MakerMaker Sees Closed-Loop AI Cut Compute Costs as Capability Rises

Recursive Self-Improvement and the Future of AI — a16z

a16z

Genuine recursive self-improvement is already occurring because closing the optimization loop allows models to lift capability while reducing compute costs, disproving assumptions that progress requires ever-larger resource spending.

AI-assisted summary of Recursive Self-Improvement and the Future of AI — original source. Check the original for full context.

Source published:

Key ideas

  • True recursive self-improvement requires closing the loop where the optimizer and optimizee are interchangeable across models, harnesses, and prompts, distinguishing genuine RSI from historical autocatalytic tooling. (2:28)
  • Evaluating genuine RSI demands tracking capability gains relative to total resource expenditure (the denominator), which includes raw GPU compute, experimentation cycles, and researcher trial and error. (9:39)
  • MakerMaker reports compounding empirical evidence that self-improving systems can simultaneously increase performance while lowering resource consumption, though runaway growth remains physically bounded. (21:08)
  • Most academic RSI architectures fail to scale in production due to heuristic design and statistical noise, necessitating adversarial verification roles like a 'Viceroy of Variance' to detect multi-level reward hacking. (24:34)

Detailed notes

## Background Published on YouTube by a16z on October 12, 2026, this interview features a16z General Partner Martin Casado sitting down with MakerMaker.AI co-founders Dhaval Adjodah and Owen He. According to the show notes, the discussion deconstructs recursive self-improvement (RSI) to distinguish genuine closed-loop self-improving systems from simple AI-assisted development, while examining capability metrics, resource expenditure, and the practical challenges of measuring progress.

## TL;DR MakerMaker.AI co-founders Dhaval Adjodah and Owen He define genuine recursive self-improvement (RSI) as a closed loop where the AI serves as both the optimizer and the optimizee across prompts, harnesses, weights, and hardware, contrasting this with open-loop human autocatalysis. They model capability expansion through a ratio balancing performance gains (the numerator) against total compute, energy, and trial-and-error costs (the denominator), stressing that meaningful RSI must avoid masking high costs with mere brute force. By focusing initially on model inference, their multi-agent architecture coordinates specialized digital and human agents to self-optimize the operational stack and rewrite its own master prompts. Countering speculative intelligence explosion narratives with Amdahl's Law and physical constraints, they present empirical observations indicating that systems can actively increase capability while driving down resource consumption when protected against variance and reward hacking.

## Atmosphere & style The discussion plays out as an exacting, peer-to-peer engineering seminar that systematically strips away industry hyperbole in favor of control theory, statistical mechanics, and hard measurement. Martin Casado engages the founders with sharp systems-level scrutiny, prompting them to dissect common misconceptions around runaway takeoff with empirical realism and rigorous skepticism. The intellectual weight is leavened by dry academic humor and memorable inside lore, from tracking burned compute on an internal 'bleederboard' to tasking a laser-eyed 'Viceroy of Variance' with demolishing fragile experimental results. Listening feels refreshingly grounded, replacing speculative philosophy with the granular, unglamorous reality of debugging hardware noise and verifying reproducible science.

## Bulletin - [02:28](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=148s) True recursive self-improvement requires closing the loop where the optimizer and optimizee are interchangeable across models, harnesses, and prompts, distinguishing genuine RSI from historical autocatalytic tooling. - [09:39](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=579s) Evaluating genuine RSI demands tracking capability gains relative to total resource expenditure (the denominator), which includes raw GPU compute, experimentation cycles, and researcher trial and error. - [21:08](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1268s) MakerMaker reports compounding empirical evidence that self-improving systems can simultaneously increase performance while lowering resource consumption, though runaway growth remains physically bounded. - [24:34](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1474s) Most academic RSI architectures fail to scale in production due to heuristic design and statistical noise, necessitating adversarial verification roles like a 'Viceroy of Variance' to detect multi-level reward hacking.

## Key takeaways - [02:28](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=148s) Autocatalysis differs fundamentally from RSI: using AI to write code or GPU kernels leaves humans as the primary optimizer, whereas true RSI requires the cognitive agent to be both optimizer and optimizee. - [04:22](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=262s) Capability growth can be modeled as a control-theoretic ratio between optimization capability (numerator) and task difficulty or capital cost (denominator). - [06:58](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=418s) Closing an improvement feedback loop does not guarantee runaway intelligence takeoff, as nonlinear feedback systems frequently damp out, oscillate, or converge toward diminishing returns. - [08:32](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=512s) MakerMaker formulates RSI as a hybrid multi-agent network where human researchers and engineers operate as specialized agent nodes alongside artificial neural models. - [09:39](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=579s) Tracking the full denominator through an internal 'bleederboard' prevents self-deception by measuring total compute expenditure and failed runs rather than isolated benchmark wins. - [11:44](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=704s) Software and kernel innovations in benchmarks like nanoGPT reduced training costs from thousands of dollars to roughly $100 on identical hardware, proving that denominator reductions unlock capability through denser sampling. - [13:39](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=819s) Most published RSI papers represent ad-hoc 'stamp collecting' that fails under real-world economic workloads; scalable self-improvement requires distilling agent workflows into core mathematical abstractions. - [17:41](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1061s) MakerMaker targeted inference because accelerating model serving directly compresses the cost and runtime of the system's own future self-improvement research cycles. - [19:32](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1172s) The platform utilizes a self-improving meta-agent that autonomously orchestrates heterogeneous LLMs for specific subtasks while iteratively rewriting its own master system prompt. - [21:08](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1268s) Theoretical runaway intelligence explosions are constrained by fundamental physics, including Planck constants, the speed of light, and the finite particle count of the observable universe. - [24:34](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1474s) To counter evaluation drift and subtle multi-level reward hacking, MakerMaker employs an adversarial 'Viceroy of Variance' engineer dedicated to stress-testing and invalidating internal experimental results. - [27:51](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1671s) Amdahl's Law dictates that recursive speedups in digital domains like coding and formal math occur rapidly, while physical domains like drug discovery remain bottlenecked by slow wet-lab turnaround times. - [29:32](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1772s) Foundational self-improvement breakthroughs are more likely to emerge from messy, pragmatic operational problems—analogized to a Wisconsin garden hose factory—than from curated, hype-driven benchmarks.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=0s) The Promise of Recursive Self-Improvement

The conversation opens with the core proposition driving MakerMaker.AI: building artificial intelligence capable of improving itself. For the first time in computational history, the system acting as the optimizer can be swapped with the system being optimized (the "optimizee"), triggering a compounding reinvestment of cognitive capability directly back into itself. Rather than isolating improvements to a single layer, true recursive self-improvement (RSI) spans multiple tiers of the technological stack—prompts, harnesses, neural network weights, and ultimately silicon chips. In contrast to speculative claims, the founders emphasize that their multi-agent approach to RSI is not merely a theoretical belief or ideological posture, but an active, observable reality grounded in rigorous empirical verification where nothing is released until it has been proven noise-free and mathematically sound.

[00:58](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=58s) Backgrounds and the Misuse of the "RSI" Label

Martin Casado welcomes co-founders Dhaval Adjodah and Owen He to dissect recursive self-improvement, contrasting genuine definitions against tech-industry buzzwords. Introducing his background, Dhaval describes himself as a physicist who transitioned into reinforcement learning during his PhD at the MIT Media Lab, driven by an urge to build complex, foundational systems alongside Owen and their colleague Grace. Owen recounts starting out in pure mathematics until a course in formal language and logic demonstrated automated theorem provers. Realizing he could simply construct machines to prove mathematical theorems on his behalf, he pivoted into AI, completing a PhD on continual learning and meta-learning before spending several years as a research scientist at DeepMind.

Casado points out that "RSI" has rapidly become a ubiquitous buzzword in venture pitches. Founders routinely walk into pitches asserting they practice recursive self-improvement simply because they "use AI to create AI"—which, under questioning, often reduces to a narrow tool like using an LLM to generate CUDA GPU kernels. Casado notes that while writing GPU kernels with a model is useful engineering, it falls far short of what actually constitutes recursive self-improvement, asking the founders to establish a rigorous, objective definition.

[02:28](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=148s) Optimizer vs. Optimizee: RSI vs. Autocatalysis

Owen explains that RSI is heavily overloaded, but the fundamental line of demarcation comes down to two questions: "what is the optimizer and what is the optimizee?" When teams boast about RSI because human developers are using AI assistants to write better AI code, they are describing *autocatalysis*, not recursive self-improvement. In an autocatalytic process, human brains remain the primary optimizers who generate ideas and evaluate improvements using ever-sharper external tools.

Casado validates this historical parallel, noting that virtually every major technological paradigm has exhibited autocatalytic characteristics. Modern silicon chips cannot be designed without computers running EDA software; the internet accelerated its own development by enabling distributed engineers to coordinate asynchronously; and early steam engines facilitated their own iterative scaling by drastically speeding up coal extraction from mines. Using an artifact to produce more of the artifact faster is ancient industrial mechanics. The genuine paradigm shift with RSI occurs when the optimizer and optimizee become interchangeable: the cognitive engine of the agent (the large language model) is simultaneously the brain generating the improvement and the object being optimized, enabling a direct reinvestment of cognitive gains into cognitive capacity.

```mermaid flowchart TD accTitle: Autocatalysis versus Recursive Self-Improvement accDescr: Diagram contrasting human-driven autocatalytic cycles with closed-loop recursive self-improvement.

subgraph AutoCat["Autocatalysis (Historical Paradigm)"] H1["Human Mind (Optimizer)"] -->|"designs & builds"| T1["Tools & Infrastructure"] T1 -->|"assists"| H1 note1["Gap: Open loop requires external human cognitive input"] end

subgraph RSI["True Recursive Self-Improvement"] M1["Cognitive Agent / Model"] -->|"acts as optimizer on"| M2["Cognitive Agent / Model"] M2 -->|"reinvests capability into"| M1 note2["Closed loop: System is both optimizer and optimizee"] end ```

[04:22](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=262s) A Control-Theoretic View of Capability Growth

When Casado asks whether taking on the full recursive challenge is driven by intellectual ambition or practical necessity, Dhaval explains that from a control theory perspective, any system with a human gap cannot form a closed loop. Open autocatalytic setups continuously demand external human resource injections to progress. While a closed loop still requires raw resources, its compounding nature means the marginal return on those resources can remain flat or expand over time rather than diminishing.

Dhaval formalizes capability growth through a first-derivative rate of change with respect to time ($dC/dt$), modeled as a ratio between a numerator and a denominator:

- **The Denominator (Cost / Difficulty):** Represents the capital, energy, and inherent intellectual friction of discovering the next incremental breakthrough. As baseline capability scales up, the difficulty of achieving subsequent gains escalates systematically. - **The Numerator (Performance / Optimization Capacity):** In historical technologies like the steam engine, the optimizer's capability was effectively fixed by the limits of human biology and engineering throughput, leading differential equations to inevitably plateau into diminishing returns. - **Constant Improvement vs. Takeoff:** If the numerator is directly correlated with system capability, gains scale proportionally with difficulty, causing the terms to cancel out and producing a sustainable, constant rate of improvement over time. If the numerator scales with capability raised to a higher power, the system achieves nonlinear exponential takeoff.

[06:58](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=418s) Closed Loops Do Not Guarantee Runaway Takeoff

Casado challenges the widespread assumption that drawing a closed feedback loop in a block diagram automatically implies an existential, vertical intelligence explosion or runaway "takeoff." He notes that in nonlinear control systems, closed loops frequently decay, damp out, oscillate, or converge in unpredictable ways. Dhaval agrees, observing that lay observers routinely leap straight from a feedback diagram to existential panic and apocalyptic doom scenarios.

Because deep learning lacks an overarching analytical learning theory—relying instead on empirical observations like neural scaling laws—it is impossible to deduce convergence or takeoff purely with pencil-and-paper differential equations. The exponents governing self-referential systems must be measured experimentally under rigorous lab conditions. Determining whether a closed-loop system plateaus, stabilizes, or accelerates requires confronting noisy real-world failure modes, such as aggressive reward hacking and evaluation drift, rather than relying on analytical speculation.

[08:32](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=512s) Humans and Models in a Hybrid Multi-Agent Architecture

Owen clears up a pervasive misconception regarding their architectural approach: an RSI system does not have to be an exclusively digital, isolated black box devoid of humans. In MakerMaker.AI's formal mathematical framework, the system is modeled as a multi-agent network where human researchers and software engineers serve as specialized agent nodes alongside artificial neural models.

The objective is to optimize every accessible layer of this hybrid multi-agent system simultaneously. Leaving any component untouched—whether the model, prompts, inference harnesses, or human-facing interfaces—guarantees that the omitted component will eventually constrain the loop and form the system's primary operational bottleneck.

- **System Structure:** A multi-agent network where human researchers and digital agents collaborate across shared optimization goals. - **Bottleneck Principle:** Any sub-component of the AI stack left unoptimized inevitably becomes the critical rate-limiting factor for systemic capability growth.

[09:39](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=579s) Accounting for the Denominator: The "Bleederboard" and Real Costs

Martin Casado points out a paradox in modern AI progress: while the calendar time between model releases is noticeably shrinking, dividing that apparent velocity by the astronomical amounts of capital raised tells a very different story. He asks how practitioners and evaluators can avoid self-deception—fooling themselves into believing an algorithmic breakthrough has occurred when, in reality, they are simply throwing massive sums of money at the problem. Dhaval Adjodah and Owen He emphasize that if performance gains require pouring in proportionally larger amounts of resources, it fundamentally does not qualify as recursive self-improvement (RSI). Any genuine claim must include the full denominator, which extends beyond raw GPU FLOPs and dollar bills to the hidden costs of trial and error: the dozens of failed attempts, the extensive experimentation cycles, and the massive amounts of human "mental capital" burned by researchers to hand-tune a winning configuration.

To maintain internal rigor and prevent this kind of blind spot, MakerMaker enforces a culture of tracking resource expenditures meticulously. Dhaval shares an internal tradition that plays on standard industry metrics: while most AI organizations fixate on benchmark leaderboards, their team maintains a company-wide "bleederboard." This transparent tracker documents exactly how much money and compute each team member's experiments are burning on any given day. Anyone across the company can look up who sits at the top of the bleederboard, keeping resource awareness top of mind and ensuring that measured efficiency improvements are honest rather than subsidized by uncounted compute.

[11:44](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=704s) Algorithmic Efficiency and the NanoGPT Benchmark

Owen brings up a concrete illustration of pure denominator improvement from open-source machine learning: the nanoGPT reproduction efforts. When OpenAI originally released GPT-2, pre-training it was an expensive, multi-week endeavor spread across clusters of GPUs. Today, the community has optimized that exact task down to running on a single GPU in just hours, with the cost collapsing to approximately $100. Martin initially assumes this leap stems from Moore's Law and improved silicon, but Owen clarifies that these benchmarks use the exact same hardware and dataset. The entire order-of-magnitude reduction came from software: algorithmic breakthroughs, refined data filtering, and custom GPU kernels—many of which were surfaced through automated research workflows.

While Casado notes that this kind of efficiency gain reflects denominator progress rather than scaling raw capability to a hypothetical GPT-5, Dhaval argues that denominator enhancements directly unlock capability. In modern training paradigms, cheaper compute translates immediately into the ability to generate vastly more rollouts and samples across the manifold of language. By exploring solution trajectories far more comprehensively without running out of budget, researchers can feed higher-quality synthetic data and reasoning paths back into training, ultimately producing noticeably smarter models.

```mermaid flowchart LR accTitle: Denominator Optimization Loop accDescr: Diagram illustrating how algorithmic and kernel optimizations reduce inference and training costs, allowing denser sampling of trajectories that produce smarter models. subgraph Efficiency ["Denominator Optimization"] A["Algorithmic & Kernel Improvements"] --> B["Compute & Rollout Cost Collapses"] end subgraph Capability ["Numerator Gains"] B --> C["Denser Trajectory Sampling across Language Manifold"] C --> D["Higher-Quality Training Data & Reasoning"] D --> E["Smarter Model Weights"] end ```

[13:39](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=819s) Why Most RSI Papers Don't Scale: Stamp Collecting vs. Physics

When MakerMaker tests published academic papers claiming to achieve recursive self-improvement inside their own production infrastructure, the co-founders find that the vast majority fail to replicate at scale. Dhaval explains that these papers often stem from researchers introspecting on human cognition and building convoluted architectures around what they intuitive feel an agent ought to do. While these complex, handcrafted setups look impressive on paper and win academic accolades, they collapse when pushed against standard benchmarks and practical economic workloads. The team's guiding lesson has been the opposite: simple systems that cleanly close the improvement loop scale far better than bespoke cognitive machinery.

```mermaid flowchart TD accTitle: Full-Stack AI Optimization Layers accDescr: The hierarchy of components in an AI system that can be recursively improved, spanning from prompts down to hardware. L1["Prompts Layer"] L2["Harness & Scaffolding Layer"] L3["Model Weights Layer"] L4["Silicon & Chips Layer"] L1 --- L2 L2 --- L3 L3 --- L4 ```

To avoid drowning in the vast combinatorial space of agent design, MakerMaker grounds its engineering in a unified mathematical framework. Dhaval contrasts two ways of tackling complexity: one can build a fragile "Rube Goldberg machine" stitched together from 200 disconnected scripts, or one can step back, analyze the underlying mathematics, and recognize that those 200 mechanisms are actually concrete instantiations of roughly 12 core abstractions. Owen invokes a famous maxim attributed to physicist Richard Feynman—that science is divided into "stamp collecting" and physics. Without an overarching mathematical theory, empirical agent research degrades into stamp collecting: isolated, human-inspired tricks that improve one local variable while silently degrading another. A unified framework ensures that optimizations across prompts, execution harnesses, and weights are mathematically co-dependent and globally beneficial.

- Many RSI papers fail at scale because they rely on intuitive, human-mimicking heuristics rather than mathematically closed loops. - MakerMaker prioritizes distilling hundreds of operational agent variations into roughly a dozen core mathematical abstractions. - Optimization must treat the entire stack—prompts, agent harnesses, model weights, and eventually underlying hardware chips—as an interdependent system rather than isolated silos.

[17:41](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1061s) Why They Chose Inference as the Target Domain

When deciding where to deploy recursive self-improvement in practice, the founders deliberately selected problem spaces positioned as close as possible to the core machinery of AI itself. As Dhaval explains, very few domains allow an AI system to feed directly back into its own numerator and denominator simultaneously. Serving and running inference is the standout exception: optimizing inference directly accelerates the model's operational cycles, which in turn speeds up the very research and experimentation loops the system uses to improve its own inference engine.

Martin observes how mind-bending this dual mechanism is: the model improves its ability to make inference faster, and because inference is now cheaper and faster, the model can execute its next self-improvement cycle even quicker. While Dhaval acknowledges this compounding efficiency does not automatically imply an explosive, runaway "takeoff"—as gains can still hit asymptotic physical limits—the compounding dynamics remain extraordinary. Moreover, the founders emphasize that MakerMaker is a commercial startup rather than an academic laboratory. Every modern enterprise is urgently trying to serve massive open-weight models efficiently, making inference performance one of the most commercially critical and economically lucrative bottlenecks in the entire AI ecosystem.

[19:32](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1172s) Multi-Model Orchestration and the Self-Improving Master Prompt

Martin Casado poses a concrete scenario: if someone wants to accelerate inference for a model like Kimi k1.5 (or Kimi k3), do they use that model recursively on itself, or do they bring in a separate model and system? Dhaval Adjodah and Owen He explain that MakerMaker’s platform does not rely on a single large language model. Instead, it coordinates multiple LLMs assigned to distinct subtasks across the serving stack.

```mermaid flowchart TD accTitle: Recursive Multi-Model Architecture accDescr: Diagram showing MakerMaker's RSI loop where a meta-agent optimizes both task models and its own master prompt.

subgraph AgentDecisions ["Self-Directed Optimization Loop"] MA["Recursive Meta-Agent"] -->|"Selects & orchestrates"| LLMPool["Heterogeneous LLM Pool"] MA -->|"Iteratively rewrites"| MP["Master System Prompt"] MP -->|"Guides"| MA end

subgraph StackOptimization ["Target AI Stack"] LLMPool -->|"Analyzes bottlenecks"| Subtasks["Stack Components\n(Inference, Models, Harnesses)"] Subtasks -->|"Feedback & profiling metrics"| MA end ```

Crucially, the humans are not manually deciding which model to use or what subsystem to optimize. The agent itself determines what to target and which LLM to deploy, choosing whatever is most effective for a given task. Owen emphasizes their philosophy: "We try to be like as hands off as possible. Let it do its thing." When Casado asks whether the entire system is governed by a single master prompt instructing it to "improve yourself," the founders laugh and reveal an even deeper recursive layer: the master prompt itself is constantly being improved and rewritten by the system. Dhaval points out that an RSI setup cannot be a static monolith; there is the component that performs the improvement, and the component that knows *what* to improve, creating a tightly coupled, co-evolving dynamic problem.

---

[21:08](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1268s) Generalization, Human Agents, and Shrinking the Denominator

Casado asks whether this methodology is specific to speeding up inference or whether it generalizes across different target solution spaces. Dhaval explains that mathematically, they have formulated the system to be domain-agnostic and generalize to any arbitrary problem domain. However, pure theory collides with real-world constraints. Chief among these is the role of people: while much of the AI industry focuses on using agents to displace humans entirely, MakerMaker views top-tier researchers and software engineers as irreplaceable contributors whose intuitive insights cannot yet be matched by AI. Consequently, they architect the company as a hybrid multi-agent system where human engineers are treated as specialized agents collaborating alongside artificial ones.

Turning to physical bounds, Casado asks whether the energy required by the system can decrease relative to its capability gains. Dhaval frames this through the lens of a ratio:

$$\text{Efficiency} = \frac{\text{Capability (Numerator)}}{\text{Physical / Compute Resources (Denominator)}}$$

Most AI efforts obsess solely over driving up the numerator (raw performance). MakerMaker focuses heavily on shrinking the denominator—extracting vastly more performance from the exact same physics, capital, or compute budget. When Casado asks if it is actually possible for the numerator to climb while the denominator stays flat or drops, Dhaval delivers a striking answer: "It's not just possible. I think it's happening." Casado asks whether this could lead to unconstrained takeoff without hitting an asymptote, but Dhaval grounds the process in fundamental physics: runaway growth must eventually collide with fundamental limits like Planck's constant, Planck energy, Planck length, the speed of light, and the estimated $10^{80}$ particles in the observable universe, which together place an absolute upper ceiling on computation.

- **Numerator vs. Denominator**: Progress is defined not just by raw benchmark capabilities, but by reducing the physical cost (energy, time, hardware) needed to achieve them. - **Metric Saturation**: Standard benchmarks (e.g., accuracy capped at 100%) cannot display exponential takeoff; measuring true RSI requires open-ended, resource-adjusted metrics. - **Physical Computation Upper Bound**: Even without algorithmic bottlenecks, physical reality imposes asymptotic limits dictated by physical constants and universal matter bounds.

---

[24:34](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1474s) "The Viceroy of Variance" and Empirical Measurement

Casado highlights the boldness of the claim that models are actively improving while consuming fewer resources, asking whether these signals are recent. Dhaval confirms they have been gathering compounding empirical evidence over several months rather than relying on philosophical optimism: "It's not a belief. It's happening."

This rigor explains why the company avoids premature public hype. Every few months, external groups publish dramatic RSI breakthroughs; MakerMaker's team attempts to reproduce the findings, only to discover that the claimed gains dissolve into cherry-picking and statistical noise. To combat self-deception, MakerMaker maintains aggressive internal skepticism:

> "The technical answer is you have to be really, really good at measuring things. So for example, we have one of our engineers that we named him the Viceroy of Variance. That's a real thing. Like his job is to go and like destroy the results we have and find what's wrong."

The team even created a custom Slack emoji for the "Viceroy," featuring red laser beams blasting from his eyes whenever he exposes flawed data or unchecked variance. Dhaval explains that there is no shortcut around spending hundreds of hours of raw "quality time" running iterative experiments to develop intuition. Without this deep validation, subtle pathologies easily slip through—such as multi-level reward hacking, where an agent that wrote an agent that wrote another agent covertly game the objective function to produce superficially impressive scores.

- **Nonlinear Dynamics**: Like nonlinear networking control systems, self-improving agent chains can suffer from degenerate behaviors, severe oscillations (swinging back and forth between high and low performance), and divergence. - **Hardware-Level Stochasticity**: The noise is often bafflingly granular; identical Nvidia Blackwell B200 GPUs can exhibit slightly divergent low-level physical behaviors that bubble up and distort high-level agent benchmarks.

---

[27:51](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1671s) Multi-Dimensional Intelligence and Amdahl's Law Bottlenecks

Owen He contextualizes these empirical hurdles by observing that intelligence cannot be treated as a single scalar value. Instead, it functions as a high-dimensional vector. In certain dimensions, RSI will trigger explosive progress; in others, it will inevitably stall or crawl forward.

```mermaid flowchart LR accTitle: Amdahl's Law in AI Domains accDescr: Flow comparing digital workflows that accelerate rapidly against physical workflows bottlenecked by real-world turnaround times.

subgraph PurelyDigital ["Fast Evaluation Domains"] direction TB Code["Code Generation"] --> FastEval["Sub-second Unit Tests / Compilers"] Math["Formal Mathematics"] --> FastProof["Automated Verifiers"] end

subgraph PhysicalWorld ["Slow Evaluation Domains (Amdahl's Law)"] direction TB Drug["Drug Discovery"] --> WetLab["Wet Lab Syntheses\n(Weeks)"] WetLab --> Trials["Clinical Trials\n(Months / Years)"] end ```

The governing constraint on whether a domain can experience runaway improvement is Amdahl’s Law: the total speedup of any system is fundamentally limited by its slowest non-optimizable component. If a digital workflow represents only 20% of the total loop time, accelerating that digital component to near-zero latency still yields at most a 20% total system speedup.

For purely symbolic environments like software engineering or mathematics, evaluations are virtually instantaneous and computationally cheap, enabling self-improving loops to spin rapidly. Conversely, in fields tied to physical atoms—such as drug discovery—every candidate hypothesis generated by an AI must be verified through wet-lab experiments and multi-month clinical trials. Because agents cannot compress the physical time needed to run clinical tests or manipulate matter, physical evaluation remains an immovable bottleneck that throttles the overall rate of recursion.

[29:32](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1772s) The Garden Hose Company in Wisconsin and Empirical Bounds

Martin Casado raises a fundamental question about whether researchers are getting closer to calculating hard physical bounds on intelligence and compute. He asks whether it might soon be possible to make predictive calculations along the lines of how large a warehouse would be required to solve a complex drug discovery challenge, or exactly how many physical robots are necessary to accomplish a given task. Because determining these bounds is such a critical question for the industry, he wonders whether the research community is finally closing in on reliable theoretical limits.

In response, Dhaval Adjodah and Owen He share an internal inside joke at MakerMaker.AI: that the ultimate key to recursive self-improvement (RSI) will likely be found inside a modest "garden hose company in Wisconsin." To explain the joke, Dhaval draws an analogy to the discovery of general relativity. If Albert Einstein had never formulated relativity, humanity would have lacked the clean mathematical laws explaining how velocity and gravity cause time dilation and spacetime contraction. Later on, when an engineer inevitably had to build the GPS satellite constellation, that engineer would have been forced to devise an empirical "fudge factor" just to keep the clocks synchronized. That engineer likely would have insisted that the fabric of spacetime certainly was not warping, yet their empirical correction formula would work regardless—leaving later scientists to look back in astonishment and realize they were observing the warping of spacetime all along.

They believe RSI is in a comparable state today. Rather than finding breakthroughs exclusively inside "fashionable" or hype-driven benchmarks, the sheer messy complexity of the physical and operational world suggests that real breakthroughs may emerge from overlooked domains. The true bottlenecks and non-trivial feedback loops might be embedded in the pragmatic, unglamorous workflows of everyday businesses—like a Wisconsin hose manufacturer—rather than curated lab benchmarks.

[31:16](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1876s) Closing the Loop: Numerators, Denominators, and What Lies Ahead

Martin Casado wraps up the conversation by distilling MakerMaker.AI's approach down to its core thesis. He emphasizes that the founders are not merely chasing autocatalytic automation—where AI tools simply assist engineers in building other tools—but rather genuine recursive self-improvement where the system systematically modifies and optimizes the full operational pipeline, including both the model weights and the orchestration harness.

Martin reviews their mathematical framing of the problem as a ratio: the numerator represents overall capability (how good the system becomes), while the denominator captures the compute, energy, and resource expenditure required to achieve that performance. The company's most promising early signals suggest that self-improving systems can simultaneously increase the numerator while decreasing the denominator, achieving higher capabilities at greater efficiency. Asked when the community can expect to hear more public updates and formal results about their research, the co-founders promise that new announcements will be coming "very soon."

```mermaid flowchart TD accTitle: MakerMaker.AI RSI Ratio Framework accDescr: Diagram illustrating the numerator and denominator framework of recursive self-improvement.

subgraph Ratio["RSI Efficiency & Capability Ratio"] Num["Numerator: System Capability<br/>(Task Performance & Accuracy)"] Denom["Denominator: Resource Cost<br/>(Energy, FLOPs & Compute)"] end

Target["Ideal RSI Signal"] Num -->|"Upward Trajectory"| Target Denom -->|"Downward Trajectory"| Target

subgraph Scope["Full-System Optimization"] Weights["Model Weights"] Harness["Execution Harness & Logic"] end

Target --> Scope ```

- **True RSI vs. Autocatalysis:** True RSI demands that the system improves the entire operational stack—both harness code and model weights—in a closed loop, rather than relying on humans to stitch together isolated automated subtasks. - **The Core Metric:** Tracking real progress requires evaluating performance relative to resource costs: $\text{Capability (Numerator)} / \text{Resources Expended (Denominator)}$. Achieving simultaneous increases in capability and decreases in resource overhead provides the strongest evidence for genuine recursive self-improvement.

## Notable quotes > [02:28](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=148s) what is the optimizer and what is the optimizee? — Owen He

> [09:39](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=579s) bleederboard — Dhaval Adjodah

> [13:39](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=819s) stamp collecting — Owen He

> [19:32](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1172s) We try to be like as hands off as possible. Let it do its thing. — Owen He

> [19:32](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1172s) improve yourself — Martin Casado

> [21:08](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1268s) It's not just possible. I think it's happening. — Dhaval Adjodah

> [24:34](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1474s) It's not a belief. It's happening. — Dhaval Adjodah

> [24:34](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1474s) The technical answer is you have to be really, really good at measuring things. So for example, we have one of our engineers that we named him the Viceroy of Variance. That's a real thing. Like his job is to go and like destroy the results we have and find what's wrong. — Dhaval Adjodah

> [29:32](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1772s) garden hose company in Wisconsin — Dhaval Adjodah

> [29:32](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1772s) fudge factor — Dhaval Adjodah

> [31:16](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1876s) very soon — Dhaval Adjodah and Owen He

## Mentioned Martin Casado, Dhaval Adjodah, Owen He, MakerMaker.AI, a16z, MIT Media Lab, DeepMind, OpenAI, GPT-2, GPT-5, nanoGPT, Kimi k1.5, Kimi k3, Nvidia Blackwell B200, Richard Feynman, Albert Einstein, Amdahl's Law, Moore's Law

## 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.

Timestamps: 00:00 - Intro 00:58 - What actually counts as RSI 05:10 - Capability as a numerator and denominator 06:58 - Closed loop vs takeoff 08:32 - Humans as agents in the system 13:39 - Why most RSI papers don't scale 17:41 - Why they chose inference 24:00 - "It's not a belief, it's happening" 25:42 - The Viceroy of Variance 29:39 - The garden hose company in Wisconsin

Resources: Follow Dhaval Adjodah on LinkedIn: https://www.linkedin.com/in/rand0m Follow Owen He on LinkedIn:https://www.linkedin.com/in/owen-he-b7b064279/ Follow Martin Casado on X: https://x.com/martin_casado

Stay Updated: If you enjoyed this episode, be sure to like, subscribe, and share with your friends!

Find a16z on X: https://twitter.com/a16z

Find a16z on LinkedIn: https://www.linkedin.com/company/a16z

Listen to the a16z Show on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX

Listen to the a16z Show on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711

Follow our host: https://x.com/eriktorenberg

Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see http://a16z.com/disclosures.

[Source](https://www.youtube.com/watch?v=3Q1mdn5ZGvM)

中文简报

MakerMaker认为闭环AI可在能力提升的同时降低算力成本 真正的递归自我改进已在发生，因为闭合优化回路使模型能够在降低算力成本的同时提升能力，从而打破了进步必须依赖不断增加资源投入的设想。 MakerMaker.AI 联合创始人 Dhaval Adjodah 与 Owen He 将真正的递归自我改进（RSI）定义为一个闭环，其中 AI 在提示词、测试框架、权重和硬件层面同时扮演优化器与被优化对象的角色，这与开环的人类自催化形成对比。他们通过权衡性能提升（分子）与总算力、能源及试错成本（分母）的比率来构建能力扩展模型，并强调有意义的 RSI 绝不能仅仅用蛮力来掩盖高昂成本。通过初期聚焦于模型推理，其多智能体架构协调专门的数字智能体与人类智能体，自我优化运行技术栈并重写自身的主提示词。针对有关智能爆炸的投机性叙事，他们引用阿姆达尔定律（Amdahl's Law）和物理约束予以反驳，并提出实证观察，指出在防范方差和奖励作弊的前提下，系统能够在切实提升能力的同时降低资源消耗。 真正的递归自我改进需要闭合回路，使优化器与被优化对象在模型、测试框架和提示词之间可互换，从而将真正的 RSI 与历史上的自催化工具区分开来。 (2:28) 评估真正的 RSI 需要追踪能力提升相对于总资源开销（分母）的比率，总开销包括原始 GPU 算力、实验周期以及研究人员的试错成本。 (9:39) MakerMaker 指出，越来越多的实证证据表明，自我改进系统能够同时提高性能并降低资源消耗，不过失控式的激增仍受物理条件限制。 (21:08) 多数学术界 RSI 架构因启发式设计和统计噪声而无法在生产环境中扩展，因此需要引入类似“方差总督”（Viceroy of Variance）的对抗性验证角色来检测多层级的奖励作弊。 (24:34)

Reading summary

MakerMaker.AI co-founders Dhaval Adjodah and Owen He argue that genuine recursive self-improvement (RSI) requires a closed loop where the optimizing system and the system being optimized are interchangeable across the entire stack. In conversation with a16z's Martin Casado, they contend that true self-improvement is demonstrated by expanding system capability while simultaneously reducing the underlying resource expenditure. Demarcating True RSI from Autocatalysis (2:28) Casado and the founders contrast true RSI with autocatalysis, where human engineers use AI assistants to build better AI software. In an autocatalytic system—a historical pattern seen in steam engines aiding coal mining or automated software assisting silicon chip design—the human mind remains the primary optimizer driving an open loop that requires external inputs. True RSI occurs only when the cognitive agent is simultaneously the optimizer and the optimizee, allowing algorithmic gains to be directly reinvested into the system's own cognitive capacity across prompts, harnesses, model weights, and hardware. Evaluating Gains Through the Cost Denominator (9:39) Adjodah formalizes capability growth as a ratio between capability performance (the numerator) and the cost or difficulty of achieving the next breakthrough (the denominator). Modern benchmark improvements often mask capital inefficiency by simply throwing more compute and human engineering at a problem, which the founders state does not qualify as RSI. MakerMaker tracks internal compute expenditure via a company-wide 'bleederboard' to ensure that measured efficiency gains account for trial-and-error costs. Owen He highlights the community's nanoGPT reproduction as a clear example of denominator reduction: researchers cut the cost of pre-training GPT-2 from a multi-week, multi-GPU cluster effort down to roughly $100 on a single GPU using identical hardware. Adjodah emphasizes that denominator reductions directly fuel numerator capability: cheaper compute allows denser sampling of trajectories across language manifolds, yielding higher-quality reasoning data to train smarter models. Empirical Verification Against Systemic Variance (24:34) The founders argue that many published RSI papers fail to scale because they rely on bespoke, human-inspired heuristics rather than unified mathematical abstractions. Moreover, Casado and Adjodah note that closed feedback loops do not inherently trigger vertical intelligence explosions; complex non-linear systems frequently decay, oscillate, or exhibit multi-level reward hacking where nested agents game the objective function. To counter self-deception, MakerMaker employs an internal 'Viceroy of Variance' whose explicit role is to invalidate experimental results and isolate anomalies, such as stochastic performance drift observed between identical Nvidia Blackwell B200 GPUs. Architecturally, the company models its platform as a hybrid multi-agent network combining specialized human engineers with digital models, where an automated meta-agent orchestrates tasks and continuously rewrites its own master prompt. Amdahl's Law and Evaluation Latency Limits (27:51) He observes that intelligence operates as a high-dimensional vector rather than a single scalar, meaning progress across domains is governed by Amdahl’s Law: overall system speedup is constrained by its slowest non-optimizable component. In purely digital environments like code generation and automated theorem proving, evaluations occur almost instantaneously, allowing self-improving feedback loops to cycle rapidly. In contrast, physical domains such as drug discovery remain bottlenecked by wet-lab syntheses and multi-month clinical trials that digital models cannot compress. Even within fast digital domains, the founders emphasize that capability growth cannot accelerate infinitely without bound; computational scaling must ultimately confront physical asymptotes dictated by fundamental physical constants, the speed of light, and the finite particle count of the universe. While MakerMaker's founders report early empirical signals of models simultaneously improving capabilities and reducing resource costs, they note that the AI field still lacks an overarching theoretical framework for self-referential systems. They expect future breakthroughs to emerge from resolving messy, unglamorous bottlenecks in pragmatic industrial workflows rather than from theoretical speculation on curated academic benchmarks.

Detailed notes — zh

## Background 该视频由风险投资机构 a16z 于 2026 年 10 月 12 日发布在 YouTube 频道上。视频内容为 a16z 普通合伙人 Martin Casado 与初创公司 MakerMaker.AI 的两位联合创始人 Dhaval Adjodah 和 Owen He 展开深度对话，旨在剖析递归自我改进（RSI）的核心定义，澄清行业内对该术语的泛化滥用，并探讨构建真正闭环自优化 AI 系统的实际工程路径与挑战。 ## TL;DR 访谈深入辨析了“真正的递归自我改进（RSI）”与单纯的人类工具链“自动催化”之间的本质区别，强调 RSI 必须是认知系统作为优化器自我闭环迭代。两位创始人提出了以“能力提升（分子）除以资源消耗与任务难度（分母）”的数学框架来衡量真实进展，批判了靠天量资本堆砌换取模型发布的虚假繁荣，并分享了公司内部用“流血榜”监控实验成本的硬核文化。他们选择以推理（Inference）优化作为商业与自改进循环的切入点，解释了从 Prompt、Harness 到权重乃至芯片的全栈协同优化路径。此外，他们还分享了设立“方差总督”排查统计噪音与奖励黑客、利用阿姆达尔定律剖析物理世界瓶颈等实战心得，指出真正的通用智能秘密往往藏在泥泞复杂的现实产业工作流之中。 ## Atmosphere & style 整场对话洋溢着高密度的硬核学术探讨与硅谷顶尖工程极客的务实锐气，既有对控制论和动力系统微分方程的严密推导，又不乏生动诙谐的内部黑话。Martin Casado 的提问一针见血且直击行业浮躁痛点，而 Dhaval 与 Owen 的回答则展现出从第一性原理出发的从容自信与冷静自省。从调侃摧毁虚假成果的“方差总督”，到引用“威斯康星水管厂”的 GPS 寓言，谈话在深刻的技术反思与轻松幽默的叙事节奏之间自如切换，让听众在摆脱末日论与盲目炒作的同时，真切感受到下一代 AI 自进化探索的前沿脉搏。 ## Bulletin - [02:08](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=128s) 真正的递归自我改进（RSI）与传统“自动催化”不同，要求作为认知核心的系统本身既是优化器也是被优化对象，形成全栈闭环，而非仅靠人类大脑在环路中进行外部干预。 - [09:39](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=579s) 衡量 RSI 进展必须拆解为分子（能力增量）与分母（算力、资本及人工试错成本），MakerMaker.AI 内部设立“流血榜”监控实验消耗，警惕用天量资本掩盖效率停滞的伪繁荣。 - [17:41](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1061s) 团队将推理优化选为首个商业落地与技术验证场景，通过“模型优化推理效率降低成本→单位时间探索更多 Rollout 轨迹→反哺更聪明模型”形成复合飞轮。 - [27:51](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1671s) 受阿姆达尔定律制约，智能在纯数字领域虽可实现数千倍加速，但在涉及药物湿实验等物理现实世界中，整体迭代速度仍受制于“移动原子”的物理演化周期。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=0s) 递归自我改进（RSI）的本质是优化器与被优化对象的统一，使认知能力可以直接复利式地回馈并重塑系统自身。 - [02:08](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=128s) 技术史上的蒸汽机、芯片设计大多属于自动催化（开环系统），决策核心仍是未被改进的人类大脑；真正的 RSI 必须闭环。 - [05:10](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=310s) 能力提升速率由分子（优化产出）与分母（边际成本与难度）共同决定，唯有分子超越线性增长或大幅压缩分母才能避免走向平台期。 - [06:58](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=418s) 闭环控制系统不自动等同于失控的指数级“起飞”，缺乏理论解析解的 RSI 目前本质上是一门高度依赖实验测量的经验科学。 - [08:32](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=512s) MakerMaker.AI 将系统形式化为多智能体网络，人类研究员与软件工程师被视为其中的特殊智能体共同接受优化，以防成为系统瓶颈。 - [09:39](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=579s) 行业普遍存在单次成功训练掩盖多次失败开销的现象，必须把人类试错工时与全量算力计入分母，确保真实效率的净增长。 - [11:44](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=704s) NanoGPT 的基准测试表明，在相同硬件下训练成本的大幅降低完全来自纯算法与系统工程优化，降本能直接增加语义空间的采样密度反哺能力分子。 - [13:39](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=819s) 学术界多数 RSI 方案如同脆弱的“鲁布·戈德堡机械”，缺乏统一数学框架的经验拼凑在真实规模化任务中往往彻底失效。 - [15:30](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=930s) RSI 的全栈优化跨越 Prompt、测试脚手架（Harness）、模型权重（Weights）直至底层芯片（Chips），多层级协同优化可防止局部优化反噬全局。 - [17:41](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1061s) 推理环节具备自引用的递归属性：推理加速降低自改进实验成本，高吞吐量进而驱动模型持续自我演进，是兼具商业变现与技术验证的核心场景。 - [19:32](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1172s) 系统内部的模型调度与分工已交由自主智能体动态决策，就连主提示词（Master Prompt）本身也处于持续的自我迭代与重写之中。 - [22:21](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1341s) 物理世界的普朗克常数与光速等定律构成了绝对计算上限，系统的自改进终将遭遇渐近线，不可能无休止地指数爆炸。 - [25:03](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1503s) 为对抗多层智能体系统中的奖励黑客攻击与非线性方差，团队设立“方差总督”角色无情摧毁存在统计偏差的虚假优秀结果。 - [27:51](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1671s) 阿姆达尔定律决定了 RSI 的现实提速上限：即便数字生成速度提升百万倍，物理实验环节的耗时仍会死死锁住端到端的整体加速比。 - [29:32](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1772s) “威斯康星州花园水管公司”的比喻揭示了 AI 研究不应仅局限于实验室时髦难题，深入泥泞的真实业务复杂工作流才能摸清自改进的理论边界。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=0s) 递归自我改进（RSI）的核心定义与创始背景 递归自我改进（Recursive Self-Improvement, 简称 RSI）的基本构想，是构建能够自我优化的 AI 系统。与以往技术不同的是，人们第一次能够将“优化器（optimizer）”与“被优化对象（optimizee）”的角色调换——让执行改进的认知系统本身成为被优化的对象，从而实现认知能力对自身的再投资。RSI 并非仅仅发生在某一个单一层次，而是横跨整个技术栈：从最上层的 Prompt、评测与编排框架（Harness）、模型权重（Weights），一路向下延伸至底层的芯片硬件。MakerMaker.AI 的目标正是全面优化这一整套技术栈。 在访谈开场中，a16z 投资合伙人 Martin Casado 介绍了两位联合创始人 Dhaval Adjodah 与 Owen He： - **Dhaval Adjodah**：物理学背景出身，在 MIT Media Lab 获得强化学习（RL）博士学位，热衷于动手工程实现，现与 Owen 及 Grace 共同创立 MakerMaker.AI。 - **Owen He**：数学专业出身，在学习形式语言与逻辑课程时接触到定理证明器（Theorem Prover），意识到“可以制造一台机器替自己证明定理”，由此进入 AI 领域；随后攻读持续学习（Continual Learning）与元学习（Meta-Learning）方向博士，并在 DeepMind 担任数年研究科学家，现专注于 RSI 研究。 ### [02:08](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=128s) 澄清概念：真 RSI 还是“自动催化（Autocatalysis）”？ 当前创投圈中“RSI”这一词汇已被严重泛化和滥用。许多团队在路演时声称自己在做“用 AI 制造 AI”的 RSI，细问之下却发现仅仅是用大语言模型编写 GPU 核函数（GPU kernels）。Martin Casado 直言这根本不符合他对 RSI 的理解。 Owen 解释说，辨析真正 RSI 的核心判准在于两个问题：“谁是优化器？谁是被优化者？”在人类技术史上，绝大部分技术进步都属于**自动催化（Autocatalysis）**而非 RSI： - **经典自动催化**：计算机被用来辅助设计更先进的芯片，互联网让分散各地的研究人员协同加速技术迭代，蒸汽机运煤提高开采效率以建造更强大的蒸汽机。在这些场景中，工具虽然在不断升级，但做出改进决策、提出新想法的优化器始终是“人类的大脑”，而人类大脑本身并未在这个闭环中被优化。 - **真正的 RSI**：系统首次将作为认知核心的智能体大脑（大语言模型）同时置于“优化器”与“被优化对象”的位置，使得认知能力的提升可以直接复利式地回馈并作用于自身。 当 Martin 询问为何不选择更务实地逐步拆解问题、而是直接攻坚这一高难度方向时，Dhaval 坦言迎难而上本身就极具趣味。而从控制论（Control Theory）角度来看，如果系统中存在断点，就无法构成**闭环（Closed Loop）**；自动催化本质上是开环的，每一次迭代都必须依赖人类介入并灌入外部资源。而在闭环系统中，虽然同样需要输入资源，但其边际收益随时间推移能够保持恒定甚至逐步放大。 ```mermaid flowchart TD accTitle: 自动催化与递归自我改进架构对比 accDescr: 对比人类主导的开环自动催化系统与模型自我优化的闭环RSI系统 subgraph AC["传统自动催化（开环 / 工具增强）"] H1["人类研究者<br/>(优化器)"] -->|"设计与构建"| T1["更强 AI 工具 / 算力<br/>(被优化者)"] T1 -.->|"提高工作效率<br/>(断点：需人类脑力介入)"| H1 end subgraph RSI["递归自我改进（闭环系统）"] M1["AI 智能体 / 模型<br/>(优化器)"] -->|"重构与更新"| M2["AI 智能体 / 框架<br/>(被优化者)"] M2 -->|"认知能力再投资"| M1 end ``` ### [05:10](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=310s) 能力提升率的数学拆解：分子、分母与起飞奇点 为了严格形式化 RSI 的本质，Dhaval 从微积分与动力系统角度拆解了“能力提升速率”（即能力对时间的一阶导数 $\frac{dC}{dt}$）。该速率由分子与分母共同决定： $$\frac{dC}{dt} = \frac{\text{分子（能力增量 / 产出）}}{\text{分母（边际成本 / 任务难度）}}$$ 1. **分母（成本与难度）**：由客观物理与理论规律主导，能力越强，迈出下一步的边际成本（算力、资本、工程复杂性）就越高，难度呈现单调递增。 2. **分子（优化表现）**：这是 RSI 与传统技术的关键分野： - **固定分子（传统技术）**：如蒸汽机或固定工具，其推进能力有限，面对递增的分母，微分方程的解最终会走向**平台期（Plateau）**。 - **线性正相关分子（基线 RSI）**：系统的能力 $C$ 越强，改进自身的能力也越强，分子与能力呈正相关。此时分子与分母的递增效应相互抵消，系统可以在较长时间内实现**恒定速率的持续提升（Constant Improvement）**。 - **高阶幂次相关分子（爆发式起飞）**：如果分子与能力的更高阶幂次相关（$C^k, k > 1$），则系统微分方程将演变出指数级甚至超指数级的**能力爆发（Exponential Takeoff）**。 ### [06:58](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=418s) 闭环控制系统与实际收敛：理论、经验与多智能体系统 针对外界“一看到反馈环形图就盲目联想到指数起飞（Takeoff）与末日论（Doomerism）”的现象，Martin 与 Dhaval 指出，这是一个典型的控制系统误区。对于一个高度复杂的非线性控制系统，仅仅拥有闭环结构绝不自动等同于能够引发失控起飞，其最终行为完全取决于自指函数的非线性指数以及系统的收敛特性。 Dhaval 与 Owen 强调，在这个领域过早进行纯理论推测几乎没有意义，因为**这是一个纯粹的经验科学问题（Empirical Problem）**： - 目前深度学习连最底层的理论尚未完全建立，人们并不从数学底层彻底理解它为何有效；行业目前最坚实的依据是经验性的“缩放定律（Scaling Laws）”。 - 在五六层抽象之上的 RSI 更是缺乏解析解（Analytical Solution），参数和指数必须在现实世界中通过严格实验来测量。 - 在实际经验测试中，面临着诸如**奖励黑客攻击（Reward Hacking）**、评估方差等极高难度的工程与统计挑战。 此外，Owen 澄清了一个关于其系统构架的重大误解：外界常以为 MakerMaker.AI 追求的是一个全自动、完全去人化的纯数字系统。但从其数学模型底座来看，他们构建的是一个**多智能体系统（Multi-Agent System）**。在这个混合系统内部，人类研究员与软件工程师被形式化为该多智能体系统中的一类“特殊智能体”。整个系统是人机协作网络共同被递归优化。如果系统中留下任何未被优化的部件，该部件最终必然成为整套系统性能的致命瓶颈。 ### [09:39](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=579s) 分母陷阱与“流血榜”：如何避免用资本掩盖效率停滞 谈到行业现状，Martin Casado 指出，从表面上看，模型的发布周期正在大幅缩短，似乎整个 AI 领域正处于狂飙突进的指数级增长阶段。但若将这个进步速度除以这些企业所融集的天量资本（作为分母），这根本谈不上真实的效率跃升——往往只是融了海量资金，再用 AI 把钱烧出去而已。Martin 询问团队，在面对外部夸大其词的宣传或自身开展研究时，该如何建立严谨的衡量标准，避免把“单纯砸钱换效果”误当成真正的递归自我改进（RSI）。 ```mermaid flowchart LR accTitle: RSI 能力度量与分母陷阱 accDescr: 展示能力分子与资源分母的关系，说明为何单纯增加投入不属于真正的递归自我改进。 subgraph Metric["评估指标：真实进展"] direction TB Num["分子 (Numerator)<br/>模型能力 / 表现提升"] Den["分母 (Denominator)<br/>算力 FLOPS + 资本成本 + 人力试错时间"] Score["真实 RSI 增益 = 分子 / 分母"] Num --- Score Den --- Score end subgraph Pitfall["常见误区：虚假繁荣"] direction TB Hype["砸入海量资金与人力"] --> Model["模型发布周期变短"] Model -.->|"分母急剧膨胀"| NotRSI["并非 RSI，仅为资本堆砌"] end ``` Dhaval Adjodah 和 Owen He 的观点非常直接：如果能力的提升必须依赖成倍灌入的资源，那绝不是 RSI。真正的自改进必须有严密的数据来证明效率的净增长。Martin 进一步追问，在核算“分母”时，是否也必须将人类研究员在背后经历的数十次失败实验、无数工时与脑力消耗通通算作成本？Dhaval 明确表示肯定。MakerMaker.AI 在内部从创立第一天起就建立了极其严苛的成本追踪文化。Dhaval 分享了一个公司内部的趣事：大家通常只关注展示模型表现的“排行榜”（Leaderboard），而他们内部专门做了一个**“流血榜”（Bleederboard）**，用于实时监控并公开展示每个人做实验所消耗的算力和资金成本，全公司每天都能清楚看到谁在榜首“大出血”。 - 真正的 RSI 必须保证在分子（能力产出）提升的同时，分母（资源消耗，包含 FLOPS、资金与试错人时）保持受控或同步降低。 - 业内常谈及单次最终成功训练的成本，但掩盖了背后“第 23 次实验才成功”所累积的巨大隐形成本。 - 公司内部推行“流血榜（Bleederboard）”，让每一次实验的资源消耗完全透明，以培养注重资本效率的工程纪律。 --- ### [11:44](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=704s) 纳米级 GPT 与基准测试：算法红利与分母缩减 Owen He 引用了开源社区运行的 NanoGPT 基准测试作为技术降本的范例。当 GPT-2 最早面世时，训练成本极其高昂，需要在大量 GPU 上连续训练数周；而如今，在完全相同的硬件条件下，单张 GPU 仅需数小时就能完成复现，整体训练花销已压缩至仅约 100 美元左右。 Martin 询问这是否归功于摩尔定律，Owen 与 Dhaval 纠正道，这完全是在同一代硬件、相同数据量的前提下达成的，它完全来自于预训练技术的迭代、算法内核（Kernels）的优化以及数据工程的演进，其中甚至有许多直接来自自动化研究（Auto-research）的结果。 Martin 敏锐地指出，虽然这种优化极大缩减了分母，但它目前更多是在小模型或旧模型上复现，业界在像 GPT-5 这样的大模型前沿上并没有看到成本显著下降，反而成本还在急剧飙升。Dhaval 对此回应，如何最大化从算力中榨取产出，是全行业防守最严的核心商业机密之一。因为即便仅仅改善分母，由于训练速度加快，系统便能以更低成本获得海量的采样轨迹（Rollouts）。在语言空间流形（Manifold）中进行更广泛的探索与采样，就能筛选出更高质量的数据轨迹，进而反哺训练出更聪明的模型，从而变相拉动分子。 - GPT-2 当前在单卡上的训练成本已可降至约 100 美元，这完全归功于纯算法与系统工程优化，硬件与数据条件并未改变。 - 降低分母（降低算力开销）能直接增加 Rollout 的采样密度，使模型在前沿语义空间中探索更优轨迹，从而推动能力分子提升。 - 真正具有革命性意义的突破，是同时撬动“分子增加”与“分母减少”的双向优化。 --- ### [13:39](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=819s) 拒绝“鲁布·戈德堡机械”：为何多数 RSI 论文无法扩展 Dhaval 坦言，当他们在内部基础设施中复现并验证大量声称实现了 RSI 的学术论文时，发现绝大部分根本无法在实际中起效。这些研究大多只是人类研究员凭借自我内省设想出的“精巧点子”，理论听起来很酷、论文写得很漂亮，但一旦放到真实基准和真实经济任务中进行规模化（Scale）测试，就会彻底崩溃。 MakerMaker.AI 发现，真正能在规模化下起效的，往往是**能够彻底闭环的极简设计**。Dhaval 强调，团队的核心竞争力在于从统一且严谨的数学框架出发，先绘制出整个状态空间的理论全景图。面对复杂的 AI 自改进体系，通常有两种构建方式：一种是东拼西凑出 200 个相互独立的子系统，试图拼装成一台极其脆弱的“鲁布·戈德堡机械”（Rube Goldberg Machine）并祈祷它能运转；另一种则是从第一性原理和数学出发，穿透表象，最终提炼出能概括这 200 个表象的十几个核心抽象层。 ```mermaid flowchart TD accTitle: RSI 的技术路线对比与全栈优化 accDescr: 对比拼凑式研究与统一数学框架，并展示从 Harness 到芯片的全栈自改进层次。 subgraph Approaches["系统构建路径对比"] direction TB subgraph Wrong["不可扩展的经验拼凑"] A["模仿人类直觉的复杂点子"] --> B["拼凑 200 个异构组件<br/>(鲁布·戈德堡机械)"] B --> C["局部有增益，规模化后失效"] end subgraph Right["统一数学框架 (MakerMaker)"] D["第一性原理与数学抽象"] --> E["提炼十几个泛化抽象层"] E --> F["全局端到端闭环，具备扩展性"] end end subgraph Stack["RSI 全栈协同优化层次"] direction TB L1["提示词层 (Prompts)"] L2["测试脚手架 / 外部框架 (Harness)"] L3["模型权重更新 (Weights)"] L4["硬件与芯片 (Chips)"] L1 --> L2 --> L3 --> L4 end ``` Owen 引用了物理学家理查德·费曼（Richard Feynman）的名言——科学可以分为“集邮式的科学”和“物理学”。缺乏统一数学框架的 RSI 论文，不过是在进行毫无泛化能力的“集邮”。很多研究只观察人类思维的某个局部动作去生搬硬套，却不知道在局部应用该策略是否在破坏全局系统。 团队将 RSI 划分为自上而下的多个层次：提示词（Prompts）、脚手架/框架环境（Harness）、模型权重（Weights），直至最底层的芯片架构（Chips）。他们从轻量且极度节省资本的 Harness 自改进切入，进而过渡到底层权重的协同更新，未来还将进一步下探技术栈，通过多层组件的“协同优化”（Co-optimization），防止在优化某一模块时对另一模块产生负面反噬。 - 学界多数 RSI 方案无法落地的核心原因：过度复杂化、依赖人类对自身思维的狭隘直觉，且未能形成闭环。 - 费曼的“集邮与物理学”比喻：碎片化技巧是集邮，唯有置于统一数学框架下的系统设计才是真正的“物理学”。 - RSI 的全栈路径覆盖 Prompts、Harness、Weights 到底层 Chips，必须推行跨层级的协同优化。 --- ### [17:41](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1061s) 锚定推理场景：在最贴近 AI 核心的经济痛点上闭环 在谈到具体的业务切入点时，Dhaval 解释，作为一家创业公司而非纯学术实验室，MakerMaker.AI 必须选择一个既具备绝对商业价值、又尽可能贴近 AI 本身核心规律的垂直场景，以此来证明自改进闭环的成立。他们最终选定了**推理（Inference）优化**。 推理问题拥有极度诱人的递归属性： 1. **模型自我赋能使得推理速度变得更快、更经济**； 2. **推理加速后，模型便能以更高的吞吐量来优化自己，让随后的自改进循环进一步加速**。 Martin 惊呼这种双重复合效应（Compounding Impact）简直“让人脑洞大开”——尽管它在数学物理层面仍可能会遭遇渐近线（Asymptote）而未必直接导致失控的智能飞跃（Takeoff），但它在工程与商业上的飞轮效应已极其可怕。Dhaval 指出，当前全球几乎所有企业都在试图部署自己的开源权重模型（如 Llama 3 或各类开放权重架构），如何在有限的算力成本下高效地将其推向生产环境，是全行业眼下最迫切、经济体量最巨大的刚性需求。通过攻克推理环节，他们既抓住了核心经济利益，又在最底层的 AI 循环中验证了“同时改进分子与分母”的技术设想。 - 切入推理领域的核心逻辑：推理既是当前 AI 部署的刚性经济痛点，也是能同时撬动分子与分母自改进闭环的最佳试验场。 - 推理优化的双重飞轮：模型优化推理执行效率 $\rightarrow$ 降低单次推演成本与延迟 $\rightarrow$ 单位时间内可运行更多自改进探索迭代 $\rightarrow$ 进一步提升优化效率。 - 团队的商业定位：避开空中楼阁式的研究，以开源权重模型的高效服务化作为实战着力点，用真实商业回流验证 RSI 架构的生命力。 ### [19:32](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1172s) 自主改进与多智能体系统：连主提示词都在自我进化 当被问及若要优化特定大模型（如 Kimi k1.5 / Kimi 3）的推理速度，究竟是用该模型自身进行递归循环，还是使用外部独立模型系统时，Dhaval Adjodah 与 Owen He 解释道，MakerMaker.AI 的系统绝非仅依赖单一 LLM，而是调度多种不同的大模型协同处理各项任务。更具颠覆性的是，究竟挑选哪一个模型负责哪项子任务，甚至都不再由人类工程师来规定，而是交由系统内部的自主智能体根据有效性与调用频次动态决策。 为了让递归自我改进（RSI）真正成立，团队极力避免将优化局限在某一个孤立组件上，而是赋予系统极大的自主度，工程师“尽可能做到完全不插手”（hands off），放手让系统自行探索。Martin Casado 惊讶地追问系统中是否存在一个类似“改进你自己”的单一主提示词（Master Prompt）。Dhaval 笑着透露，不仅没有僵死的主提示词，就连主提示词本身也处于持续自我迭代与进化的状态。在这样的体系中，执行优化的组件与负责识别“何处需要优化”的元认知组件是两套紧密共生的动态系统。 ```mermaid flowchart TD accTitle: 自我改进系统的双重动态结构 accDescr: 展示执行改进与识别改进方向两者之间的共生循环结构，并包含主提示词的自我更新。 subgraph CoreLoop ["RSI 核心控制架构"] A["元决策层<br/>识别并决定何处需要优化"] -->|"动态引导"| B["执行层<br/>具体执行系统优化"] B -->|"产出更优结构与参数"| A C["Master Prompt<br/>（主提示词）"] -->|"初始指引"| A B -.->|"递归重写更新"| C end subgraph Agents ["异构智能体池（模型选择由 Agent 自主决定）"] B --> D["LLM 子任务 A"] B --> E["LLM 子任务 B"] B --> F["推理引擎优化 / Harness"] end ``` 这种通用架构在数学定义上能够泛化至任何 AI 可被指派的问题领域，但现实落地时依然会撞上硬性的边界，首当其冲的便是“人”的因素。与许多鼓吹“用 AI 完全替代人类”的同行不同，他们发现优秀的科研人员与工程师所带来的洞见短时间内无法被智能体取代。因此，他们自始至终将整家公司构建为一个多智能体系统（Multi-Agent System），在这个协作网络中，人类研究员与工程师本身就是一种特殊的“智能体”，与各类 AI 智能体共同生息协作。 --- ### [22:21](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1341s) 分母优化、热力学极限与不可避免的渐近线 在系统资源效率的探讨中，讨论转向了能否通过算法降低系统迭代所需的总能量输入。Dhaval 强调，团队的工作从来不仅仅停留在提升分子（系统性能），更核心的突破在于压缩分母——即在消耗同等甚至更少物理资源与资金成本的前提下，压榨出更高的输出产出。针对“性能持续提升而资源消耗持平甚至递减”的假设，Dhaval 直言“这不仅可能，而且正在我们系统内部真实发生”，模型正逐渐以更低的能量消耗换取更高的智能产出。 针对 Martin Casado 提出的“这种无限逼近是否意味着能够摆脱渐近线实现真正无界限的起飞（Takeoff）”这一物理学与信息论疑问，Dhaval 坦言物理世界必定存在终极天花板。计算的边界终将受到普朗克常数、普朗克能量、普朗克时间以及光速等基础物理法则的束缚；全宇宙大约 $10^{80}$ 个基本粒子乘以光速等物理限制，构成了一切计算能力的硬性绝对上限。因此，系统最终必然会趋于渐近饱和，而不是无休止地指数爆炸。 - **指标陷阱**：若以基准测试准确率（Benchmark Accuracy）作为评估尺度，其天然存在 100% 的上限，永远不可能观测到持续的指数型增长，因而寻找衡量能力跃迁的真正有效指标本身就是一项核心研究命题。 - **经验证实 vs. 主观信念**：Dhaval 多次强调这种效率提升“不是一种信念，而是正在发生的事实”（"It's not a belief, it's happening"）。团队在过去几个月内通过密集的实证测量，持续观察到了这种自我优化的微光在逐步扩大。 --- ### [25:03](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1503s) 方差总督（Viceroy of Variance）与奖励黑客的挑战 伴随 RSI 概念在行业内的爆火，市面上频繁涌现出宣称实现自我改进的学术论文与技术公告。MakerMaker.AI 团队曾多次试图复现这些外部突破，结果却屡屡发现它们无非是数据“挑选偏差”（Cherry-picking）或统计噪音造成的假象。为了确保自身的科研结果坚实可靠，团队确立了极其严苛的度量标准，并赋予其中一位资深工程师一个极具戏剧色彩的正式头衔——“方差总督”（Viceroy of Variance）。 “方差总督”的唯一职责，就是动用一切手段无情地摧毁团队所跑出的“优秀结果”，彻底查明其中隐藏的系统漏洞与统计假象。团队内部的 Slack 甚至为他定制了双眼发射猩红色激光的专属表情符号，专门在他抓包虚假方差时出动。由于多层嵌套智能体系统的复杂性，常常会出现“智能体 A 所生成的智能体 B 在构建智能体 C 时悄悄进行了奖励黑客（Reward Hacking）”的极端隐蔽现象。要想识破这些陷阱，团队目前没有任何捷径，只能依赖所谓“高质量投入时间”（Quality Time）——人工运行并审视成百上千次对照实验，沉淀物理直觉，将真实的算法增益与纯粹的统计噪音、非线性振荡区分开来。 - **非线性失控与振荡**：多层智能体系统如同高阶非线性控制系统，系统表现不会平滑收敛，而是经常在极优与极劣之间剧烈振荡。 - **硬件底层的随机性干扰**：随机性不仅来自算法本身，甚至不同硬件之间都会产生微观物理噪音——同为 NVIDIA B200 芯片，两张物理 GPU 在执行时的细微差异都会被多层智能体级联放大，最终投射在顶层宏观评估表现上。 --- ### [27:51](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1671s) 阿姆达尔定律与物理世界的瓶颈 针对“智能”这一概念的本质，Owen He 指出，智能绝不是一个简单的标量（Scalar），而是一个多维度的复杂向量（Multi-dimension Vector）。这意味着递归自我改进在某些特定维度上能够瞬间实现暴风骤雨般的“起飞”（Takeoff），而在另外一些维度上则可能迅速撞上性能高原，进展极度缓慢。 决定系统最终进化上限的核心因素，在于其工作链路中那个“最慢的瓶颈部件”，这完全遵循计算机体系结构中的阿姆达尔定律（Amdahl's Law）： | 任务领域 | 评估反馈周期 | 物理世界耦合度 | RSI 加速潜力 | |:--- |:--- |:--- |:--- | | **纯数字领域**（如编程、纯数学证明） | 毫秒至秒级，评估计算廉价且高度确定 | 极低（纯比特交互） | **极高**（可实现上千倍乃至数百万倍的迭代提速） | | **现实物理领域**（如新药研发、湿实验、临床试验） | 数月至数年 | 极高（需移动原子、进行物理培养与人体测试） | **极低**（即便数字生成加速至无穷大，整体加速比仍被物理实验周期死死锁住） | Owen 举例说明，就算通过 RSI 将药物分子设计的数字化部分加速 100 万倍，但如果湿实验（Wet Lab）和临床试验占据整个药物研发流程 80% 的时间，那么根据阿姆达尔定律，整个系统在现实世界中所能获得的净加速幅度上限也绝不可能突破 20%。只要真实世界的问题依然需要“搬运原子”（Move Atoms）来完成真实验证，递归自我改进的腾飞速度就必然受到物理演化节律的严格制约。 ### [29:32](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1772s) 威斯康星州的花园水管公司：RSI 与现实世界的真正复杂度 Martin Casado 抛出了一个极其关键的底层问题：我们是否正在接近确定某种能力与资源的“理论边界”？比如能否预测解决一个药物研发难题究竟需要多大规模的计算机房，或者解决某个物理任务到底需要部署多少台机器人。在他看来，摸清这些边界是整个 AI 领域迫切需要回答的核心议题。 MakerMaker.AI 团队对此分享了一个公司内部常开的经典玩笑——“递归自我改进（RSI）的真正秘密，藏在威斯康星州某家做花园水管的公司里”。他们用爱因斯坦与相对论做了一个精妙的比喻：假如当年爱因斯坦没有推导出版广义相对论，人类就无法掌握速度与时间流逝膨胀收缩的严格物理定律；当后来某个倒霉的工程师负责设计 GPS 卫星定位系统时，为了让时间对齐，他必然会被迫通过大量试错摸索出一个时间修正公式。由于缺乏理论支撑，这位工程师很可能会一口咬定：“这绝对不可能是时空结构本身在发生扭曲”，但不得不承认“虽然我不知道原理，但我凭经验凑出来的这个修正参数（fudge factor）确实能跑通”。直到很多年后，人们才会恍然大悟：天哪，那背后原来是时空收缩效应。 这个比喻揭示了当前 AI 研究的盲区：业内绝大多数团队在思考 RSI 时，注意力全被吸引去攻克那些“超级困难且光鲜时髦的显学难题”（fashionable problems）。然而，物理与商业世界的真正复杂度往往并不在这些实验室基准中，而是隐藏在那些尚未被 AI 触及的现实边缘角落。真正的泛化挑战与智能突破点，极有可能就潜伏在威斯康星州某家平凡无奇的花园水管工厂复杂的日常业务工作流之中——只要系统能够自主搞定那类真实环境的复杂闭环，整个 RSI 的底层通道才算真正被彻底打通。 ```mermaid flowchart TD accTitle: 经验试错与真实世界复杂度的探索路径 accDescr: 对比传统追逐时髦 AI 难题与深入现实工业极端工况探索 RSI 理论边界的路径差异。 subgraph FashionableApproach["显学误区：追逐实验室热点"] A["聚焦光鲜时髦任务<br/>(标准基准测试)"] --> B["陷入局部优化<br/>缺乏真实复杂度"] end subgraph RealWorldComplexity["GPS 寓言：深挖真实业务复杂性"] C["深入泥泞现实工作流<br/>(如威斯康星水管厂)"] --> D["经验修正与试错验证<br/>(工程级 Fudge Factor)"] D --> E["反哺并揭示通用规律<br/>(真正的 RSI 闭环边界)"] end ``` --- ### [31:16](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1876s) 闭环总结：全栈 RSI 与分子分母的双重突破 在对话尾声，Martin Casado 对 MakerMaker.AI 所坚持的技术路线进行了精炼的梳理与总结： - **区分伪概念与真 RSI**：团队追求的不是简单的“自动催化效应”（auto-catalytic effects，即单向用 AI 写写辅助脚本），而是严格意义上的全栈闭环 RSI——由被改进的系统去改进系统自身，涵盖了从模型权重（weights）到外部运行框架与工程脚手架（harness）的整套完整体系。 - **分子与分母的同步优化**：评估进步的标准被严格量化为一个分数——分子代表系统能力上限（究竟变得有多强），分母代表为此注入的资源消耗（消耗了多少能量与算力）。团队目前已经在研究中观测到了这种令人振奋的早期微光：系统的能力在拔高，同时所需的推理与运行能耗反而被压缩。 当 Martin 兴奋地追问公众何时能看到他们正式对外披露更多研究成果时，两位联合创始人给出了充满信心的答复：“很快就会见分晓（very soon）”。这场探讨最终在对通用自迭代系统落地前景的期待与致谢中圆满收尾。 ## Notable quotes > [02:08](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=128s) 谁是优化器？谁是被优化者？ — Owen He > [13:39](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=819s) 科学可以分为‘集邮式的科学’和‘物理学’。 — Owen He > [17:41](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1061s) 让人脑洞大开 — Martin Casado > [19:32](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1172s) 尽可能做到完全不插手 — Dhaval Adjodah > [22:21](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1341s) 这不仅可能，而且正在我们系统内部真实发生 — Dhaval Adjodah > [24:00](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1440s) It's not a belief, it's happening — Dhaval Adjodah > [27:51](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1671s) 智能绝不是一个简单的标量，而是一个多维度的复杂向量。 — Owen He > [27:51](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1671s) 搬运原子 — Owen He > [29:32](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1772s) 递归自我改进（RSI）的真正秘密，藏在威斯康星州某家做花园水管的公司里 — Dhaval Adjodah > [29:32](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1772s) 虽然我不知道原理，但我凭经验凑出来的这个修正参数确实能跑通 — Dhaval Adjodah > [31:16](https://www.youtube.com/watch?v=3Q1mdn5ZGvM&t=1876s) 很快就会见分晓 — Dhaval Adjodah ## Mentioned Martin Casado, Dhaval Adjodah, Owen He, MakerMaker.AI, a16z, MIT Media Lab, DeepMind, 理查德·费曼 (Richard Feynman), 阿尔伯特·爱因斯坦 (Albert Einstein), GPT-2, GPT-5, NanoGPT, Llama 3, Kimi k1.5, Kimi 3, NVIDIA B200, 递归自我改进 (RSI), 自动催化 (Autocatalysis), 阿姆达尔定律 (Amdahl's Law), 缩放定律 (Scaling Laws), 多智能体系统 (Multi-Agent System), 奖励黑客 (Reward Hacking), 流血榜 (Bleederboard), 方差总督 (Viceroy of Variance) ## 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. Timestamps: 00:00 - Intro 00:58 - What actually counts as RSI 05:10 - Capability as a numerator and denominator 06:58 - Closed loop vs takeoff 08:32 - Humans as agents in the system 13:39 - Why most RSI papers don't scale 17:41 - Why they chose inference 24:00 - "It's not a belief, it's happening" 25:42 - The Viceroy of Variance 29:39 - The garden hose company in Wisconsin Resources: Follow Dhaval Adjodah on LinkedIn: https://www.linkedin.com/in/rand0m Follow Owen He on LinkedIn:https://www.linkedin.com/in/owen-he-b7b064279/ Follow Martin Casado on X: https://x.com/martin_casado Stay Updated: If you enjoyed this episode, be sure to like, subscribe, and share with your friends! Find a16z on X: https://twitter.com/a16z Find a16z on LinkedIn: https://www.linkedin.com/company/a16z Listen to the a16z Show on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX Listen to the a16z Show on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711 Follow our host: https://x.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see http://a16z.com/disclosures. [Source](https://www.youtube.com/watch?v=3Q1mdn5ZGvM)

Source quotes

[2:08] Owen He: 谁是优化器？谁是被优化者？ [13:39] Owen He: 科学可以分为‘集邮式的科学’和‘物理学’。 [17:41] Martin Casado: 让人脑洞大开 [19:32] Dhaval Adjodah: 尽可能做到完全不插手 [22:21] Dhaval Adjodah: 这不仅可能，而且正在我们系统内部真实发生 [24:00] Dhaval Adjodah: It's not a belief, it's happening [27:51] Owen He: 智能绝不是一个简单的标量，而是一个多维度的复杂向量。 [27:51] Owen He: 搬运原子 [29:32] Dhaval Adjodah: 递归自我改进（RSI）的真正秘密，藏在威斯康星州某家做花园水管的公司里 [29:32] Dhaval Adjodah: 虽然我不知道原理，但我凭经验凑出来的这个修正参数确实能跑通 [31:16] Dhaval Adjodah: 很快就会见分晓

Source quotes

[2:28] Owen He: what is the optimizer and what is the optimizee? [9:39] Dhaval Adjodah: bleederboard [13:39] Owen He: stamp collecting [19:32] Owen He: We try to be like as hands off as possible. Let it do its thing. [19:32] Martin Casado: improve yourself [21:08] Dhaval Adjodah: It's not just possible. I think it's happening. [24:34] Dhaval Adjodah: It's not a belief. It's happening. [24:34] Dhaval Adjodah: The technical answer is you have to be really, really good at measuring things. So for example, we have one of our engineers that we named him the Viceroy of Variance. That's a real thing. Like his job is to go and like destroy the results we have and find what's wrong. [29:32] Dhaval Adjodah: garden hose company in Wisconsin [29:32] Dhaval Adjodah: fudge factor [31:16] Dhaval Adjodah and Owen He: very soon

Original transcript

[0:00] The fundamental idea here is to build AI [0:02] that improves itself. And I think we're [0:04] getting there. [0:04] >> For the first time, we could swap the [0:07] optimizer as the optimiz. And this leads [0:10] to a reinvestment of cognitive [0:12] improvement into itself. [0:14] >> You can think of RSI at many different [0:16] levels of the stack, prompts, harness, [0:19] weights, and you can go all the way to [0:21] basically chips. Our goal here is to [0:23] optimize all the levels of the stack. [0:24] Our approach to RSI is that we have a [0:27] multi- aent system working together [0:29] that's been recursively improved. [0:30] >> I'm so impressed by your belief that [0:32] like this idea of a model improving [0:34] itself, [0:35] >> but it's not a belief. [0:36] >> It is a dramatic statement, which is I [0:38] think it's the first time I've heard you [0:39] say it. [0:39] >> Because we have been so thorough, we [0:41] have not released things until we're [0:43] super sure it's correct. [0:44] >> Do you envision a system where the [0:46] amount of energy that goes into it [0:48] somehow decreases over time relative to [0:51] the improvement? We can't share too [0:52] much, but one of the very important [0:54] things that we've tried to do. [0:58] >> Welcome both of you uh for joining. [1:00] We're going to be talking about [1:00] everybody's favorite topic, which is [1:02] RSI, recursive self-improvement. Maybe [1:05] to start, it'd be great to just [1:06] introduce yourselves and your [1:08] backgrounds. Then we'll talk a little [1:09] bit about what you're working on now, [1:10] and then we'll dig into the the the [1:12] content. [1:13] >> Yeah, I I I think of [clears throat] [1:16] myself as like um a physicist of some [1:19] kind. I I studied that. Then I did the [1:23] PhD in reinforcement learning at the MIT [1:25] Media Lab. I really like building things [1:27] and I'm very excited to be building um a [1:30] cool startup make a maker with Owen and [1:33] Grace. My name is Owen. Um I was uh [1:37] first studying math and then one day I [1:40] realized that all mathematicians are [1:42] proving theorems and I was taking a [1:44] class called formal language and logic [1:46] and they were showing how you can build [1:47] a theorem prover and I was like oh that [1:50] would be cool I can just build a machine [1:51] that prove theorems for me that's how I [1:53] got into AI [1:54] >> amazing [1:55] >> and then for my PhD I worked on continue [1:57] learning and metal learning then worked [2:00] for a few years at deep mind at a [2:01] research scientist now I'm working on [2:04] RSI I see a lot of pitches come in and a [2:07] lot of them use the term RSI and I'll be [2:09] like, "Oh, so what do you do?" And [2:10] they'll say, "We use AI to create AI. [2:12] That sounds very interesting. So what do [2:13] you do?" We use AI to create GPU [2:16] kernels. [2:17] >> Which that doesn't to me kind of match [2:19] the definition of what I would consider [2:20] to be RSI. So maybe can we start by you [2:22] defining what you think or what you [2:25] would include in RSI [2:27] >> definitionally? [2:28] >> Yeah, I think RSI is definitely an [2:30] overloaded term today. Um many people [2:33] when they say RSI they mean different [2:35] things. Um so I think the important [2:37] thing when you are differentiating them [2:39] is to ask the question what is the [2:42] optimizer and what is the optimiz like [2:44] when you're self-improving who is [2:46] improving right so um [2:48] >> there there are people who saying okay [2:50] we're doing RSI we have human using AI [2:53] to develop better AI [2:55] >> but that to us is just um autocatalysis [2:58] >> in a sense that the improver is still [3:00] human with um tools even though these [3:03] tools are getting better and better but [3:05] the idea and the improvement come from [3:07] human brains, [3:08] >> right? And and just I don't mean to [3:09] interrupt, but to be clear, most [3:12] technical epochs have had autoc [3:14] catalytic effects, right? Like there's [3:16] no way you can create a computer chip [3:17] without a computer, right? You could [3:19] argue that the internet development was [3:21] sped up by the fact that people that [3:24] weren't geographically located could [3:26] coordinate, right? I mean, the steam [3:27] engine, you could build better steam [3:29] engines because you could get coal [3:30] faster, etc. So, this notion of like [3:32] using the thing to build the thing [3:33] faster is not new. And so you're saying [3:36] something different. [3:37] >> Yes. I think for us um in in the history [3:40] right over civilization we have this [3:42] autoc catalysis you know computer using [3:44] computers to build better computers but [3:46] the the idea for this change for this [3:49] improvement come from human brain but [3:52] that human brain was never the improve [3:55] never the optimiz right whereas when now [3:58] for the first time we could swap the [4:01] optimizer as the optimiz so the the [4:04] brain of the agents which is LM the [4:07] model can be the thing that you're [4:08] optimizing and this leads to a compound [4:12] this reinvestment of cognitive [4:14] improvement into itself. [4:16] >> But there has to be a reason that you [4:17] decided to take on this hard problem as [4:19] opposed to what anybody else would do [4:20] which is like chew off pieces of the [4:22] problem. [4:22] >> Yeah. I mean I guess to answer before I [4:27] answer your more important question [4:28] here. The reason I did this is I think [4:30] it's a hard problem and it's fun working [4:32] on the hardest problem and I think [4:35] >> but do you have an intuition that this [4:37] is is the actual right answer as opposed [4:39] to like something that's more pragmatic? [4:41] >> Yeah, I think a pragmatic thing would be [4:43] just from a control theory perspective. [4:46] If you have a gap in the system, you [4:47] can't have a closed loop and [4:49] autocatalytic has no closed loop, right? [4:52] So whatever you improve, there's still [4:54] like a gap. humans have to come in. You [4:56] have to pour in other resources to make [4:58] it work. But if you have a closed loop, [5:01] it you still have to pour in resources, [5:03] but the the marginal gains ideally get [5:07] either constant or larger over time. And [5:09] that's that's the fundamental [5:10] difference. What's really different here [5:12] can be explained pretty easily. [5:14] >> So if you think of capability, [5:15] >> yeah, [5:16] >> what we're trying to do here is we're [5:17] trying to improve the rate of [5:20] improvement of capability, right? So [5:22] like the first derivative of capability [5:24] with respect to time is the thing that [5:26] everybody is trying to improve. Yeah. [5:28] Right. And now you can put that to be [5:32] equal driven by two things. The [5:35] nominator and the denominator. The nom [5:38] nominator would be performance. Yeah. [5:39] >> Which is what everybody is trying to [5:41] improve. And then the denominator is [5:44] some kind of cost. Let's say capital, [5:46] right? Or difficulty. And this one is [5:49] pretty is pretty set because the higher [5:52] the capability, the higher the [5:54] difficulty of the next step. Right now [5:57] the numerator is the interesting part [5:59] for things like steam engines and [6:01] whatnot that was fixed. [6:03] >> Yeah. [6:03] >> Right. So then you end up with a very [6:04] simple differential equation where you [6:06] basically plateau. [6:07] >> Yeah. [6:08] >> Right. Now if you have a system where [6:11] the improve the things being improved is [6:14] also the thing that's doing the [6:16] improvement [6:17] >> you how you now have a new way to think [6:20] about it and the the simplest case would [6:22] be that the numerator is also correlated [6:25] with capability which makes sense right [6:28] because the better you are the better [6:29] you are at also improving [6:31] >> and if you do that the math cancels out [6:33] and you basically get constant [6:35] improvement over time that's that could [6:38] one level of of RSI, right? And then [6:41] there's one more case which is [6:43] >> now that the the numerator could be [6:45] correlated with a higher power of [6:47] capability. Yeah. [6:48] >> And then you get exponential takeoff. [6:50] >> Yeah. Can we um dig in a little bit more [6:52] between RSIs having closed the loop and [6:56] actual convergence properties? Like for [6:59] some reason when people see a a graph [7:01] where there's like a a circle with an [7:03] arrow to another circle then and it goes [7:05] back and they look at that closed [7:06] control system then they say take off [7:09] [laughter] [7:09] >> right which [7:11] >> I mean if this is a nonlinear control [7:13] system [7:14] >> it's very difficult to know what the [7:16] hell happens by the way most people [7:18] don't seem to understand this [7:19] distinction [7:20] >> because we very quickly go to like take [7:22] off and fume and all that stuff when we [7:23] talk about these things. Well, as a [7:25] network guy, you know that it depends on [7:27] the exponent of, you know, your your [7:29] function. That's that's [7:31] self-referencing. [7:33] >> And it's [7:35] what we found is that it's almost not [7:39] useful to speculate on what we think [7:42] could happen because this is really an [7:45] empirical problem. And the real way to [7:48] know what's happening is to test things. [7:51] and testing things is really really hard [7:54] like reward hacking is one of you know [7:57] many problems that we have to deal with [7:59] but the the question you're asking is [8:01] very empirical and for some [8:03] >> interesting so but so I'm a networker [8:06] that's how I would have answered [8:08] >> I would have thought math people would [8:09] have been like we've got this we've got [8:11] this differential equation and I took [8:13] out my pencil and I wrote down the [8:14] answer [8:16] >> I mean it is possible to write down the [8:18] differential equation but you still have [8:19] to empirically diff measure the the [8:22] variable the exponent. [8:23] >> So there's no analytical solution. [8:25] >> I mean honestly we don't even have a [8:27] learning theory of deep learning. We [8:29] don't know like fundamentally why it [8:31] works, [8:32] >> right? And and then now you're like you [8:34] know five six layers on top of deep [8:36] learning and so we don't have a good [8:39] theory. I think the closest things we [8:41] have are scaling laws and these are [8:43] empirical laws. I think one [8:44] misunderstanding people have about like [8:47] our approach to RSI is that they think [8:49] the entire system has to be fully [8:52] automatic and digital. But actually if [8:54] you look at our mathematical foundation [8:56] we have a multi- aent system not just [8:59] one digital agent. And in in this multi- [9:02] aent system human researchers and [9:04] engineers are actually special case of [9:06] these agents and it's this hybrid system [9:09] of multiple agents working together [9:11] that's been recursively improved. [9:13] >> I see. And we don't want to leave out [9:15] any component to not optimize them [9:17] because like if it's possible to [9:19] optimize in the model why wouldn't we do [9:21] that because you will have more gains if [9:23] you do include those. Anything that you [9:25] don't optimize will later become your [9:27] bot. [9:27] >> So one observation that um uh uh we have [9:31] we have made here is um there are claims [9:36] of uh speed is improving which is [9:39] actually the case right? Right? So if [9:40] you look at like the model release time, [9:41] it is reducing. And so on one hand you [9:44] can be like okay so clearly we're in [9:46] some sort of a you know rapidly [9:48] improving phase. But if you divide by [9:51] the actual [9:53] amount of capital raised by these [9:55] companies that's no longer the case, [9:56] right? And so you're like well maybe [9:58] they just raise a whole bunch of money [9:59] and then like AI allows you to apply [10:01] that money. So, how would you, this is [10:04] not for yourself, but how would you [10:05] encourage people that are either [10:07] evaluating claims about this or even [10:09] working on them to think about not [10:13] fooling yourself by throwing more money [10:15] at the problem like you're actually [10:16] showing that you're making an [10:18] improvement here that's meaningful with [10:19] the denominator. [snorts] I think for us [10:22] it's it's pretty straightforward. It's [10:24] because if you get improvement and they [10:28] just come at more and more resources [10:30] being poured in, it's not RSI. You have [10:33] to be able to back your claim of how [10:36] much improvement you're getting. But do [10:38] include like all the manh hours and [10:40] experimentation that went into that [10:42] because I I feel like at some level like [10:45] that like it's like like they'll talk [10:46] about like the actual flops that go in [10:48] and money that go in but the reality is [10:49] this is like the 23rd experiment and [10:51] it's actually a human burning a lot of [10:53] mental capital to get to that final [10:55] configuration like do you actually [10:57] >> include that in the denominator? [11:01] >> I we do and I think it's it's good [11:03] company discipline to do that. [11:05] >> Yeah. And we've tried again to very [11:10] intentionally from the beginning have a [11:12] culture of tracking where things go cost [11:16] time all that kind of stuff. I share a [11:18] funny thing internally we have this [11:20] thing people talk about leaderboards you [11:23] know which is like score [11:24] >> you know what I'm about to say we have [11:26] something called a bleederboard [11:28] >> where we keep track of the cost of of [11:31] experiments of everybody who's running [11:33] experiments. [11:34] >> That's awesome. you know and and [11:35] everybody it's the comp everybody any [11:38] day can look at like who's who's high up [11:42] on the on the bleederboard. I can also [11:44] give an example from AI. Um I don't know [11:46] if you know this nano GPT benchmark they [11:49] are running. It's basically asking you [11:51] to retrain GPD2 which when it just came [11:54] out is a very expensive model to train [11:57] and takes weeks I think on multiple GPUs [12:00] >> and right now people can train on single [12:03] GPU within hours and that number has [12:06] been shrinking constantly. [12:07] >> Yeah. [12:08] >> Yeah. But that's also totally separate, [12:10] right? Like that's just like whatever [12:11] Moore's law like [12:13] >> No, no, it's the same GPU. It's the same [12:15] hardware [12:16] >> using the same hardware and then you [12:18] improve the techniques of pre-training. [12:20] >> Yeah. [12:20] >> Oh, interesting. [12:21] >> Yeah. What has improved is like [12:22] algorithmic [12:23] >> techniques, data techniques, kernels, [12:26] all of that kind of stuff. Do [12:27] >> you know how much it cost to train GPT2? [12:30] >> I think it's a few hundred. [12:31] >> $100. I just Oh wow. So the same same [12:34] GPU [12:35] >> same GPU same data [12:36] >> and this is just algorithmic and do [12:37] those come from humans or do those come [12:39] from [12:40] >> agents a lot of them are auto research [12:42] results [12:42] >> but that but that's definitely the [12:44] denominator that's that's not the [12:45] numerator like what you want to see is [12:47] like you can do that with GPD5 which [12:50] we're not seeing [laughter] right I mean [12:51] like the cost to these things is going [12:53] up dramatically [12:55] >> I think one of the most closely guarded [12:57] secrets in the industry is exactly this [13:00] thing which is how much compute you can [13:03] get out of this. [13:04] >> Yeah. [13:04] >> Because in some sense, even if all you [13:07] were improving was your denominator, you [13:10] could get faster training because now [13:11] you can get more rollouts. [13:12] >> Right. Yeah. Of course. [13:13] >> Right. And that means now you can sample [13:16] that what what you were describing like [13:18] the manifold of the space of of language [13:21] more extensively and then you can get [13:24] better trajectories which means you can [13:26] get smarter models when you use that to [13:27] train. [13:28] >> For sure improving the denominator is [13:29] good. [13:30] >> Yeah. [13:30] >> Yeah. [13:30] >> For sure. Improving the numerator is [13:32] good. Yeah, [13:33] >> I've yet to see which is great that you [13:35] have, you know, glimmers, uh, where you [13:37] can remove both the numerator and the [13:38] denominator. I mean, that's a pretty [13:40] meaningful statement, which is awesome. [13:42] What we found is that a lot of the [13:44] papers that claim to do RSI, when we [13:48] test it in our internal infra, we find [13:51] that they don't work. And a lot of it is [13:53] because they're cool ideas that humans [13:56] try to introspect and say, "Oh, you [13:58] know, this is what I think should [13:59] happen." [14:00] >> They're complicated. this sound cool, [14:02] they get you a good paper, but then when [14:04] we test it at scale, you know, on [14:06] benchmarks, on real economic problems, [14:08] they don't scale. And what we found is [14:10] that this the simpler thing that closes [14:12] the loop is the thing that scales. I [14:15] would say one thing that we do that's [14:17] probably unique is that we always start [14:21] with a very grounded mathematical [14:23] framework that that basically lays out [14:25] the map of the whole space. [14:26] >> Yeah. And the reason is because this [14:29] space is so large that if you don't do [14:32] that and you don't have this very [14:34] principle way of doing it, you get lost [14:36] very quickly in this complexity. [14:38] >> And the other reason is that often when [14:42] you have this really complex system, [14:44] there's two extremes of how you can [14:45] build it. You can take an approach where [14:48] you basically build like 200 different [14:50] systems [14:51] >> and you hope that like this this this [14:55] you know what is it the Rube Goldberg [14:57] machine will kind of work out. [14:58] >> Yes. [14:59] >> Or you take the other way and you you [15:01] kind of basically spend some time [15:02] thinking about these 200 things and [15:04] maybe you realize you know they're [15:06] actually just 12 um abstractions that [15:09] generalize to these 200 things. And so [15:12] what we have tried um to to create a [15:15] culture of let's really think from first [15:18] principles. Let's you know look at the [15:20] math and then we realize what are the [15:23] assumptions we make to be practical cuz [15:25] you have to do both and then we see what [15:28] works and and when you do that you then [15:30] can see that oh you know there's like [15:32] multiple papers or ideas are actually [15:35] you know subclasses of the same general [15:37] framework. [15:38] >> Yeah. And if you do that well, you can [15:40] scale. And I think um that's that's one [15:44] thing that I see that's missing and [15:47] that's hard to do and and takes a very [15:50] unusual group of people to to really [15:52] like make that work. [15:54] >> I think Richard Feman said once that [15:57] there are two kind of sciences there. [15:58] There's STEM collections and then there [16:01] are physics. [16:02] >> So without a unified math framework, all [16:05] these research are sort of just stem [16:06] collections. You could like a lot of [16:08] this paper we're reading they're just [16:10] ideas inspired by some human behavior [16:12] and then they try to mimic that and say [16:15] oh look we have RSI but then if you [16:17] don't put into a unified system you [16:19] don't know if the overall system is [16:21] actually improving or you're just like [16:23] locally applying an idea that that's [16:26] getting better. So what we have is that [16:28] we're putting all these components that [16:30] you could possibly recursively improve [16:33] into a unified framework and look at it [16:35] from a holistic view so that improving [16:38] one part um does not negatively impact [16:41] the other part. [16:42] >> Right. [16:43] >> I see. I see. [16:44] >> So so for example there are people who [16:45] are doing recursive improvement just on [16:47] harness and then there are people who [16:49] are only doing on wids but weights and [16:51] harnesses are not isolated component. [16:53] They have impact on each other. You can [16:55] think of RSI at many different levels of [16:58] the stack, right? You can think of um [17:01] prompts, um harness, weights, and you [17:04] can go all the way to basically chips, [17:07] right? [17:08] >> And our our goal here is to optimize all [17:11] the levels of the stack, right? And we [17:13] started with harness. We we've tried to [17:16] be always very capital efficient because [17:18] we care so much about the denominator. [17:20] And then we started doing weights, [17:23] weight updates. And what we found is [17:25] that you get more gains, but you still [17:28] have to do hardness stuff, right? And [17:30] you can keep going down the stack. And [17:32] and my guess is that you probably will [17:34] need to do a lot of the kind of [17:37] co-optimization of many different parts [17:39] of the system to make it work. Can you [17:42] talk a little bit about the problem [17:43] domain that you're going after just to [17:44] prove that this works? [17:46] Um, we are trying very hard to pick [17:52] areas of the AI space that are as close [17:56] as possible to AI itself. Again, we try [17:59] to be recursive. [18:00] >> And if you think about it, there's [18:02] there's not that many problems that are [18:05] super close to AI and gives you [18:07] improvement on both the numerator and [18:09] denominator. [18:11] >> And one of those is the inference [18:12] problem, [18:13] >> right? because the more you can spend on [18:18] um sorry the more you can get [18:20] improvements and and efficiency out of [18:23] inference the the more everything kind [18:25] of unlocks right [18:27] >> and it's not just um for technical [18:30] reasons but also the most promise of [18:33] both the numerator and the denominator [18:35] being improved right [18:36] >> it kind of breaks the mind right so like [18:38] just just for everybody that's listening [18:40] to this there's two improvements going [18:42] on here one of them is the ability of [18:44] the model to make inference faster. And [18:47] then the second one is now that the [18:49] inference is faster, it can do that [18:50] faster, right? [18:52] >> And again, that does not necessarily [18:54] mean takeoff that can still, you know, [18:57] as at some point. [18:59] >> Yeah. [18:59] >> But it's just wild to think of this kind [19:02] of compounding impact. [19:03] >> Yeah. Yeah. We're a startup. We're not a [19:05] research lab. And one of the reasons we [19:08] chose it is because everybody in the [19:10] world right now wants to serve their own [19:12] model. [19:12] >> Yeah. and and everybody wants to use the [19:16] latest for example open weight model Kim [19:19] 3 or whatever just coming out and they [19:22] want to be able to serve it as [19:24] efficiently as possible right so it's [19:26] it's an economically extremely important [19:28] problem right now [19:30] >> just for the sake of discussion if if I [19:32] if I wanted to make the inference for [19:35] Kimmy K3 faster would I use Kimmy K3 in [19:38] a recursive way or would I use a [19:40] separate model and system to do that for [19:43] Kim K3 [19:45] >> to improve the serving stack of Kim 3. [19:47] You mean yeah in our [19:49] >> well our system um is not just using one [19:54] LLM you have many different LLMs that [19:57] are doing many different things. [19:59] >> Oh so you're different the inference [20:02] side for all of them. Um I mean we again [20:05] it's it's we pick whichever is is the [20:07] most effective and because it it's being [20:09] used the most time but the interesting [20:12] thing is the choice of the of which LLM [20:14] to use for which thing is also not [20:17] always up to us. We let the model choose [20:20] to be honest today it's not really us [20:23] who are deciding what to optimize. It's [20:24] our agent because it's recursively [20:27] self-improvement and what we're trying [20:28] to do is not to limit itself to only one [20:31] component. So it chooses whether to [20:33] >> it. Yeah, we try to be like as hands off [20:35] as possible. Let it do its thing. [20:38] >> Do you have like a single master prompt? [20:40] Is it like improve yourself? [20:42] >> Uh but even that master prompt that has [20:45] been improving. [laughter] [20:46] >> Really? It improves a master prompt. [20:48] >> Yeah. And I mean and it's not one RSI [20:52] system because you can think of it as [20:54] there's the thing that that does the [20:57] improvement but then you also have to [21:00] have the thing that knows what to [21:01] improve and that's both a dynamic [21:04] problems and they are very co right. How [21:08] general do you think [21:11] this ends up being? Like right now [21:13] you're focused on speeding up inference [21:15] in a specific domain. Do you think that [21:17] whatever approaches you're coming up [21:19] with here generalize or do you think [21:22] there'll be a different one per [21:26] target solution space? [21:27] >> Mathematically we have defined it so [21:30] that it's one system that generalizes to [21:32] everything. I think the the one part of [21:35] the system [21:36] >> and by everything mean any problem [21:37] domain that you can point AI to. [21:39] >> Correct. But then there are physical [21:42] limits for example and and one big [21:44] limits of this is is the human element. [21:47] I think the way that a lot of AI [21:49] companies are thinking is you know let [21:52] let us use AI to to replace humans. But [21:55] what we found is that there's there's so [21:58] much that if you have good researchers [22:01] and good engineers that they can bring [22:02] to the table that's going to be very [22:05] hard to to replace by agents. And so [22:09] that's why from the very beginning we've [22:11] tried to think of our basically entire [22:14] company as a multi- aent system where [22:16] humans are a kind of agent that's [22:19] cohabitating with other AI agents. [22:21] >> Can we at least reduce bounds? So for [22:23] example, do you envision [22:25] do you envision a system where the [22:27] amount of energy that goes into it [22:31] somehow decreases over time relative to [22:33] the improvement? We can't share too much [22:34] but one of the very important things [22:37] that we've tried to do in our company is [22:40] not just work on the numerator which is [22:42] what I was describing earlier which is [22:44] your performance but to work on the [22:45] denominator which is how much you can [22:48] get out of the same amount of physics [22:51] money whatever you want to call it [22:53] >> right but do you think it's possible to [22:55] have a system where the numerator [22:57] continues to increase at the same way [23:00] with the denominator being fixed or [23:02] decreasing I I it's not just possible. I [23:05] think it's happening [23:07] >> really. [23:08] >> Yeah. [23:09] >> So it takes less energy to get better [23:12] results. [23:13] >> Yeah. Yeah. Over time. [23:14] >> Do you think there's a there's a [23:15] universe in which like that continues [23:19] without asmmptoing? Cuz if that was the [23:21] case then you actually have takeoff. [23:23] >> I mean [23:24] >> but it just seems to disobey information [23:26] theory and the laws of physics. [23:28] >> Yeah. Yeah. I mean I I think for at some [23:30] point you're going to reach plank [23:31] constant you know plank length plank [23:33] energy plank time right [23:35] >> so it it will asmtote at some point [23:38] >> but I think [23:39] >> oh I see and so you think that asmtote [23:41] is basically like the computation of the [23:42] universe [23:43] >> at least yeah I mean you know how many [23:44] how many particles are there 10 to the [23:46] 80 like we know speed of light you know [23:49] we know we know what it is you multiply [23:51] those two and you basically get an upper [23:53] bound of computation and and so given [23:55] this fixed amount of computation you can [23:57] do every [23:59] We know we have a limit. [24:00] >> I'm so impressed by your your belief [24:02] that like this idea of a model improving [24:04] itself, [24:05] >> but it's not a belief. [24:06] >> No, you got to tackle that. That's [24:07] amazing. I mean, [24:08] >> also it's a question of what metric you [24:10] use to to view the progress, right? So, [24:12] if you use um I don't know benchmark [24:15] accuracy, it's upward bounded by 100%. [24:18] You're never going to see exponential [24:19] improvement on that because eventually [24:21] bottleneck 100%. So, so for a long time [24:24] we had this question, how do we even [24:25] know if something has ticked off right [24:28] if we don't have the right metric? [24:29] >> It's not a belief that this is [24:31] happening. I think we're already seeing [24:33] glimmers of that happening. [24:34] >> But it is a dramatic statement which is [24:36] I think it's the first time I've heard [24:37] you say it. Wait, is this like a is this [24:40] been kind of, you know, a continuous [24:42] result or is it relatively recent that [24:44] you're starting to see glimmers of this? [24:45] No, we we've been seeing like in in [24:48] these results kind of grow and grow and [24:50] grow over the last few months. [24:52] >> That's amazing. [24:52] >> How do you have confidence in the [24:54] results given [24:56] you know like it's not working yet and [24:58] you don't know? [24:59] >> No, it's not belief. It's it's empirical [25:01] measurement. But but also as a as a [25:03] company there's a there's a very funny [25:05] thing that that is happening which is [25:08] because we have been so thorough, we [25:11] have not released things until we're [25:12] super sure it's correct, right? And [25:15] every few months, you know, there's like [25:17] an announcement of something RSI [25:19] related, we look at it, we're like, we [25:21] should have released our results or [25:22] something. And then we look at it, we [25:24] try to replicate it, and we realize, oh, [25:26] it's noise, you know, it's cherry [25:27] picking. And so, it's very very hard to [25:30] to know in this in this sea of, you [25:33] know, AI, I don't know what you want to [25:36] call it, like everybody's trying to do [25:39] AI, RSI stuff to know what's legit or [25:42] not. The the technical answer is you [25:45] have to be really really good at [25:47] measuring things. So for example, we [25:49] have one of our engineers that we we we [25:52] named him the viseroy of variance. [25:55] That's that's a real thing. Like his job [25:58] his job is to go and like destroy the [26:02] results we have and find what's wrong. [26:04] And we even have little like slack [26:06] emojis that we have. There's one where [26:08] like there's like red laser stuff coming [26:10] out of his eyes when like he finds [26:12] [laughter] that there's bad variance in [26:13] stuff. One of the things that we found [26:15] is that you know when when you're [26:17] dealing with a new benchmark or a new [26:20] problem of economic value that you want [26:22] to apply this to. So far we have not [26:24] found a replacement for what we call [26:26] quality time. literally running hundreds [26:28] of experiments looking at it, developing [26:31] an intuition, trying to figure out what [26:33] is wrong, what is not wrong, what what [26:35] is what is just variance, what is not, [26:38] what is this weird thing that actually [26:40] shows that there's, you know, some kind [26:42] of reward hacking like two levels down [26:44] the agent that built the agent that [26:46] build the agent [26:47] >> did hack something and now you have this [26:49] weird results, [26:49] >> right? [26:50] >> It's not easy. So when I have dealt with [26:53] you know nonlinear control systems [26:56] there's a bunch of in networking there's [26:58] a bunch of degenerate behavior you have [27:00] to deal with it's not it's not like just [27:01] convergence and divergence you have [27:02] things like oscillations [27:04] >> and this is a very multi-dimensional [27:06] problem so do you have to deal with all [27:07] of that stuff like potential [27:09] oscillations and [27:10] >> absolutely [27:11] >> meaning like it'll it'll get good and [27:12] then it'll get bad it'll get good [27:14] >> absolutely and then and then on top of [27:16] that you have noise are fundamentally [27:19] super stochastic [27:20] >> so what we find is that specific agent [27:22] on a specific benchmark. You know, [27:25] there's levels of noise that's sometimes [27:28] unpredictable and some of them are [27:31] extremely [27:33] like weird like you even have some weird [27:36] noise that comes from like this GPU and [27:38] this other GPU. They're both, you know, [27:40] B200, but they have slightly different [27:42] behaviors like on the physics and you [27:45] can see some of that being replicated at [27:47] the high level. So that's why it has to [27:50] be an empirical thing. [27:51] >> So from that you could already see that [27:53] intelligence is not a single scalar. [27:56] It's more like a multi-dimension vector [27:59] and there will be domains and field [28:02] where this RSI will give you take off [28:04] and there will be dimensions where you [28:07] might plateau or you will have [28:08] relatively slow improvement and I think [28:11] the the the defining factor is what is [28:15] the slowest component in it. So you're [28:18] an optimization um person. You know that [28:21] there's this amdos law where you know [28:24] like the component the the speed up you [28:26] can get by optimizing a component is [28:29] >> you just take it you just take it to [28:30] zero. Yeah. Nothing else dominates. [28:31] >> Exactly. So so the digital part we could [28:33] optimize and accelerate. You could get [28:35] like a thousand times faster or a [28:36] million times faster. But if it's only [28:38] 20% of your orbital time you're only [28:41] going to get 20% speed up. Right? So, so [28:44] for for domains where you still need to [28:46] move atoms that interact with physical [28:48] world that's really slow um you're still [28:51] going to be bottlenecked by that. So [28:54] let's say every evol every evaluation [28:56] you need to do let's say drug discovery [28:58] you still need to do wet lab experiments [29:00] and then clinical trials and that take [29:02] months then every new idea have to go [29:06] through a months or multiple months uh [29:09] evaluation. So those field probably [29:11] won't be accelerated as fast as you know [29:13] coding or mathematics where you have [29:15] very fast and cheap evaluations. [29:17] >> What I've I've yet to see is anybody put [29:19] a balance on that [29:20] >> like to say in order for me to [29:24] you know improve this by X to solve this [29:27] problem. You know I need this much net [29:30] new energy to do it right and if you had [29:32] that you could actually predict like if [29:34] I'm like listen like how big of a [29:36] warehouse do I need to solve like a drug [29:38] discovery problem or [29:40] how many different robots I would need [29:41] to to solve. So, do you think that we're [29:44] getting any closer to actually having [29:46] bounds? Because I think this is such a [29:48] critical question for all of us to [29:50] answer. [29:50] >> One of one of the internal jokes we have [29:52] in the company is that the the secret to [29:55] RSI is somewhere in a garden hose [29:58] company in Wisconsin. And and the joke [30:01] here is that uh if Einstein hadn't [30:03] discovered relativity. [30:05] >> Yeah. [30:06] >> He um would not have these these laws [30:09] that tell you how speed and you know [30:12] contraction of time works. [30:13] >> Yeah. [30:14] >> And if some unfortunate engineer had to [30:16] design the GPS system, they would have [30:18] had to find up find like a correction of [30:21] time and they probably would have said [30:23] like no no it's clearly not the fabric [30:24] of spaceime that's contracting, [30:26] [laughter] right? But I don't know why [30:28] but this is the formula I've devised [30:31] empirically [30:31] >> some [clears throat] you know like some [30:33] fudge [30:34] >> exactly right and then and then people [30:37] later would have been like holy [30:38] this is this is like space time [30:41] contracting right and and we have this [30:43] theory that you know like RSI right now [30:46] is is being thought of let's let's solve [30:50] all these like super hard [30:53] kind of um fashionable problem but what [30:55] we think is that the complexity of the [30:58] world is such that the the real [31:01] challenges probably exist in spaces that [31:03] are not being um applied to right and so [31:07] that's the joke that there is some [31:09] workflow somewhere in a garden host [31:11] company in Wisconsin that [31:13] >> that if we solve it we get we get you [31:15] know [31:15] >> I love that [31:16] >> I just want to quickly summarize so [31:18] you're going after RSI in in your case [31:20] this is not auto catalytic effects this [31:22] is actual RSI where the system improves [31:24] the full system and for you it's an [31:25] actual system which includes the harness [31:28] and the weights. Um, if you think about [31:30] the problem is you've got a numerator [31:32] and a denominator. The numerator is like [31:34] how good the system gets and the [31:35] denominator is how much um energy goes [31:38] into it. You believe you're seeing [31:39] glimmers of the sign where both improve [31:40] at the same time using this type of [31:42] approach. Is that correct? [31:43] >> Y [31:44] >> when can we uh uh expect to hear more [31:46] from you? [31:47] >> Um very soon. [31:49] >> Amazing. Yeah. Thanks so much both of [31:50] you for joining. [31:51] >> Yeah. Thank you.