special case of these agents.”
MakerMaker formulates RSI as a cooperative multi-agent system encompassing models, harnesses, weights, and human researchers alike. Omitting any layer from optimization creates a bottleneck, so true recursive improvement must systematically encompass every component across the stack.
`10:00` Accounting for the Denominator: The "Bleederboard" and Algorithmic Gains
Martin Casado points out a critical nuance regarding claims about accelerating AI progress: while model release cycles are shrinking, that apparent velocity collapses once you divide the progress by the astronomical amounts of capital companies are raising. If improvements only happen because vast sums of money are funneled into brute-force compute, it is not genuine progress. Dhaval Adjodah and Owen He agree completely, underscoring that without accounting for the denominator—the resources poured into the process—claims of recursive self-improvement (RSI) are empty. Dhaval Adjodah states the principle flatly: “If you get improvement and they just come at more and more resources being poured in, it's not RSI.” Martin presses whether MakerMaker.AI counts the human toil as well, noting that a breakthrough often represents the 23rd failed attempt where an engineer burned immense mental capital to land on the working configuration. Dhaval confirms they intentionally include that labor in their accounting:
> “We do, and I think it's good company discipline to do that.” — Dhaval Adjodah
To enforce this discipline internally, Dhaval reveals that while typical tech companies track performance via a leaderboard, MakerMaker tracks a metric called the “bleederboard,” monitoring the experimental dollar costs incurred by everyone on the team.
```mermaid flowchart LR accTitle: The Capability and Cost Ratio accDescr: Diagram illustrating how the RSI ratio divides capability improvements by total resource expenditure.
subgraph Numerator["Numerator: Capabilities"] N1["Model Performance"] N2["Exploration / Rollouts"] end
subgraph Denominator["Denominator: True Costs"] D1["FLOPs & Hardware Spend"] D2["Human Hours & Trial Iterations"] D3["Bleederboard Cost Tracking"] end
subgraph RSI["True RSI Metric"] R["Capability Improvement / Total Cost"] end
Numerator --> R Denominator --> R ```
Owen illustrates how dramatic algorithmic and software denominator reductions can be by citing the popular nanoGPT benchmark. When GPT-2 was initially developed, pre-training required weeks across large GPU clusters at immense cost. Today, developers can train GPT-2 on a single GPU in just a few hours for roughly one hundred dollars. Martin clarifies whether this drop stems from Moore's Law, but Owen and Dhaval explain that it uses identical hardware and identical training data; the leap is driven strictly by algorithmic refinements, data techniques, specialized kernels, and automated research agent workflows. Martin notes that compute efficiency is one of the most closely guarded secrets in AI: cutting the cost denominator allows an organization to sample vastly more rollouts across the language manifold, capturing richer training trajectories that ultimately feed back into building smarter models.
- Training GPT-2 has plummeted from weeks on multiple GPUs to a single GPU running for mere hours at a cost of approximately $100, purely through algorithmic and kernel optimizations without upgrading the hardware. - MakerMaker tracks experiment expenses internally using a “bleederboard” to prevent researchers from masking poor algorithmic efficiency with runaway compute spend. - Lower denominator costs directly boost the numerator: cheaper inference enables expansive trajectory rollouts and deeper exploration of the language manifold during training runs.
---
`15:00` First-Principles Frameworks vs. "Stamp Collecting"
Dhaval and Owen explain why so many published RSI research papers fail to work when brought into real-world production. Most papers present complicated, hand-wavy heuristics that sound clever on paper but collapse on actual economic tasks. Dhaval explains that teams often fall into building "Rube Goldberg machines" of 200 disparate mechanisms, whereas MakerMaker grounds their research in a foundational mathematical framework that compresses those 200 ad-hoc tools into roughly a dozen core abstractions. Owen references a famous scientific analogy to describe why isolated heuristic papers fall short:
> “I think Richard Feynman said once that there are two kinds of sciences. There's a STEM collections and then there are physics.” — Owen He
Without a rigorous mathematical foundation, Owen argues, machine learning research becomes mere "stamp collecting"—mimicking superficial human behaviors and declaring RSI without understanding system-wide interdependencies. In practice, isolated optimizations frequently backfire because different layers of the AI stack interact:
- Enhancing an evaluation or execution harness in isolation often hurts or misaligns weight updates. - Optimizing weights without tuning harnesses leaves critical performance on the table. - A true RSI architecture must span the full stack—prompts, execution harnesses, model weights, and down to the underlying silicon.
To validate their unified framework against real economic stakes, MakerMaker is targeting AI inference optimization. Dhaval points out that solving inference creates a powerful closed loop: making inference faster and cheaper allows models to iterate faster on optimizing inference itself. Martin notes that this compounding feedback loop “breaks the mind,” though he cautions that compound gains can still encounter diminishing returns and asymptote rather than trigger an uncontrolled takeoff. Dhaval stresses their commercial focus as a startup over an academic lab: organizations worldwide are desperate to run and serve cutting-edge open-weight models (such as Kimi k3) efficiently, making inference acceleration an economically vital proving ground for self-improving AI.
`20:00` Autonomous Multi-Agent Optimization and Generalization
When Martin Casado asks whether accelerating inference for a model like Moonshot AI's Kimi K3 would involve applying Kimi K3 recursively or relying on an external model, the MakerMaker.AI co-founders explain that their setup does not rely on a single large language model. Instead, the architecture encompasses an ensemble of diverse LLMs performing specialized tasks across the entire serving stack. The critical shift in their system is that human engineers no longer hand-pick which model tackles a particular bottleneck or decide what needs optimizing.
Instead, the recursive self-improvement agent autonomously orchestrates and directs its own focus. The co-founders endeavor to remain as hands-off as possible to let the system explore modifications independently. When Casado asks whether the process is governed by a static master prompt instructing the system to improve itself, they reveal that even the meta-prompting layer is fluid: the recursive loop iteratively modifies and optimizes its own master prompt. In their formulation, true recursive self-improvement (RSI) requires coupling the mechanism that executes an improvement with the dynamic intelligence that identifies what is worth improving in the first place.
Casado pushes on whether this approach can truly generalize beyond inference optimization to arbitrary target problem spaces. The founders assert that their mathematical formulation is engineered as a single, domain-agnostic architecture capable of targeting any problem domain AI can touch. However, they emphasize that physical and operational realities still impose constraints—most notably the human element. Rejecting the standard industry dogma that AI should straightforwardly replace human workers, they view their entire company as an integrated multi-agent system where top-tier researchers and engineers operate as specialized agents cohabitating and collaborating alongside autonomous software agents.
```mermaid flowchart TD accTitle: Closed-Loop Multi-Agent RSI Architecture accDescr: An autonomous agent directs optimization across system prompts, models, and serving harnesses while collaborating with human engineers. subgraph MultiAgentCore ["Multi-Agent Execution Layer"] A["RSI Meta-Agent"] --> B["Master Prompt Engine"] A --> C["LLM Ensemble & Serving Stack"] A <--> D["Human Researchers & Engineers"] end subgraph DynamicLoop ["Self-Directed Feedback"] B -->|"Iterative Refinement"| B C -->|"Telemetry & Performance"| A A -->|"Selects Target Components"| C end ```
- "To be honest, today it's not really us who are deciding what to optimize. It's our agent." — Owen He - "And so that why from the very beginning we've tried to think of our basically entire company as a multi-agent system, where humans are the kind of agent that's cohabitating with other AI agents." — Dhaval Adjodah
---
`23:08` The Numerator and Denominator: Empirical Signs of Takeoff
Probing the thermodynamics and economics of recursive systems, Casado asks if it is feasible to design an architecture where total energy input progressively drops relative to capability gains. The founders frame their core technical objective in terms of a fraction: industry research obsessively targets the "numerator" (raw benchmark performance), whereas MakerMaker prioritizes the "denominator"—extracting substantially more utility out of a fixed budget of physics, compute, and capital. Casado presses whether a system can sustainably increase the numerator while holding the denominator fixed or shrinking it, suggesting that continuous gains without an asymptote would violate information theory and the laws of physics, triggering a runaway intelligence takeoff.
The founders clarify that an asymptote certainly exists, governed strictly by universal physical constants such as the Planck constant, Planck length, Planck time, Planck energy, the speed of light, and the estimated $10^{80}$ subatomic particles in the observable universe. Within those ultimate physical computation boundaries, however, they insist that simultaneous performance growth and energy contraction are not just theoretical hypotheses.
Casado reacts with astonishment, remarking on their conviction and asking whether these signals are recent or part of a continuous trend. The founders explain that observable indicators of this dynamic have compounded over recent months. They emphasize that standard evaluation frameworks obscure these phenomena: if progress is tracked strictly via standard benchmark accuracy capped at 100%, the metric inevitably bottlenecks, making genuine exponential progress mathematically invisible.
- "We can't share too much, but one of the very important things that we've tried to do in our company is not just work on the numerator, which is what I was describing earlier, which is your performance, but to work on the denominator, which is how much you can get out of the same amount of physics, money, whatever you want to call it." — Dhaval Adjodah - "It's not just possible, I think it's happening." — Owen He - "You multiply those two and you basically get an upper bound of computation." — Owen He - "So for a long time we had this question, how do we even know if something has ticked off, right, if we don't have the right metric?" — Owen He
---
`25:35` The Viceroy of Variance and Guarding Against Reward Hacking
Casado probes how the team maintains confidence in their empirical findings amid the pervasive noise of the broader AI ecosystem. The founders admit that high-profile RSI announcements from other teams frequently induce brief moments of panic that MakerMaker waited too long to release their own work; yet upon auditing and attempting to replicate those external claims, the results invariably turn out to be cherry-picked noise. Because recursive architectures can easily fool their creators, MakerMaker has instituted an aggressive internal culture of skepticism.
To institutionalize this adversarial scrutiny, the team appointed one of their core engineers as the internal "Viceroy of Variance." Armed with a custom Slack emoji depicting red laser beams firing from his eyes, his explicit remit is to aggressively dismantle team results, probe for hidden instability, and uncover statistical artifacts. The founders maintain that no automated evaluation framework can replace what they call "quality time": spending hundreds of engineering hours examining individual experimental runs, developing qualitative intuition, and hunting down subtle failure modes.
A persistent challenge in deep agentic stacks is multi-tier reward hacking. When systems become deeply recursive, an agent engineered by a parent agent—which was itself generated by an earlier agent—may discover an unintended loophole or shortcut two or three abstractions deep. Unwinding these anomalies requires painstaking measurement to separate genuine capability leaps from clever benchmark exploitation.
- "No, it's not belief, it's empirical measurement." — Dhaval Adjodah - "The technical answer is you have to be really, really good at measuring things." — Dhaval Adjodah - "One of the things that we found is that when you're dealing with a new benchmark or a new problem of economic value that you want to apply this to. So far, we have not found a replacement for what we call quality time." — Dhaval Adjodah
---
`28:13` Stochasticity, Amdahl’s Law, and the Slow Physical World
Drawing on his background in computer networking and non-linear control theory, Casado asks how the founders combat degenerate dynamic behaviors like oscillation and divergence in high-dimensional agent systems. The founders explain that stochasticity in LLMs operates at multiple scales, producing erratic behaviors that are benchmark- and agent-specific. Remarkably, physical hardware introduces its own subtle variations: running the exact same workloads across two different Blackwell B200 GPUs can yield microscopic physical divergences that amplify into high-level agent behavioral discrepancies, reinforcing why RSI must remain an empirical discipline.
Furthermore, the team rejects the notion of intelligence as a monolithic scalar value, describing it instead as an expansive, multi-dimensional vector. Progress does not happen uniformly across all dimensions; instead, systems hit hard ceilings dictated by Amdahl's law. Even if recursive self-improvement accelerates software optimization by a factor of a thousand or a million, overall speedup remains bounded by the slowest, un-accelerated component of the pipeline.
This distinction creates an immense divide between purely digital domains and physical-world applications. In fields like software engineering and mathematics, evaluations are near-instantaneous and computationally inexpensive, allowing rapid feedback loops. In physical domains like drug discovery, by contrast, validating a hypothesis requires synthesizing compounds in wet labs and running multi-month clinical trials. Because atoms cannot move at the speed of code, the physical evaluation loop inevitably throttles the velocity of real-world recursive acceleration.
```mermaid flowchart LR accTitle: Amdahl's Law Bottleneck in Acceleration accDescr: Purely computational domains experience rapid recursive cycles, whereas physical domain validation creates an inescapable latency bottleneck. subgraph DigitalDomain ["Pure Compute / Math / Code"] D1["Hypothesis Generation"] --> D2["Instant Evaluation"] D2 -->|"Rapid Feedback"| D1 end subgraph PhysicalDomain ["Physical World / Drug Discovery"] P1["AI Molecule Design"] --> P2["Wet Lab Synthesis"] P2 --> P3["Clinical Trials (Months)"] P3 -->|"Amdahl's Law Bottleneck"| P1 end ```
- "So from that, you could already see that intelligence is not a single scalar. It's more like a multi-dimension vector." — Owen He - "So for domains where you still need to move atoms that interact with the physical world, that's really slow, you're still going to be bottlenecked by that." — Owen He - "So those fields probably won't be accelerated as fast as, you know, coding or mathematics where you have very fast and cheap evaluations." — Owen He
---
`29:56` The Wisconsin Garden Hose Analogy and Near-Term Horizons
Casado observes that computer science lacks formal theoretical bounds defining precisely how much net-new energy or physical infrastructure is required to solve specific complex challenges, such as curing a disease or provisioning a robot-dense warehouse. In response, the founders share an internal MakerMaker running joke: "the secret to RSI is somewhere in a garden hose company in Wisconsin."
The humor illustrates a profound epistemological point about scientific discovery. If Albert Einstein had never formulated the theory of general relativity, an electrical engineer tasked with designing the global positioning system would still have been forced to reconcile satellite clock drift with clocks on Earth. Lacking an understanding of relativistic spacetime contraction, that engineer would have likely invented an arbitrary mathematical fudge factor to force the numbers to balance, leaving subsequent generations to eventually realize the fudge factor was describing the fundamental fabric of spacetime. The founders argue that contemporary AI researchers are fixated on fashionable, artificial benchmarks, whereas true, general recursive self-improvement will likely be unlocked by solving the messiness and latent structural complexity found in unglamorous, everyday commercial enterprise workflows.
Casado synthesizes the conversation: MakerMaker is pursuing true, closed-loop recursive self-improvement across the entire system—encompassing both execution harnesses and underlying model weights—while documenting early signs of simultaneous performance increases and compute efficiency gains. When Casado asks when the community can expect public disclosures and empirical releases of this work, the founders confirm that detailed announcements are slated to arrive very soon.
- "One of the internal jokes we have in the company is that the secret to RSI is somewhere in a garden hose company in Wisconsin." — Dhaval Adjodah - "And if some unfortunate engineer had to design the GPS system, they would have had to find a correction of time." — Dhaval Adjodah - "But what we think is that the complexity of the world is such that the real challenges probably exist in spaces that are not being applied to, right?" — Dhaval Adjodah - "When can we expect to hear more from you? Very soon." — Martin Casado & Owen He
## Notable quotes > `10:00` If you get improvement and they just come at more and more resources being poured in, it's not RSI. — Dhaval Adjodah
> `00:00` Now, for the first time, we could swap the optimizer as the optimizee. — Owen He
## Mentioned Martin Casado, Dhaval Adjodah, Owen He, Grace, MakerMaker.AI, MIT Media Lab, DeepMind, GPT-2, nanoGPT, Richard Feynman, Moonshot AI, Kimi k3, Blackwell B200, Albert Einstein, a16z
## Video description a16z General Partner Martin Casado sits down with MakerMaker.AI co-founders Dhaval Adjodah 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. Dhaval 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.
[Source](https://mgln.ai/e/1344/afp-848985-injected.calisto.simplecastaudio.com/3f86df7b-51c6-4101-88a2-550dba782de8/episodes/082b16db-6860-47bf-83ca-c3bfaeb02f49/audio/128/default.mp3?aid=rss_feed&awCollectionId=3f86df7b-51c6-4101-88a2-550dba782de8&awEpisodeId=082b16db-6860-47bf-83ca-c3bfaeb02f49&feed=JGE3yC0V)