WebAI CEO Says Monolithic LLMs Face Netscape-Style Commoditization

AI as a Civilization: Networks of Expert Models, Not One Model to Rule Them All With webAI — Ai4

Ai4

Frontier foundation models will commoditize into low-margin utilities like early web browsers, shifting enduring enterprise value to decentralized networks of fine-tuned edge models tied to operational outcomes.

AI-assisted summary of AI as a Civilization: Networks of Expert Models, Not One Model to Rule Them All With webAI — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 24, 2025, this presentation features David Stout, Founder and CEO of webAI. According to the video description, the session explores moving away from monolithic AI models and vanity usage metrics toward a distributed 'civilization' of sovereign, on-device expert models engineered for measurable enterprise ROI, latency reduction, and data privacy.

## TL;DR David Stout argues that the AI industry is currently stalled by treating massive foundation models as end products and relying on unsustainable unit economics and vanity metrics like Daily Active Users. Drawing parallels to distributed aviation networks and the commoditization of web browsers, he proposes that true superintelligence is achieved through decentralized networks of specialized, fine-tuned models rather than one all-encompassing system. Stout outlines webAI's low-level infrastructure primitives and asynchronous discovery layer, which enable heterogeneous edge models to collaborate securely without cloud dependencies. This architectural shift transitions enterprise AI from high-friction, prompt-and-extract queries to an ambient, proactive operating system.

## Atmosphere & style The talk carries an incisive, pragmatic keynote tone, blending understated humor with sharp critiques of prevailing Silicon Valley assumptions. Stout guides the audience through grounded, real-world analogies—from the logistics of a flight from Austin to Las Vegas to the absurdity of selling coffee at a persistent loss—to demystify distributed systems. Listening to the session feels like an urgent, clarifying briefing on enterprise software economics, trading speculative frontier hype for disciplined, ground-up systems engineering.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=4s) True superintelligence mirrors complex human systems like commercial aviation, relying on a distributed network of specialized expert models rather than a single monolithic brain. - [03:15](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=195s) The AI industry is overly focused on internet vanity metrics such as DAUs and query volumes instead of outcome-based metrics, resulting in high capital burn while roughly 78% of enterprise AI projects fail to reach production. - [05:31](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=331s) Frontier AI unit economics are unsustainable because prompt engineering subsidizes negative margins, and user chat prompts provide inquiries rather than the verified ground-truth answers required for domain-specific training. - [07:59](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=479s) Base foundation models are falling into the 'Netscape Trap'—rapidly commoditizing into basic utilities while enduring economic value accrues to domain-specific outcome delivery. - [10:09](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=609s) Using massive generalist models for narrow industrial tasks represents an inefficient 'elephant gun' approach, as core value concentrates above the model (specialized pipelines) and below it (runtime execution). - [10:09](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=609s) webAI builds low-level primitives compiled down to binary neural networks (BNNs) and custom GPU shaders to execute non-transformer, non-autoregressive models directly on edge hardware. - [12:31](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=751s) Crossing the critical 10% accuracy threshold for enterprise deployment requires fine-tuning on proprietary operational data rather than relying on brittle, token-heavy prompt engineering. - [15:38](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=938s) Resilient multi-model coordination requires asynchronous discovery and communication layers across edge devices rather than rigid, centralized agent loops. - [18:20](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=1100s) Enterprise architectures must evolve from load balancing raw compute to load balancing intelligence across thousands of sovereign models to eliminate single points of failure. - [18:20](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=1100s) A decentralized civilization of diverse, heterogeneous models provides a statistically sound and democratic path to artificial superintelligence compared to relying on singular monolithic models. - [20:10](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=1210s) The future of AI will transition from a high-touch 'prompt-and-extract' interrogation paradigm into an ambient, proactive 'push model' that functions as civilization's operating system.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=4s) AI as a Civilization and the Flight to Las Vegas

The prevailing paradigm in artificial intelligence often presumes a future dominated by a single, monolithic model that knows and executes everything. webAI challenges this framing with the concept of "AI as a civilization"—the belief that the future belongs not to one all-encompassing system, but to hundreds of thousands or even millions of small, sovereign, distinct models working together. To illustrate how real-world intelligence actually functions, David Stout reflects on his journey from Austin to the conference in Las Vegas. While joking that this trip marked one of the few times he realized Austin might not be the hottest place on Earth, he details the seemingly ordinary yet vastly complex process of commercial travel: hailing an Uber, passing through TSA security, waiting at the Delta Sky Club, boarding an aircraft, cruising safely through the sky in a pressurized metal tube with jet engines, landing in Nevada, and taking another car to the venue.

```mermaid flowchart TD accTitle: Distributed Intelligence in Aviation accDescription: A multi-agent breakdown of airline travel showing how distinct specialized intelligences combine to produce a single safe flight experience.

subgraph Ground["Ground Operations"] A["Rideshare Driver\n(Urban Navigation)"] --> B["TSA Security\n(Protocol & Screening)"] B --> C["Airport Logistics\n(Gate & Lounge Ops)"] end

subgraph Flight["In-Flight Coordination"] D["Flight Crew & Pilots\n(Aviation Training)"] E["Airframe Avionics\n(Esoteric Flight Tech)"] F["Air Traffic Control\n(~15 Sector Hand-offs)"] D <--> E D <--> F end

C --> Flight Flight --> G["Outcome:\nSingle Safe, Super-Intelligent Journey"] ```

This journey relies on an intricate stack of distributed, specialized intelligences rather than a single master brain. Drivers possess deep navigation instincts; TSA officers are trained on safety protocols; flight crews and pilots undergo thousands of hours of specialized instruction to manage safety and mission completion; esoteric avionics software runs specialized onboard systems; and air traffic control hands off airspace management across roughly 15 distinct tower transitions during a single flight between Texas and Nevada. None of these elements is singular, yet together they compose a single, super-intelligent outcome. Stout argues that current AI initiatives are largely failing to construct super-intelligent outcomes because the industry remains obsessed with creating monolithic experiences rather than coordinating specialized problem solvers.

- **The Core Analogy**: True super-intelligence mirrors modern transportation—a network of specialized human and technological experts coordinating seamlessly. - **Airspace Hand-offs**: A routine domestic flight from Austin to Las Vegas transitions through nearly 15 separate air traffic control sectors to guarantee airspace safety.

---

[03:15](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=195s) Counting Cars Instead of Building Cities: Vanity Metrics in AI

Much of today's AI ecosystem is focused on "counting cars" rather than "building cities." When city planners design urban environments, their objective is never to maximize traffic congestion, slow commutes, and inflate transit costs. Instead, successful cities are engineered around specific desired outcomes, establishing clear metrics to ensure scalable, smooth flow. Today's prominent AI organizations, however, have inherited misleading metrics from consumer internet platforms—most notably Daily Active Users (DAU)—which fail to translate into healthy unit economics for AI workloads.

Under current business models, AI companies with the largest user bases often burn the most capital. While 78% of organizations reported using AI (a significant jump from 55% the prior year), approximately 78% to 79% of those enterprises have failed to deploy a functional use case into live production environments. This discrepancy occurs because organizations chase raw usage and "congestion stats" rather than defining the specific, measurable enterprise outcomes they need to deliver.

| Metric Type | Typical Focus | Core Problem in AI Deployments | |:--- |:--- |:--- | | **Internet Vanity Metrics** | Daily Active Users (DAU), query volumes, prompt counts | Subsidizes high compute costs without delivering profitable enterprise solutions ("Prompts don't pay the bills"). | | **Outcome-Based Metrics** | Production deployment rate, latency reduction, task accuracy | Connects model behavior directly to specific business problems and acute operational bottlenecks. |

---

[05:31](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=331s) Flawed Unit Economics and the Myth of User Data

The assertion that "prompts don't pay the bills" underlines a disconnect from fundamental business principles. If an entrepreneur approached Chase Bank seeking a small business loan for a hometown coffee shop, promising to sell cups of coffee for $6 while spending $10 to brew each one, no loan officer would approve the financing on the promise that "eventually the coffee will be so good everyone will want it, and then we'll change the price." Yet this exact subsidization logic dominates frontier AI conversations.

Proponents often defend negative margins by claiming that raw chat interactions generate proprietary data to train subsequent models. However, everyday users chatting with LLMs ask questions rather than providing verified answers. Because training robust question-and-answer (Q&A) datasets requires accurate ground-truth pairs, conversational traffic provides only half the equation. Consequently, enterprise clients find that massive foundational models—while scientifically impressive—cannot simply be prompt-engineered into solving acute, highly regulated domain challenges, such as diagnosing and servicing Pratt & Whitney jet engines.

- **The $6 vs. $10 Unit Margin**: Selling inferences at a loss under the expectation of future pricing power defies basic enterprise economics. - **The QA Data Deficit**: Prompts capture user inquiries, not the authoritative, specialized domain answers needed to train expert systems. - **Prototyping vs. Production Gap**: Prompt engineering produces functional demos, but fails when exposed to mission-critical, edge-case enterprise operations.

---

[07:59](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=479s) The Netscape Trap: Base Models Are Not the Product

The current race among frontier lab providers mirrors the early browser wars of the internet era. When Marc Andreessen and Netscape built Navigator, the prevailing belief was that by building the primary gateway to the web, they would own the economics of the internet itself. While creating the browser was an immense technical and scientific feat, browsing software quickly commoditized into a free utility. The enduring value was captured not by the software facilitating access, but by companies like Amazon that harnessed the protocol to deliver real-world outcomes, such as selling and shipping books online.

```mermaid flowchart LR accTitle: The Netscape Historical Parallel accDescription: Comparing Netscape's browser commoditization with foundation models versus outcome-oriented businesses like Amazon.

subgraph InternetEra["1990s Internet Era"] N["Netscape Navigator\n(Access Tool / Browser)"] -.->|Commoditized into Free Utility| N_Loss["Low Value Capture"] A_Old["Amazon\n(E-Commerce Outcome: Books)"] -->|Built on Protocol| A_Win["Massive Value Creation"] end

subgraph AIEra["Modern AI Era"] M["Monolithic Base Models\n(Foundational Commodity)"] -.->|Rapidly Commoditizing| M_Loss["Shrinking Margin Moat"] E["Specialized Expert Networks\n(Enterprise Business Outcomes)"] -->|Solves Acute Industry Needs| E_Win["High Enterprise ROI"] end ```

Foundation models are undergoing an identical commoditization curve at an exponential rate. Treating a raw foundation model as the end product conflates access technology with the actual business solution. Just as few people today pay a subscription simply to run a web browser, foundational models will serve as the plumbing, while enduring economic value will accrue to networks of sovereign models designed specifically to execute concrete enterprise tasks.

- **Historical Parallels**: Netscape pioneered web access but captured negligible value compared to the transaction platforms built on top of the web. - **Commoditization of Base Weights**: Frontier models are rapidly matching capabilities across providers, eroding the defensibility of generalized base LLMs. - **Core Principle**: "A base model is not your product"—value lives in delivering reliable, measurable domain outcomes.

[10:09](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=609s) The Elephant Gun Problem and WebAI's Primitive Stack

Many enterprises adopting AI make the fundamental mistake of treating a massive foundation model as their core product, attempting to control it from the outside using steering guidelines, rigid guardrails, and extensive prompt engineering. In practice, using these monolithic, generalized models for specialized industrial tasks is like bringing an "elephant gun" to a job that requires surgical precision. For specific operational workflows—such as automated quality assurance on a manufacturing conveyor belt—a narrow, focused solution is far more effective. The long-term enterprise market will not be dominated by a single "one model to rule them all" accessed via API; instead, specific end-product requirements will dictate model architecture. Value does not reside within the monolithic model itself, but rather *above* it (in specialized training pipelines and domain-specific vertical applications, like world-class medical models) and *below* it (in low-level runtime execution environments).

```mermaid flowchart TD accTitle: Monolithic vs Distributed Civilization Architecture subgraph Mono["Monolithic Approach (Fragile)"] A["Massive General Model"] -->|"Prompt Engineering & Steering"| B["Single API Bottleneck"] B --> C["Generic Output"] end

subgraph Civ["Civilization of Experts (Resilient)"] direction LR D["F1 Suspension Model"] <--> E["Discovery & Async Mesh"] E <--> F["Conveyor QA Model"] E <--> G["Aircraft Engine Model"] E <--> H["Local Medical Model"] end ```

Addressing this lower-level opportunity, webAI positions itself not as a model vendor, but as a tooling and infrastructure company. Rather than merely quantizing and compressing existing models, the team engineered a ground-up primitive stack that compiles directly down to binary neural networks (BNNs) and custom GPU shaders. Their runtime departs from standard autoregressive transformer architectures to enable non-transformer execution distributed across heterogeneous hardware environments, specifically targeting mission-critical edge deployments where cloud connectivity is absent.

- **Infrastructure layer**: WebAI writes low-level primitives directly to BNNs and hardware shaders to run non-transformer, non-autoregressive models on heterogeneous edge devices without cloud connectivity. - **Value distribution**: Real enterprise value resides below the model (runtime/execution infrastructure) and above the model (domain-specific data pipelines), commoditizing the monolithic center.

---

[12:31](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=751s) The 10% Deployment Threshold: Prompt Engineering vs. Inherent Inference

In production environments, closing a seemingly modest 10% accuracy gap—elevating performance from 60–70% up to 80–90%—is the exact tipping point that dictates whether an enterprise can deploy an AI system into the wild. While developers eagerly adopt tools like Cursor because technical users are "low hanging fruit" willing to tolerate rough edges and guide the system, non-technical users cannot bridge that gap; handing such a tool to an everyday user will not magically generate a functioning website without deep domain intervention. Achieving that crucial accuracy jump requires tying AI performance directly to tangible enterprise outcomes rather than vanity metrics. In high-stakes environments—such as deploying on-device AI in a Formula 1 car to dynamically tune suspension dampening through corners for faster lap times—the metric of success is unambiguous. Training models on generic internet scrapings like Reddit cannot solve these challenges because the critical operational data only exists inside enterprise boundaries.

Partnering directly with enterprises unlocks the proprietary data needed to achieve this 10% boost through fine-tuning rather than brittle prompt engineering. Prompt engineering is suitable for quick prototypes, but commercially inefficient: every single query requires prepending a hidden system prompt to repeatedly convince a massive model of its identity and constraints. When fine-tuned, a model inherently knows its operational directive without needing constant runtime reminders. This drastically reduces inference costs, minimizes latency, increases reliability, and eliminates the absurd practice of taking a giant model trained on the entire internet and trying to persuade it to ignore poetry and finances to focus solely on checking a conveyor belt or adjusting an engine.

- **Outcome-driven evaluation**: Real-world utility is measured by hard metrics (e.g., F1 lap time reductions via real-time suspension tuning) rather than generalized benchmark scores. - **Prompt engineering overhead**: Prepending long system prompts forces general models into narrow roles on every call, inflating token counts and introducing behavioral drift. - **Dedicated inference**: Purpose-built models deliver faster execution, lower compute costs, and predictable behavior without conversational scaffolding.

---

[15:38](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=938s) Building a Resilient Civilization: Async Communication and Discovery

The architectural metaphor of human civilization highlights a vital lesson: historical civilizations collapse when they become rigid or overly centralized, meaning longevity is never guaranteed. To build enduring systems, AI architectures must mirror successful societies by distributing ownership and intelligence downstream rather than centralizing control. A true "civilization" of narrow expert models—ranging from aircraft maintenance models to on-device driving assistants—cannot coordinate using simplistic agentic loops or rigid, hard-coded `for` loops. Human civilization functions through asynchronous communication triggered by evolving scenarios and outcomes, and machine networks must operate with the same organic adaptability.

To solve this inter-model communication problem, webAI recruited Netflix's former head of serverless systems—the engineer who designed the seamless zero-touch authentication protocols used to log users into hotel smart TVs. By applying those principles to distributed intelligence, webAI built a discovery layer where small, discrete models automatically discover each other, evaluate available edge hardware, and collaborate asynchronously. This approach allows small models running locally on mobile phones or edge hardware to collectively orchestrate complex physical systems, like commercial aircraft, demonstrating that true superintelligence arises from collective networks rather than isolated monoliths.

> "Everything in nature, everything in science tells us that the collective is how we drive super intelligent outcomes. Even your body isn't singular."

---

[18:20](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=1100s) Load Balancing Intelligence and the Democratic Path to Superintelligence

Modern cloud architectures focus heavily on load balancing raw compute, but enterprise systems must transition to "load balancing intelligence." Distributing 3,000 unique, sovereign models across an entire organization creates structural resilience: if individual nodes or network connections fail, the remaining edge models continue communicating, resolving tasks, and maintaining operations locally without taking down the enterprise.

```mermaid flowchart LR accTitle: Load Balancing Compute vs Intelligence subgraph Traditional["Load Balancing Compute"] direction TB LB1["Load Balancer"] --> API1["Central Monolith Replica A"] LB1 --> API2["Central Monolith Replica B"] end

subgraph WebAI["Load Balancing Intelligence"] direction TB M1["Model 1 (Finance)"] <--> M2["Model 2 (Edge QA)"] M2 <--> M3["Model 3 (Logistics)"] M3 <--> M1 end ```

Furthermore, diverse model ensembles produce superior reasoning outcomes, mirroring how diverse teams outperform solitary decision-makers in human organizations. Pursuing artificial superintelligence through a single monolithic model relies on improbable statistical anomalies. A distributed civilization of expert models offers a statistically sound and democratic path forward, ensuring humanity actively participates in creating sovereign intelligence rather than remaining passive consumers of centralized infrastructure:

- **Resilience through distribution**: Deploying thousands of localized expert models eliminates single points of failure across enterprise workflows. - **Diversity of modeling**: Combining heterogeneous models yields richer, more accurate problem-solving than relying on a single uniform architecture. - **Participation over consumption**: A decentralized model network allows every enterprise and developer to contribute sovereign intelligence to the broader ecosystem. - **Invisible AI**: The ultimate indicator of mature, successful AI deployment is when the technology seamlessly integrates into infrastructure and disappears from the user's conscious view.

[20:10](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=1210s) From Prompt-and-Extract to Civilization's Operating System

David Stout concludes his presentation by contrasting the current state of artificial intelligence with the future webAI is building. Today's paradigm is defined by friction and high-touch interactions, relying on a rigid "prompt extract model" where users must constantly question, interrogate, and manually coax machines into delivering the results they need. As the paradigm shifts toward a decentralized network of expert models, artificial intelligence will transition from a passive, query-based tool into a true proactive partner.

Instead of waiting to be prompted, this architecture enables AI to function as "civilization's operating system"—moving to a dynamic "push model" that continuously guides, assists, and pushes relevant insights forward across environments without requiring human interrogation. Stout wraps up the session by inviting attendees to visit the webAI booth to continue the conversation, thanking the audience for their time and support.

- **Current Paradigm vs. Future State**: Moving away from high-touch, "prompt extract" interrogations to an ambient "push model" that actively guides workflows. - **AI as Civilization's Operating System**: Small, sovereign expert models working collaboratively in the background to serve as an intelligent, proactive foundation for society.

```mermaid flowchart LR accTitle: Paradigm Shift from Prompt-Extract to Proactive Push AI subgraph Current ["Current AI Paradigm"] A["User Interrogation"] --> B["Monolithic Model"] B --> C["Prompt & Extract Output"] end subgraph Future ["Civilization's Operating System"] D["Network of Expert Models"] --> E["Proactive Push & Guidance"] E --> F["Autonomous Outcomes"] end ```

## Notable quotes > [00:04](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=4s) True super-intelligence mirrors modern transportation—a network of specialized human and technological experts coordinating seamlessly. — David Stout

> [03:15](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=195s) Prompts don't pay the bills. — David Stout

> [07:59](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=479s) A base model is not your product. — David Stout

> [10:09](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=609s) one model to rule them all — David Stout

> [15:38](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=938s) Everything in nature, everything in science tells us that the collective is how we drive super intelligent outcomes. Even your body isn't singular. — David Stout

> [18:20](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=1100s) load balancing intelligence — David Stout

> [20:10](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=1210s) civilization's operating system — David Stout

> [20:10](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=1210s) prompt extract model — David Stout

## Mentioned David Stout, webAI, Ai4, Uber, TSA, Delta Sky Club, Chase Bank, Pratt & Whitney, Marc Andreessen, Netscape Navigator, Amazon, Cursor, Reddit, Formula 1, Netflix

## Video description David Stout, Founder & CEO, webAI

Abstracts

Imagine AI not as a single monolithic model, but as a civilization of experts: small, sovereign models, each great at one thing, coordinating across devices, plants, clinics, and trading floors.

That’s the future we’re building. This session will reframe how we design, deploy, and measure AI: from usage vanity metrics to outcomes per watt, risk reduced per query, and lives improved per inference. David will share how organizations are standing up distributed, on-device intelligence that protects data, crushes latency, and drives measurable ROI, without sacrificing interoperability or control.

Attendees will get a concrete blueprint for assembling expert swarms: how to place them, connect them, govern them, and prove their value.

[Source](https://www.youtube.com/watch?v=cK3EvKdHEGo)

中文简报

WebAI CEO称单体大模型正面临“网景式”商品化 前沿基础模型将如同早期的网页浏览器一样商品化为低利润的基础设施，而持久的企业价值将转向与业务成果紧密挂钩的微调边缘模型去中心化网络。 David Stout认为，AI行业目前因将庞大的基础模型视作最终产品，并依赖不可持续的单位经济效益以及日活跃用户（DAU）等虚荣指标而陷入停滞。他以分布式航空网络和网页浏览器的商品化为例，提出真正的超级智能应通过专用的微调模型去中心化网络来实现，而非依赖单一包罗万象的庞大系统。Stout概述了webAI的底层基础设施原语与异步发现层，使异构边缘模型在无需云端依赖的情况下实现安全协作。这一架构转向推动企业级AI从高摩擦的“提示-提取”式查询演进为环境感知型的主动式操作系统。 webAI 提出“AI 即文明”的理念，主张超级智能并非源于单一通晓一切的庞大模型，而是由大量专注于特定领域的小型主权模型协同构建，如同民用航空系统依赖精密分布式的专家分工体系。 (0:04) 当前 AI 行业过度聚焦 DAU 与查询量等互联网虚荣指标，导致高昂资本消耗且约 78% 至 79% 的企业项目无法上线；通用基座模型正在陷入 Netscape 式的快速商品化陷阱，长远价值将流向解决行业具体痛点的垂直网络。 (3:15) 在工业专有任务中使用庞大通用模型如同低效的“大象枪”；webAI 构建了直接编译为二进制神经网络（BNN）与定制 GPU 着色器的底层运行环境，支持在离线边缘硬件上高效运行非 Transformer 模型。 (10:09) 跨越企业落地关键的 10% 准确率门槛，必须基于企业专有运营数据进行微调，而非依赖脆弱且大幅推高 Token 成本的 Prompt Engineering，让模型在推理时天然具备特定场景的专业指令。 (12:31) 企业架构亟需从传统的“算力负载均衡”转向“智能负载均衡”，借助异步发现与通信层组织数千个端侧模型协同运作，消除单点故障并将 AI 从被动提问模式演进为主动赋能的“文明操作系统”。 (18:20)

Reading summary

webAI's David Stout argues that the artificial intelligence industry's reliance on centralized, monolithic models is both economically unsustainable and architecturally flawed. Instead, he contends that true superintelligence will function as a decentralized civilization of specialized, sovereign models collaborating across enterprise and edge environments. Flawed Economics and Base Model Commoditization (3:15) Stout compares commercial aviation—where a safe flight from Austin to Las Vegas relies on coordinated specialists including ground crews, avionics, and roughly 15 air traffic control sector hand-offs—to how artificial intelligence should function. In contrast, current industry initiatives chase consumer vanity metrics like Daily Active Users rather than measurable operational outcomes. Stout highlights that while 78% of organizations report experimenting with AI, between 78% and 79% have failed to deploy functional use cases into live enterprise production. This gap is compounded by negative unit economics. Stout likens current inference subsidization to a business selling $6 cups of coffee that cost $10 to brew, relying on an unproven hope of future pricing power. Furthermore, public conversational prompts supply user questions rather than verified ground-truth data, leaving foundational models ill-prepared for acute industry challenges. Drawing a historical parallel, Stout argues base models will suffer the fate of Netscape Navigator: pioneering access tools that commoditized into free utilities while lasting economic value accrued to platforms built on top. Specialized Training Versus Brittle Prompt Engineering (10:09) Enterprises often treat massive foundational models as end products, attempting to steer them with conversational guidelines and prompt engineering. Stout characterizes this as bringing an "elephant gun" to surgical tasks, arguing that generalized architectures are ill-suited for narrow industrial workflows like conveyor-belt quality assurance. Sustainable value, he argues, resides not in the base model itself, but above it in specialized domain data and below it in execution runtimes. Closing the 10% accuracy gap needed to elevate models from 70% to 80–90% production viability requires internal enterprise data rather than internet scrapings. While developers using coding assistants tolerate rough edges, non-technical enterprise operators cannot bridge that threshold. Stout explains that prompt engineering inflates costs and latency because long system prompts must repeatedly convince a massive model to ignore unrelated internet data. Dedicated fine-tuning produces purpose-built models that inherently understand their directives, such as tuning suspension dampening in real time for Formula 1 cars. Asynchronous Networks and Load Balancing Intelligence (15:38) Drawing on historical patterns where centralized societies collapsed due to rigidity, Stout argues that resilient AI architectures must distribute intelligence downstream. webAI implements this through an asynchronous discovery layer—architected by Netflix's former head of serverless systems—allowing sovereign models on local devices to discover each other and coordinate dynamically without brittle, hard-coded loops. To enable edge execution without cloud reliance, webAI built a primitive stack compiling directly to binary neural networks and custom GPU shaders for non-transformer models. Stout contrasts traditional cloud architectures that load balance compute with the concept of "load balancing intelligence" across thousands of heterogeneous models. Distributing localized expert models eliminates single points of failure, mirrors the reasoning advantages of diverse human teams, and broadens ecosystem participation. Ultimately, Stout forecasts a shift away from high-friction "prompt-and-extract" user interactions toward an ambient "push" operating system that proactively delivers guidance within enterprise workflows. webAI's thesis rests on the premise that base foundation models will rapidly commoditize and that edge runtimes can effectively bypass standard transformer architectures. However, adopting this civilization model assumes enterprises have the operational capacity and specialized data pipelines required to fine-tune, coordinate, and maintain thousands of autonomous local models without relying on centralized cloud providers.

精炼阅读摘要

webAI发言人David Stout提出，人工智能的未来并非由单一全知全能的单体大模型主导，而是由数以百万计的小型、自主专业模型构成的“AI文明”网络。他认为，当前行业过度沉迷于消费级互联网的虚荣指标与通用大模型，忽视了底层推理经济学与企业具体成果，真正的超级智能必须依托分布式专家网络的协同协作与边缘执行。 分布式智能的本质与单体架构局限 (0:04) Stout指出，当前AI领域普遍假定未来将由掌握一切知识的单体模型掌控，但现实世界复杂系统的运作逻辑完全不同。他以从奥斯汀前往拉斯维加斯参加会议的商业飞行为例：整个旅程不仅涉及网约车司机的城市路况判断、TSA安检人员的筛查规程、机场地勤运作，还高度依赖飞行员受训数千小时的专业技能、机载航空电子软件，以及得州至内华达州航线中经历的约15次不同空管塔台的无缝交接。这一系列各司其职、高度分散的专业智能共同作用，才保障了单次超智能飞行体验的安全实现。 Stout认为，目前的AI项目之所以难以构建真正的超级智能成果，正是因为行业执着于打造单一、庞大的通用体验，而非协调专门的问题解决者。真实世界中没有任何单一实体可以独揽全局，将解决工业级难题的希望寄托于单一通用模型，背离了现代复杂社会与自然演进的协同运作规律。 虚荣指标与网景陷阱下的经济学脱节 (3:15) 在商业模式层面，Stout批评当今AI生态过度关注“数车流”而非“建城市”。许多AI机构沿袭了消费级互联网的日活跃用户数（DAU）和提示词调用量等虚荣指标，但提示词本身无法偿付高昂的算力账单。调查数据显示，尽管78%的企业表示正在使用AI（较上一年的55%大幅攀升），但其中约78%至79%的企业未能在实际生产环境中上线可运行的用例。企业若将推理成本高企的通用模型以亏本方式提供给用户，寄希望于未来提价，如同以10美元成本制作咖啡却仅售6美元的小微贷款方案，在商业底层逻辑上无法持续。 与此同时，寄希望于从用户日常对话中提取数据以训练后续模型的设想也存在缺陷。日常对话大多由用户提出问题，缺乏专业核验的真实基准答案，无法弥补构建专家系统所需的“问答数据赤字”。这使得通用基座模型陷入类似早期Netscape浏览器的商品化陷阱：尽管技术突破显著，但单纯充当互联网访问入口的软件迅速沦为免费公用设施，核心经济价值最终被亚马逊等专注于实际交易成果的企业捕获。基础大模型权重的快速同质化意味着单体通用模型并非真正的终端产品，其利润护城河正在加速萎缩。 底层硬件原语与专用模型微调 (10:09) Stout用“猎象枪”比喻企业用庞大基础模型处理精细工业任务的错配。在制造业传送带质检或喷气发动机检修等受监管的垂直场景中，试图用冗长提示词和护栏在外部牵引通用大模型，不仅导致每一次API调用都因重复声明角色设定而增加Token成本与延迟，还会引入不可预测的行为漂移。相比之下，跨越生产级部署的关键通常在于将准确率从60%至70%提升至80%至90%这一关键的10%门槛。这一跃升无法通过网络抓取数据实现，必须深入企业内部边界，利用专有数据进行微调，使模型天然掌握执行指令，摆脱对话式脚手架。 为支撑离线与边缘关键任务，webAI将自身定位为工具与基础设施提供商，而非通用模型厂商。其底层技术堆栈并未局限于对现有模型进行量化压缩，而是从零构建了能够直接编译为二进制神经网络（BNN）与定制GPU着色器的硬件原语堆栈。该运行时脱离了传统的自回归Transformer架构，支持非自回归模型在异构边缘硬件上离线运行。Stout指出，企业AI的真实价值链正在向两端转移：一端在模型下方，即底层的运行时执行基础设施；另一端在模型上方，即垂直领域的专有数据管线，而处在中间的通用单体模型则面临商品化。 智能负载均衡与主动推送式操作系统 (15:38) 在网络协作机制上，Stout借鉴人类文明演进的教训强调，过度集权与僵化的系统终将走向崩溃。一个由数十万或数百万个小型模型构成的AI文明，不能依赖简单的Agent循环或硬编码逻辑，而需采取类似现代分布式系统的异步通信与自主发现机制。webAI引入前Netflix无服务器系统架构师设计无感认证协议的思路，构建了让离线边缘小模型在手机或工业硬件上自动感知硬件资源并异步协作的发现层。 这种架构将传统的“计算负载均衡”升级为“智能负载均衡”。通过在企业各业务节点分布数千个主权模型，即便部分节点或网络连接中断，其余边缘模型仍能本地通信并解决任务，避免单点故障导致全盘瘫痪。此外，多样化模型组合的推理鲁棒性显著优于单一孤立决策者。Stout预测，AI系统最终将告别目前用户反复提问、审问并诱导机器的“提示-抽取”摩擦模式，演进为静默融入基础设施的“文明操作系统”；它将在后台主动分析情境并向外推送精准决策建议，让AI技术在深度嵌入业务流程后隐于无形。 David Stout倡导以分散、异构且深度细分的专业模型网络取代对集中式全能单体模型的盲目追求，主张唯有将AI锚定于具体的企业业务指标、专有微调数据与本地边缘执行基础设施，技术才能实现从高摩擦演示到商业级投产的跨越。不过，该论述由webAI高管呈现，主要代表了该团队推广其边缘底层堆栈与分布式网络架构的技术主张与商业愿景，并未深入探讨超大规模跨模型异步协同网络在实际落地中可能面临的数据一致性治理、端到端延迟开销以及去中心化安全审计等现实工程挑战。

Detailed notes — zh

## Background 本内容由 Ai4 频道于 2025 年 8 月 24 日在 YouTube 平台发布，记录了 webAI 创始人兼 CEO David Stout 的主题演讲。根据页面介绍，该演讲主张将 AI 重构为小型主权专家模型协同合作的“文明网络”，而非依赖单体大模型，旨在实现保护数据、降低延迟与高投资回报率的分布式端侧智能。 ## TL;DR David Stout 在演讲中批判了当前 AI 行业对单体基础大模型和 DAU 等虚荣指标的盲目追逐，指出基础模型正迅速大宗商品化，而商业价值真正沉淀在上下游与具体业务结果中。他以现代航空复杂协作系统为隐喻，提出“AI 即文明”的架构理念，倡导用非 Transformer 架构与底层算子重构轻量化端侧小模型。通过类似 Serverless 的设备自发现与异步协同机制，企业能够实现“智能负载均衡”，告别低效的提示词工程，推动 AI 从被动的“问答提取”迈向主动自治的“文明级操作系统”。 ## Atmosphere & style 演讲氛围务实、犀利且极具启发性，David Stout 摒弃了云端 AGI 的宏大叙事，以接地气的生活案例（如开咖啡馆、乘坐航班、生儿育女）切入，生动拆解了技术狂热背后的商业常识漏洞。整场分享节奏明快，语言兼具工程严谨性与幽默感，听众仿佛在聆听一位深谙工业落地的老练架构师直面房间里的大象，将复杂的分布式系统哲学讲得通透而令人信服。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=4s) AI 即文明的协同范式：未来的 AI 绝非由单一全能大模型统治一切，而是由成千上万个小型垂直主权模型在端侧与云端协同构成的去中心化网络，犹如现代复杂航空系统的分工运转。 - [03:15](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=195s) 虚荣指标与落地鸿沟：虽然 2024 年已有 78% 的企业声称使用 AI，但其中近 79% 的项目无法在实际生产中上线；盲目追逐 DAU 等流量指标掩盖了高昂的推理亏损与交付缺失。 - [05:31](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=331s) 提示词无法打破商业基本规律：C 端交互产生的大多是单向的“提问”而非标准“答案”，缺乏闭环数据；仅靠通用模型打补丁和 Prompt Engineering 无法解决高门槛严苛工业难题。 - [08:01](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=481s) 基础大模型的大宗商品化：正如当年网景浏览器迅速商品化而真正价值流向亚马逊等垂直业务公司一样，基础模型本身正在快速商品化，真正的企业级价值流向交付具体结果的专业模型网络。 - [10:09](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=609s) 拒绝“大炮打苍蝇”，价值在模型上下游：通用巨型模型在垂直场景效率低下，商业壁垒存在于“模型下方”（跨设备端侧高效运行的算子与基础设施）与“模型上方”（深度垂直的领域训练工作流）。 - [11:41](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=701s) 突破工业落地的 10% 精度生死线：webAI 深入底层硬件指令集与 GPU 着色器重写算子栈，摆脱传统 Transformer 自回归架构，实现异构边缘设备在无网络环境下的高效离线推理与高精度交付。 - [13:25](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=805s) 以确定性微调取代提示词工程：在工业场景中用长提示词约束大模型不仅成本高且容易产生幻觉，针对具体业务指标微调的专用小模型能够原生锁定目标并保障确定性产出。 - [15:38](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=938s) 借鉴无服务器架构的自发异步协同：利用设备与服务自发现机制，让散落在各边缘端的小模型能够互相感知与异步交互，形成类似人类社会的事件驱动群体智能协作。 - [18:21](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=1101s) 从算力负载均衡到“智能负载均衡”：在企业全球业务边缘部署数千个专业小模型，不仅能消除单点故障与提升系统工程韧性，更通过思想多元性（Diversity of Thought）大幅优化复杂决策质量。 - [20:10](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=1210s) 从“被动提取”演进为“主动推送”的文明操作系统：AI 交互正在告别需要人类反复质询的被动模式，转向由分布式专家模型主动预测与协同引导的文明级底层智能设施。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=4s) AI 即文明：从奥斯汀到拉斯维加斯的航空协作隐喻 在演讲开场时，webAI 的创始人兼 CEO David Stout 提出了一个核心理念——“AI 即文明”（AI as a Civilization）。行业目前普遍迷恋“由一个大一统的单一模型统治一切”的愿景，但真实的未来必然是由成千上万、乃至数百万个代表着不同专业形态的小型主权模型相互协作组成的网络。为了形象地阐释这一观点，他以自己从奥斯汀前往拉斯维加斯参会的旅程作为切入点：从家门口打 Uber 到机场、体验不需要脱鞋的 TSA 快速安检、在达美航空贵宾室（Delta Lounge）候机、登机飞越天空，最后乘车到达酒店。人们往往把现代出行视作理所当然，但把人类装进带发动机的金属圆筒并抛向万米高空，背后依赖的是极其庞大且分散的专业系统。 这段旅程的安全性与顺畅度绝非由某种“单一力量”保障。从熟悉路况和安全驾驶的网约车司机，到严格执行机场安全协议的安检人员；从经过海量专业飞行与应急训练的机组与机长，再到机身内高度垂直、专属于航空领域的硬核航电技术；甚至飞机在航程中需要切换近 15 次空管塔台，由不同区域的空管员确认空域安全。整个流程中没有任何一个环节是单一集中的，而是无数垂直领域的专业智能层层堆叠，才最终拼凑出一个“超级智能”的端到端体验。David 指出，当前 AI 行业迫切需要厘清“追求虚幻的超级模型体验”与“实现真正超级智能的现实业务结果”之间的本质区别。 ```mermaid flowchart LR accTitle: AI as a Civilization Aviation Analogy subgraph Departure["出发与地面交通"] A["网约车司机<br/>(路径导航/地面安全)"] --> B["TSA 安检协议<br/>(人身与违禁品安全)"] end subgraph InFlight["空中飞行协同"] C["专业机组与机长<br/>(海量飞行规程与应急训练)"] D["机载航电系统<br/>(专用航空软硬件)"] E["多区域空管塔台<br/>(约15次接力调度)"] C --- D --- E end subgraph Destination["到达交付"] F["超级智能的出行结果<br/>(安全准时抵达目的地)"] end Departure --> InFlight --> Destination ``` --- ### [03:15](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=195s) 数汽车还是建城市：虚荣指标与企业落地困境 David 指出，当下的 AI 行业正陷入一种“只顾着数路上的汽车，却忘记了如何建设城市”的盲目状态。优秀的城市规划师绝不会把“制造拥堵的交通流量”作为规划目标，更不会希望城市又慢、又贵、让人倍感挫败；相反，真正的城市设计始终以结果为导向：先明确希望实现的目标，确立衡量标准，再思考如何在大规模下稳定交付。然而，当前许多最风光的 AI 企业却依然在使用传统互联网公司的虚荣指标（Vanity Metrics）来评估 AI，例如将日活跃用户数（DAU）奉为圭臬。 将 DAU 视作 AI 核心指标是一个巨大的危险信号。现实情况是，拥有最多 C 端用户的 AI 公司往往也在承受最巨额的亏损。这种现象反映出它们本质上只是“套壳使用 AI 的互联网公司”，而非真正以 AI 交付硬核业务价值的“原生 AI 公司”。宏观数据也印证了这一断层：尽管 2024 年已有 78% 的组织声称在使用 AI（高于前一年的 55%），但在这些声称使用的企业中，有将近 79% 的项目依然未能成功部署到真实生产环境（in the wild）。全行业都在追逐没有实质意义的“拥堵数据”，却很少有人真正追问：这些方案究竟解决了哪些棘手问题？是否催生了能够解决“十亿美元级真实痛点”的产业实体？ - **组织采用率与落地落差**：企业声称使用 AI 的比例从 2023 年的 55% 飙升至 2024 年的 78%，但在使用 AI 的企业中，仍有接近 79% 的方案根本无法在实际生产业务中成功上线。 - **虚荣指标的陷阱**：高 DAU 掩盖了巨大的推理与运营亏损，以流量拥堵（Congestion Stats）代替实际产出（Outcomes）。 --- ### [08:01](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=481s) 历史的镜像：Netscape 浏览器的陷阱与基础模型的商品化 将当下的基础模型狂热与互联网早期的演变进行对比，可以发现历史惊人的相似性。回顾网景（Netscape）时代，创始人 Marc Andreessen 当时的商业逻辑是：网景打造了人们访问互联网的基础工具（浏览器），因此凭借这个浏览器就能掌控整个互联网的入口与价值。如今 AI 领域的论调如出一辙——“只要你拥有一个知晓一切的底层基础模型，你就赢得了 AI 战争”。但企业端与前沿科研的实际反馈却表明，基础大模型正以指数级的速度走向“大宗商品化”（Commodity）。 ```mermaid flowchart TD accTitle: Internet Era vs AI Era Commodity Comparison subgraph InternetEra["互联网时代 (1990s)"] N["网景浏览器 (Netscape)<br/>基础访问工具 / 核心技术极难"] -->|迅速大宗商品化/免费| B["工具层价值衰减"] A["亚马逊 (Amazon)<br/>利用网络卖书 / 交付实际业务结果"] -->|重塑商业闭环| E["捕获万亿美元真实价值"] end subgraph AIEra["当前 AI 时代"] FM["单一基础大模型 (Base Model)<br/>通用底座 / 耗资巨大但趋同"] -->|迅速大宗商品化| C["模型即商品 (并非最终产品)"] SW["专业化模型协作网络<br/>解决垂直产业硬核痛点"] -->|闭环业务结果| V["捕获最终企业级价值"] end ``` 在网景开发浏览器的同时，一家通过互联网线上卖书的小公司（亚马逊）也在悄然起步。浏览器确实能带用户访问网站，研发浏览器本身的计算机底层技术也非常艰深且不可或缺，但长远来看，真正的巨额商业价值并没有停留在浏览器本身，而是流向了利用底层技术促成实际业务结果的企业。如今的浏览器早已完全免费并高度普及，提供基础浏览工具本身无法支撑可持续的暴利。同理，“基础模型不是你的最终产品”，将基础模型本身当成终极护城河，很可能重蹈网景当年的覆辙。 --- ### [05:31](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=331s) 提示词不会替你买单：商业本质与问答数据的缺陷 David 直言：“提示词不会替你买单”（Prompts don't pay the bills）。商业的基本规律并未因 AI 的诞生而发生丝毫改变。他打了一个通俗的比喻：如果你打算在家乡开一家咖啡店，跑去大通银行（Chase）申请贷款，却对信贷经理说：“我打算每杯咖啡卖 6 美元，但目前每做一杯的成本是 10 美元；不过别担心，等我以后把咖啡做得足够好、所有人都离不开时，我再提价。”银行显然会立刻把你轰出去。然而，这恰恰是当前许多 AI 公司向市场和投资者讲述的荒谬故事。 支持这一模式的人通常辩解称，不断涌入的用户和 Prompt 交互能为公司提供海量宝贵数据，从而构建数据护城河。但这种逻辑存在致命缺陷：用户在使用对话式 AI 时，是在向模型“提问”，而不是在给模型“提供标准答案”。机器学习需要完整的“问题-答案”（QA）数据集来训练，仅凭单向的问题输入只有半边数据，根本无法构成闭环的高质量数据集。在企业级场景中，各大企业在经过大量 Prompt Engineering（提示工程）和原型制作后发现，仅靠通用基础模型打补丁根本无法稳定交付实际生产所需的精准结果。基础模型本身的技术突破固然令人惊叹，但它绝不可能直接解决普惠公司（Pratt & Whitney）航空发动机维修等高度垂直、极其严苛的现实工业难题。 - **咖啡店单位经济学谬误**：以 10 美元单杯成本制造、以 6 美元补贴销售，寄希望于规模扩大后再提价的商业假设，违背了商业常识。 - **用户交互数据的半边陷阱**：C 端交互产生的大多是单向的问题（Question），缺乏专业标注与验证的真实业务答案（Answer），难以单独支撑高价值闭环模型的持续进化。 - **提示工程的局限性**：在涉及精密制造、工业诊断等严苛领域，通用底座模型配合 Prompt Engineering 无法填补面向具体业务结果的专业落地鸿沟。 ### [10:09](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=609s) 拒绝“大炮打苍蝇”：价值在模型的上下游 很多企业在落地 AI 时往往陷入一个误区：试图将整个业务系统死死套在一个通用大模型周围，仿佛这个模型本身就是最终产品。为了让这些庞大而难以驾驭的模型听话，团队不得不投入海量精力去编写复杂的系统提示词（System Prompts）、制定繁琐的安全护栏与引导规则。然而在实际工业场景中，这种巨型通用模型就像是一把“打大象的猎枪”（Elephant Gun）——如果你的实际需求只是在工厂装配线上做流水线质检（QA），杀鸡用牛刀不仅成本高昂，效果也难以保证。相比于幻想用一个“君临天下”的全能模型（One Model to Rule Them All）并通过单一 API 统领一切，更切实际且已被商业验证的路径是由具体的产品场景反向驱动模型选型。 在 David Stout 看来，单体大模型本身正在快速泛指化与平庸化，真正的商业价值其实存在于“模型的上方与下方”。如果一家公司深耕“模型下方”，打造底层基础设施与工具链，让各类模型能够极致高效地在任意异构边缘设备上随处运行，这将产生巨大的底层赋能价值；而如果聚焦于“模型上方”，专注于构建最顶级的领域训练系统，打磨出在特定细分赛道（例如医疗诊断）无人能及的专业模型，同样能建立起深厚的行业壁垒。 ```mermaid flowchart TD accTitle: 传统单体架构与专业模型网络对比 subgraph Monolithic["传统单体模式 (大炮打苍蝇)"] direction TB M1["单一巨型通用模型 (API)"] --> P1["繁琐提示词工程 / 护栏控制"] P1 --> U1["各行业通用任务 (流水线质检 / 医疗 / 办公)"] end subgraph Network["webAI 文明网络模式 (专业协同)"] direction TB T1["上层：专业训练与领域工作流"] B1["底层：端侧轻量计算与着色器优化"] subgraph Experts["分布式专业小模型群"] E1["装配线视觉质检模型"] E2["F1 悬挂动态调节模型"] E3["航发动能分析模型"] end T1 --> Experts B1 --> Experts end ``` --- ### [11:41](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=701s) 10% 的质变门槛与端侧轻量化架构 谈及 webAI 的自身定位，David 强调他们并不是一家兜售现成模型的公司，而是一家深耕底层基础架构与运行工具链的企业。webAI 自研了运行端侧 AI 的核心库，其技术路径绝非简单的模型剪枝或低比特量化（Quantization），而是直接深入底层硬件指令集，面向 BNN（二进制神经网络）与 GPU 着色器（Shaders）重写了整套算子原语栈。他们主动摒弃了传统的 Transformer 与自回归（Autoregressive）生成架构，探索非 Transformer 的高效推理范式，支持模型在无网络连接、算力异构的极端边缘环境中独立运行。 在工业与企业级部署中，关键往往在于跨越那至关重要的“10% 精度跃升”。目前像 Cursor 这类 AI 代码辅助工具之所以能迅速风靡，是因为软件工程师本身就是最容易尝鲜的群体，只要工具能带来哪怕一丁点效率提升，工程师就愿意买单；但如果把 Cursor 递给一位完全没有编程背景的普通人（比如 David 自己的妻子），她不仅无法凭空搭出一个理想网站，反而会大失所望，因为用户必须具备深厚的专业背景才能驾驭当前的工具。 - **商业部署的分水岭**：模型表现若仅停留在 60% 至 70% 的准确率，在工业界根本无法投产；唯有将其推进至 80% 到 90%，AI 才能真正跨过商业化落地的生死线。 - **异构硬件与离线运行**：通过重构底层算子栈，webAI 实现了将开源或定制的小模型直接分发到端侧，即使在网络完全瘫痪的无连接场景下依然可以满血运行。 --- ### [13:25](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=805s) 走出 Reddit 象牙塔：以确定性结果取代提示词工程 要实现这关键的 10% 性能飞跃，脱离实际业务指标的空谈毫无意义，必须紧扣企业最核心的“交付成果”（Outcome）。例如在 F1 赛车项目中，部署在赛车本地的 AI 模型需要根据实时转弯姿态动态调节悬挂系统的阻尼，其终极度量标准极其纯粹——单圈圈速究竟有没有缩短、AI 的动作响应是否精准无误。这种严苛工业场景下的核心工程数据在 Reddit 或公开互联网上根本无迹可寻，唯一的突破口在于深入实体企业内部，挖掘那些真正沉睡在业务一线的核心数据资产。 David 直言，目前行业内盛行的“提示词工程”（Prompt Engineering）只是一套适合做 Demo 原型的权宜之计，在严肃的商业化部署中漏洞百出。每一次用户向大模型提问，系统都要在后台拼接一长串隐蔽的系统级指令，竭力告诫它“记住你的角色、遵守你的指令”，这种做法本质上是在“强行把一个庞大的模型说服成一个小模型”。相比之下，针对特定任务进行微调与蒸馏的垂直模型原生具备明确的职责认知，无需在每次上下文切换时被反复提醒。 > “我们现在的做法就像在拼命规劝大模型：‘嘿，我知道你上知天文下知地理，但我求求你清醒一点，我现在根本不需要你写一首关于青蛙的诗，我只需要你帮我修好这台引擎、做好传送带上的质检、或者开好这辆赛车。’” - **提示词工程的弊端**：推理延迟高、上下文冗长、每次交互均需隐式重复系统前缀，推理成本居高不下且偶发幻觉。 - **端到端微调的优势**：模型天生锁定特定目标，无需每次从头唤醒角色设定，推理延迟极低、运行成本低廉且输出高度稳定。 --- ### [15:38](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=938s) “死去的文明”与模型的自发异步网络 当 David 在新加坡演讲展示“文明模型”幻灯片时，曾有观众当场尖锐发问：“你为什么要在胶片上放一张已经覆灭的古代文明遗迹照片？”这句质问恰恰切中了该架构的核心哲学：在人类历史上，并非所有文明都能存续至今；同样，AI 演进路径如果押注单一庞然大物，其脆弱性不言而喻。webAI 倡导的分布式文明体系，旨在将模型做得更轻小、更敏捷，把模型的自主权与算力所有权彻底下放给各行各业，构建由千百个各司其职的小模型共同组成的去中心化网络。 为了解决多模型之间的协同难题，webAI 招募了前 Netflix 无服务器架构（Serverless）核心负责人——此人正是当年在酒店电视上一键扫码无缝登录系统的发明者。webAI 借鉴了这种无状态的设备与服务自发现机制，让散落在手机、车载电脑或车间机床上的每一个模型都具备“自我发现、硬件识别、互相感知”的能力。模型之间的通讯绝非简单僵硬的 `for` 循环或固化串行的 Agent 工作流，而是像人类社会一样，仅在特定场景与结果触发时进行异步的事件驱动交互。 > “大自然与科学界的一切法则都在告诉我们：超凡的智慧从来只诞生于群体协作。你自己的身体都不是单一结构运作的——我自己这辈子单打独斗从没做出过任何称得上‘超级智能’的事，最接近的可能就是生了两个可爱的孩子，但就算是生孩子，你也绝不可能一个人完成。” --- ### [18:21](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=1101s) 负载均衡智能：通往超级智能的众创之路 云计算行业过去数十年讨论的焦点始终围绕着“算力负载均衡”（Load Balance Compute），而 webAI 正在重新定义这一维度，提出“智能负载均衡”（Load Balance Intelligence）。想象一下，一家大型跨国企业内部不再依赖统一调用云端某个单点大模型 API，而是在全球业务边缘部署了 3000 个各具专长、独立运行的定制小模型。一旦某条专线中断或某个节点宕机，其余模型依然能够在本地独立思考、异步通信并继续分发工作流，从而在根本上杜绝单点故障，带来前所未有的工程韧性。 学术研究与实践表明，这种由海量模型构成的“思想多元性”（Diversity of Thought）能够大幅改善复杂任务的决策表现，正如一场高规格会议绝不该只有一个人唱独角戏，多元视角的博弈才是催生高质量决策的土壤。从宏观演进来看，单体大模型试图毕其功于一役地通往 AGI，在统计学与工程资源上充满了巨大的不确定性；而网状协同的专家模型群，不仅符合人类社会文明演化的规律，更让每个人都能以模型所有者的身份“参与构建超级智能”，而非沦为云端超级智能的被动消费者。正如 David 在演讲尾声所总结的那样：当技术真正融入无形、用户甚至感觉不到 AI 存在的时候，才证明系统正在以最高效的状态运转。 ### [20:10](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=1210s) 从“提示词提取”转向“主动推送”：AI 作为文明级操作系统 当前人们与人工智能之间的交互普遍属于高触点、高介入（high-touch）的模式。David Stout 总结指出，随着向分布式专家模型网络的迁移，人工智能将不再局限于传统的“提示词—提取（prompt-extract）”模式，而是真正演变为支撑人类社会运转的“文明操作系统（civilization's operating system）”。 在未来的交互架构中，AI 将转型为一种主动的“推送模式（push model）”。它能够主动预测需求、提供引导并自主协同，而不再需要人类用户反复去“盘问与审讯（question and interrogate）”才能获取期望的结果。在演讲的最后，David 热情邀请参会者前往 webAI 的展台进行深入探讨，并对大家的聆听与到场致以真诚的感谢。 ```mermaid flowchart LR accTitle: Comparison of AI Interaction Models subgraph OldModel["传统交互模式 (Prompt-Extract)"] direction TB A["用户反复质询 / Interrogate"] --> B["被动响应提取"] end subgraph NewModel["未来网络范式 (Civilization OS)"] direction TB C["分布式专家网络"] --> D["主动推送与引导 / Push Model"] end ``` - **交互范式迭代**：彻底改变需要人类频繁介入提问的被动局面，转向自主协调与前置引导的主动服务架构。 - **文明底层设施**：由专业化的小模型网络取代单一庞大模型，构建真正无缝嵌入设备与场景的智能底座。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=4s) 真实的未来必然是由成千上万、乃至数百万个代表着不同专业形态的小型主权模型相互协作组成的网络。 — David Stout > [03:15](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=195s) 当下的 AI 行业正陷入一种只顾着数路上的汽车，却忘记了如何建设城市的盲目状态。 — David Stout > [05:31](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=331s) 提示词不会替你买单（Prompts don't pay the bills）。 — David Stout > [08:01](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=481s) 基础模型不是你的最终产品，将基础模型本身当成终极护城河，很可能重蹈网景当年的覆辙。 — David Stout > [10:09](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=609s) 单体大模型本身正在快速泛指化与平庸化，真正的商业价值其实存在于模型的上方与下方。 — David Stout > [11:41](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=701s) 模型表现若仅停留在 60% 至 70% 的准确率，在工业界根本无法投产；唯有将其推进至 80% 到 90%，AI 才能真正跨过商业化落地的生死线。 — David Stout > [13:25](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=805s) 我们现在的做法就像在拼命规劝大模型：‘嘿，我知道你上知天文下知地理，但我求求你清醒一点，我现在根本不需要你写一首关于青蛙的诗，我只需要你帮我修好这台引擎、做好传送带上的质检、或者开好这辆赛车。’ — David Stout > [15:38](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=938s) 大自然与科学界的一切法则都在告诉我们：超凡的智慧从来只诞生于群体协作。你自己的身体都不是单一结构运作的——我自己这辈子单打独斗从没做出过任何称得上‘超级智能’的事，最接近的可能就是生了两个可爱的孩子，但就算是生孩子，你也绝不可能一个人完成。 — David Stout > [18:21](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=1101s) 当技术真正融入无形、用户甚至感觉不到 AI 存在的时候，才证明系统正在以最高效的状态运转。 — David Stout > [20:10](https://www.youtube.com/watch?v=cK3EvKdHEGo&t=1210s) 人工智能将不再局限于传统的‘提示词—提取’模式，而是真正演变为支撑人类社会运转的‘文明操作系统’。 — David Stout ## Mentioned David Stout, webAI, Marc Andreessen, Netscape (网景), Amazon (亚马逊), Chase (大通银行), Pratt & Whitney (普惠公司), Cursor, Netflix, Delta Air Lines (达美航空), Uber, TSA (美国运输安全管理局), Reddit, F1 (一级方程式赛车) ## Video description David Stout, Founder & CEO, webAI Abstracts Imagine AI not as a single monolithic model, but as a civilization of experts: small, sovereign models, each great at one thing, coordinating across devices, plants, clinics, and trading floors. That’s the future we’re building. This session will reframe how we design, deploy, and measure AI: from usage vanity metrics to outcomes per watt, risk reduced per query, and lives improved per inference. David will share how organizations are standing up distributed, on-device intelligence that protects data, crushes latency, and drives measurable ROI, without sacrificing interoperability or control. Attendees will get a concrete blueprint for assembling expert swarms: how to place them, connect them, govern them, and prove their value. [Source](https://www.youtube.com/watch?v=cK3EvKdHEGo)

Source quotes

[0:04] David Stout: 真实的未来必然是由成千上万、乃至数百万个代表着不同专业形态的小型主权模型相互协作组成的网络。 [3:15] David Stout: 当下的 AI 行业正陷入一种只顾着数路上的汽车，却忘记了如何建设城市的盲目状态。 [5:31] David Stout: 提示词不会替你买单（Prompts don't pay the bills）。 [8:01] David Stout: 基础模型不是你的最终产品，将基础模型本身当成终极护城河，很可能重蹈网景当年的覆辙。 [10:09] David Stout: 单体大模型本身正在快速泛指化与平庸化，真正的商业价值其实存在于模型的上方与下方。 [11:41] David Stout: 模型表现若仅停留在 60% 至 70% 的准确率，在工业界根本无法投产；唯有将其推进至 80% 到 90%，AI 才能真正跨过商业化落地的生死线。 [13:25] David Stout: 我们现在的做法就像在拼命规劝大模型：‘嘿，我知道你上知天文下知地理，但我求求你清醒一点，我现在根本不需要你写一首关于青蛙的诗，我只需要你帮我修好这台引擎、做好传送带上的质检、或者开好这辆赛车。’ [15:38] David Stout: 大自然与科学界的一切法则都在告诉我们：超凡的智慧从来只诞生于群体协作。你自己的身体都不是单一结构运作的——我自己这辈子单打独斗从没做出过任何称得上‘超级智能’的事，最接近的可能就是生了两个可爱的孩子，但就算是生孩子，你也绝不可能一个人完成。 [18:21] David Stout: 当技术真正融入无形、用户甚至感觉不到 AI 存在的时候，才证明系统正在以最高效的状态运转。 [20:10] David Stout: 人工智能将不再局限于传统的‘提示词—提取’模式，而是真正演变为支撑人类社会运转的‘文明操作系统’。

Source quotes

[0:04] David Stout: True super-intelligence mirrors modern transportation—a network of specialized human and technological experts coordinating seamlessly. [3:15] David Stout: Prompts don't pay the bills. [7:59] David Stout: A base model is not your product. [10:09] David Stout: one model to rule them all [15:38] David Stout: Everything in nature, everything in science tells us that the collective is how we drive super intelligent outcomes. Even your body isn't singular. [18:20] David Stout: load balancing intelligence [20:10] David Stout: civilization's operating system [20:10] David Stout: prompt extract model

Original transcript

[0:04] I think this is going to be a really [0:05] interesting conversation. It is to me. [0:08] Um, we've gotten to share this around [0:11] the world uh at a few events and we [0:14] found that the feedback um creates a lot [0:16] of great conversation and hopefully we [0:18] walk away [0:20] identifying some new problems but also [0:22] uh seeing clear pathways to address them [0:25] and create maybe AI that works for more [0:28] people, AI that works for more [0:30] industries and AI that solves problems [0:32] that are more acute to everyday [0:34] industry. So to frame this, [0:37] AI as a civilization is something that [0:39] we use a lot at WebAI, which is this [0:40] idea that there's not just going to be [0:43] this scenario where there's one model [0:45] that's prevalent. Um, rather that [0:47] there's going to be [0:49] hundreds of thousands of millions of [0:51] models that represent distinct forms of [0:54] intelligence. And a good way to frame [0:55] this to start is to think about how you [0:58] all got here. Um, I'm not from Vegas. [1:01] I'm from Austin. And this is now one of [1:03] the first times in my life I'm convinced [1:05] I didn't choose to live in the hottest [1:06] place in the world. Um uh but I got on a [1:09] plane to get here and um I took an Uber [1:12] from my house and got to the airport and [1:17] uh went through TSA. I get to keep my [1:19] shoes on now, which is pretty awesome. [1:21] Uh uh made it to my gate, went to the [1:24] Delta Lounge and waited for my flight. [1:27] got on the plane, successfully made it [1:29] to Vegas, got in another car, made it to [1:31] my hotel. We we take this process for [1:35] granted, but it's fairly complex [1:37] process. We're putting people in a metal [1:38] tube with engines and we're throwing [1:40] them across the sky and then along the [1:43] way there's all this technology. There's [1:45] all of this knowhow. Um, but it's not [1:48] singular, not even in the slightest way. [1:51] Um, you have people that know how to [1:54] drive you to the airport, which also is [1:57] something we take for granted, but [1:58] people that know how to navigate and do [1:59] that safely. Um, we have people that [2:02] manage security and that know all the [2:04] protocols required to get you through [2:06] the airport one safely, um, but also so [2:09] you can be on a plane and be safe on the [2:11] aircraft. Then when you get on the [2:13] plane, there's a crew that's there to [2:15] help you be safe along that whole [2:17] journey. So whether it's the flight crew [2:19] or the pilots, they've all been trained [2:22] on an incredible amount of information [2:24] to keep you safe and ultimately complete [2:26] their mission. [2:29] And then not to mention the tech on the [2:31] aircraft, right? There's incredible [2:32] amounts of tech that's esoteric specific [2:34] to just flying. Uh and then when a plane [2:37] is traveling from somewhere to Austin to [2:38] Vegas, uh you're changing, you know, [2:41] your tower controls close to 15 times. [2:44] So, different towers switching, telling [2:46] you if your airspace is safe or not. [2:50] There's nothing about that process [2:52] that's singular. There's all of these [2:54] different types of intelligence stacked [2:56] throughout that to make one super [2:58] intelligent experience. Just one. And [3:01] today, we're going to talk about what [3:03] we're doing in the market today [3:06] with the hopes of making super [3:07] intelligent experiences and then what I [3:08] think we actually can do to have [3:11] outcomes that are super intelligent. And [3:13] I think those are two different things. [3:15] So today we see a lot of people [3:20] counting cars in my opinion instead of [3:23] building cities. And why I use this [3:26] analogy is get to travel a lot of great [3:28] cities all over the world. And if you [3:30] can imagine city planners don't design [3:33] cities based on the outcome of high [3:35] traffic. They're not like, "Hey, let's [3:36] build a city so we have a ton of traffic [3:38] and it's super frustrating to use and [3:42] it's slow and it's cost intensive and [3:45] all of these things. [3:47] Great cities are designed on outcomes [3:49] and saying what do we want this outcome [3:50] to look like? How are we going to [3:53] measure that outcome? And then how do we [3:54] ensure that happens at scale?" [3:59] I'm going to talk a little bit about the [4:01] internet era and why it's relevant to AI [4:03] specifically, but this is one of the [4:05] first points that I think is overlooked [4:08] today. If you look at the most [4:10] successful AI companies, they're [4:11] counting internet company metrics, and [4:13] that's a big false flag for me. Daily [4:16] active users, not a really important [4:19] metric for AI. It's just not. The [4:22] companies with the most users lose the [4:23] most money. So, why is that? [4:26] I mean, we can have tons of [4:27] conversations around that and there's a [4:29] lot of people that like to defend it, a [4:30] lot of people that def it. I'm just [4:32] saying it's a problem and I think both [4:33] sides will agree it's a problem. Um, and [4:36] I think that typically we're seeing that [4:37] in companies that represent more of [4:39] internet companies that are using AI [4:41] rather than AI native companies that are [4:44] delivering an AI product. [4:46] So today, for example, 78% of [4:49] organizations say they use AI, which is [4:52] a lot. That's up from 55% last year in [4:54] 2024. [4:56] But 78% [4:59] of that 78, so it's closer to 79 in that [5:01] number um have still failed to apply AI [5:04] into a use case in the wild. So tons of [5:07] people are using it, but we're not [5:09] seeing it deployed. Why? I don't think [5:10] we're focused on the outcome and we're [5:13] we're chasing these congestion stats in [5:15] my opinion that are kind of irrelevant. [5:17] I don't care how many users these [5:18] companies have. What problems have you [5:20] solved? What billion dollar companies [5:21] have you made because of your solution? [5:24] Um, and not just billion dollars based [5:25] on tech valuations, but like a billion [5:27] dollar problem. [5:31] Prompts don't pay the bills. [5:34] Uh, this is going to take us into the [5:36] internet era here. So, the fundamentals [5:39] of starting a business still haven't [5:41] changed. Just because AI exists, the [5:42] fundamentals haven't changed. If you [5:43] went to the bank and you had this [5:45] business idea, let's say you want to [5:47] launch a coffee shop [5:50] and it's really simple and you're going [5:51] to you're going to you want to launch it [5:52] in your hometown and you know the type [5:54] of machines you want and the the [5:56] audience you're going to have and you're [5:58] you think that you can build a real [6:00] business and if you went to the bank and [6:02] you told Chase, "Hey, I'm I'm going to [6:04] serve coffees for $6, but it's going to [6:08] cost me $10 to make them." [6:11] How excited would Chase be to give you a [6:13] loan? And then you said, "Well, don't [6:15] worry. Eventually, I'm gonna make coffee [6:18] that's so good everyone wants it. Then [6:20] I'm going to change the price." [6:23] Um, really, I mean, this is just what [6:25] people are saying, and it's analogous, [6:27] obviously, but I don't think it actually [6:29] works. So, there's a few reasons. They [6:32] would say, "Well, prompts do pay the [6:34] bills because users matter. Because [6:36] users are giving us valuable data." [6:38] Well, that sounds awfully internet [6:40] company of you because last time I was [6:42] aware, users that go chat with AI aren't [6:45] giving AI answers. They're asking [6:47] questions. [6:49] Does AI train on questions? I mean, I'm [6:52] sure it does. Obviously, it does. We [6:53] know that. But does it just train on [6:56] questions or does it need answers to [6:57] have a robust QA data set? QA stands for [7:00] question answer, by the way. [7:02] So, we have half of the data, but we [7:05] don't have the whole picture yet. We're [7:07] saying that users matter, interactions [7:09] matter, they're massively valuable, [7:12] and we're going to build these large [7:13] closed data sets that are really [7:15] valuable to drive these really powerful [7:17] singular models or [7:21] in enterprise. We talk with a lot of [7:23] different enterprises all over the [7:24] world. What we're hearing is these [7:26] models aren't solving their problem. [7:28] So they can prompt engineer them. they [7:31] do a lot of prototyping and then when it [7:33] comes to deployment a prompt engineer [7:35] model doesn't really reflect what they [7:37] need in the outcome. So there's still [7:39] this gap from I have a really great [7:40] foundational model that's incredibly [7:42] impressive by the way and needed to be [7:44] done and that model work is fundamental [7:47] but we can't also say that that [7:48] foundational model is going to fix Pratt [7:50] and Wendy engines like it's just not and [7:53] there's different issues for these [7:55] companies that are highly acute um that [7:57] are unsolved [7:59] and this goes into I think the paradigm [8:01] piece of today's talk which is important [8:04] is AI right now a lot of the winners are [8:07] like the big AI companies are modeling [8:10] something that we saw in the early [8:11] internet era and we should tread lightly [8:13] on that and track that. [8:17] So, Netscape, if you know the company, [8:20] um I've done a little bit of study on [8:21] them. I unfortunately was not alive when [8:24] this was going on and I would read read [8:27] a bunch on it, talked with people who [8:28] were and [8:31] I found some really interesting [8:32] interviews that date back with with Mark [8:34] Andre uh talking about this and what he [8:37] would say and how they framed the [8:38] business was what we did is we built the [8:41] tool that lets people access the [8:43] internet and because we have this [8:46] browser we're going to own the internet [8:49] and what I'm hearing a lot in AI today [8:52] is well if you have the foundational [8:54] model that knows everything you win AI [8:57] you have the best solution you have the [8:58] model but what I'm seeing in enterprise [9:02] and across this the board in science is [9:04] that that's that's not true the model [9:05] appears to be a commodity um and it's [9:08] happening at a a pretty exponential rate [9:11] um and if that's true the model is more [9:13] like the browser and at the same time [9:16] Netscape was being built there was a [9:18] small company that was selling books [9:19] online that was being built and the [9:23] browser could get you to that website, [9:24] but what was more valuable? Well, it was [9:27] the company that was building the [9:28] outcome and facilitating an outcome with [9:31] that technology. [9:33] And this core science is important. [9:35] Building a browser is hard. I'm not [9:37] denying that it's hard. It was really [9:39] important and hard. But was that where [9:41] the value actually lived or was it a [9:42] false flag? [9:44] I think it was a false flag. And I think [9:46] there's a lot of data that said it was. [9:47] Now, there are browser companies today [9:48] that are really valuable. Um, [9:51] but not a ton of money in providing a [9:54] browser. I don't know many people that [9:55] are paying to use a browser today. It's [9:58] pretty standard to be free. There's a [10:00] lot of different options. [10:03] So, [10:05] a base model is not your product. [10:08] So see a lot of companies, see a lot of [10:09] enterprises adopting AI that are trying [10:11] to wrap their system around a model as [10:14] if this model is their product and now [10:16] they need to to find a way to control [10:18] this model with with steering and all [10:20] these guidelines and prompt engineering [10:23] and and find a way to control these [10:26] these massive models that they think are [10:28] really valuable. But a lot of times what [10:29] we're seeing is the model's an elephant [10:30] gun. this model that can do essentially [10:34] all these general tasks isn't [10:35] necessarily what you want to use. Uh if [10:38] you're working in an assembly line and [10:41] you're doing QA, [10:43] there's probably a much better solution [10:45] for that that's that's focused and [10:46] narrow. This points back to that [10:48] civilization model is does every is [10:50] there going to be one model to rule them [10:52] all that we all API against or run [10:55] against or is it more likely that the [10:59] product is going to drive the model [11:00] that's being used [11:03] and that's what we're seeing and that's [11:04] what we're seeing success with. So the [11:06] value lives in [11:08] above and below the model in my opinion. [11:11] So if you can if you're a company that [11:13] can make technology that makes these [11:14] models work really effectively and run [11:16] anywhere, I'd say there's value for you. [11:19] If you're above that model and you're [11:20] you're focusing all your energy on how [11:23] do I make the best system to train [11:25] models and how do I make it so I can [11:27] offer the best medical model in the [11:29] world [11:31] that solves this problem. I think [11:33] there's a ton of value there. um not [11:35] seeing a ton of value in this one model, [11:38] one-sizefits-all scenario. In in regards [11:41] to that medical example, we're seeing [11:43] this like 10% gain. Uh we work in lots [11:46] of different markets. To give you an [11:47] idea of what Web AI does, just at a high [11:49] level, we're a tooling like [11:51] infrastructure company. So we we created [11:52] the core AI library for running models [11:56] on device. And we're not just shrinking [11:58] and quantizing models. We wrote our own [12:00] primitive stack directly to the BNNs and [12:02] shaders. And we change the way these [12:04] models work. We're getting away from [12:05] transformer architectures and working on [12:07] non-transformer non-auto reggressive [12:09] models. We're working on uh core [12:11] executions to distribute these models [12:13] across heterogeneous hardware because we [12:15] want to bring models into spaces where [12:17] connectivity is gone. And that's really [12:19] our business model. And then in that we [12:21] have some things that stack above those [12:22] models. But we're not really a model [12:23] company. We're like a tooling company. [12:25] We slap things into models like that are [12:26] open source or we spin up custom models [12:29] for users or customers based on certain [12:31] size. And what we're seeing is this 10% [12:34] gain at a minimum drastically changes [12:36] the ability to deploy this model into [12:38] the wild especially with enterprises [12:40] because we see a lot of success in like [12:42] coding tools and it's frankly as an [12:44] engineer it's because we we're low [12:47] hanging fruit we'll take anything that's [12:48] a little bit better but if I handed my [12:50] wife cursor who's not a developer she'd [12:53] be very disappointed just because it's [12:54] not going to build her website for her. [12:56] It's she's gonna say what she wants and [12:58] it's not going to work because you need [12:59] to be a developer to know how to use it. [13:01] Um, and that's what I mean by that 10% [13:03] like if we can find a way to get from 60 [13:06] 70% accuracy to like an 80 90 [13:11] it changes the ability for these to be [13:12] deployed and how do we do that? Um, we [13:14] need to start measuring what actually [13:16] moves the needle for these enterprises [13:17] and for these use cases. So the metrics [13:20] matter and the users matter but what [13:23] what matters even more is what's the [13:24] outcome you're trying to create. [13:26] So, if you're if you're working on a [13:29] Formula 1 car and you want an AI model [13:32] that's running on the car to adjust the [13:35] dampening and the suspension on turns so [13:37] you can get a better lap time, your [13:39] outcomes really measured. And we know [13:41] this because we're working with some of [13:42] those groups. And if you want to see an [13:46] outcome, it's very measurable. Did the [13:48] AI do the right thing? Did we get a [13:49] better lap time? Um, and the only way [13:52] you get that and even start to attack [13:55] this problem, you can't train a model on [13:56] Reddit to figure this out. The data is [13:59] not there. You have to go to where the [14:01] data actually lives, which is with these [14:02] enterprises. [14:05] And what we're seeing is these [14:07] enterprises hold the crown jewel to [14:09] these 10% gains. So there's tons of [14:12] great open source models, but partnering [14:15] deep with these enterprises, you you you [14:17] get that 10% plus gain, which means the [14:20] difference from having to prompt [14:22] engineer a system to just infer it. What [14:24] I mean is prompt engineering is very [14:26] effective for prototypes. It's not great [14:28] for commercial deployment because what [14:30] you're doing is you're prepending a [14:31] prompt. You're trying to convince the [14:33] every time you send a question, there's [14:34] this invisible query before your prompt [14:37] that's saying, "Hey, remember this is [14:39] your goal. This is your directive." [14:41] whatever I told system prompt told you [14:42] to do. Um, but when you fine-tune and [14:44] adjust these models, that never happens. [14:46] The model always knows its task, always [14:48] knows its job. It doesn't need to be [14:50] reminded every question it circumstance [14:53] or chat history. Um, so inference is [14:56] more effective. It's lower cost to run, [14:58] faster inference times, more reliable. [15:01] You're not trying to convince the model [15:02] that it's a narrow model. And what we're [15:04] doing today is we're trying to convince [15:05] really big models that they're small. [15:08] like, "Hey, hey, I know you trained on [15:10] all this stuff, but let me remind you [15:13] what actually matters to me. I don't [15:14] want you to even I don't want you to [15:16] write a poem about frogs. [15:18] I want you to fix this engine. I want [15:21] you to do the Q&A on my conveyor belt. I [15:24] want you to drive this car. [15:27] I don't want you to write a poem. I [15:29] don't want you to understand finances. I [15:30] don't want you to pretend to be a doctor [15:32] right now." And that's what prompt [15:34] engineering is doing. There's a better [15:35] way to build these systems, though. [15:38] And I think that is the civil [15:39] civilization model. So [15:43] I was in Singapore. I was talking about [15:44] this and someone right away said, "Why [15:46] are you showing a picture of a dead [15:47] civilization?" [15:49] And it actually is really important [15:52] because not all civilizations are [15:54] successful. [15:55] Um, we're all in this room, so [15:58] historically our civilizations have been [16:00] successful to an extent, [16:03] but the assumption that every [16:04] civilization lives on is a is a big [16:07] mistake. [16:08] And I think we have a really big [16:10] opportunity as well as a big [16:12] responsibility to build AI in a way [16:14] that's reflective to everyone. And when [16:17] we distribute these models downstream, [16:19] so when you make smaller models that are [16:21] more capable and more nimble, what [16:23] happens is you can distribute it to [16:25] ownership and you can distribute share [16:27] of intelligence. [16:30] So if we have a model now that is going [16:32] to be really good at driving a Formula 1 [16:34] car, helping the helping the driver, now [16:35] we have a model that's really good at [16:36] Q&A. We have a model that's really good [16:38] at aircraft engines. [16:40] We have a bunch of different models. But [16:41] how do those models talk? Oh, that's the [16:44] that's the web in web AI. Looks like our [16:47] slides died. Um, [16:51] civilization communicates and it's not [16:53] agentics. It's not for loops. It's not [16:55] saying when you get to this point, make [16:57] sure you tell this model what you did. [17:01] Civilization communicates async when [17:04] required based on scenarios, based on [17:06] outcomes. And not all of those can be [17:08] programmed. It can't be rigid. So webai, [17:12] the web in web AI as a local inference [17:14] company was how do we create a network [17:16] of these for these models to talk and we [17:19] recruited the head of serverless systems [17:20] from Netflix. If you've ever, you've [17:22] probably done it since you're at a [17:23] hotel. If you've ever like hit a button [17:24] and you automatically sign into the TV, [17:26] he invented that. And we thought, what [17:29] happens if we do that and every model [17:33] could discover itself, discover others, [17:36] discover hardware, [17:38] and act as one, but all but while [17:42] knowing these are massively discreet [17:44] different models. Well, you start to [17:47] have super intelligent outcomes. [17:50] And you're running models on things like [17:51] phones, but you're driving use cases [17:53] that control things like planes. [17:57] And that's how we do things. I've never [18:01] singular ever done something super [18:03] intelligent ever. The closest thing is [18:06] having kids, but you can't even do that [18:08] by yourself. Like my I have two kids. [18:12] They're amazing. [18:14] Everything in nature, everything in [18:16] science tells us that the collective is [18:19] how we drive super intelligent outcomes. [18:20] Even your body isn't singular. [18:32] models give us resilience. We talk a lot [18:35] about load balancing compute. Something [18:37] Web AI is trying to coin and push is how [18:39] do we load balance intelligence? [18:42] So instead of having an API that you're [18:44] hitting, what if we have 3,000 different [18:46] unique models across your organization [18:48] and now you've load balanced [18:50] intelligence. So if something goes down, [18:52] those models are still talking, working, [18:54] distributing tasks. [18:57] Also, this diversity of thought and [18:58] modeling, we've done some research. [19:01] there's really compelling results on the [19:03] diversity of thought in the outcome. [19:05] It's similar to like you don't want just [19:07] one person in a room in a meeting. You [19:09] want different minds, different groups, [19:13] different representations of thought to [19:14] get to powerful outcomes. [19:17] And what I think is exciting about this [19:19] is it's one resilient for [19:22] deploying AI, it's resilient for [19:24] humanity because it lets more people be [19:26] involved in that story. Uh but it also [19:30] is the most clear rich path towards [19:32] super intelligence that I've personally [19:35] like heard of or witnessed or have [19:38] engaged with. There's statistical [19:40] anomalies that would be required for [19:41] this monolithic path to work or the path [19:44] to work. Um this is the most ripe path [19:48] for us to approach. It's it's the most [19:49] ripe path for civilization as well [19:51] because it's one that we all get to [19:52] participate in. Um, your model [19:55] contributes to super intelligence uh [19:57] versus you are a consumer of super [20:00] intelligence. Um, [20:03] and lastly, I know we're at time is I [20:06] think when you don't see AI, you know [20:08] it's working. And I think there's a lot [20:10] of hightouch interactions with AI today. [20:13] And when we move towards this model, [20:15] we'll start to see AI being [20:17] civilization's operating system rather [20:19] than this prompt extract model. we'll [20:23] start to see AI be a push model, pushing [20:26] and guiding rather than something we [20:28] have to question and interrogate to get [20:30] results we want. Um, we have a booth. [20:32] We'd love to talk to you all. Um, thank [20:34] you so much for your time. Thanks for [20:35] coming and saying hi. Um, really [20:37] appreciate it. Take care.