Naveen Rao Unveils 4D Dynamical Chip to Cut AI Power Draw 1,000x

Naveen Rao: 4D Computing, AI's Energy Wall & Beating Biology — All-In Podcast

All-In Podcast

Replacing digital matrix multiplication with continuous oscillator physics eliminates memory-shuffling bottlenecks, enabling a 1,000-fold efficiency gain to prevent AI inference from crashing into a grid energy wall within three years.

AI-assisted summary of Naveen Rao: 4D Computing, AI's Energy Wall & Beating Biology — original source. Check the original for full context.

Source published:

Key ideas

  • Unconventional AI taped out the world's first physical dynamical computer in five months, demonstrating working image generation at approximately 500 nanojoules per generation—orders of magnitude below the millijoule range required by conventional GPUs. (15:28)
  • Naveen Rao warns that AI infrastructure will hit a hard electrical power constraint within three years, noting that electricity already drives roughly 50% of operational token costs while total US data center draw remains capped near 40 gigawatts. (4:45)
  • The company aims to achieve a 1,000x efficiency gain within three and a half years by replacing Von Neumann architectures and matrix multiplications with '4D computing' that exploits continuous oscillator physics. (3:23)
  • A commercial turnkey data center rack is slated for delivery within two years, designed to accept standard network tokens and convert existing frontier models at the whole-model layer via native Python abstractions. (19:31)

Detailed notes

## Background Published on YouTube by the All-In Podcast on September 21, 2026, this video features a presentation by deep-tech entrepreneur Naveen Rao followed by a fireside discussion with Chamath Palihapitiya. According to the show description, the talk examines AI's looming energy wall, 4D computing, the replacement of traditional computing abstractions with dynamical systems, and the commercial path toward productizing an entirely new class of hardware.

## TL;DR Naveen Rao, former founder of Nervana Systems and MosaicML, outlines his new venture Unconventional AI, which aims to break through AI's impending energy bottleneck by pursuing a 1,000x gain in power efficiency. Contrasting legacy von Neumann architectures—which waste massive amounts of energy endlessly shuffling bits—with biological brains operating on mere milliwatts, Rao advocates for eliminating digital abstractions in favor of continuous dynamical systems. His team demonstrated this paradigm with their open-source image model UNO and a taped-out physical 4D dynamical chip that slashes generation energy to 500 nanojoules without native matrix multiplication. Chamath Palihapitiya joins the stage to probe the commercial realities, revealing a two-year roadmap toward a turnkey data center rack programmed through high-level Python libraries rather than traditional CUDA kernels.

## Atmosphere & style The presentation carries a calm, intellectually commanding, and unapologetically ambitious atmosphere rooted in first-principles physics rather than standard tech hype. Rao speaks with quiet, seasoned authority, methodically contrasting the absurd energy draw of modern data centers against the elegant eight-milliwatt efficiency of a squirrel. When Chamath steps onstage, the dynamic shifts into an energized, inquisitive exchange; Chamath's visible astonishment at the audacity of 'beating biology' meets Rao's measured, pragmatic explanations of system integration and team dynamics.

## Bulletin - [15:28](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=928s) Unconventional AI taped out the world's first physical dynamical computer in five months, demonstrating working image generation at approximately 500 nanojoules per generation—orders of magnitude below the millijoule range required by conventional GPUs. - [04:45](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=285s) Naveen Rao warns that AI infrastructure will hit a hard electrical power constraint within three years, noting that electricity already drives roughly 50% of operational token costs while total US data center draw remains capped near 40 gigawatts. - [03:23](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=203s) The company aims to achieve a 1,000x efficiency gain within three and a half years by replacing Von Neumann architectures and matrix multiplications with '4D computing' that exploits continuous oscillator physics. - [19:31](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1171s) A commercial turnkey data center rack is slated for delivery within two years, designed to accept standard network tokens and convert existing frontier models at the whole-model layer via native Python abstractions.

## Key takeaways - [04:45](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=285s) AI scaling will hit an unavoidable energy wall within approximately three years because energy generation scales linearly while computational demand expands exponentially. - [04:45](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=285s) Data center constraints have shifted from square footage, networking, and GPU availability directly to power contracts, with electricity representing roughly 50% of the operating cost of serving a token. - [07:04](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=424s) Biological systems prove massive computational efficiency is physically possible: the human brain operates on 20 watts, while a squirrel executes real-time 3D spatial navigation on just 8 milliwatts. - [08:05](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=485s) The primary cause of AI compute inefficiency is data movement rather than calculation: while the human cortex transfers only 16 Gbps across 13–14 billion neurons, a high-end GPU shuffles 30 Tbps across external memory interfaces. - [10:24](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=624s) Digital computers waste immense energy by forcing continuous transistor physics into binary abstractions, linear algebra routines, and floating-point matrix multiplications. - [10:24](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=624s) Physical dynamical systems—such as mechanically coupled metronomes synchronizing across a movable surface—demonstrate that complex computation naturally emerges from simple, local physical rules without centralized control. - [13:05](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=785s) Testing with the UNO image generation model proved that introducing structural sparsity into coupled oscillator networks prevents quadratic connection scaling and dramatically improves training stability. - [15:28](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=928s) Unconventional AI fabricated a physical dynamical prototype in five months that completed image generation consuming only 500 nanojoules, drastically undercutting GPU energy consumption. - [15:28](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=928s) '4D computing' eliminates the Von Neumann memory-compute bus by collocating state and processing across three physical spatial dimensions alongside continuous temporal state dynamics. - [17:55](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1075s) Modern computing operates roughly 10 billion times (10¹⁰) away from thermodynamic physical limits, whereas mammalian biology operates within just one to two orders of magnitude of physical limits. - [17:55](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1075s) Under Jevons paradox, reducing inference energy costs by 1,000x will not contract industry revenue, but will instead dramatically expand overall demand and unlock planetary-scale autonomous compute. - [19:31](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1171s) The planned commercial release is a turnkey data center rack ready within two years, providing a drop-in token interface while executing internal operations through dynamical state transitions rather than native matrix multiplications. - [19:31](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1171s) Porting existing neural network models to dynamical hardware is done via an upfront transformation at the macro model layer rather than by mapping discrete arithmetic kernels. - [21:37](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1297s) Software programmability is delivered via high-level Python libraries tailored for time-varying states and physical stochastic dynamics, bypassing the need for a low-level CUDA-like instruction set.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=0s) Introduction and Naveen Rao's Background

Naveen Rao is introduced as an "outlier founder" who has successfully built and exited two major deep-tech companies, having previously scaled an enterprise from roughly $20 million to $700–$800 million. Taking the stage, Rao frames AI as the next fundamental evolution of humanity and expresses his enthusiasm for speaking at what he calls an "anti-doomer conference."

Reflecting on his journey, Rao shares that his interest in intelligent machines began in childhood. He started programming around the age of five or six after his family got a computer in 1978, viewing coding like solving puzzles. Driven by science fiction and a fascination with building intelligent machines, he trained as an electrical engineer. After an initial career designing traditional computers, he went back to academia to earn a PhD in neuroscience to study biological intelligence firsthand, aiming to apply those principles back to computing.

In 2014, long before artificial intelligence became a mainstream buzzword, Rao co-founded Nervana Systems, one of the earliest dedicated AI chip companies. Although he feels he sold Nervana to Intel "way too early," he went on to build and lead Intel's AI products group. By 2020, seeing the imminent rise of large language models, he turned his focus toward the scaling infrastructure needed to train them, co-founding MosaicML. MosaicML streamlined GPU platformization and erupted into high demand after ChatGPT launched in late 2022. Databricks acquired MosaicML in 2023, where its technology quickly grew to generate roughly one-quarter of Databricks' total revenue.

- **Nervana Systems**: Founded in 2014 as an early pioneer in AI hardware accelerators; acquired by Intel. - **MosaicML**: Founded in 2020 to build scalable infrastructure for training large AI models; acquired by Databricks in 2023. - **Databricks Revenue Impact**: The platform built by MosaicML contributes approximately 25% of Databricks' overall revenue.

---

[03:23](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=203s) Rethinking the Machine: Unconventional AI

Rao introduces his latest venture, Unconventional AI, which is built on the belief that traditional computing architectures must be rethought from first principles to break through the energy wall. While computers have operated fundamentally the same way for decades, Unconventional AI is pursuing a singular objective: achieving a 1,000x improvement in power efficiency. Rao originally set a five-year target for this milestone, but has accelerated the roadmap to three and a half years because AI-driven tools have enabled the team to solve deep scientific hurdles far faster than anticipated.

To achieve this, the company operates as a tightly integrated, "top-to-bottom" multidisciplinary organization:

```mermaid flowchart TD accTitle: Unconventional AI Organizational Stack accDescr: A top-to-bottom flow showing theorists feeding into modeling, circuit design, and final product packaging.

subgraph TheoryAndModeling["Theory & Validation"] A["Theorists<br>(Math & Theoretical Neuroscience PhDs)"] --> B["Model Builders<br>(Real Data & Benchmarks)"] end

subgraph HardwareExecution["Physical Implementation"] B --> C["Circuit Architects & Designers<br>(Physical Circuit Simulation)"] C --> D["Systems Engineering<br>(Boards, Systems & End Products)"] end ```

The stack begins with theorists—holders of doctorates in mathematics and theoretical neuroscience—who devise mathematical abstractions that inherently minimize data movement. Next, model builders validate these mathematical concepts against real datasets and empirical criteria. Once proven, physical circuit architects and chip designers implement and simulate these algorithms in custom silicon circuits, which systems engineers ultimately package into boards, complete server systems, and commercial products.

---

[04:45](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=285s) AI's Looming Energy Crisis

Rao argues that AI's expansion faces a hard physical constraint: available electrical power. Using Google as a reference because of its public disclosures, he highlights that Google processes over 3.2 quadrillion tokens every month. Assuming a conservative efficiency baseline of roughly 10 joules per token, that scale of inference consumes approximately 12 gigawatts of power.

``` Total Global Data Center Capacity: < 100 GW US Data Center Allocation: ~ 40 GW Google AI Token Processing Alone: 12 GW (at 10 J/token across 3.2Q tokens/mo) ```

Because total US data center draw is around 40 gigawatts—representing roughly half of global data center capacity (<100 GW)—a single hyperscaler's inference workload can theoretically demand a substantial fraction of the country's entire data center grid. With larger frontier models demanding more energy and user demand growing exponentially, Rao warns that the computing sector is on track to hit a hard energy wall within about three years. While market value is projected to reach $1 trillion or more by 2030, energy generation scales only linearly. Technology must bridge that widening divide.

This dynamic has fundamentally changed how data center infrastructure is developed and valued: 1. **Floor & Rack Space**: Early data centers were bottlenecked primarily by physical square footage and real estate. 2. **Networking**: The bottleneck shifted toward high-speed inter-rack switches and interconnect fabrics. 3. **Silicon (GPUs)**: Compute clusters were constrained by the supply and delivery of high-end GPUs. 4. **Power Contracts (Present Day)**: Operators secure multi-megawatt energy contracts first, and only then design racks and hardware to monetize every individual watt.

Today, electricity alone accounts for approximately 50% of the operational cost of serving an AI token, with the remainder split across capital expenditure (amortized hardware) and physical facilities. Unconventional AI's core economic proposition is simple: enabling data center operators to monetize their power contracts at 1,000x greater output per watt than legacy hardware.

---

[07:04](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=424s) Biology as Proof: The Efficiency of the Brain

To prove that orders-of-magnitude gains in computational efficiency are physically possible, Rao points to evolutionary biology. While modern clusters burn megawatts to run large models, the human brain performs general intelligence on just 20 watts of power. Rao highlights that the scaling across animal brains is even more astonishing:

- **Monkeys**: Operate on roughly 1 watt—roughly equivalent to the power draw of an average smartphone. - **Rats and Bats**: Function entirely within milliwatt power budgets. - **Squirrels**: Run their entire central nervous system on a mere 8 milliwatts of power.

Rao emphasizes the squirrel as a prime example of real-time physical precision: a squirrel can execute complex, high-speed spatial navigation between tree branches flawlessly a thousand times out of a thousand, powered by 8 milliwatts. A modern smartphone battery could theoretically power more than 100 complete squirrel brains simultaneously. Biology proves that a physical substrate can generate real-time intelligent behavior at microscopic energy costs.

---

[08:05](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=485s) The Cost of Moving Bits

Echoing the philosophy that "I don't feel like we truly understand something until we can create it," Rao explains that today's artificial intelligence systems are fundamentally inefficient because of how computers were originally designed. The overwhelming majority of energy in modern computers is spent not on computation itself, but on moving information back and forth between memory and compute units.

| System | Processing Elements | Data Moved In/Out of Memory | |:--- |:--- |:--- | | **Human Cortex** | ~13–14 billion neurons | ~16 billion bits per second (16 Gbps) | | **Modern High-End GPU** | Billions of transistors / streaming multiprocessors | ~30 trillion bits per second (30 Tbps) |

While the human cortex moves only 16 gigabits per second across its roughly 13 to 14 billion neurons, a single high-performance accelerator or GPU shuffles nearly 30 terabits per second across off-chip memory interfaces alone—with on-chip data movement running 10x to 100x higher still. This relentless shuffling of bits is the root cause of the AI energy wall, an artifact of digital architectures established in the 1930s and 1940s that modern silicon still largely follows.

[09:19](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=559s) The Limits of Von Neumann Architecture and Moore's Law

The foundational architecture of computing has not experienced a fundamental paradigm shift since its inception. Traditional machines rely on a structure where memory sits on the outside, separated from the computing core, requiring bits to be shuffled back and forth continuously. While this design yields very high execution speeds, it demands massive data movement. Computers were historically optimized exclusively for raw speed rather than energy efficiency. For example, the ENIAC—built in 1945 to calculate artillery trajectories—was created solely to outperform human "computers" who previously performed the arithmetic by hand. Today, computers are still marketed on speed metrics (such as one machine being twice as fast as another), without accounting for energy consumption.

For decades, the industry relied on packing more transistors onto silicon, but Dennard scaling halted frequency gains and single-thread performance growth, and efficiency gains have now stalled. Moore's Law—the predictable shrinkage of transistors—has largely reached its end, meaning engineers can no longer count on smaller nodes to deliver automatic efficiency improvements.

- **Historical Precedent**: ENIAC (1945) replaced human calculation teams purely to increase calculation speed. - **Physical Roadblocks**: Dennard scaling and single-thread scaling plateaued years ago, and the end of Moore's Law now limits physical efficiency improvements.

---

[10:24](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=624s) Cutting Out the Middleman: Physics, Abstractions, and Dynamical Systems

To bypass the energy wall, computing must "cut out the middleman" by stripping away layers of lossy abstraction. Digital logic—reducing the physical world strictly to ones and zeros—is itself an artificial construct; a physical transistor naturally occupies continuous intermediate states between high and low voltages, but engineers force it to behave binarily. Stacking higher-level software, neural networks, and linear algebra routines on top of this binary layer introduces compounding inefficiencies. In contrast, the human brain performs intelligence without floating-point math or matrix multiplication; cognitive processing emerges directly from the continuous physics of interacting neurons.

```mermaid flowchart TD accTitle: Conventional vs Dynamical Computing Stack accDescr: Flow diagram contrasting the layered abstractions of traditional computing with the direct physics approach of dynamical computing.

subgraph Conv["Traditional Digital AI Stack"] direction TB P1["Semiconductor Physics"] --> D1["Digital Abstraction: 0s & 1s"] D1 --> M1["Linear Algebra & Floating-Point"] M1 --> N1["Neural Networks & Generative AI"] end

subgraph Dynamic["Dynamical Computing Stack"] direction TB P2["Semiconductor Physics & Dynamics"] --> N2["Direct Neural Computation: UNO"] end ```

Computation exists throughout nature via dynamical systems theory, where complex, emergent behaviors arise from simple local rules without centralized instruction. Flocking birds evaluate only their immediate left and right neighbors to maintain coordinated flight, and ant colonies solve complex logistical problems without centralized oversight.

- **The Metronome Analogy**: If multiple out-of-phase mechanical metronomes are placed onto a shared rigid plank that rolls freely, their tiny physical forces push through the plank until every metronome locks into identical phase synchronization. - **System Variations**: Depending on system parameters and initial states, oscillators can settle into split-phase states or synchronized clusters, demonstrating self-organizing physical computation.

---

[13:05](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=785s) Model UNO and the Power of Sparsity

To demonstrate that physical oscillator networks could run generative AI, a model named UNO was released and open-sourced. UNO is an image generation model that simulates networks of coupled oscillators. When conditioned on target prompts (such as generating an airplane, car, or bird), the network settles along distinct state-space trajectories—mapping how the phases of the coupled oscillators evolve across time to yield the final generated visual output.

A central engineering challenge in coupled systems is scaling. Fully interconnecting elements results in quadratic growth:

$$\mathcal{O}(n^2)$$

Connecting 10 elements requires 100 connections ($10 \times 10$), while scaling to 1,000 elements explodes to 1,000,000 connections ($1,000 \times 1,000$). To solve this bottleneck, sparsity was introduced. Rather than degrading performance, pruning the vast majority of physical connections actually stabilized training dynamics and improved system performance. The resulting sparse architecture proved more scalable, more efficient, and easier to train in both software simulations and hardware implementations.

---

[15:28](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=928s) The First Physical Dynamical Computer and 4D Architecture

This research culminated in the development of the world's first physical dynamical computer. The company began operating in earnest in January without a complete team, yet taped out the physical chip design to the semiconductor foundry on June 1st—delivering a working prototype in just five months. The fabricated silicon returned to the lab and successfully generated its first functional images.

```mermaid flowchart LR accTitle: Von Neumann Architecture vs 4D Dynamical Computing accDescr: Architectural comparison between traditional Von Neumann memory transfer and collocated 4D computing.

subgraph VN["Von Neumann GPU / CPU"] direction LR Mem["External Memory"] <-->|"High-Energy Bus Shuffling"| Comp["Compute ALU"] end

subgraph FourD["4D Dynamical Machine"] direction TB Node["Unified Compute + Memory Element"] Time["Temporal Evolution: State Dynamics"] Node --- Time end ```

The physical prototype achieves unmatched power reduction:

- **Energy Consumption**: Generating an image requires roughly 500 nanojoules ($500\text{ nJ}$, where $1\text{ nJ} = 10^{-9}\text{ J}$). - **Comparison to Conventional GPUs**: Standard GPUs consume energy on the order of millijoules ($10^{-3}\text{ J}$) per equivalent generation task, making the dynamical chip multiple orders of magnitude more energy-efficient. - **Unified Architecture**: Unlike CPUs and GPUs—which adhere to the Von Neumann bottleneck by moving data across a high-energy memory bus—the dynamical computer fuses memory and compute into each individual element. - **"4D Computing" Defined**: The architecture harnesses all three physical spatial dimensions through planar routing and vertical die-stacking (3D), combined with the continuous time dimension of the dynamical state evolution (1D).

This architecture establishes a path toward building computing systems that approach fundamental thermodynamic limits, focusing directly on maximizing "intelligence per watt."

[17:55](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1075s) Thermodynamic Limits, Beating Biology, and Jevons Paradox

Naveen Rao frames the physical horizon of computing by contrasting biological evolution with silicon engineering. While mammalian brains operate within just one or two orders of magnitude of the ultimate thermodynamic limits of physics, modern electronic computing sits squarely on the far left of the efficiency curve—roughly 10 billion times ($10^{10}$) less efficient. Rao projects that within three and a half years, his company can push silicon to the physical limits of 2D lithography. The foundational thesis is audacious: "The overarching goal of this company is to beat biology." Reaching that threshold would unlock ubiquitous intelligence, allowing compute to permeate everything from decentralized infrastructure to dynamic swarms of billions of autonomous robots that assemble on demand to tackle planetary-scale challenges.

Such a drastic leap in efficiency would fundamentally restructure the physical footprint of artificial intelligence. Instead of constructing monolithic, gigawatt-scale data center facilities that stress power grids, computing could decentralize into smaller, adaptive, and environmentally friendly local nodes. Addressing the economic implications, Rao invokes Jevons paradox to show why a radical collapse in inference costs will not shrink industry revenues:

> "There's a concept called Jevons paradox where when you drop the underlying cost of an asset, you actually consume more than the drop of that asset. So, if you make something half the price, you'll consume more than 2x. If you make something 1,000th the price, you'll consume more than 1,000th of it. And I think this will create the largest market that humanity's ever seen."

```mermaid flowchart TD accTitle: The Path to Beating Biological Efficiency accDescr: Progression from current computing inefficiency to the thermodynamic limits of biology and beyond.

subgraph Efficiency["Physical Computing Horizons"] A["Current 2D Computing<br/>10¹⁰ from thermodynamic limit"] --> B["3.5-Year Target<br/>Limits of 2D lithography"] B --> C["Mammalian Brains<br/>1–2 orders of magnitude away"] C --> D["Ultimate Goal<br/>'Beat Biology'"] end

subgraph Impact["Systemic Ramifications"] D --> E["Decentralized Local Compute<br/>Replaces gigawatt mega-facilities"] D --> F["Jevons Paradox Unleashed<br/>1000x cost drop expands total spend"] end ```

- **Thermodynamic Gap:** Mammalian brains run roughly 10 to 100 times above physical limits, whereas modern silicon chips lag by a factor of 10,000,000,000. - **2D Lithography Roadmap:** The engineering plan aims to reach the absolute limits of two-dimensional lithography within 3.5 years. - **Decentralization Trend:** Compute architecture is expected to transition away from centralized gigawatt facilities into widespread, localized, and resilient data centers.

---

[19:31](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1171s) Chamath Joins: The Data Center Rack and Abandoning MatMul

Chamath Palihapitiya joins the stage, visibly taken aback by the ambition of the presentation ("I mean that was extremely unexpected. I got to say that was pretty amazing"). He immediately presses Rao on the commercial reality of bringing such an architecture to market, pointing out that Jensen Huang and NVIDIA rely on a massive, coordinated web of foundries, packagers, and system integrators. Chamath asks what it will realistically take in time and complexity to turn this into a usable product. Rao responds that they are within two years of delivering a complete product: a turnkey data center rack system. From the customer's perspective, the interface remains straightforward—"tokens in, tokens out through a network cable"—yet the computational mechanics inside the chassis represent an absolute departure from von Neumann and GPU designs.

Chamath pushes further on backward compatibility, asking whether the tech can accommodate prevailing industry architectures that have spent years hyper-optimizing around KV caching and mechanically reductive abstractions. Rao acknowledges that while switching always entails a tradeoff between efficiency gains and migration friction, they designed the system to make adoption irresistible. Crucially, the porting process does not happen at the micro-operations or kernel level; instead, existing models are converted at the higher model layer, a procedure that requires a dedicated upfront compute pass. When Chamath asks whether elemental primitives like standard matrix multiplications (MatMul) exist on the silicon, Rao reveals they do not exist natively:

> "You can characterize it as MatMul, but it doesn't implement it as MatMul... It implements it as a sort of time-varying behavior. But each one of those time steps you can analyze as basically a matrix of the current state times a transition matrix."

- **Time to Market:** Projected full product availability within two years. - **Form Factor:** Delivered as an integrated data center rack system serving tokens directly over high-speed networking. - **Abstraction Level:** Legacy models port at the whole-model layer rather than compiling discrete operation kernels. - **Non-MatMul Execution:** Replaces traditional arithmetic logic units with continuous dynamical state transitions ($S_{t+1} = S_t \times T$).

---

[21:37](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1297s) Assembling the Team and Building the Software Bridge

Chamath asks about the human capital required to build a machine so detached from standard computer architecture: "When you're building a team like this, who are these people? These are biologists plus physicists plus what?" Rao explains that the core challenge lies in marrying two disciplines that have historically lived in total isolation: theorists specializing in dynamical systems theory—a mathematical field spanning over a century—and hard-nosed silicon chip designers. Getting pure mathematicians and physical hardware engineers to align, communicate, and construct a unified system has been one of the steepest organizational hurdles inside the company.

Chamath inquires about the software translation layer, asking what serves as their CUDA equivalent to bridge abstract model definitions down to this radical silicon. Rao shares that rather than inventing a convoluted low-level instruction set or a proprietary variant of CUDA, they have built high-level runtime libraries in Python. This software layer provides researchers with a familiar environment while offering specialized primitives designed to express time-varying system states and stochastic physical behaviors directly. Chamath closes the session by praising the raw scope of the undertaking, calling the initiative "incredibly impressive" and "so ambitious."

- **Cross-Disciplinary Team:** Blends 100-year-old mathematical dynamical systems theory with modern semiconductor silicon design. - **Programmability:** High-level Python libraries replace low-level CUDA kernels, enabling developers to declare stochastic behaviors and continuous time steps directly.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=0s) anti-doomer conference — Naveen Rao

> [00:00](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=0s) way too early — Naveen Rao

> [08:05](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=485s) I don't feel like we truly understand something until we can create it — Naveen Rao

> [17:55](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1075s) The overarching goal of this company is to beat biology. — Naveen Rao

> [17:55](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1075s) There's a concept called Jevons paradox where when you drop the underlying cost of an asset, you actually consume more than the drop of that asset. So, if you make something half the price, you'll consume more than 2x. If you make something 1,000th the price, you'll consume more than 1,000th of it. And I think this will create the largest market that humanity's ever seen. — Naveen Rao

> [19:31](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1171s) I mean that was extremely unexpected. I got to say that was pretty amazing — Chamath Palihapitiya

> [19:31](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1171s) tokens in, tokens out through a network cable — Naveen Rao

> [19:31](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1171s) You can characterize it as MatMul, but it doesn't implement it as MatMul... It implements it as a sort of time-varying behavior. But each one of those time steps you can analyze as basically a matrix of the current state times a transition matrix. — Naveen Rao

> [21:37](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1297s) When you're building a team like this, who are these people? These are biologists plus physicists plus what? — Chamath Palihapitiya

> [21:37](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1297s) incredibly impressive — Chamath Palihapitiya

> [21:37](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1297s) so ambitious — Chamath Palihapitiya

## Mentioned Naveen Rao, Chamath Palihapitiya, Jensen Huang, Nervana Systems, Intel, MosaicML, Databricks, ChatGPT, Unconventional AI, Google, ENIAC, UNO, NVIDIA, CUDA, Python

## Video description (0:00) Welcome Naveen Rao (4:45) Is energy really the problem? The cost of a token, power contracts & the gap to close (10:24) Cutting out the middleman: abstractions, dynamical systems & a new kind of machine (19:40) Chamath joins: the path to product, porting existing models & building the team

Thanks to our partners for making this possible!

IREN is a vertically integrated AI Cloud platform, delivering data centers, compute and software for AI training and inference. https://iren.com/

Oracle connects the data, applications, and infrastructure that turn AI into business outcomes—with the flexibility, choice, and control to optimize as AI evolves. http://oracle.com/ai EY helps tech innovators scale from startup to exit to megacap. You build the future. We’ll handle the rest. http://www.ey.com

Meta believes the future is for everyone. We're focused on giving every person the tools to reach their full potential and making sure the benefits of technology are distributed to all. http://www.meta.com

Keel Infrastructure owns the power, land, and connectivity that HPC and AI run on - backed by secured energy assets and established grid interconnections across North America. https://keelinfra.com/

Airwallex - Agentic Global Business Accounts. Open local accounts in 70+ countries to accept payments, earn yield, pay globally, and manage spend. http://airwallex.com

PayPal has been revolutionizing commerce globally for more than 25 years. Creating innovative experiences that make moving money, selling, and shopping simple, personalized, and secure, PayPal empowers consumers and businesses in approximately 200 markets to join and thrive in the global economy. For more information, visit https://www.paypal.com

Google for Startups connects founders with the right people, products, and best practices to help startups build faster and go further. https://startup.google.com/

Explore ideas, industries, and technologies worth understanding with Chamath every week on Learn with Me: https://research.socialcapital.com/allin

Follow the besties: https://x.com/chamath https://x.com/Jason https://x.com/DavidSacks https://x.com/friedberg

Follow on X: https://x.com/theallinpod

Follow on Instagram: https://www.instagram.com/theallinpod

Follow on TikTok: https://www.tiktok.com/@allin

Follow on LinkedIn: https://www.linkedin.com/company/allinpod

Intro Music Credit: https://rb.gy/tppkzl https://x.com/yung_spielburg

#allin #tech #news

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

中文简报

Naveen Rao公布4D动力学芯片，拟将AI能耗降低1000倍 以连续振荡器物理学替代数字矩阵乘法消除了内存数据搬运瓶颈，实现1000倍能效提升，避免AI推理在三年内撞上电网能源之墙。 Nervana Systems与MosaicML的前创始人Naveen Rao介绍了其新创办的Unconventional AI，该公司旨在通过追求1000倍的能效提升，突破AI即将面临的能源瓶颈。Rao将因不断搬运比特而浪费大量能源的传统冯·诺依曼架构与仅以毫瓦级功耗运行的生物大脑进行了对比，主张摒弃数字抽象，转而采用连续动力学系统。其团队通过开源图像模型UNO以及一款已流片的物理4D动力学芯片展示了这一范式，该芯片在无需原生矩阵乘法的情况下，将单次生成的能耗大幅削减至500纳焦耳。Chamath Palihapitiya随后登台探讨其商业落地现实，透露了一份为期两年的路线图，计划推出可通过高级Python库而非传统CUDA内核进行编程的开箱即用型数据中心机架。 Unconventional AI在五个月内完成了全球首款物理动力学计算机的流片，实现了单次生成约500纳焦耳的可用图像生成能耗，比传统GPU所需的毫焦耳级别低了数个数量级。 (15:28) Naveen Rao警告称，AI基础设施将在三年内遭遇严峻的电力硬约束，并指出电力已占token运营成本的大约50%，而全美数据中心总耗电量仍被限制在约40吉瓦。 (4:45) 该公司计划在三年半内通过利用连续振荡器物理特性的“4D计算”来取代冯·诺依曼架构和矩阵乘法，从而实现1000倍的能效提升。 (3:23) 一款商业化的开箱即用型数据中心机架计划在两年内交付，其设计旨在接收标准网络token，并通过原生Python抽象在全模型层面上对现有前沿模型进行转换。 (19:31)

Reading summary

Naveen Rao argues that artificial intelligence is on track to hit an insurmountable electrical power constraint within three years unless computing abandons legacy Von Neumann architectures. To overcome this limit, his venture Unconventional AI is building physical dynamical systems that merge memory and continuous physics, targeting a 1,000x improvement in energy efficiency to match and eventually exceed biological intelligence. The Impending Grid Limit on AI Scaling (4:45) AI expansion faces an imminent physical bottleneck in electrical generation. Rao notes that Google alone processes 3.2 quadrillion tokens monthly, which requires an estimated 12 gigawatts under a conservative baseline of 10 joules per token. Given that total US data center draw is approximately 40 gigawatts—out of less than 100 gigawatts globally—a single hyperscaler's inference workload can theoretically consume a large share of national capacity. Because grid capacity scales only linearly while AI demand grows exponentially, Rao forecasts a hard energy wall within about three years. This shift has fundamentally altered infrastructure economics. Data center bottlenecks have evolved from physical real estate and networking fabrics to GPU allocation, and ultimately to power contracts. Electricity now represents approximately 50% of the operational cost of serving an AI token, forcing operators to prioritize revenue per watt over raw hardware counts. The Von Neumann Data Movement Bottleneck (8:05) Modern silicon inefficiency stems from foundational computer design choices made in the 1930s and 1940s, where architectures like the ENIAC prioritized pure calculation speed over energy efficiency. Conventional Von Neumann systems separate processing cores from memory, consuming the vast majority of energy simply shuttling bits across interconnects rather than performing mathematical operations. To illustrate this disparity, Rao contrasts digital processors with biology: the human cortex coordinates roughly 13 to 14 billion neurons while moving only 16 gigabits per second, whereas a single high-performance GPU transfers nearly 30 terabits per second across off-chip memory interfaces alone. With Dennard scaling halted and Moore's Law largely exhausted, engineers can no longer rely on standard semiconductor shrinkage to resolve this physical barrier. Dynamical 4D Systems and Physics-Based Compute (10:24) Biology proves that low-power intelligence is physically achievable: the human brain operates on 20 watts, and a squirrel executes precise spatial navigation on just 8 milliwatts. Rao contends that conventional computing loses efficiency by imposing artificial digital abstractions—forcing continuous transistor voltages into binary ones and zeros, layered beneath floating-point matrix multiplications. In contrast, natural systems compute via dynamical systems theory, where complex behaviors emerge from local interactions, such as coupled metronomes synchronizing on a movable surface. Unconventional AI validated this approach with UNO, an image model simulating sparse networks of coupled oscillators settling into state-space trajectories. Translating this into silicon, the company taped out a prototype in five months. The chip fuses memory and processing natively across three physical dimensions and continuous time (termed 4D computing). The fabricated prototype consumes roughly 500 nanojoules per image generation task, compared to millijoules on conventional GPUs. Commercial Rack Deployment and Jevons Paradox (19:31) Addressing commercial feasibility, Rao projects shipping a turnkey data center rack within two years that will interface with existing infrastructure via standard network connections ('tokens in, tokens out'). The silicon does not execute native matrix multiplications (MatMul); instead, computation occurs through continuous dynamical state transitions ($S_{t+1} = S_t \times T$). Existing models must undergo a high-level conversion pass, programmed through high-level Python libraries rather than low-level CUDA kernels. Rao claims silicon can reach the limits of 2D lithography within three and a half years and ultimately operate closer to thermodynamic boundaries than mammalian brains. Invoking Jevons paradox, Rao argues that reducing token delivery costs by 1,000x will not diminish sector revenues; rather, cheaper compute will spur hyper-proportional demand, decentralizing monolithic data centers into ubiquitous, distributed applications. While Unconventional AI demonstrated initial image generation on silicon in five months, the commercial viability of the platform hinges on whether high-level model translation can seamlessly offset the abandonment of native MatMul execution and displace entrenched CUDA ecosystems within a two-year commercialization window.

精炼阅读摘要

芯片与系统创业者Naveen Rao指出，现代人工智能正逼近传统电子计算与数据中心电网的物理硬墙；为此，其创办的Unconventional AI主张抛弃冯·诺伊曼架构与传统矩阵乘法，采用直接利用连续物理规律的“四维动力学计算”架构，目标在数年内将计算能效提升1000倍并最终超越生物系统的能效极限。 电力墙与数据搬运的结构性危机 (4:45) Rao强调，人工智能的持续扩张正面临不可逾越的供电物理瓶颈。根据公开数据，Google每月处理超过3.2万亿（3.2 quadrillion）个Token，即便按每个Token消耗10焦耳的保守基线推算，单是其AI推理工作负载就需消耗约12吉瓦（GW）电力。目前全美数据中心总电力容量约为40吉瓦，全球亦不足100吉瓦，这意味着单一超大规模服务商的推理需求即可占去全美数据中心电网的相当份额。在电力供给仅能线性增长的现状下，Rao预计计算行业将在约三年内撞上严重的“能源墙”，而电力成本已攀升至单Token运营支出的约50%，促使数据中心的核心竞争从早期争夺占地面积和GPU硬件，全面转变为对兆瓦级电力合同的每瓦变现能力。 这一危机的根源并不在算术运算本身，而在于现代芯片所依赖的冯·诺伊曼体系。自1945年ENIAC问世以来，计算机始终将内存置于外部，必须在计算核心与存储单元之间频繁往返搬运信息。数据显示，拥有130亿至140亿神经元的人类大脑皮层每秒仅移动约16吉比特（16 Gbps）数据，而单颗高端GPU仅在片外接口搬运的数据就逼近每秒30太比特（30 Tbps），片内搬运更超出数十至数百倍。随着丹纳德缩放定律的失效和摩尔定律的停滞，单纯依靠缩小晶体管尺寸已无法带来自动化能效提升，海量比特的低效移动彻底卡死了能效上限。 生物能效参照与动力学系统计算 (7:04) 为证明大幅降低能耗并非天方夜谭，Rao以演化生物学作为物理证据：人类大脑运行通用智能仅消耗约20瓦功率；猴子大脑仅耗电1瓦，相当于普通智能手机的功耗；而松鼠仅凭8毫瓦即可驱动中枢神经系统，在千百次树枝跳跃中保持零失误的实时高精度空间导航。相比之下，哺乳动物大脑的工作能效距离物理学绝对热力学极限仅差1至2个数量级，而现存电子计算机的能效与该极限相差达100亿倍（10^10），说明利用物理基质以极低微能量生成智能行为在物理上完全可行。 Rao认为破解能效壁垒的关键在于“剔除中间环节”，即绕开强制将连续电压抽象为二进制0和1的数字逻辑，以及构建于其上的浮点矩阵数学层。自然界的动态系统（如依靠局部感知协作飞行的鸟群，或放在滚动平板上借助微小振动实现相位自同步的机械节拍器）展示了无须集中指令即可涌现全局计算的特征。团队据此开源了名为UNO的图像生成模型，通过模拟耦合振荡器网络在状态空间中的相位随时间演化生成画面。针对振荡器全连接带来的O(n^2)连接数爆炸难题，团队引入稀疏化设计，证实剪枝绝大多数物理连接不仅不会削弱性能，反而能稳定训练并显著改善系统扩展能力。 四维芯片原型与原生非矩阵架构 (15:28) 在工程实现上，Unconventional AI仅耗时5个月便完成物理芯片设计并流片。其实验室原型硅片成功跑通了图像生成，单次生成图像仅消耗约500纳焦耳（500 nJ）能量，相较传统GPU毫焦耳（mJ）级别的同类生成消耗实现了数个数量级的下降。该架构消除了传统内存总线的高功耗传输，将存储与计算完全融合进同一微观单元，被定义为“4D计算”——即结合3D物理实体（芯片平面布线与垂直晶圆堆叠）与1D连续时间维度的动力学状态演化。 在底层执行逻辑上，该芯片彻底抛弃了传统算术逻辑单元所必需的矩阵乘法（MatMul）。系统原生并不执行矩阵相乘，而是依赖物理状态随时间的动力学连续变迁（分析上等效于当前状态矩阵乘以转移矩阵）。面对商业化落地的兼容性质疑，Rao表示系统定位为交钥匙式数据中心机架，对外保持标准的以太网输入输出Token接口，且不重写类似CUDA的底层内核，而是通过Python高级运行时库表达时变与随机物理状态，并在全模型层面完成上层既有模型的静态转换与适配。 商业重构、物理极限与杰文斯悖论 (17:55) 针对技术推广的经济影响，Rao援引经济学中的“杰文斯悖论”（Jevons Paradox）反驳了推理成本下降会削减行业收入的担忧。他认为，当基础计算资产的使用成本呈断崖式下跌（例如降至原有成本的千分之一）时，市场对其消费量的激发将远超降价倍数，从而催生人类历史上规模空前的计算市场空间。同时，极端能效的达成将深刻改变AI物理设施的分布结构，使算力摆脱对集中式吉瓦级超大电网的依赖，走向本地化、高弹性的边缘节点乃至可即时编队的数十亿自主机器人集群。 在技术路线规划上，Rao声称其工程目标是在三年半内将硅基芯片推进至2D光刻技术的物理终局，并在两年内交付完整的数据中心机架产品，最终愿景是制造出能效“超越生物学”的机器。这一跨界工程面临着巨大的团队协同挑战，其研发体系不仅需要数学家与理论神经科学家构建最小化数据搬运的抽象模型，还需要将其与具有百年历史的动力学系统理论同硬核芯片物理设计深度融合。 Rao勾勒的技术路线图极为激进，但其能否在宣称的两年至三年半窗口期内兑现仍面临多重未决挑战：将百年动力学理论与现代半导体制造工艺融合的跨学科协作难度极高；上层既有模型在非MatMul硬件上的转换编译仍需要昂贵的前期计算投入；同时，在缺乏像CUDA那样历经十数年沉淀的庞大软件生态与开发者惯性面前，其Python抽象库与专用机架能否顺利说服超大规模客户承受架构切换的摩擦成本，依然是未经验证的商业假设。

Detailed notes — zh

## Background 本期视频由知名科技与创投播客 All-In Podcast 于 2026 年 9 月 21 日发布在 YouTube 平台。视频简介显示，本期访谈围绕 Unconventional AI 创始人 Naveen Rao 的全新探索展开，深入探讨了 4D 计算、AI 所面临的能源瓶颈、动力系统计算如何打破传统抽象层，以及将现有模型迁移至新型硬件并打造商用系统的实际路径。 ## TL;DR 连续成功创办 Nervana 与 MosaicML 的芯片与 AI 先驱 Naveen Rao 带来了他的第三次创业突破——Unconventional AI。面对 AI 发展因数据搬运与冯·诺依曼架构局限而即将撞上的全球能源高墙，团队跳出传统的数字逻辑与浮点矩阵抽象，从动力系统理论（Dynamical Systems Theory）出发，在短短 5 个月内成功研制出全球首款物理动力学芯片。该芯片通过三维物理结构与连续时间演化实现“4D 计算”，生成单张图像仅耗费约 500 纳焦耳，实现了千倍级别的能效跨越，团队计划在两年内以整机柜形式交付这一直面热力学极限并力求“超越生物系统”的颠覆性算力架构。 ## Atmosphere & style 全场对话洋溢着纯粹的硬核极客精神与强烈的反“AI 毁灭论”乐观主义，既有从神经科学与物理第一性原理切入的深邃理论推演，又有短兵相接、数月内完成芯片流片的工程魄力。Naveen 举重若轻地用松鼠大脑能耗、节拍器共振等鲜活意象拆解深奥的前沿科学，台风沉稳而笃定。随着 Chamath 登台就芯片落地生态与底层 MatMul 替代机制展开层层追问，现场气氛在务实的商业交锋与宏大的技术愿景间激荡，让观众切实感受到一种正在亲眼见证半导体计算范式被彻底重构的智力震颤与澎湃雄心。 ## Bulletin - [15:36](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=936s) Naveen Rao 创办的 Unconventional AI 仅历时 5 个月便完成设计并点亮首款物理动力学芯片原型，单张图像生成能耗低至约 500 纳焦耳，相比传统高端 GPU 实现了千倍以上的能效跃迁。 - [04:45](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=285s) AI 算力正面临严峻的物理能源高墙：仅 Google 每月推理的 3.2 千万亿 Token 若按每 Token 10 焦耳测算便需 12 GW 电力，占全美数据中心总负荷近三分之一，行业预计将在 3 年内撞上电力供给上限。 - [16:50](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1010s) 团队提出“4D 计算”新范式，将 3D 立体晶圆堆叠与 1D 连续时间动力学演化相结合，从物理底层彻底消除分立内存接口与数据搬运损耗。 - [19:40](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1180s) 商业落地方面，Unconventional AI 计划在 2 年内推出面向数据中心的完整机柜系统，提供“Token 输入、Token 输出”的标准接口，并通过 Python 抽象层与模型层编译降低生态迁移成本。 ## Key takeaways - [04:45](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=285s) AI 产业的核心矛盾已演变为指数级爆发的算力需求与线性爬升的物理电网之间的剪刀差，数据中心建设的最前置瓶颈已全面从土地和芯片转向电力合同配额。 - [04:45](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=285s) 在单个 Token 的综合成本构成中，电力开销已占据 50%，推动算力行业的核心评估标准从纯粹的运算速度转向“每瓦特智能”（Intelligence per Watt）。 - [07:04](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=424s) 生物大脑展示了智能的高效物理载体，松鼠大脑仅凭 8 毫瓦功率即可完成高难度空间导航，而人类大脑皮层的数据搬运吞吐量（16 Gbps）远低于现代顶配 GPU 的片外传输量（30 Tbps）。 - [09:19](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=559s) 传统数字系统陷入能耗泥潭的根源在于冯·诺依曼架构的分离设计，以及层层人为叠加的有损抽象（离散 0/1 逻辑、浮点运算与线性代数），加之摩尔定律物理微缩红利已基本终结。 - [10:24](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=624s) 动力系统理论（Dynamical Systems Theory）提供了一种“砍掉中间抽象层”的途径，利用半导体器件连续的物理特性（如振荡器相互作用）直接涌现智能，无需执行显式矩阵乘法。 - [12:04](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=724s) 开源模型 UNO 证明了动力学计算的可行性：通过在状态空间轨迹中自然收敛至动力学吸引子，耦合振荡器系统能够在没有传统逐层数字计算的情况下稳定完成图像生成任务。 - [14:11](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=851s) 稀疏性（Sparsity）理论突破解决了物理器件全互联带来的 N^2 扩展难题，剪除绝大部分物理连线不仅降低了功耗，反而提升了系统的可训练性与全局表现。 - [15:36](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=936s) Unconventional AI 首款物理动力学芯片仅用 5 个月即完成流片与回片验证，单图生成能耗降至约 500 纳焦耳，相比 GPU 的毫焦耳级开销实现了数量级跨越。 - [16:50](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1010s) “4D 计算”彻底抛弃了外置内存接口与高频数字时钟，利用 3D 立体芯片堆叠承载紧密网络，将时间作为微分方程演化的内生维度，实现存算天然合一。 - [17:55](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1075s) 现代数字计算距物理热力学极限差了 100 亿倍（10^10），而生物大脑仅差 1 到 2 个数量级；团队目标在 3.5 年内逼近 2D 光刻物理极限，最终实现“超越生物系统”。 - [17:55](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1075s) 根据杰文斯悖论，当单位 AI 算力成本下降至千分之一时，下游应用需求将呈指数级井喷，推动集中式吉瓦级机房向分布式自适应微型数据中心和百亿协作机器人网络演进。 - [19:40](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1180s) 商用落地定于 2 年内推出整机柜数据中心方案，对外保持标准网络接口，内部完全替换传统计算硬件，在模型层进行统一编译迁移而无需开发者手工重写底层算子。 - [21:37](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1297s) 为弥合动力学理论家与芯片硬件工程团队的认知断层，团队基于 Python 打造了专属软件抽象层，允许算法开发者以熟悉的高级语法描述具有随机特性的时间演化单元。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=0s) 开场介绍与 Naveen Rao 的个人历程：从神经科学到 AI 芯片先驱 大会在热烈的掌声中迎来了 Unconventional AI 的联合创始人兼首席执行官 Naveen Rao。主持人将他形容为一位“定义级别的特例创始人”（definitionally outlier founder），以连续打造并成功出售两家深科技公司而闻名业界。Naveen 带着强烈的工程师务实精神登台，他信奉一句名言——“唯有亲手造出一样东西，你才算真正理解了它”，并直言自己是当下科技圈“AI 灭绝论者”（AI doomer）的反面。在他看来，人工智能是人类进化的下一个里程碑，而当前行业的核心瓶颈在于：要想实现真正的智能，必须在底层硬件物理载体（hardware substrate）上展开颠覆式创新。 面对台下观众，Naveen 笑着坦言，今天来到这个“反 AI 毁灭论”（anti-doomer）的舞台让他倍感兴奋。回顾人生轨迹，他感叹自己如今恰好站在了童年梦寐以求的坐标上。大屏幕上打出了一张他五六岁时的老照片——那是 1978 年前后，他的家庭极早地购置了一台计算机。到了 80 年代初，年幼的 Naveen 就将编程视作解谜游戏，乐此不疲。出于对科幻小说的痴迷以及“如何制造一台智能机器”的执念，他最初攻读了电气工程学位，投身计算机硬件研发多年；随后，为了探寻大脑运转的奥秘，他又重返校园拿下了神经科学博士学位。令他欣慰的是，数十年来整个技术世界的演进浪潮，最终与他个人探索的交汇点不谋而合。 在创业历程中，Naveen 始终是敢于吃螃蟹的人。早在 2014 年，绝大多数普通大众甚至科技从业者对“人工智能”一词还毫无概念时，他就创立了全球首家专用 AI 芯片初创公司 Nervana Systems。当时要说服投资人和行业相信 AI 的前景并为其定制硬件极其艰难，最终 Nervana 被英特尔（Intel）收购，Naveen 随后一手创立并掌管了英特尔的人工智能事业部。2020 年离开英特尔后，他敏锐地洞察到下一阶段的核心是构建超大规模模型基础设施，于是创办了 MosaicML，专注于 GPU 平台化与超大规模训练拓展。2022 年底 ChatGPT 爆发后，MosaicML 瞬间成为业界构建自研大模型的首选平台，并在 2023 年被数据分析巨头 Databricks 以约 13 亿美元收购。如今，这套 AI 基础设施产品已占据 Databricks 整体营收的四分之一。 ### [03:22](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=202s) 创立 Unconventional AI：从第一性原理重构计算架构，挑战千倍能效 在带领 MosaicML 取得商业成功后，Naveen 并未止步，而是选择回归第一性原理，创办了新公司 Unconventional AI。现代计算机体系结构沿用数十年未变，而他的核心目标是用全新的物理与数学思维设计一种全新的机器，唯一的专注方向就是极致的能效比。团队最初设立的目标是在 5 年内实现相比现有计算平台 1000 倍的能效提升。令人振奋的是，这一时间表已被 Naveen 提前至 3.5 年，因为团队在 AI 自身辅助科研的加速下，突破底层深奥科学难题的速度远超此前预期。 在研发架构上，Unconventional AI 组建了一支贯穿“顶层理论到物理硅片”的垂直全栈团队。研发链条的最前端是由数学博士与理论神经科学专家构成的理论组，他们的任务是从减少信息移动的本质出发，推导能大幅缩减能耗的全新计算范式；算法团队随即接棒，将这些抽象概念转化为可以在真实数据集上训练与测试的算法模型；接下来，芯片架构与电路设计团队将这些数学逻辑物理化为实际电路并进行严谨仿真；最终由系统工程团队制造出真实的板卡与计算系统，交付可商用的产品。 ```mermaid flowchart LR accTitle: Unconventional AI 全栈研发流程 accDescr: 展示从理论研究到芯片物理落地的垂直整合过程 subgraph Theory["理论研究层"] A["理论学者 (数学 / 理论神经科学)"] --> B["推导极低搬运能耗的计算范式"] end subgraph Model["模型验证层"] B --> C["真实数据训练与端到端算法验证"] end subgraph Hardware["物理落地层"] C --> D["电路架构设计与仿真验证"] D --> E["物理芯片、硬件板卡与整机系统"] end ``` ### [04:45](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=285s) AI 的“能耗高墙”：从 Token 成本到全网电力危机 谈及行业现状，Naveen 抛出了一个核心拷问：“能源真的已经成为 AI 的死穴了吗？”为了让抽象的能耗具有体感，他援引了 Google 公布的公开运营数据进行了量化推演： - 仅 Google 一家公司，每月推理处理的 Token 规模就已突破 3.2 千万亿（3.2 quadrillion）之巨。 - 即便按照非常保守的模型低端能耗指标——每个 Token 消耗 10 焦耳（Joules）计算，单纯维持这一服务所需的电力就高达 12 吉瓦（GW）。 - 作为对比，当前全美所有数据中心的总受电容量约为 40 吉瓦，而全美数据中心容量占到了全球的一半左右，这意味着全球数据中心总功耗目前尚不足 100 吉瓦。仅一家巨头的一项 AI 服务，理论算力功耗就吃掉了全美数据中心总电力的近三分之一。 Naveen 强调，随着大模型参数体量持续膨胀，以及全社会调用需求的几何级爆发，二者相乘将引发灾难性的电力赤字。摆在行业面前的是一个残酷的剪刀差：一边是到 2030 年预计突破万亿美元并保持指数级狂飙的 AI 市场需求，另一边则是受限于电网基建与发电设备、只能线性缓慢爬升的物理能源供给。按照目前的消耗速率，他断言整个产业将在大约 3 年内彻底撞上能源供给的坚硬墙壁。 这种能源瓶颈彻底颠覆了数据中心的商业运营逻辑。早年间建立数据中心，核心瓶颈是机房占地面积（floor space）；随后演变为机架空间与网络设备；前几年则是疯抢 GPU 卡；而如今，数据中心圈地的最前置条件已全面转向“电力合同”（power contract）。企业必须先锁定供电配额，再考虑如何往机房填充硬件。在如今生成或调用一个 Token 的综合成本构成中，50% 直接来自电力开销，剩下的 50% 才是芯片等硬件的资本支出（CapEx）与机房折旧。因此，当下算力运营商的核心诉求变成了“如何让每一瓦特电力变现出最大价值”，而 Unconventional AI 瞄准的商业支点，正是将单位瓦特的算力变现效率提升 1000 倍。 ### [07:04](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=424s) 师法自然与能耗根源：松鼠脑的 8 毫瓦奇迹与数据搬运之困 在探索超高效计算的过程中，大自然早已给出了存在的证明——生物大脑。成年人类大脑拥有约 860 亿个神经元，其运转总功率仅为极其惊人的 20 瓦左右；而动物界展现出的能耗比更令工程学汗颜： - 一只猴子的大脑功率仅约 1 瓦，功耗与每个人口袋中的智能手机相当。 - 大鼠、蝙蝠等小型哺乳动物的大脑消耗仅在毫瓦（mW）级别。 - 最具代表性的是人们生活中常见的松鼠。松鼠在树枝间跳跃穿梭，连续上千次都能做到分毫不差的精准落点与平衡控制，而支配这一高难度空间感知与运动控制的大脑，运行功率仅有 8 毫瓦（mW）。这意味着一部手机的供电预算，足以同时驱动 100 多个高度精准的松鼠大脑。 生物神经系统展示了何为承载智能的最佳物理介质。对比之下，人造数字系统之所以能效低下，病根在于“数据搬运”（moving information around）的系统架构浪费。Naveen 用了一组极具冲击力的对比数据： - 人类大脑皮层（承担高级认知功能的外层组织）虽然容纳了约 130 亿到 140 亿个神经元，但每秒在神经回路间搬运的数据量仅约 160 亿比特（16 Gbps）。 - 一台现代顶配 GPU 或高性能计算系统，光是在芯片与片外内存之间每秒往复搬运的数据量就接近 30 万亿比特（30 Tbps）；如果算上芯片内部各个缓存与运算单元之间的数据搬移，数字还要再翻 10 到 100 倍。 现代数字计算将绝大多数能量毫无意义地耗费在了比特的总线穿梭上。从几百年前的机械差分机，到上世纪初的模拟计算机，再到 1930、1940 年代冯·诺依曼架构奠定的现代数字计算机基石，人类计算系统的底层运转范式历经八十年未发生本质改变。而要打破 AI 的能源死锁，就必须从这种传统数字范式中突围。 ### [09:19](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=559s) 传统冯·诺依曼架构的瓶颈与摩尔定律的终结 现代计算机自诞生以来，其底层工作范式其实从未发生过颠覆性的改变。计算机将内存置于外部，在计算核心与内存之间不断来回搬运比特数据。这种设计虽然造就了极致的运算速度，但也带来了一台必须频繁进行海量数据搬运的“搬运机器”。Naveen Rao 回顾道，人类设计计算机的初衷就是为了追求绝对的速度——早在 1945 年制造出的第一代电子计算机 ENIAC，其最初目标是计算火炮弹道轨迹，击败由人类计算员手动推演的传统方式。到了今天，计算机比拼的标尺仍然是“这台机器比上一台快两倍”，整个行业售卖算力的方式依然建立在速度比较上，却从来没有把“能效比”作为底层核心考量。 然而，半导体行业曾经依赖的红利早已难以为继。长期以来，芯片上的晶体管数量不断增加，但频率却无法持续提高，单线程性能也陷入瓶颈。如今，整个行业甚至无法再通过微缩工艺来提升能效——“摩尔定律，正如大家所熟知的，通过缩小晶体管尺寸来获取收益的时代已经基本结束。”单纯依靠先进制程缩小晶体管，已无法换来系统性的能耗改善，这倒逼着从业者必须彻底重构计算的物理底层。 - **ENIAC 的历史对照**：1945 年研制的 ENIAC 初衷是取代人工计算火炮弹道，确立了以“速度胜过替代方案”为主导的研发哲学。 - **摩尔定律停滞的影响**：登纳德缩放定律失效后，单线程性能、主频以及单位计算的能耗提升均陷入停滞。 --- ### [10:24](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=624s) 消除抽象“中间商”：从数字逻辑到物理动力系统 面对能效困境，Naveen Rao 提出的核心解决直觉是“砍掉中间商”（cut out the middleman）。传统计算机层层叠加了人为构建的抽象层，最典型的一层便是“数字逻辑（0 和 1）”。但在物理世界中，根本不存在绝对的 0 和 1，晶体管本身具有连续的中间态，工程师们耗费巨大的能量和电路代价去强制它呈现离散状态。在这些数字硬件抽象之上，业界又构建了线性代数、浮点运算库，最后再在其上搭建人工神经网络和学习算法。每一个抽象层都是有损的、低效的，它们在屏蔽底层物理复杂度的同时，也白白浪费了惊人的能量。 新思路的核心在于寻找一种能够直接将半导体物理特性与神经网络连接起来的全新抽象。Naveen 引导观众反观人类大脑：“如果观察人类大脑，里面布满了神经元，但这里面根本没有线性代数，也没有浮点数学运算，正是神经元自身的物理特性自然涌现出了智能。”大自然中处处存在着没有集中调度的物理计算，例如鸟群掠过天空时，单只鸟只需要观察左右两边邻居的简单动作，就能涌现出极具复杂度和美感的群集协同飞行；蚁群仅遵循极为简单的本能规则，也能组织起高度智能的集体行为。在数学与物理学中，这被称为“动力系统理论”（Dynamical Systems Theory）——研究如何通过个体单元的极简交互自然涌现出宏观复杂智能。 ```mermaid flowchart TD accTitle: Abstraction Layers Comparison accDescr: Comparing traditional deep abstraction stack with the direct physics-to-intelligence dynamical stack. subgraph Trad["传统计算堆栈（层层抽象，能耗极高）"] A["物理半导体状态"] --> B["数字抽象（强制离散化为 0 和 1）"] B --> C["浮点算术与线性代数（BLAS / 矩阵乘法）"] C --> D["深度学习与人工神经网络"] end subgraph Dyna["动力系统计算堆栈（砍掉中间商）"] E["连续半导体物理特性（如振荡器、阻抗）"] --> F["非线性动力学交互（相空间自然演化）"] F --> G["直接涌现生成智能与表征能力"] end ``` --- ### [12:04](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=724s) 节拍器共振实验与 UNO 生成式模型 为了形象地解释动力系统的计算潜力，Naveen Rao 举了一个生动的物理实验：在钢琴旁常见的机械节拍器。如果将多个频率相近的摆动节拍器摆在一块可以左右自由滚动的平木板上，初始状态下它们各自以不同相位杂乱摆动；但随着木板受到每一次微小摆动的物理反作用力，物理系统内部会自发相互调谐，最终所有节拍器都会自发锁定在完全一致的相位上实现同步共振。这种物理自组织可以扩展到成百上千个节拍器，甚至可以构造出更复杂的耦合结构——例如一部分保持同相、另一部分保持反相或特定周期模式。 物理系统仅凭部件之间内在的物理互联就能完成复杂的动力学演化，这让团队提出了一个前沿构想：能否直接利用这种“节拍器系统”的物理震荡来进行生成式 AI 的计算？为此，团队在前期公开发布了一个名为 **UNO** 的开源图像生成模型。该模型基于一组耦合振荡器构建并在软件上进行了完整仿真，证明了利用连续振荡器动力学系统不仅能进行大规模训练，还能稳定地生成有用图像。 - **相空间轨迹控制（State Space Trajectory）**：系统在时间维度上的演化可以用多维振荡器相位的状态空间轨迹来刻画。当人类输入“生成一架飞机”、“生成一辆汽车”或“生成一只鸟”作为条件时，动力系统便会在特定的动力学吸引子与状态空间轨迹中自然收敛至目标图像特征。 - **开源验证**：UNO 模型首次向公众证实，脱离传统逐层数字计算，纯振荡器动力学系统同样具备参数可扩展性与复杂生成式任务的处理能力。 --- ### [14:11](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=851s) 稀疏性理论突破：解决 $N^2$ 互联扩展瓶颈 在随后几个月深入的基础科学研发中，团队攻克了物理网络互联的规模扩展瓶颈——全互联拓扑带来的 $N^2$ 爆炸问题。当系统包含 10 个振荡元件时，全互联需要 $10 \times 10 = 100$ 条物理连接；但当扩展到 1,000 个元件时，连接数会暴增至 1,000,000 条。如果硬件系统中每个计算单元都与其他所有单元物理相连，芯片布线和信号干扰将彻底锁死可扩展性。 团队将目光投向了“稀疏性”（Sparsity）理论。研究人员提出了一个关键假设：如果物理剪除绝大多数连接通路，整个系统的动力学行为和生成能力能否被挽救？实验和理论推导给出了令人振奋的结论——稀疏连接不仅没有牺牲网络能力，反而大幅改善了系统的可训练性与全局表现。这造就了工程界罕见的“圣杯”状态：同时获得了更低的能量消耗、更优异的物理扩展性以及更强大的任务性能。这一突破为动力学系统从理论仿真走向实体芯片铺平了道路。 --- ### [15:36](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=936s) 5 个月造出首款物理动力学芯片：500 纳焦耳的能效飞跃 现场迎来了全场最重磅的首次公开宣布：Naveen Rao 揭晓了团队在过去几个月内秘密研发的全球首款物理动力学计算机（Physical Dynamical Computer）。这一工程成果是在极其紧迫的时间线中完成的——公司在当年 1 月才真正全职启动，当时甚至连完整的芯片研发团队都尚未组建齐全。然而团队在 1 月立下誓言，必须在年内制造出实体原型硬件。 仅仅历时 5 个月，团队便在 6 月 1 日完成了芯片设计并正式送往晶圆厂流片（Tape-out）。如今测试芯片已返回实验室，并在实体硬件上成功跑通，输出了全球首批由真实物理动力学芯片直接生成的图像。更具颠覆性的是其能效表现：该芯片生成单张图像仅耗费约 **500 纳焦耳（nJ）**的能量。相比之下，主流商用 GPU 完成单次同类计算通常需要消耗毫焦耳（mJ）级别的能量。由于纳焦耳处于 $10^{-9}$ 焦耳量级，动力学芯片在物理效率上直接实现了数个数量级（千倍以上）的巨大跃迁，其本质就在于它彻底免去了在存储与运算单元之间反复搬运数据所带来的寄生功耗。 | 指标维度 | 传统高端 GPU 计算架构 | 首代物理动力学芯片原型 | |:--- |:--- |:--- | | **基础计算范式** | 冯·诺依曼架构（计算与存储分离） | 连续动力学演化（计算与存储完全合一） | | **单张图像生成能耗** | 毫焦耳级（mJ，约 $10^{-3}\text{ J}$） | 约 500 纳焦耳（nJ，约 $5 \times 10^{-7}\text{ J}$） | | **能效差距** | 基准线（受制于内存墙与数据搬运功耗） | **提升数个数量级（超 1,000 倍能效比）** | | **研发执行节奏** | 传统芯片公司通常需数年设计周期 | 1月团队破土动工，6月1日流片，回片即成功点亮 | --- ### [16:50](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1010s) 4D 计算：颠覆冯·诺依曼架构与每瓦智能的物理极限 从硬件演进史来看，计算架构经历了从单核 CPU 到高度并行的 GPU，再到内存与计算更为紧密分布的演变，但万变不离其宗，它们全都是“冯·诺依曼架构”——计算单元与存储器始终分立，必须依靠高频总线不断来回传输比特。而 Naveen 团队打造的动力学计算机从根源上消除了内存接口（Memory Interface），其内部每一个基本的动力学计算元件本身就是物理状态的承载记忆体，存算完全天然合一。 团队将这一全新架构命名为“**4D 计算**”（4D Computing）： 1. **物理三维空间（3D）**：利用先进的垂直晶圆/裸片立体堆叠（Die Stacking）与芯片平面布线，在物理的三维空间内排布致密的动力学互联网络。 2. **时间演化维度（1D）**：打破传统依靠高频外部时钟步进的数字节拍，将时间作为动力学系统微分方程连续演化展开的内生计算维度。 三维立体物理结构结合内生时间维度，构成了真正意义上的 4D 计算系统。Naveen Rao 总结道，如果能将 AI 算力的能耗效率提升一千倍以上，整个科技世界的边界将被彻底重塑。在人工智能时代，真正值得行业全神贯注追求的根本指标是“每瓦智能”（Intelligence per Watt），而动力学计算将直面自然界的热力学终极极限，开启一条超越传统硅基数字芯片的全新演进路径。 ### [17:55](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1075s) 逼近热力学极限与“超越生物系统”的愿景 Naveen Rao 深入剖析了当前计算架构与物理学极限之间的巨大鸿沟。物理学中存在不可逾越的热力学极限，而哺乳动物等动物的大脑在能量效率上仅落后该极限一到两个数量级。相比之下，当今的数字计算系统处于图表的最左端，距离热力学极限足足差了 100 亿倍（$10^{10}$）。Naveen 预测，在未来约三年半的时间内，他们的架构有望触及传统 2D 光刻技术的物理极限。该公司的核心目标正是“超越生物系统（beat biology）”，打造能效更高、形态更自由的新型计算范式，使普适计算在未来十年内融入全新形态的机器人等实体系统之中。 在这种技术变革下，全球 AI 基础设施的形态也将发生根本转变。计算将不再拘泥于消耗数吉瓦（GW）电力的巨型集中式数据中心，而是演变成遍布各地的分布式小型数据中心。这种架构不仅更加环境友好，还具备极强的局部适应性。更进一步，这种微型化、超高能效的算力将赋能数以百亿计的机器人，让它们能够根据现实世界的任务动态协作。 在市场规模方面，Naveen 引用了著名的“杰文斯悖论”（Jevons Paradox）：当一种底层资产的获取成本大幅降低时，人类对其消耗的总量往往会呈指数级爆发，远超成本下降的幅度。若将 AI 算力的成本降低到千分之一，其消耗量绝不会仅仅增加一千倍，这将直接催生人类有史以来最大的技术市场。 - **能效差距**：当前数字计算机与热力学极限相差 100 亿倍（1 后面跟 10 个 0），而生物大脑距离该极限仅 1~2 个数量级。 - **时间节点**：预计在 3.5 年内逼近 2D 光刻工艺极限；未来十年内实现算力向多样化机器人形态的深度渗透。 - **产业范式转移**：从单体吉瓦级（Gigawatt）大型数据中心转向环境友好、本地化部署的分布式微型数据中心。 - **杰文斯悖论效应**：若 AI 计算成本下降 1,000 倍，需求端激增将打破现有万亿美元市场的上限。 ```mermaid flowchart TD accTitle: 计算能效演进与市场扩张逻辑 accDescr: 展示从当前数字芯片经由全新架构逼近生物大脑与热力学极限，并经杰文斯悖论引发算力市场爆发的路径。 subgraph Efficiency["计算能效演变"] A["现代数字计算\n(差距: 10^10)"] -->|三年半目标| B["突破 2D 光刻极限\n(新型动力系统芯片)"] B -->|终极愿景| C["超越生物大脑\n(逼近热力学极限)"] end subgraph Impact["基础设施与市场重塑"] C --> D["小型分布式自适应数据中心\n+ 百亿级动态协作机器人"] D --> E["单位算力成本降至 1/1000"] E --> F["杰文斯悖论驱动\n人类历史最大算力消费市场"] end ``` ### [19:40](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1180s) 产品落地路径与现有模型架构的迁移 Chamath Palihapitiya 登台加入讨论，直接切入核心商业与落地问题：芯片要真正交付到客户手中，需要庞大的制造代工厂（Fab）与封装生态支持，从早期概念到商用产品的实际路线与周期究竟是怎样的。Naveen 明确表示，他们计划在两年之内推出完整的商用产品。与常规的单芯片销售或云端虚机（VM）租用不同，他们交付的将是面向数据中心的完整机柜系统架构。对外接口遵循标准网络协议，实现“输入 Token，输出 Token”，但在系统机柜内部，其底层硬件与运行逻辑与传统冯·诺依曼计算机截然不同。 针对业界关心的生态兼容性问题——现有的大语言模型（如依赖 KV Cache 等机制的 Transformer 架构）是否需要重构，Naveen 指出，在技术跃升的收益与生态迁移的阵痛之间始终存在权衡。为了吸引开发者，他们选择提供极具颠覆性的能效诱因，同时尽量降低迁移成本。用户不需要在算子（Operations）层面进行繁琐的手工重写，而是在“模型层（Model layer）”进行跨架构迁移与编译，虽然这一迁移过程本身需要消耗相当规模的算力。 当 Chamath 追问新硬件中是否存在像矩阵乘法（MatMul）这样的基础指令集时，Naveen 直言系统中并不存在物理意义上的 MatMul 指令。该芯片并非直接执行矩阵乘法运算，而是将计算建模为连续的时间变化行为；若从分析角度拆解每一个离散时间步，系统的状态演化在数学上等价于“当前状态向量与状态转移矩阵的相乘”。 - **产品周期**：商用产品预计在 2 年内落地。 - **产品形态**：数据中心整机柜系统；外部通过网络缆线进行 Token 的输入与输出，内部完全摒弃传统计算机架构。 - **兼容策略**：在模型层实现迁移，而非底层算子指令层；无需用户手写底层矩阵优化代码。 - **计算本质**：底层依靠连续的动力学系统时间演化，而非硬编码的离散矩阵乘法单元（MatMul）。 ### [21:37](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1297s) 交叉团队构建与基于 Python 的抽象层 打造这种前所未有的计算平台，面临的最大瓶颈之一在于人才团队的跨界与磨合。Chamath 好奇询问 Naveen 麾下究竟是由哪些学科的专家组成。Naveen 表示，研发团队由两类长期处于平行世界、彼此互不相通的人群构成：一类是研究动力系统理论（Dynamical Systems Theory）的理论科学家（该数学理论已发展了一个世纪），另一类则是芯片设计与硬件工程专家。让这两拨语言体系截然不同的人坐在一起，协同构建同一套系统架构，是公司运营中最具挑战性的工作。 为了连接高层的数学理论与底层的动力系统芯片，团队必须打造类似英伟达 CUDA 抽象层的开发环境。Naveen 透露，他们并没有另造一套晦涩的底层语言，而是基于 Python 构建了一套高度定制的软件库。该库允许算法工程师以熟悉的 Python 语法来直接描述具有随机性行为（Stochastic behavior）的时间演化动态单元，从而架起了高维理论模型与新型底层硬件之间的沟通桥梁。Chamath 对这一极为宏大而前沿的工程尝试给予了高度赞赏。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=0s) 唯有亲手造出一样东西，你才算真正理解了它 — Naveen Rao > [09:19](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=559s) 摩尔定律，正如大家所熟知的，通过缩小晶体管尺寸来获取收益的时代已经基本结束。 — Naveen Rao > [10:24](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=624s) 砍掉中间商 — Naveen Rao > [10:24](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=624s) 如果观察人类大脑，里面布满了神经元，但这里面根本没有线性代数，也没有浮点数学运算，正是神经元自身的物理特性自然涌现出了智能。 — Naveen Rao > [16:50](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1010s) 4D 计算 — Naveen Rao > [16:50](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1010s) 每瓦智能 — Naveen Rao > [17:55](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1075s) 超越生物系统 — Naveen Rao > [19:40](https://www.youtube.com/watch?v=yAsrMA_ADPc&t=1180s) 输入 Token，输出 Token — Naveen Rao ## Mentioned Naveen Rao, Chamath Palihapitiya, Unconventional AI, Nervana Systems, Intel, MosaicML, Databricks, Google, ENIAC, UNO, Python, CUDA, Jevons Paradox ## Video description (0:00) Welcome Naveen Rao (4:45) Is energy really the problem? The cost of a token, power contracts & the gap to close (10:24) Cutting out the middleman: abstractions, dynamical systems & a new kind of machine (19:40) Chamath joins: the path to product, porting existing models & building the team Thanks to our partners for making this possible! IREN is a vertically integrated AI Cloud platform, delivering data centers, compute and software for AI training and inference. https://iren.com/ Oracle connects the data, applications, and infrastructure that turn AI into business outcomes—with the flexibility, choice, and control to optimize as AI evolves. http://oracle.com/ai EY helps tech innovators scale from startup to exit to megacap. You build the future. We’ll handle the rest. http://www.ey.com Meta believes the future is for everyone. We're focused on giving every person the tools to reach their full potential and making sure the benefits of technology are distributed to all. http://www.meta.com Keel Infrastructure owns the power, land, and connectivity that HPC and AI run on - backed by secured energy assets and established grid interconnections across North America. https://keelinfra.com/ Airwallex - Agentic Global Business Accounts. Open local accounts in 70+ countries to accept payments, earn yield, pay globally, and manage spend. http://airwallex.com PayPal has been revolutionizing commerce globally for more than 25 years. Creating innovative experiences that make moving money, selling, and shopping simple, personalized, and secure, PayPal empowers consumers and businesses in approximately 200 markets to join and thrive in the global economy. For more information, visit https://www.paypal.com Google for Startups connects founders with the right people, products, and best practices to help startups build faster and go further. https://startup.google.com/ Explore ideas, industries, and technologies worth understanding with Chamath every week on Learn with Me: https://research.socialcapital.com/allin Follow the besties: https://x.com/chamath https://x.com/Jason https://x.com/DavidSacks https://x.com/friedberg Follow on X: https://x.com/theallinpod Follow on Instagram: https://www.instagram.com/theallinpod Follow on TikTok: https://www.tiktok.com/@allin Follow on LinkedIn: https://www.linkedin.com/company/allinpod Intro Music Credit: https://rb.gy/tppkzl https://x.com/yung_spielburg #allin #tech #news [Source](https://www.youtube.com/watch?v=yAsrMA_ADPc)

Source quotes

[0:00] Naveen Rao: 唯有亲手造出一样东西，你才算真正理解了它 [9:19] Naveen Rao: 摩尔定律，正如大家所熟知的，通过缩小晶体管尺寸来获取收益的时代已经基本结束。 [10:24] Naveen Rao: 砍掉中间商 [10:24] Naveen Rao: 如果观察人类大脑，里面布满了神经元，但这里面根本没有线性代数，也没有浮点数学运算，正是神经元自身的物理特性自然涌现出了智能。 [16:50] Naveen Rao: 4D 计算 [16:50] Naveen Rao: 每瓦智能 [17:55] Naveen Rao: 超越生物系统 [19:40] Naveen Rao: 输入 Token，输出 Token

Source quotes

[0:00] Naveen Rao: anti-doomer conference [0:00] Naveen Rao: way too early [8:05] Naveen Rao: I don't feel like we truly understand something until we can create it [17:55] Naveen Rao: The overarching goal of this company is to beat biology. [17:55] Naveen Rao: There's a concept called Jevons paradox where when you drop the underlying cost of an asset, you actually consume more than the drop of that asset. So, if you make something half the price, you'll consume more than 2x. If you make something 1,000th the price, you'll consume more than 1,000th of it. And I think this will create the largest market that humanity's ever seen. [19:31] Chamath Palihapitiya: I mean that was extremely unexpected. I got to say that was pretty amazing [19:31] Naveen Rao: tokens in, tokens out through a network cable [19:31] Naveen Rao: You can characterize it as MatMul, but it doesn't implement it as MatMul... It implements it as a sort of time-varying behavior. But each one of those time steps you can analyze as basically a matrix of the current state times a transition matrix. [21:37] Chamath Palihapitiya: When you're building a team like this, who are these people? These are biologists plus physicists plus what? [21:37] Chamath Palihapitiya: incredibly impressive [21:37] Chamath Palihapitiya: so ambitious

Original transcript

[0:00] Naveen Ralph, co-founder and CEO of [0:03] Unconventional AI, which is an AI chip [0:05] startup. [0:06] >> Best known for building and selling two [0:08] deep tech companies. Naveen is kind of [0:10] definitionally outlier founder. When I [0:12] came there, we had about a $20 million [0:14] business and it was, you know, 7 or 800 [0:16] million when I left. [0:17] >> I don't think you really understand [0:18] something until you can build it. [0:19] >> Just because something is tried does not [0:20] mean it's wrong. [0:21] >> I'm the opposite of an AI goomer. I [0:23] think AI is the next evolution of [0:25] humanity. We need innovation on the [0:27] hardware substrate to actually build [0:30] true intelligence. [0:31] >> Please welcome Naveen Ralph. [0:44] Hey everyone, great to be here. Um, you [0:47] know, switching gears a little bit to AI [0:49] now, which you may have heard a little [0:50] bit about. Um, it's super exciting to be [0:53] at this conference specifically because [0:55] as was said in the uh intro, I'm the [0:58] opposite of a doomer. I think AI is one [1:00] of the most transformational [1:01] technologies that humanity's ever [1:03] created and will enable us to get to [1:05] that next level of evolution which I'm [1:07] here for and uh this is sort of the [1:09] anti-doomer conference. So, let's go. [1:14] Um, so before we get going, I'll tell [1:16] you a little bit about myself. Um, you [1:18] know, I'm it's kind of weird. I'm really [1:20] right where I wanted to be my whole [1:21] life. Uh this was me at about five or [1:23] six years old, something like that. We [1:25] had a computer very early on, so I'll [1:27] date myself, but uh this was in 1978. We [1:30] got a computer. This is probably in the [1:32] early 80s. Uh learned to program when I [1:34] was a little kid. I just thought it was [1:35] like a puzzle. Um you know, learned to I [1:38] became an electrical engineer. Um really [1:40] because I enjoyed sci-fi and always [1:42] wanted to think about how I could make a [1:44] intelligent machine. And uh you know [1:46] then after a career in uh building [1:50] computers I went back to school and got [1:52] a PhD in neuroscience and the idea was [1:54] like let's go back to that thing how do [1:55] we make computers intelligent and [1:58] fortunately the whole world kind of [1:59] moved in this direction. So you know as [2:01] a technologist it's sort of the dream [2:03] right now. [2:05] Um a little bit about me tech uh from a [2:08] company entrepreneurship standpoint like [2:11] I actually founded the first AI chip [2:13] company called Nirvana Systems. So um [2:15] this was in 2014 if anyone remembers [2:18] back then there was no AI or at least [2:20] not in the common vernacular and uh it [2:23] was really hard to actually convince [2:24] people that this is important much less [2:25] to build hardware around it. Now you [2:27] heard from Jensen up here like the [2:29] largest company in the world a hardware [2:30] company because of AI. So we were early [2:33] on um I think I sold the company way too [2:35] early uh to Intel but I uh I ran I [2:38] started and ran the AI group at Intel. [2:41] Um, after I was done with that in 2020, [2:43] I actually started thinking about the [2:44] next problem, which was how do we build [2:46] bigger models like the large language [2:48] models we talk about today. And it was [2:50] about how do I build the infrastructure [2:52] to build those models? And so we started [2:54] platformizing GPUs and enabling it to [2:57] scale and making that easy to use for [2:59] other people. And after chat GBT [3:00] happened in 20 2022, we were kind of the [3:03] best game in town for people to start [3:05] building their own models. So took off [3:07] really fast. We decided to actually join [3:09] forces with data bricks. um that was in [3:11] 2023 and and uh actually that's a [3:15] quarter of the total revenue of of data [3:16] bricks today. So um a lot of fun doing [3:19] that whole thing with uh Alli and team [3:21] at data bricks. And now I want to tell [3:23] you about unconventional AI which is [3:25] rethinking the foundations of how a [3:26] computer works. So there we're going [3:30] back to first principles here to really [3:31] trying to build a new machine. Computers [3:33] have worked a certain way for a long [3:35] time. We want to rethink that for the [3:37] singular purpose of making something [3:38] very power efficient and the goal has [3:42] been within it was initially within 5 [3:44] years to get to a 1000x power [3:46] efficiency. I've actually revised this [3:48] to three and a half years because things [3:49] have gone faster than we anticipated. [3:51] We've actually solved very deep [3:52] scientific problems quicker because of [3:55] AI interestingly enough. So just a [3:58] little bit how are we organized? Like [4:00] we're truly a toptobottom company. We [4:02] start with theorists. These are people [4:03] with like math PhDs and um you know [4:06] theoretical neuroscience that kind of [4:08] thing. They come up with you know [4:09] concepts that we think would effectively [4:12] give us more power efficiency from the [4:13] perspective of moving less information [4:15] around and we then translate that into [4:18] models and that do real things trained [4:20] on real data and evaluated against real [4:22] criteria. So it's kind of the rubber [4:24] hitting the road of these concepts. Then [4:27] eventually we have to actually build [4:28] something physical. So these are people [4:30] who architect a physical circuit um [4:33] actually design those circuits uh model [4:35] them and see if it actually works and we [4:37] try to connect this whole stack [4:38] together. Um then eventually we have to [4:40] build a system and and a board and all [4:42] that kind of stuff and build a product. [4:45] So is energy really a problem? Uh I'm [4:48] not sure how much everyone in this [4:49] audience has thought about this but [4:50] interestingly enough I'll give you some [4:52] some data points here. This is one [4:53] company. This is just Google. I'm using [4:55] Google because Google has actually [4:56] talked about this publicly. um per month [4:59] they cross 3.2 quadrillion tokens. It's [5:02] a crazy number. I never even I never [5:04] think in quadrillions but that's the [5:06] world we're in today. And if I just take [5:09] 10 jewels per token of energy. This is a [5:11] actually on the lower end of uh of the [5:13] energy spectrum for models. But let's [5:15] just take that number and multiply it [5:17] out. This is 12 gawatt. The US puts [5:20] about 40 gawatts of energy into data [5:22] centers today. And we're about half of [5:25] the data center capacity of the world. [5:27] And so, you know, we're under a 100red [5:29] gawatts of of data center energy in the [5:32] world today. 12 gigawatts is going into [5:34] one company just for AI services. So, [5:37] you can imagine if models get bigger [5:40] that that energy goes up and if demand [5:42] grows, which it is, that energy goes up. [5:45] So, we're going to run out of energy [5:46] pretty fast in like 3 years or so is my [5:48] estimate. So, really just putting it [5:50] graphically, this is what we have. You [5:52] know, we have this huge market that's [5:54] growing exponentially. call it a [5:56] trillion dollar market in 2030. Maybe [5:57] it's bigger than that. And then we have [5:59] this kind of linearized energy at the [6:02] bottom. And so you've heard a lot about [6:04] this today. Uh but this gap is the [6:07] problem. And we want to solve that gap [6:08] with technology. [6:13] So um I don't know if people are aware [6:15] of this, but the way we think about data [6:17] centers has shifted over the last [6:18] several years. It used to be about floor [6:20] space. Can I get the floor space? Can I [6:22] get the rack space? Then it was about [6:23] networking equipment. Then it became [6:25] about GPUs. Today it's about energy. [6:27] First you think about energy. I get the [6:28] energy contract and then I have to [6:30] figure out how to fill it and and [6:32] basically create u u um uh [6:35] infrastructure out of out of GPUs and [6:37] things like this about 50% of the cost [6:40] of serving a token. So every time you [6:42] try something on chat GBT 50% of that [6:44] cost is energy. The rest of it is the [6:47] capex of the hardware and the floor [6:49] space and all that kind of stuff. So [6:51] today we sort of think about oops we [6:53] kind of think about it as I get I get a [6:56] a power contract I need to monetize [6:57] every watt and simply put our business [6:59] case is pretty easy we're going to [7:01] monetize that a thousandx better than [7:02] existing hardware. [7:04] So then the question becomes, okay, [7:06] great, that all makes sense, but can we [7:08] actually do it? How do we build a [7:09] better, more efficient computer? Well, [7:11] biology actually provides some proof for [7:13] us here. So the human brain, you may [7:15] have heard this, runs on about 20 watts [7:16] of energy. And what's even more [7:18] remarkable to me is actually animal [7:20] brain. So that red number is how many [7:22] neurons are in the brain. And if you [7:24] kind of scale it linearly down to like a [7:26] a monkeykey's brain, it runs on one [7:28] watt. To put that in perspective, the [7:30] cell phone in your pocket runs on about [7:32] one watt. And other animals like you [7:35] know um rats and bats and things like [7:37] this. They run on millowatts of energy. [7:40] So just something that's pretty [7:41] relatable to everyone is a squirrel. You [7:43] probably watched how accurate they can [7:44] be. They jump between branches and they [7:47] do it perfectly a thousand times out of [7:48] a thousand. Their brain runs on 8 mill [7:51] of energy. You could run over a 100 [7:53] squirrel brains on your phone and it it [7:56] has very precise and accurate behavior. [7:58] So biology created something quite [8:00] incredible. In fact, it's the right kind [8:01] of physical substrate for intelligence. [8:05] So, this is a quote I love. I don't feel [8:08] like we truly understand something until [8:10] we can create it. We've gotten a lot [8:11] better creating intelligence systems. [8:14] However, they do it in a kind of [8:15] inefficient way. What kind of [8:17] inefficiency is there? Well, as I kind [8:19] of hinted at the beginning, most of the [8:22] energy in a computing system goes into [8:23] moving information around. Just to put [8:25] some numbers on it, the human cortex, [8:28] the squiggly part of your brain, the [8:29] outside of it, um, only moves about 16 [8:32] billion bits per second. That's actually [8:34] kind of a small number if you think [8:35] about it because there's, you know, some [8:37] 13 or 14 billion neurons in that cortex. [8:40] Um, a GPU or, you know, a high-end [8:42] computing system moves nearly 30 [8:45] trillion bits in and out of memory per [8:47] second. That's outside the chip. Inside [8:49] the chip, it's probably, you know, 10 [8:52] 100x more than that. So we're moving a [8:54] lot more bits in these synthetic systems [8:56] than the brain does. And that's actually [8:58] what drives the energy demand. And you [9:02] know how do we get here? You know [9:03] computers have been around for hundreds [9:05] of years actually in some form. They [9:07] were mechanical. They became analog [9:09] around the turn of the century and they [9:10] became digital back in the 1930s 1940s. [9:13] So this the operation of that computer [9:15] in 1940 1945 is actually very similar to [9:19] how they operate today. There's not a [9:20] huge paradigm shift. actually have this [9:22] memory on the outside. You have some [9:24] kind of computing and you move bits back [9:25] and forth. That operation creates a [9:28] machine that just requires a lot of [9:30] movement but it's very fast. We built [9:32] computers to be fast. They were always [9:34] faster than the alternative. Uh [9:36] incidentally that computer in 1945 that [9:38] was built called Eniac was built to do [9:41] artillery trajectory calculation and it [9:43] was built to do it faster than the [9:44] alternative. The alternative were humans [9:46] who actually did the calculations. Now [9:48] the alternative typically is some other [9:50] computer. this this computer is twice [9:52] the speed of that computer. That's how [9:53] we sell computers, but it's doesn't [9:55] contemplate energy efficiency. And [9:57] that's what we're changing. [9:59] And you know, these have been trends [10:00] going on for a long time where the [10:02] number of transistors kept going up, but [10:04] we couldn't keep scaling the frequency. [10:06] We could couldn't keep scaling uh single [10:08] thread performance. And now we're [10:10] actually not scaling efficiency any [10:12] longer. Moore's law, if you may have [10:13] heard of this, it's like making [10:14] transistors smaller has largely ended. [10:17] So we're not seeing efficiency gains [10:19] just from making transistors smaller. So [10:21] we need to rethink the problem a bit. So [10:24] how do we do it? Well, a good intuition [10:26] is that we cut out the middleman. So [10:28] computers have been built up with these [10:30] abstractions. So I mentioned digital. [10:32] Digital means one and zero. That itself [10:34] is an abstraction of the physical world. [10:36] Like we don't actually have systems that [10:39] behave as one and zero. A transistor [10:41] actually has states in the middle. But [10:43] we engineer it to behave that way. And [10:45] that that's an abstraction. So we kept [10:47] building these abstractions up and [10:48] eventually we started creating neural [10:50] networks and learning machines on top of [10:52] it. Each one of these abstractions [10:54] actually is lossy. It means it's [10:56] inefficient. It it it doesn't [10:57] contemplate all the complexity [10:59] underneath it. That's why it's an [11:00] abstraction. So what we're doing is kind [11:02] of simplifying this in some sense. We [11:04] find an abstraction of the physics of [11:07] the semiconductor and connect that to [11:09] the neural network. If you think about [11:10] your brain for a moment has a bunch of [11:12] neurons in it, but there's no linear [11:14] algebra. There's no uh floatingoint [11:16] math. It's actually the physics of the [11:18] neurons give rise to intelligence. We [11:20] want to kind of mimic some of that uh [11:22] with a semiconductor. [11:26] And this is also not a new concept. [11:28] There's actually computation all [11:29] throughout nature. Uh birds in a flock. [11:32] You may have seen things like this where [11:34] you know a bird does a simple behavior. [11:35] It looks left, looks right, and figures [11:37] out where the next guy's going. And when [11:39] they do that, they actually create these [11:41] interesting flocking behaviors, this [11:42] emergent behavior. And we see that with [11:44] ant colonies. Ant colonies actually do [11:46] intelligent things just by very simple [11:48] rules. And uh this study is called [11:51] dynamical systems theory. It's basically [11:53] how I get these emergent properties from [11:55] very simple behaviors of individual [11:56] components. Our brain actually works [11:58] this way as well. So we're taking these [12:01] ideas and actually starting to build [12:02] circuits out of them. So let me give you [12:04] an example of such a such a system. So [12:07] everyone here probably knows what a [12:08] metronome is. just, you know, when [12:10] you're learning to play the piano or [12:11] something like that, it's just tick tock [12:13] tick- tock physical thing moving back [12:15] and forth. So, this is kind of cool. If [12:17] you put multiple ones of them on a rigid [12:19] plank and that plank can just roll back [12:21] and forth, you'll actually see them [12:22] start to synchronize. And that's just [12:24] due to the physics of the system. They [12:26] each push against the plank just a [12:28] little bit. And even if they're off by a [12:30] little bit from each other, they'll all [12:31] synchronize to exactly the same phase. [12:32] You can actually scale this up to [12:34] hundreds of of metronomes on a physical [12:36] system and they'll all synchronize. So [12:38] this is actually a form of a dynamical [12:40] system that goes through you know some [12:42] starting point of all these different [12:44] phases and always synchronizes. You can [12:46] imagine a slightly more complicated [12:48] version of this where maybe they don't [12:49] all synchronize. Maybe half of them are [12:51] synchronized with each other and other [12:52] half are synchronized uh in an opposite [12:54] pattern or something like this. But the [12:56] idea is that this is a physical system [12:58] that that kind of uh behaves the way it [13:00] does just by by the the inherent [13:02] interconnection of the system itself. [13:05] Uh so we asked the question can we [13:08] actually use such a system like like [13:10] that metronome system to do computation. [13:13] We want to connect that to generative AI [13:15] that's what we really care about. So we [13:16] released a model we called UNO which [13:18] actually demonstrated this. It's an [13:20] image generation model that's built on a [13:22] set of oscillators like that. And uh we [13:25] simulated it and made it open source so [13:26] people can play with it. But this was a [13:28] f the first demonstration that I can [13:30] actually scale something up, train it [13:32] and actually get useful output like [13:34] image generation. [13:37] And uh you know this is kind of some of [13:39] the analysis that we uh we we provided [13:41] in that uh in that write up. What you're [13:43] seeing here is um we call this state [13:45] space trajectory. It's basically how the [13:47] system evolves in time. You can imagine [13:49] if you characterize the state of the [13:51] system as all the phases of those [13:53] oscillators um and then look at how it [13:55] evolves through time and you basically [13:58] condition that on the output. You say I [14:00] want to generate an airplane or a car or [14:01] a bird. It'll actually go through [14:03] different state space trajectories. And [14:05] so that's what we're seeing here is just [14:06] an analysis of this. And these are [14:08] actual images that were generated by it. [14:11] Now it turns out we actually started to [14:13] build a lot of that fundamental science [14:15] up over the last couple of months. And [14:17] we found other things that enable this [14:19] to work even better. So this is a [14:21] concept of what we call sparsity. [14:22] Sparsity means that you know if I have [14:25] uh let's say I have a bunch of elements [14:27] all connect to each other like we have [14:28] on the left hand side there. So you know [14:30] if I have uh 10 elements I want to [14:32] connect them all to each other. I have [14:34] 10 * 10 elements. So I have a 100 [14:36] connections. That's okay. But if I have [14:38] a thousand now I have thousand* 1,000 [14:41] which becomes a million. So this doesn't [14:43] scale very well. We call this n squared [14:45] scaling. So the more I add, it actually [14:47] becomes way more hard to scale, right? [14:49] So that's not great. So sparsity allows [14:51] us to actually say, well, can I throw [14:53] away some of those connections? If I [14:54] throw them away now, can I actually [14:56] rescue the behavior of the whole system? [14:58] And it turns out you can actually not [15:00] only throw away some of the connections, [15:02] but you can actually get better behavior [15:04] out of the whole system. It actually [15:05] becomes more trainable. Uh there's a lot [15:07] of theoretical reasons for this, but we [15:09] actually are able to do this in um in [15:11] not only simulated systems, but actually [15:13] real physical systems. So, it's one of [15:15] these rare things where you get [15:16] something that's more efficient, uh, [15:18] that's actually more scalable and even [15:19] gives you more performance. So, this is [15:21] kind of a holy grail. It's been it's [15:23] been a problem for a long time, but we [15:24] had to frame the problem the right way [15:26] to actually find this solution. And so, [15:28] this is actually the first time I'm [15:29] talking about this publicly. This is uh [15:32] I wanted to do it at this venue because [15:33] I think uh um it's a really big deal. [15:36] This is actually the first physical [15:39] dynamical computer ever built. Uh, we [15:42] did this in 5 months. This company [15:44] started [15:46] Thank you. [15:48] started in earnest in January. We didn't [15:50] even have a team, but we said we're [15:52] going to build this first physical [15:53] prototype and do it, you know, this [15:55] year. And actually, we taped out the [15:56] design, meaning we sent it to the fab uh [15:58] in June on June 1st. The chip is back in [16:00] our lab, and we actually have results [16:02] from it. So, these are the first ever [16:04] images generated from such a computer. [16:07] Now, great, cool. But does it do [16:09] anything useful uh beyond just um uh [16:12] images? you can actually do any kind of [16:14] a task like sequence modeling or or [16:16] language models. But the interesting [16:17] thing here is it's only five 500 or so [16:20] nano per image. So to put that in [16:23] perspective a a normal computer like a [16:26] GPU is on the order of millig. So nanog [16:30] is 10 theus 9th. It's really really [16:32] small. So it's many orders of magnitude [16:35] more efficient than a standard computer [16:37] and it's because it just doesn't move [16:38] information around. This is proof [16:40] positive this works. So this is [16:42] literally the first time we're talking [16:43] about it publicly. So um yeah, thank [16:46] you. [16:50] So what's cool here is this is really [16:52] the emergence of something new. So [16:53] computers have gone from CPUs to GPUs [16:56] going more and more parallel to compute [16:58] and memory which is even more parallel [17:00] and fine grain. Um but all of these are [17:02] what we call vonoyman architecture. They [17:04] have a memory and compute and we move [17:06] information back and forth. What we [17:08] built is what's called a dynamical [17:09] computer which actually has compute and [17:11] memory in one thing. We don't have a [17:13] memory interface. Each individual [17:15] computing element is a memory. So it's a [17:17] completely different way to look at the [17:18] problem. [17:20] And so we call this 4D computing where [17:22] we use the time dimension in the [17:23] dynamics and we actually use the [17:25] physical three dimensions of die [17:27] stacking putting things vertically as [17:29] well as in a in a planer form. So we [17:31] have three dimensions from from the [17:32] physical and we have one dimension in [17:34] time. So this is actually kind of a new [17:36] way of thinking about a computer and it [17:37] it really is proving to work uh for [17:39] efficiency sake. So now what are the [17:42] implications if we build something [17:43] that's a thousand times more power [17:45] efficient? I think this is actually [17:46] pretty cool. So intelligence per watt is [17:49] what we care about. Can we optimize this [17:51] and make it better over time? There is [17:53] actually a thermodynamic limit which you [17:55] can never exceed. And uh the mamillian [17:58] brains though animal brains are [17:59] somewhere within one or two orders of [18:01] magnitude of that. Today we're on the [18:03] far left of this graph and we're about [18:06] 10 billion times away. That's 10 one [18:08] with 10 zeros after it uh from that [18:11] thermodynamic limit. We think in that [18:14] three and a half years we can hit the [18:15] the limits of 2D lithography. And the [18:18] the overarching goal of this company is [18:20] to beat biology. We want to make [18:21] something better and enable you know [18:24] compute everywhere including compute in [18:26] new robotic um forms uh and things like [18:29] this in the next uh decade or so. [18:33] I think what'll be interesting is that [18:34] we'll see the shift going from like big [18:36] big data centers with gigawatts to many [18:38] small data centers all over the place. I [18:41] think this is a good thing. It actually [18:42] makes things that are more [18:43] environmentally friendly, more local, [18:45] more adaptive. [18:47] And as I said, I think enabling the [18:50] ability to build billions of robots that [18:52] kind of dynamically assemble to solve [18:55] problems in our world is something [18:57] that's I think is actually really cool. [18:58] This is something that will enable us to [19:00] think about bigger problems and do even [19:02] more. [19:03] And you know, I talked about AI being a [19:05] trillion dollar market. Well, if we [19:07] disrupt it by a thousandx, it's there's [19:10] a there's a concept called Jieven's [19:12] paradox where when you drop the [19:13] underlying um cost of an asset, you [19:16] actually consume more more than the drop [19:19] of that asset. So, if you make something [19:20] half the price, you'll consume more than [19:21] 2x. If you cons you make something [19:24] 1,000th the price, you'll consume more [19:25] than 1/ 1,000th of it. And uh I think [19:28] this will create the largest market that [19:29] humanity's ever seen. [19:31] >> I mean that was extremely unexpected. I [19:34] got to say that was pretty amazing. [19:37] >> Let me let me start with actually [19:38] probably the thing that's on everybody's [19:40] mind which is [19:41] >> to the extent that this works Nav and [19:43] you probably saw Jensen earlier. [19:45] >> There just needs to be an entire [19:46] ecosystem of people that are beside you [19:48] and around you whether it's the fabs [19:50] packagers etc etc. [19:53] What does it take to get from this early [19:56] version [19:58] >> to something that sits in somebody's [20:00] hand or that people use and how do you [20:03] see that path in terms of time and [20:05] complexity? What does that look like? [20:06] >> Yeah, timewise we're within two years of [20:09] of getting it to a full product. I mean [20:12] >> and what is the product? [20:13] >> Yeah, [20:13] >> it's a VM that sits somewhere that you [20:15] guys manage and [20:16] >> effectively we're we're building a new [20:18] data center product for so it's a whole [20:20] rack. as a system, right? [20:21] >> And the idea is that basically we'll run [20:23] those um uh models on it. And so tokens [20:26] in, tokens out through a network cable, [20:28] but the inner guts are completely [20:29] different than an existing computer. [20:31] >> And do you expect that you'll have to [20:32] move to support the existing model [20:35] families and existing architectures? And [20:37] will this work in a world where, you [20:39] know, we've spent all of this time like, [20:41] okay, KV cache and let's all like this [20:43] all just so mechanically reductive based [20:45] on, as you said, these abstractions that [20:47] we've lived on, right? So how do you [20:48] expect the rest of us to kind of move [20:50] towards this? Because I mean I think you [20:52] see that efficiency curve. We'd all want [20:53] it. So how do we take advantage of it? [20:55] >> Yeah. So I think there's a there's a [20:56] sliding scale between you know how much [20:58] better something is and how much pain [21:00] you'll you'll take to move to it. And [21:02] you know I I basically took the the tack [21:04] of like let's make it really really [21:05] compelling to move. There is going to be [21:07] some work to port things over. We [21:09] actually don't port at the operations [21:11] layer, you port the model layer. So yes, [21:13] the existing models will work, but [21:15] there's a fair bit of compute required [21:17] to make that transition happen [21:19] >> and very basic elements like mapm does [21:22] it is that exist in your it doesn't [21:24] exist. It's [21:24] >> I mean you can you can characterize it [21:26] as map but it doesn't implement it as [21:28] maple. Okay, [21:29] >> it implements it as a sort of time [21:30] varying behavior. But each one of those [21:32] time steps you can analyze as basically [21:34] a matrix of the current state times a [21:36] transition matrix. [21:37] >> And when you're building a team like [21:38] this like who are these people? These [21:39] are biologists plus physicists plus what [21:42] are these people? [21:43] >> Yes. Um like it's sort of like theorists [21:45] that that come up with these uh uh like [21:48] dynamical systems theory has been around [21:49] for 100 years. Yeah. [21:50] >> So we got people from that world and [21:52] then we got people who actually build [21:53] chips and they don't talk to each other. [21:54] >> They don't talk to each other. [21:55] >> So we had to facilitate that. That's [21:57] actually one of the most challenging [21:58] things about this company is the span of [21:59] talents that we have is so big that [22:01] getting them to kind of all coordinate [22:03] and build one thing is actually pretty [22:04] hard. Well, and what is this like CUDA [22:06] like equivalent, if you will, just to [22:08] use a bad analogy, that allows these [22:10] people up here to talk to these people [22:11] down there. [22:12] >> Yeah. So, we actually have built a set [22:13] of libraries in Python. In Python, [22:16] >> it's it's Python. It's not CUDA, but [22:17] it's uh you know, it's it's a language [22:19] of sorts that allows you to kind of [22:20] express time varying elements that have [22:23] stoastic behavior. [22:24] >> I mean, it's incredibly impressive. It's [22:26] so ambitious. Thank you very much. It's [22:28] great to see you. Great to see you. [22:29] Amazing.