Oumi Automates In-House AI Builds to Slash Enterprise Inference Costs

The Next Era of Enterprise AI: The Shift Underway With Oumi — Ai4

Ai4

Renting commercial frontier models leaks proprietary domain intelligence to competitors and inflates operating costs, forcing enterprises to build and own specialized AI to protect margins and competitive moats.

AI-assisted summary of The Next Era of Enterprise AI: The Shift Underway With Oumi — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 14, 2026, this presentation features Manos Koukoumidis, CEO and Co-Founder of Oumi. The description outlines the enterprise shift from renting generic, cost-prohibitive frontier intelligence toward building and owning specialized, automated domain-specific AI.

## TL;DR The talk argues that enterprises face a critical strategic imperative to move from renting third-party frontier model APIs to building, specializing, and fully owning their core AI capabilities. Relying on generic APIs leads to homogenized products, severe margin compression, and the 'Reverse Information Paradox,' where proprietary business insights leak back to foundation model providers. Through its automated, Apache 2.0-based 'AI factory,' Oumi demonstrates how organizations can turn prompts into specialized, highly accurate models in hours—cutting inference costs by 5x to 100x and establishing self-improving, defensible intelligence loops.

## Atmosphere & style The presentation carries an authoritative, pragmatic executive tone that blends high-level boardroom strategy with hands-on technical validation. It delivers a sobering wake-up call about the economic and intellectual property traps of commercial API dependency before pivoting into an energetic, empowering engineering demonstration. The overall experience feels like an urgent playbook for enterprise leaders, transforming daunting AI infrastructure challenges into accessible, rapid operational wins.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=4s) Enterprise AI is undergoing a fundamental strategic shift from renting generic frontier model APIs to building, fine-tuning, and owning specialized open models. - [00:04](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=4s) Organizations must internally build and own the AI that drives their core competitive differentiators, while aggressively optimizing unit economics for high-volume tasks. - [02:15](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=135s) Relying strictly on off-the-shelf frontier APIs homogenizes product experiences, whereas training specialized models on proprietary domain knowledge creates a compounding defensive moat. - [02:15](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=135s) AI leaders like Cursor (10x lower cost) and Intercom Fin (5x cheaper) demonstrated that purpose-built specialized models can match or outperform general frontier models at a fraction of the operating cost. - [05:10](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=310s) The 'Reverse Information Paradox' means enterprises paying frontier API providers inadvertently supply the operational edge cases and domain insights used to train foundation models that benefit direct competitors. - [05:10](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=310s) Frontier models impose massive compute overhead for narrow business tasks; specialized models cut inference costs by 5x to 100x, turning variable token billing into low, fixed training expenses. - [07:42](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=462s) Technological sovereignty safeguards enterprises against platform lock-in, forced model deprecations, geopolitical cutoff risks, and regulatory compliance issues. - [09:46](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=586s) Automated 'AI factories' collapse development cycles from months to hours by automating failure diagnostics, targeted synthetic data generation, fine-tuning, and deployment. - [09:46](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=586s) Open-source foundations (such as Apache 2.0) ensure enterprises maintain complete transparency and permanent ownership over training recipes, synthetic data pipelines, and model weights. - [11:46](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=706s) Intelligent task routers can evaluate query complexity and direct simpler workloads to models 100x smaller, delivering identical outputs at one-tenth of the inference cost. - [14:40](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=880s) A continuous self-improvement loop clusters production prediction errors and synthesizes targeted training sets, enabling models to adapt continuously to real-world data drift. - [16:23](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=983s) Enterprise deployments validate the shift to owned AI, from global banks dropping serving costs to zero via on-device SLMs to CX firms slashing inference costs 33x while reducing false positives. - [16:23](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=983s) Long-term competitive advantage belongs to enterprises that build and own their operational intelligence as permanent, compounding corporate IP rather than renting commodity capability.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=4s) The Strategic Shift from Renting to Owning Enterprise AI

Enterprise AI is reaching a critical inflection point. Organizations are actively transitioning away from merely renting generic intelligence via black-box APIs—such as OpenAI's GPT models, Anthropic's Claude, or Google's Gemini—toward building, specializing, fine-tuning, and fully owning open models tailored directly to their domains and workflows. Enterprises that fail to establish their own AI loop risk ending up "on the wrong side of history."

```mermaid flowchart TD accTitle: Enterprise AI Strategic Shift subgraph S1["Phase 1: Renting Generic AI"] A["External Frontier APIs<br/>(GPT / Claude / Gemini)"] --> B["Black-Box Consumption"] B --> C["High API Costs & Margin Squeeze"] end

subgraph S2["Phase 2: Owning Specialized AI"] D["Domain Expertise & Internal Data"] --> E["Open Models + Specialization / Fine-Tuning"] E --> F["Proprietary AI Flywheel & 5-10x Cost Reduction"] end

S1 -.->|"Shift Underway"| S2 ```

Across conversations with dozens of C-level executives and senior leaders at top organizations, a central question continually arises: where is long-term business value going to accrue? If "AI is eating software," leaders want to know where to place their strategic focus and capital. The answer is not simply about building the *most* AI, but specifically building the AI that is vital to their core competitive advantage.

Two foundational rules govern this AI strategy: - **Build and own your core differentiators:** If an AI workload drives the core value proposition of the company, the enterprise must build, own, and compound that intelligence in-house. For example, while a bank can reasonably outsource general coding assistance to commercial tools like Anthropic's Claude Code (since general software engineering does not differentiate one bank from another), core capabilities such as mortgage processing or credit risk analysis represent the business's true edge and must be built and owned internally. - **Optimize unit economics for high-volume tasks:** If an AI workload materially affects operating economics or incurs high query volume, organizations should adopt or build specialized, cost-effective models rather than overpaying for bloated frontier models.

---

[02:15](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=135s) The Catalysts: Cursor, Intercom, and the Differentiation Dilemma

While 2023 through 2025 was dominated by enterprises renting generic intelligence via third-party APIs, a rapid acceleration toward self-owned intelligence has unfolded. Companies that rely heavily on AI to deliver value discovered they could no longer accept the cost and dependency risks of frontier providers. In recent months, high-profile AI-first software companies have demonstrated this shift: - **Cursor** developed its own specialized coding agent to avoid sole dependency on Claude and GPT, achieving equivalent coding quality at **10x lower cost**. - **Intercom** launched **Fin**, a customer support agent outperforming general-purpose frontier models on customer experience tasks while operating **5x cheaper**.

```mermaid flowchart LR accTitle: The Enterprise AI Flywheel subgraph Loop["Continuous Intelligence Flywheel"] A["Institutional Knowledge & Domain Data"] --> B["Specialized Custom Model"] B --> C["Production Deployment"] C --> D["User Interactions & Edge Cases"] D --> A end ```

The primary driver behind this transition is **differentiation**. When an enterprise relies on prompting the same commercial frontier models as its competitors, its product experience homogenizes. As AI takes over larger portions of product functionality, using off-the-shelf APIs causes unique differentiation to decline over time. Conversely, training and refining specialized models using internal institutional knowledge creates a proprietary data flywheel where every customer interaction continuously sharpens the enterprise's defensive moat.

- **Kirkland & Ellis Example:** As the world's highest-grossing law firm, they recognized that general-purpose generative AI merely "raises the floor for everybody." Because premier legal clients hire them for elite execution rather than baseline competency, they opted against relying purely on packaged legal AI tools running standard GPT or Claude instances, choosing instead to develop proprietary specialized intelligence alongside their top legal talent. - **Outsourcing Learning vs. Institutional Compounding:** Enterprises cannot afford to outsource their organizational learning. Competitive moats come from the collective, continuously improving intelligence shared between human experts and internal AI systems.

---

[05:10](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=310s) The Reverse Information Paradox and Economic Traps of API Renting

Renting frontier APIs introduces a hidden, long-term threat: enterprise intellectual property leakage. Even when frontier providers promise not to train directly on customer data, model providers monitor failures, track query patterns, and generate synthetic data targeting the exact edge cases their customers struggle with. As next-generation frontier models incorporate these learnings, proprietary enterprise domain insights are effectively commoditized and redistributed to competitors.

- **The Reverse Information Paradox:** Enterprises pay steep premium pricing to frontier model providers while simultaneously supplying the operational domain signal those providers need to improve their own foundation models—effectively "paying twice" to give away their intellectual property. - **Alex Karp / Palantir Perspective:** Enterprise customers are increasingly frustrated that frontier AI labs capture their proprietary business "alpha" and IP through API usage patterns. - **Contract Ambiguities:** Unless commercial vendor agreements explicitly prohibit using aggregated operational metadata and behavioral metrics to improve provider systems, domain insights inevitably bleed back into general-purpose models.

```mermaid flowchart TD accTitle: The Reverse Information Paradox A["Enterprise Workflow / Edge Cases"] -->|"API Queries & Failure Modes"| B["Frontier AI Lab"] B -->|"Synthetic Data / Model Iterations"| C["Next-Gen Frontier Model"] C -->|"Redistributed to"| D["Direct Competitors"] A -.->|"Enterprise pays premium fees"| B ```

Renting generic frontier models also creates severe economic headwinds. Frontier models are engineered to perform thousands of general tasks, forcing businesses that require only a single specific task to shoulder massive, unnecessary compute overhead. Enterprises caught in this dynamic face a losing dilemma: raise pricing and get undercut by cost-efficient competitors who own lightweight specialized models, or absorb the margins and operate at a structural loss. In contrast, purpose-built specialized models cut inference costs by 5x to 10x (or up to 100x), allowing organizations to expand their operating margins over time.

---

[07:42](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=462s) Technological Sovereignty and the AI Factory Solution

Beyond differentiation and economics, the third critical pillar driving enterprises toward owned models is **control**: - **Geopolitical and Platform Risks:** Reliance on third-party APIs leaves businesses vulnerable to sudden access cutoffs driven by geopolitical events (as seen with platforms like Fable) or strategic conflict (e.g., a frontier lab launching competing offerings in specialized sectors like drug discovery or financial services). - **Forced Deprecations & Migration Overhead:** Frontier providers regularly deprecate model checkpoints, forcing enterprise customers into costly, unpredictable migration cycles for newer models that may be twice as expensive without providing measurable performance gains for specific enterprise tasks. - **Privacy, Security, and Compliance:** In regulated industries, maintaining strict data lineage, data residency, and auditability requires absolute ownership over where models are deployed and how inferences are executed.

```mermaid flowchart LR accTitle: AI Factory Architecture subgraph Inputs["Factory Inputs"] P["Problem Definition"] D["Proprietary Data"] E["Domain Expertise"] end

subgraph Factory["AI Factory Engine"] T["Automated Training & Fine-Tuning"] V["Model Evaluation & Specialization"] end

subgraph Output["Output Asset"] M["Owned Specialized AI Model"] F["Compounding Production Loop"] end

Inputs --> Factory Factory --> Output ```

The reason many organizations have delayed this transition is operational friction: traditionally, building bespoke AI required scarce machine learning research talent and months of custom engineering per use case. To resolve this bottleneck, enterprises are adopting the concept of an **AI factory**—an automated infrastructure platform that takes raw business problem definitions, proprietary data, and domain expertise, and operationalizes the evaluation, training, and deployment of specialized models. This turns specialized AI development into a scalable, repeatable flywheel that compounds intelligence with every production interaction.

[09:46](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=586s) The Automated AI Factory and Enterprise Governance

To build a defensible competitive moat, enterprise AI systems must be easily governable, auditable, transparent, and flexible—especially within heavily regulated industries. Because existing tooling forced companies into closed ecosystems, Oumi built a fully automated "AI factory." The platform allows a team to take a standard prompt originally written for frontier models like ChatGPT, Claude, or Gemini and instruct the system: "Build a model that does whatever you put in that prompt." Oumi automates the end-to-end lifecycle, running side-by-side evaluations of open and closed models, diagnosing exact failure modes through deep qualitative analysis rather than surface-level metrics, and automatically generating synthetic training data tailored specifically to those identified failure points. It then selects appropriate training algorithms and hyperparameters, executes the fine-tuning runs, iterates until the specialized model reaches peak performance, and offers one-click deployment with GPU autoscaling or export options for on-premises hosting.

```mermaid flowchart TD accTitle: Oumi Automated AI Factory Pipeline subgraph Input ["Specification"] A["Prompt / Task Spec"] end subgraph Diagnostic ["Evaluation & Diagnostics"] B["Benchmark Open & Closed Models"] C["Deep Failure Mode Analysis"] end subgraph Optimization ["Data & Training Loop"] D["Targeted Synthetic Data Generation"] E["Automated Fine-Tuning & Hyperparameters"] end subgraph Production ["Deployment & Ownership"] F["1-Click GPU Autoscaling / On-Prem Export"] G["Exportable Apache 2.0 Recipes & Weights"] end

A --> B B --> C C --> D D --> E E --> C E --> F E --> G ```

This workflow avoids operating as an opaque black box. Retaining full control over AI assets is an essential strategic requirement—a point echoed by leaders like Satya Nadella. Every recipe, evaluation standard, synthetic data pipeline, and training configuration remains completely transparent, inspectable, and editable. The infrastructure runs on an Apache 2.0 open-source library, ensuring enterprises retain complete ownership of their data, model weights, and pipeline logic without facing vendor lock-in.

- **Automated Lifecycle**: From prompt input to evaluation, failure diagnosis, synthetic data synthesis, training, and deployment. - **Full Transparency**: All data generation pipelines and training recipes are inspectable and customizable. - **Open-Source Core**: Built on Apache 2.0-licensed software to eliminate vendor lock-in and ensure full asset ownership.

---

[11:46](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=706s) Demo: Building a Cost-Effective Model Router via Agentic AI Engineering

A demonstration presented by Stefan highlights the severe cost disparity between default frontier model usage and intelligent model routing. Sending every raw coding query directly to a frontier model like Claude burns metered tokens and severs the enterprise feedback loop. By contrast, an intelligent router can analyze an incoming query and route simpler tasks to a model 100 times smaller, delivering the identical output at one-tenth of the inference cost on hardware the organization already owns. Oumi automates the engineering heavy lifting—evaluations, test sets, data curation, failure analysis, and fine-tuning iterations—acting like "Claude Code, but for AI engineering."

The interactive session showcases building a specialized coding router from a single natural language prompt without pre-existing datasets. The prompt instructs the router to evaluate incoming coding queries by complexity, category, and requirements, routing basic tasks to inexpensive open models while reserving large models like Claude Sonnet and Claude Opus for the most complex logic. The Oumi agent engages in a clarifying dialogue to refine the specifications, agreeing on a structured JSON output with two fields: `model_chosen` and `rationale`. Once the constraints are established, the agent generates and executes a complete development plan on managed infrastructure: defining metrics, synthesizing test sets, establishing baseline benchmarks, generating synthetic training data, fine-tuning, evaluating post-training performance, and planning next iterations.

- **Base Architecture**: Fine-tuning a compact Qwen base model using synthetic training data distilled from Claude Opus. - **Model Flexibility**: Supports a broad spectrum of open-weight and proprietary closed models for data synthesis, distillation, and LLM-as-a-judge evaluations. - **Zero Initial Data**: Entire routers and task-specific models can be bootstrapped from prompt specifications alone.

---

[14:40](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=880s) Distillation Results and the Continuous Self-Improvement Loop

The resulting fine-tuned router matches the benchmark performance of Claude Opus while running on a model that is 100 times smaller and 10 times cheaper to operate. While pure distillation typically allows the student model to match the teacher's capabilities, fine-tuning with targeted external or domain-specific data regularly outperforms frontier models by 10% or more. This transitions enterprise AI economics from perpetual, variable metered API token billing to a one-time, fixed training cost of just a few dollars.

```mermaid flowchart LR accTitle: Continuous Self-Improvement Loop subgraph Production ["Production Runtime"] P1["Live Predictions"] P2["Drift & Error Detection"] end subgraph Diagnosis ["Failure Mode Analysis"] D1["Group Prediction Errors"] D2["Isolate Failure Clusters"] end subgraph Retraining ["Targeted Adaptation"] R1["Synthesize Targeted Data"] R2["Iterative Model Update"] end

P1 --> P2 P2 --> D1 D1 --> D2 D2 --> R1 R1 --> R2 R2 --> P1 ```

Oumi's "self-improvement loop" captures feedback signals before and after production rollout rather than discarding them. The platform clusters prediction errors into high-level failure modes, isolates the root causes, and automatically synthesizes a curated training set targeted squarely at those gaps. This continuous loop runs in production to adapt to real-world distribution drift and improve model reliability over time, allowing organizations to compound domain intelligence rather than giving away valuable learning signals to proprietary API providers.

- **Cost and Scale Efficiency**: Replaces open-ended token billing with predictable, low-cost training runs. - **Failure Mode Clustering**: Automatically categorizes errors and targets synthetic data generation directly at accuracy gaps. - **ROI Assessment**: Teams can quantify expected cost savings through the calculator available at `oumi.ai/calculator`.

---

[16:23](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=983s) Enterprise Results and the Imperative to Own Intelligence

Building a specialized enterprise model on Oumi requires minimal manual effort—typically just 15 minutes of an engineer's hands-on time, even if background computation takes several hours or a full day. Oumi was founded by engineers and researchers with backgrounds building Gemini at Google, Meta, Microsoft, and Apple Intelligence, alongside prominent academic advisors in AI.

Real-world enterprise deployments demonstrate substantial performance and cost advantages over standard API renting:

| Enterprise / Sector | Solution & Architecture | Accuracy & Performance Impact | Financial & Compute Impact | |:--- |:--- |:--- |:--- | | **Top-10 International Bank** | On-device SLM built in 2 hours | 66% reduction in error rates | Serving cost dropped to zero (runs locally on mobile devices with no GPU cloud fees) | | **DMG** | Custom specialized task model | Matched and exceeded baseline accuracy | 100× cheaper than prior commercial GPT usage | | **Top-10 Customer Experience (CX) Firm** | Domain-specific fine-tuned model | +3.5% absolute accuracy gain; 50% reduction in false positives | 33× lower inference cost |

The strategic winners of the next AI era will not be the companies that rent generic intelligence through third-party APIs. Competitive advantage belongs to organizations that operationalize building, specializing, and compounding proprietary intelligence into permanent corporate IP: "Own the intelligence that your business runs on. You can literally build it in hours if not in minutes and then let it compound for you into your asset forever."

## Notable quotes > [00:04](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=4s) on the wrong side of history — Manos Koukoumidis

> [00:04](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=4s) AI is eating software — Manos Koukoumidis

> [02:15](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=135s) raises the floor for everybody — Manos Koukoumidis

> [05:10](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=310s) paying twice — Manos Koukoumidis

> [09:46](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=586s) Build a model that does whatever you put in that prompt. — Manos Koukoumidis

> [11:46](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=706s) Claude Code, but for AI engineering. — Stefan

> [16:23](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=983s) Own the intelligence that your business runs on. You can literally build it in hours if not in minutes and then let it compound for you into your asset forever. — Manos Koukoumidis

> [00:04](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=4s) Build and own your core differentiators — Manos Koukoumidis

> [00:04](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=4s) Optimize unit economics for high-volume tasks — Manos Koukoumidis

## Mentioned Alex Karp, Anthropic, Apache 2.0, Apple Intelligence, ChatGPT, Claude, Claude Code, Claude Opus, Claude Sonnet, Cursor, DMG, Fable, Fin, Gemini, Google, GPT, Intercom, Kirkland & Ellis, Manos Koukoumidis, Meta, Microsoft, OpenAI, Oumi, Palantir, Qwen, Satya Nadella, Stefan

## Video description Manos Koukoumidis - CEO and Co-Founder, Oumi

Most organizations began their AI journey by renting generic intelligence through frontier models. It was the fastest way to get started, but they quickly discovered three fundamental limitations: generic models rarely deliver the highest quality on domain-specific tasks, inference costs become difficult to justify at scale, and the intelligence powering their business remains under someone else's control.

Leading enterprises are now shifting toward specialized AI they own. Until recently, building domain-specific models required months of work and scarce AI expertise. Today, AI-native platforms automate model evaluation, data generation, training, deployment, and continuous improvement, reducing development cycles from months to hours and enabling domain experts to build production-grade AI.

If building specialized AI becomes as accessible as modern software development, how should enterprise AI strategy evolve? What competitive advantages become possible when every organization can build, own, and continuously improve intelligence tailored to its business? What will separate the winners from everyone else?

[Source](https://www.youtube.com/watch?v=j0kMTC6Lkfo)

中文简报

Oumi推动企业自建AI自动化，大幅降低推理成本 租赁商业前沿模型会导致专有领域智能泄露给竞争对手并推高运营成本，迫使企业自建并掌握专用AI，以保护利润空间和竞争壁垒。 本次演讲指出，企业正面临一项至关重要的战略抉择：从租赁第三方前沿模型API，转向自建、专业化定制并完全掌控自身的核心AI能力。依赖通用API会导致产品同质化、利润率严重受挤压，并引发“逆向信息悖论”——即企业的专有业务洞察反向泄露给基础模型提供商。通过其基于Apache 2.0协议的自动化“AI工厂”，Oumi展示了机构如何在数小时内将提示词转化为高精度的专业模型，从而将推理成本降低5倍至100倍，并建立起具备防御能力且可自我完善的智能闭环。 企业 AI 正在发生根本性战略转变，逐步从租用通用前沿模型 API 转向构建、微调并完全拥有专用的开源模型，以掌握核心竞争壁垒。 (0:04) Cursor（成本降至 1/10）与 Intercom Fin（成本降低 5 倍）等标杆案例表明，针对特定场景定制的专用模型在性能上完全可以匹敌甚至超越通用大模型，且推理成本大幅下降。 (2:15) “逆向信息悖论”导致企业在支付高昂 API 费用的同时，无意中向基础模型提供商输送了业务边缘案例与领域认知，进而削弱了自身的防御壁垒。 (5:10) 拥有自主模型可规避平台锁定、强制弃用和地缘断供风险，同时借助 Oumi 等自动化“AI 工厂”，将模型评估、合成数据生成与微调部署周期从数月压缩至数小时。 (7:42) 真实企业落地案例显示，定制化专用模型使某排名前十的国际银行错误率下降 66% 且实现零云端推理成本，同时帮助企业将长期的浮动 Token 计费转变为低廉且固定的训练资产。 (16:23)

Reading summary

Enterprises are transitioning from renting generic frontier AI models via third-party APIs toward building, fine-tuning, and owning specialized models to protect core differentiation, reduce inference costs, and prevent intellectual property leakage. Core Differentiation and the Limits of Third-Party APIs (0:04) Enterprises are establishing strict criteria for AI investments: workflows that represent a company's core value proposition must be built and owned internally, whereas non-differentiating tasks can be outsourced. For instance, while a financial institution might use external tools for general software development, core workflows such as mortgage underwriting and credit risk analysis require proprietary intelligence. Similarly, Kirkland & Ellis avoided relying solely on packaged legal AI tools running standard foundation models, concluding that commercial APIs merely establish a baseline competency floor rather than delivering elite execution. The Reverse Information Paradox and Economic Headwinds (5:10) Frontier models are engineered for thousands of generalized tasks, forcing enterprises with narrow use cases to pay for unnecessary compute overhead. To escape these margin squeezes, companies like Cursor and Intercom developed proprietary specialized agents, cutting costs by 10x and 5x, respectively. Without custom models, enterprises face a dilemma: raise customer prices and risk being undercut by leaner competitors, or absorb steep API fees at a structural loss. Renting APIs also introduces the 'reverse information paradox.' Even when providers agree not to train directly on customer inputs, operational metadata, query failures, and edge cases are monitored to craft synthetic data for future foundation models. As highlighted by Palantir's Alex Karp, enterprises effectively pay premium rates while surrendering the domain intelligence that providers commoditize and redistribute to competitors. Additional vulnerabilities include forced checkpoint deprecations, unexpected migration expenses, and geopolitical platform cutoffs. Automating Model Distillation via AI Factories (9:46) To lower the engineering barrier of building bespoke architectures, organizations are turning to automated 'AI factories' that transform natural language prompts into specialized, deployable models. Frameworks such as Oumi use an Apache 2.0 open-source core to ensure companies retain full ownership of their recipes, synthetic datasets, and model weights without vendor lock-in. The platform conducts side-by-side evaluations, identifies failure modes, synthesizes targeted training data, and optimizes hyperparameters automatically. In one demonstration, an intelligent coding router was built from a single prompt to categorize query complexity. By distilling Claude Opus into a compact Qwen base model, the specialized router matched baseline performance on a footprint 100 times smaller, operating at one-tenth the cost. Production feedback loops then cluster real-world prediction errors and synthesize targeted updates, allowing models to continuously adapt to distribution drift. Measured Performance Gains in Production Deployments (16:23) Real-world production implementations show substantial operational and cost improvements when replacing generic APIs with dedicated models. A top-10 international bank developed an on-device small language model in two hours, cutting error rates by 66% and reducing serving costs to zero by executing inferences locally on mobile hardware rather than cloud GPUs. Similar gains appear across other data-heavy industries. DMG deployed a specialized model that matched baseline accuracy while running 100 times cheaper than its previous commercial GPT deployment. Additionally, a top-10 customer experience firm achieved a 33-fold reduction in inference costs, an absolute accuracy gain of 3.5%, and a 50% drop in false positives by shifting to a domain-specific fine-tuned model. The strategic argument concludes that long-term enterprise advantage relies on converting transient API token expenses into permanent, compounding corporate IP. However, achieving these gains requires organizations to adopt automated evaluation and retraining pipelines to ensure smaller, specialized models reliably handle real-world drift and edge cases.

精炼阅读摘要

企业级人工智能正在经历从租用通用黑盒API向自主构建、微调并拥有垂直专用模型的战略转折。演讲者指出，企业若无法建立内部AI闭环与专有数据飞轮，不仅会面临推理成本高企与利润被侵蚀的困境，还将承受知识产权流失及丧失核心竞争壁垒的长期风险。 拥有核心差异化：企业智能的战略分流 (0:04) 企业布局人工智能的关键在于辨明长远商业价值的沉淀路径。演讲者提出两条核心策略法则：企业必须自建并完全掌控支撑自身核心竞争力的AI工作负载，而对于通用任务（例如银行使用第三方工具辅助通用软件开发）则可合理外包；若依赖现成的商用API处理核心业务，产品体验将随着同质化而丧失防御护城河。以顶级律所柯克兰律师事务所（Kirkland & Ellis）为例，他们认为通用生成式AI仅能抬高全行业基线，无法提供顶级业务所需的高阶执行力，因此选择协同专业人才自主训练专有模型。 AI原生企业已先行验证了拥有专属模型的成效。Cursor与Intercom等公司为规避对通用大模型的单一依赖，分别构建了专用编程代理与Fin客服代理，在达到同等或更优表现的同时，将推理成本缩减了5至10倍。演讲者强调，企业绝不能外包自身的组织学习，真正的竞争壁垒源自领域专家与内部AI系统在长期交互中持续演进的智慧沉淀。 反向信息悖论：API租用下的经济与数据流失 (5:10) 依赖租用前沿模型API隐藏着长期的知识产权流失风险。演讲者提出了“反向信息悖论”：即便模型供应商承诺不直接使用客户数据进行训练，但仍可通过监控调用失败模式、查询分布及边缘案例来合成针对性数据，进而优化其下一代基础模型。这意味着企业支付高昂调用费的同时，实质上也在向模型供应商输送关键业务认知，而这些认知最终会被商品化并重新分发给行业竞争对手。Palantir高管亚历克斯·卡普（Alex Karp）亦指出，企业客户日益不满前沿AI实验室通过API使用模式捕获其专有商业机密与“Alpha”。 此外，租用通用模型在经济性上面临严重挑战。前沿基础模型为覆盖数千种广泛任务而设计，迫使只需单一专用功能的企业承担巨大的冗余算力开销。这导致企业陷入被低成本自建模型的对手压制、或自行吞下微薄利润的困境。相比之下，针对特定业务微调的专用模型能将单次推理成本压缩5至10倍甚至百倍，使企业的单位经济模型回归健康。 底层技术主权：规避断供与强制弃用周期 (7:42) 技术主权与控制权是促使企业转向自建模型的另一核心支柱。过度依赖第三方平台使企业暴露于多重不可控风险之下，包括地缘政治事件引发的突发访问中断，或者供应商在特定垂直领域（如药物研发或金融科技）推出竞品所产生的商业冲突。与此同时，前沿AI厂商频繁强制弃用旧版本模型检查点，迫使企业陷入成本高昂且不可预测的模型迁移周期，而升级后的新模型往往价格翻倍，却未必能在垂直任务上带来有效性能提升。 在受到强监管约束的行业中，严格的数据血统、数据驻留及合规审计要求企业必须对模型部署环境和推理逻辑拥有绝对控制权。以往由于缺乏高精尖算法人才且单用例工程周期长达数月，企业难以推进自主研发；而“AI工厂”理念的出现，旨在通过工程化平台将业务问题、专有数据与领域知识直接转化为受控的专有生产力模型。 AI工厂架构：自动化蒸馏与生产自进化闭环 (9:46) 为兼顾高灵活性与严格治理，Oumi打造了基于Apache 2.0开源协议的“AI工厂”架构。该架构允许团队仅通过自然语言Prompt定义目标任务，平台即可自动化完成双边基准测试、深层失败模式归因诊断、定向合成训练数据，并自动调优超参数完成微调训练。所有数据生成配方和训练逻辑均保持完全透明与可编辑，支持一键式GPU弹性部署或本地化导出，彻底消除了技术黑盒与厂商锁定。 在现场构建代码路由器的演示中，系统展示了在零初始标注数据集的情况下，仅凭任务描述从Claude Opus提炼蒸馏合成数据，成功训练出紧凑的Qwen模型。该路由器性能比肩大模型，但体积缩小至百分之一，运行成本缩减至十分之一。此外，系统通过在生产环境中实时监测漂移、聚类错误根因并定向合成数据更新模型，构建出不依赖外部实验室的“自进化闭环”，将单次不可控的Token账单转变为仅需几美元的固定训练投入。 落地成效与长期复利：自有智能的资产沉淀 (16:23) 真实的行业部署案例展现了自建专用模型在性能与经济维度的确定性收益。某排名前十的国际商业银行利用该体系在2小时内构建了端侧SLM模型，错误率降低66%，且由于完全在本地移动设备运行，云端GPU服务开销降为零；DMG打造的专用任务模型在维持并超越原准确率的同时，运行成本比原先调用的商业GPT降低了100倍；一家排名前十的客户体验（CX）企业通过微调专用模型使任务准确率提升3.5%，误报率降低50%，推理成本下降33倍。 演讲者总结指出，构建专用模型的手动工程时间目前可压缩至十余分钟，企业无需持续承受昂贵且不透明的计量API账单。未来AI时代的真正赢家，必然属于那些将业务经验转化为专有模型资产、让数据飞轮在内部持续产生复利的企业，而非单纯租用公共智能的消耗者。 本内容系统论证了企业从调用公有API向自建微调模型演进的必然趋势，并以Oumi平台的自动化链路展示了将Prompt直接转化为受控端到端模型的工程方案。不过，该论述主要代表技术基础设施方案商的视角，对于在极端缺乏高质量领域输入、或任务规则极度发散的长尾场景中，合成数据与自优化闭环的泛化稳定性仍有待更长周期的生产检验。

Detailed notes — zh

## Background 本内容由 Ai4 于 2026 年 8 月 14 日发布在 YouTube 平台上，由 Oumi 联合创始人兼 CEO Manos Koukoumidis 主讲。视频页面简介指出，企业 AI 正在从租用通用模型转向拥有专有智能，随着 AI 原生平台将专用模型的开发周期从数月压缩至数小时，企业需要重新思考自身的 AI 战略与竞争护城河。 ## TL;DR 演讲探讨了企业级 AI 正经历的重大范式转变——从单纯依赖调用第三方黑盒 API 租用通用智能，转向构建并完全拥有针对核心业务深度定制的专有智能。主讲人深入剖析了这一转变背后的三大核心驱动力：捍卫企业差异化与防范“反向信息悖论”、大幅压降推理运行成本（最高可降 100 倍），以及全面掌握技术、数据合规与供应链控制权。通过 Oumi 打造的端到端全自动“AI 工厂”，企业能够以类似 Agent 的交互方式，在数小时内完成模型评测、数据合成、自动化微调与部署，并在生产环境中建立持续反哺进化的数据闭环，将专有模型转化为企业长期的核心资产。 ## Atmosphere & style 整场演讲风格沉稳而极具前瞻洞察力，主讲人凭借在谷歌等前沿 AI 团队的深厚工程底蕴，展现出兼具战略高度与技术落地的专业自信。叙事节奏紧凑洗练，从直击行业痛点的商业逻辑剖析，平滑过渡到生动的实机 Demo 与真实的标杆客户落地数据，层层递进地破除企业自建模型的工程壁垒，给听众带来一种既紧迫又豁然开朗的战略启迪感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=4s) 企业 AI 迎来历史性拐点：企业正从向外部模型实验室“租用通用智能”，全面转向构建并完全拥有自身的“专有智能”，掌控 AI 闭环成为避免被淘汰的核心能力。 - [01:00](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=60s) 企业 AI 投资的核心评估准则：凡是构成核心业务差异化优势的环节必须自建且自主掌控，而通用非核心任务可直接租用成熟外部工具；当高昂的推理成本危及商业模型时，必须专项构建轻量模型。 - [02:15](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=135s) 头部企业自建智能初见成效：Cursor 与 Intercom（Fin）等企业已自主研发专有模型或智能体，在达到或超越顶级前沿模型效果的同时，将运营成本降低了 5 至 10 倍。 - [02:40](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=160s) 捍卫差异化护城河：若所有企业调用相同的底层大模型，产品体验将不可避免地走向同质化；真正的竞争优势源于将企业独有的机构知识沉淀至专有模型中并持续复利。 - [04:10](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=250s) 防范“反向信息悖论”：单纯调用黑盒 API 存在核心商业机密泄露风险，企业支付高昂 API 费用调用服务的同时，其业务交互与报错数据正被前沿模型实验室用于升级自身模型，进而无差别赋能给所有竞争对手。 - [06:37](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=397s) 通用大模型的经济性硬伤：前沿模型为兼顾上千种任务而设计，导致企业为单一业务场景支付高出实际需求 5 到 100 倍的推理成本，极大地侵蚀了商业毛利率。 - [07:15](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=435s) 技术自主掌控防范四重外部风险：依赖第三方黑盒 API 面临地缘政治断供、模型厂商切入下游垂直领域成为直接竞对、被动承受强制版本迁移以及监管合规与隐私泄露的严重威胁。 - [08:50](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=530s) 破解工程与人才瓶颈：传统自建模型需要数月周期且极度依赖稀缺 AI 算法专家，企业亟需借助自动化平台流水线化打造自有的“AI 工厂”，实现生产交互反哺模型的智能复利飞轮。 - [09:46](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=586s) 端到端自动化 AI 工厂：Oumi 支持仅凭一个业务 Prompt，即可自动完成多模型评测基准、失败模式诊断、靶向合成训练数据、自动化微调及模型一键部署。 - [10:40](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=640s) 资产所有权与去黑盒化：从数据合成到超参数配置的全部技术配方（Recipes）对企业完全透明，支持在 Apache 2.0 开源库运行，模型权重可自由导出至本地私有化机房（On-prem）部署以杜绝厂商锁定。 - [11:46](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=706s) 代码路由器实机验证：通过 Opus 教师模型对 4B 小型开源学生模型进行微调蒸馏，成功构建出路由决策能力比肩前沿模型的专有路由器，将代码任务分流至体积小 100 倍的模型，实现 10 倍降本。 - [15:33](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=933s) 生产环境的自进化闭环：告别向 API 厂商单向流失数据的模式，系统自动捕获线上推理的失败案例并聚类分析失效模式，定向生成靶向训练数据驱动模型增量微调以应对分布漂移。 - [16:23](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=983s) 行业标杆落地实绩：全球前十大银行在 2 小时内完成模型构建，将错误率降低 66% 并实现移动端本地零 GPU 成本运行；DMG 与全球顶级 CX 企业则分别达成 100 倍与 33 倍的运营降本。 - [17:30](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=1050s) 下一代企业 AI 竞争法则：AI 时代的最终赢家不是那些长期租用外部智能的企业，而是能够将构建、定制和复利自身专有智能转化为长久核心 IP 与永续进化资产的组织。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=4s) 企业 AI 的历史拐点：从“租用通用智能”转向“拥有专有智能” 企业级 AI 正处于有史以来最重大变革的拐点时刻。各大企业正以前所未有的速度，从向外部租用通用智能，转向构建并完全拥有自身的专有智能（Specialized Intelligence）。过去，企业主要依赖于调用大型闭源模型的黑盒 API，例如 OpenAI 的 GPT 系列、Anthropic 的 Claude，以及演讲者此前在谷歌参与研发的 Gemini；而现在，越来越多的企业正转向微调开源模型，根据自身业务领域和具体任务进行深度定制，并实现 100% 的自主掌控。演讲者直言，企业如果无法掌控自身的“AI 闭环（AI loop）”，缺乏构建、拥有并持续复利自身智能的能力，终将被历史淘汰。 近两年来，演讲者与数十家大型组织的资深高管及 C-level 领袖交流，他们最常探讨的痛点是：**未来的商业价值究竟沉淀在哪里？** 行业内盛传“AI 正在吞噬软件”，那么价值究竟会流向软件体验层，还是 AI 本身？如果是 AI，企业又该将投资重心放在何处？演讲者指出，企业不需要去盲目堆砌尽可能多的 AI，而是要聚焦在对业务成败至关重要的 AI 上，并遵循两条核心准则： 1. **核心差异化原则**：如果某项 AI 能力是企业的核心业务与差异化优势所在，企业必须亲自构建、自主拥有，并使其智能在内部持续复利。相反，如果是通用型辅助工作（例如一家银行在做基础的软件工程开发），直接采用外部成熟工具（如 Anthropic 提供的 Claude Code）即可，因为写代码并非银行区别于竞争对手的核心能力；但若是风险分析（Risk Analysis）、按揭贷款审批（Mortgage Processing）等核心业务，就必须自主拥有专属的 AI 模型。 2. **经济成本驱动原则**：如果某项 AI 能力的运行成本过高，直接威胁到企业的商业经济模型，那么企业必须寻求针对该任务专项优化的模型。市场上若有合适模型则直接采用；若无，则必须亲自构建以压降成本。 ```mermaid flowchart TD accTitle: 企业 AI 决策分流架构 subgraph 评估维度 A["业务与成本评估"] --> B{"是否为业务核心差异化？"} B -- "是" --> C["自主构建与拥有专有 AI<br/>（如银行的风控/贷款审批）"] B -- "否" --> D{"推理调用成本是否威胁商业模型？"} D -- "是" --> E["构建/采用轻量专有 AI 优化成本"] D -- "否" --> F["租用成熟通用 API<br/>（如通用代码辅助）"] end ``` --- ### [02:15](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=135s) 真实企业转向先例与原因之一：捍卫差异化与防范“反向信息悖论” 回顾过去几年的轨迹：2023 年 ChatGPT 横空出世至 2025 年末，大部分企业普遍处于“租用”状态，依赖黑盒 API 快速试水。然而进入 2026 年，尤其是近两三个月，头部企业正在以极快速度脱离对基础模型 API 的单一依赖。行业内已涌现出标志性案例： - **Cursor**：公开表示不能完全受制于 Claude 或其他模型厂商，随后研发并上线了自主掌控的代码智能体（Coding Agent），其性能比肩顶级前沿模型，但运行成本足足降低了 **10 倍（10x）**。 - **Intercom**：仅仅一两周后，Intercom 也推出了自主训练的客户支持智能体 **Fin**，在客户服务（CX）场景中的表现不仅优于通用 GPT 与 Claude，综合运营成本更直接降低了 **5 倍（5x）**。 企业必须做出这种转变的**第一个核心原因在于“差异化（Differentiation）与竞争护城河”**。如果一家企业与所有竞争对手调用同一款底层模型、输入相似的 Prompt，其交付的产品体验便毫无差异可言。随着 AI 在客户产品体验中所占的比重越来越大，由于底层同质化，企业的核心差异化优势只会随时间推移被不断稀释。真正的护城河，应当是将企业独特的机构知识（Institutional Knowledge）固化并沉淀进专有模型中，通过生产环境下的海量交互持续微调升级，让智能在业务迭代中不断复利。 以一家顶尖律所为例，该律所按总营收计算高居行业前列，并非科技公司。其管理层清醒地指出：“通用生成式 AI 确实抬高了行业门槛与基线，但客户为我们付费不是为了得到平均水平的‘地板级交付’，而是为了顶级的卓越表现。”因此，他们拒绝像其他同行一样单纯依赖 Harvey 或 Legora 这类底层套壳 GPT/Claude 的现成工具，而是毅然着手训练自己的专有法律模型。 这种对自主掌控权的追求，更关乎防范自身的业务情报向基础大模型实验室泄漏。演讲者提到了业内关于“没有生态系统的 Frontier 是不稳定的”的深度讨论，强调企业绝不能外包核心学习能力；人与专有 AI 的协同学习能力才是企业的灵魂所在。单纯调用黑盒 API 存在巨大的安全隐患： - **无形的数据与逻辑泄漏**：即使 API 合同声称“不使用客户数据进行模型训练”，模型厂商依然在持续监控报错和交互流，以此构造合成数据修补自身模型缺陷。企业实质上是在把自己最前沿的业务逻辑与 Know-How 奉送给模型厂商，后者随后升级新版模型，无差别赋能给企业的所有竞争对手。 - **反向信息悖论（Reverse Information Paradox）**：微软 CEO 萨提亚·纳德拉（Satya Nadella）精准地将这一现象概括为“反向信息悖论”——企业在支付极其昂贵的 API 溢价的同时，反而在免费为模型厂商提供业务经验，帮助对方学习自己的商业机密，相当于企业为此付出了“双重代价”。 - **帕兰提尔（Palantir）CEO 亚历克斯·卡普（Alex Karp）的警告**：卡普同样直言，许多企业客户对此感到出离愤怒，因为前沿模型实验室正在系统性地窃取客户的知识产权与商业核心竞争优势（Alpha）。 --- ### [06:37](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=397s) 原因之二与之三：商业经济性与技术掌控力 ```mermaid flowchart LR accTitle: 专有模型与通用模型经济成本对比 subgraph 通用大模型调用 G1["全能基础模型<br/>（训练兼顾上千种非必要任务）"] --> G2["单次调用成本高昂<br/>（贵 5x - 100x）"] G2 --> G3["被迫涨价失去客户<br/>或承受亏损吞噬毛利"] end subgraph 垂直专有模型 S1["领域专有轻量模型<br/>（仅专注核心垂直任务）"] --> S2["极低推理运行成本<br/>（降低 5x - 10x）"] S2 --> S3["差异化提升带来溢价空间<br/>毛利率跨周期持续扩大"] end ``` **第二个转变驱动力在于经济模型（Economics）**。现有的前沿通用大模型为了能够胜任一千种复杂任务而设计，但具体到企业的单点业务场景，企业往往只需要其中的一种能力，却不得不为模型具备的其余 999 种无关能力持续买单。这导致通用 API 的推理成本往往比业务实际所需高出 **5 到 10 倍，甚至 100 倍**。 - 面对如此沉重的推理成本负担，企业若选择提高产品售价，就会被拥有 Cursor 或 Intercom 那样低成本专有方案的竞争对手在价格上彻底击败； - 若不涨价，高昂的 Token 账单则会直接侵蚀毛利，导致严重的商业亏损。 - 相比之下，从第一天就构建专有模型的企业，不仅能立即砍掉 5 到 10 倍的运营成本，还能伴随产品在实际使用中的效果迭代形成差异化壁垒，从而获得更强的定价权与不断扩张的利润率。 **第三个驱动力在于全方位的控制权（Control）**。企业将关乎身家性命的核心业务押注在第三方黑盒模型上，正面临四重致命风险： 1. **地缘政治与供应风险**：地缘政治波动或制裁政策可能瞬间导致外部 API 访问被掐断。 2. **潜在商业竞争风险**：前沿模型实验室（如 Anthropic 等）随时可能调整战略，切入生物制药、药物发现或金融服务等下游纵深领域，直接与企业客户争夺市场，并随时收回 API 访问权。 3. **强制技术迁移负担**：外部厂商经常单方面淘汰旧模型，强制企业迁移到费用翻倍但对特定业务并无显著改善的新版本上。演讲者透露，其正在服务的一家大型银行正深陷此类强制迁移所带来的沉重技术包袱中。 4. **数据隐私、合规与安全防线**：尤其在强监管行业，数据流动合规至关重要。拥有模型所有权意味着企业可以完全决定模型部署的机房位置、网络环境以及敏感数据的流转边界。 演讲者发出了尖锐的质问：“在面对直接决定业务成败的关键智能时，企业真的能够承受得起靠‘租用’过日子吗？”答案显然是否定的。 --- ### [08:50](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=530s) 跨越落地阻碍：打造企业内部的“AI 工厂”飞轮 既然自建专有智能的逻辑如此清晰且极具说服力，为何大多数企业此前依然迟疑不前？演讲者指出，核心症结在于极其严峻的**工程与运维挑战（Operational Challenges）**。 - 大量超大型企业曾向演讲者坦白：“我们非常渴望拥有自己的专有智能，但这太难了。针对每一个具体用例，往往需要耗费数月时间去训练；后续还要承担繁重的微调与模型运维；更糟糕的是，企业内部极度缺乏稀缺的顶尖 AI 专业算法人才。” 为了打破这一工程瓶颈，企业亟需的是实现专有智能构建的**工程化与流水线化（Operationalize）**，打造属于企业自己的**“AI 工厂（AI Factory）”**。这一体系应具备标准化能力：只需输入业务问题的具体定义、企业现有的自有业务数据（即使数据量有限），以及专家的领域知识（Domain Expertise），就能在平台化自动化驱动下，极速构建出属于企业自身的专有轻量化模型。 更重要的是将整套系统打造成一个自我强化的**飞轮机制（Flywheel）**。这正是微软纳德拉所强调的“智能复利”理念：让专有模型部署上线后的每一次真实生产交互，都能持续回流、反哺并优化模型。通过这一飞轮不断积攒企业的独家资产，在业务壁垒与智能化程度上快速起跑，将无法自建飞轮的竞争对手彻底甩在身后。 ### [09:46](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=586s) 全自动 AI 工厂与透明可控的资产所有权 在企业构建坚固竞争护城河的过程中，治理能力是不可或缺的关键一环。企业一方面需要紧跟 AI 创新的步伐全速前进，另一方面——尤其是在受严格监管的行业——必须确保系统完全可审计、灵活且透明。然而，市场上此前并不存在能够兼顾这几点的方案。为此，Oumi 构建了业界首个全自动化的“AI 工厂”（Fully Automated AI Factory）。用户只需提供一个原始 Prompt（无论当前使用的是 GPT、Claude 还是 Gemini），系统就能全自动完成后续的所有工作：自动对比并评估各类开源与闭源模型、深度诊断模型在何处存在短板（不仅给出量化指标，更提供深度的失败机理分析）、针对识别出的失效模式自动生成用于补短板的训练数据、推荐最佳算法与超参数并自动执行微调，最终通过多轮迭代产出最适配该任务的专用模型。训练完成后，用户既可以选择一键部署到具备 GPU 自动扩缩容的基础设施上，也可以将模型权重导出并在本地机房（On-prem）独立运行。 ```mermaid flowchart TD accTitle: Oumi 全自动 AI 工厂微调与部署流程 subgraph Phase1 ["1. 意图输入与基准评测"] A["输入业务 Prompt"] --> B["多模型并行评测与失败模式诊断"] end subgraph Phase2 ["2. 数据生成与模型迭代"] B --> C["针对性合成补短板训练数据"] C --> D["自动化微调与超参数优化"] D --> E{"性能是否达标？"} E -- 否 (迭代优化) --> C end subgraph Phase3 ["3. 自由部署与数据闭环"] E -- 是 --> F["云端一键部署 / 本地私有化导出"] F --> G["生产环境失效捕捉与持续自进化"] G -. 持续反馈.-> C end ``` 这种自动化绝非不可控的“黑盒”。正如微软 CEO 萨提亚·纳德拉（Satya Nadella）所强调的，企业必须掌控自己的技术命运，避免被单一供应商锁定。在 Oumi 平台上，每一步骤的技术配方（Recipe）——包括数据如何合成、模型如何评估、超参数如何配置——全部完全透明、可审查且支持自由修改。企业不仅拥有最终的模型权重和数据，还拥有全部构建配方，这些流程均可在基于 Apache 2.0 协议的开源库上运行，真正实现无厂商锁定的自主掌控。 - **端到端自动化**：从单个 Prompt 出发，自动完成评测对比、失效模式归因、数据合成、模型微调与多轮迭代。 - **透明度与所有权**：开放全部训练与合成配方，支持在 Apache 2.0 开源框架下复现与迁移，杜绝黑盒锁定。 - **灵活部署**：支持云端 GPU 弹性伸缩托管，也支持将权重完全导出进行 On-prem 私有化部署。 --- ### [11:46](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=706s) 实机演示：通过 Agent 智能工程构建代码路由模型 在实机演示环节，团队展示了一个具体的开发场景：构建一个代码任务路由器（Coding Model Router）。如果将所有代码查询不加区分地直接发送给最顶级的闭源模型（如 Claude），成本将极其高昂；而通过专用路由器将简单或中等难度的请求分流至更小、更便宜的开源模型，复杂任务才交由顶级大模型处理，可以在保证相同回答质量的前提下，将请求分流至体积小 100 倍的模型，实现成本降低 10 倍，并且直接运行在企业自有的硬件资产上。Oumi 平台的操作体验就像“AI 工程领域的 Claude Code”，通过对话式的“Agent 化 AI 工程”（Agentic AI Engineering）即可完成全部闭环。 ```mermaid flowchart LR accTitle: Oumi 驱动的代码模型路由器工作流 A["代码查询请求"] --> B["Oumi 驱动的专用代码路由器"] B -- "简单 / 标准任务" --> C["小型开源模型 (低成本/私有部署)"] B -- "极端复杂任务" --> D["Claude Opus / 前沿大模型"] ``` 在 Demo 中，工程师输入了路由器的设计目标：分析代码问题的复杂度、类别和需求，在保证回答质量的同时平衡成本，针对简单任务推荐经济的开源模型，仅对极复杂任务调度 Claude Sonnet 或 Opus。系统随后以智能体方式展开交互，主动询问细节（如确认 JSON 输出格式及 `model_chosen`、`rationale` 等字段），随后自动制定执行计划，涵盖定义评测指标、合成测试集、评估基线模型、合成训练数据、微调模型、后验评估以及复盘等环节。 演示中选用了 Qwen 3.5 4B 作为待微调的基础学生模型，并以 Opus 4.6 作为合成训练数据的教师模型（平台亦支持市面上主流的各类开源与闭源大模型担任此角色）。经过微调蒸馏后，这个体积仅为原模型百分之一的 4B 模型不仅大幅超越了自身的基线表现，更完全追平了 Opus 4.6 的路由决策能力，将原本持续按 Token 计费的沉没成本，变成了仅需几美元微调费用即可永久拥有的私有资产。 - **蒸馏与超越**：在纯蒸馏场景下，小型学生模型即可打平顶级教师模型的表现；而在引入企业专有外部数据微调时，专用模型的效果通常能超越通用前沿模型 10% 以上。 - **成本计算器**：Oumi 提供了 `umi.ai/calculator` 工具，用于量化对比“租用前沿模型”与“拥有自建模型”在具体业务场景下的长期 ROI 差异。 --- ### [15:33](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=933s) 生产环境中的持续自进化循环 将微调模型交付并非终点，真正的价值在于构建模型上线后的“自进化循环”（Self-Improvement Loop）。在传统调用第三方 API（如 OpenAI 或 Anthropic）的模式下，所有生产环境产生的失败案例、边界情况以及潜在的数据反馈信号都被白白丢弃；而在自建闭环中，这些数据是企业最宝贵的改进资产。 ```mermaid flowchart LR accTitle: Oumi 生产环境自进化闭环 A["生产环境推理"] --> B["捕获失败预测 / 识别数据分布漂移"] B --> C["Oumi 失效模式高维聚类"] C --> D["定向生成靶向训练集"] D --> E["增量微调与模型迭代升级"] E --> A ``` Oumi 平台能够对模型在生产环境中出现的预测错误进行高层次的分类聚类，精准剖析“失效模式”（Failure Modes）。随后，平台自动针对这些薄弱环节定向策划并补充新的训练集，实现自动化的增量微调。这一机制使得模型在部署到生产环境后能够持续学习错误案例并检测输入分布的漂移（Distribution Drift），让企业的模型智能随着业务运行不断复利增长。 - **摆脱信号浪费**：告别向 API 厂商单向贡献数据却无法留存反馈信号的局面，将线上真实流量转化为模型迭代燃料。 - **定向失效修复**：通过聚合错误预测自动反哺训练，无需人工介入繁重的数据清洗与测试集构建。 --- ### [16:23](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=983s) 行业落地实效与掌控企业专属智能的未来 Oumi 团队拥有深厚的工程底蕴，核心成员曾在 Google 参与 Gemini 等核心项目的研发（主讲人本人便是在 Google F4 开启职业生涯），并汇聚了来自 Meta、微软以及 Apple Intelligence 的核心专家，同时获得了多位 AI 领域权威学者的支持。在实际业务落地中，该平台已展现出显著的降本增效成果： | 客户类型 / 案例 | 核心成效 | 业务价值 | |:--- |:--- |:--- | | **全球前十大国际银行** | 仅耗时 **2 小时** 完成轻量化模型构建，错误率降低 **66%** | 模型完全在移动终端本地运行，GPU 推理成本彻底降为 **零** | | **DMG** | 准确率大幅提升 | 推理成本较此前使用的 GPT 模型直降 **100 倍**（100x） | | **全球前十大客户体验（CX）企业** | 准确率提升 **3.5%**，假阳性（误报）减少 **50%** | 推理运营成本降低至原来的 **1/33**（即便宜 33 倍） | 主讲人总结道，构建专有模型的实际操作时间其实极短，"仅需占用你 15 分钟的有效时间"，后台自动化运行几个小时到一天即可交付。 在演讲尾声，主讲人向全场强调了下一代企业 AI 战略的核心法则："下一代 AI 时代的赢家，绝不会是那些仅仅租用智能的企业；真正的赢家，属于那些学会将构建、专用化并持续复合自研智能转化为企业永续进化资产与 IP 的组织。"这构成了任何企业在 AI 时代最坚固的护城河。最后，主讲人邀请参会者参加下周举办的深度研讨会（Webinar），并欢迎大家前往 1341 号展位现场体验产品 Demo 并领取周边。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=4s) 企业如果无法掌控自身的“AI 闭环（AI loop）”，缺乏构建、拥有并持续复利自身智能的能力，终将被历史淘汰。 — Manos Koukoumidis > [00:04](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=4s) 如果某项 AI 能力是企业的核心业务与差异化优势所在，企业必须亲自构建、自主拥有，并使其智能在内部持续复利。 — Manos Koukoumidis > [02:15](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=135s) 通用生成式 AI 确实抬高了行业门槛与基线，但客户为我们付费不是为了得到平均水平的‘地板级交付’，而是为了顶级的卓越表现。 — 律所管理层 > [02:15](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=135s) 反向信息悖论：企业在支付极其昂贵的 API 溢价的同时，反而在免费为模型厂商提供业务经验，帮助对方学习自己的商业机密，相当于企业为此付出了“双重代价”。 — Satya Nadella > [02:15](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=135s) 前沿模型实验室正在系统性地窃取客户的知识产权与商业核心竞争优势（Alpha）。 — Alex Karp > [06:37](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=397s) 在面对直接决定业务成败的关键智能时，企业真的能够承受得起靠‘租用’过日子吗？ — Manos Koukoumidis > [09:46](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=586s) 企业必须掌控自己的技术命运，避免被单一供应商锁定。 — Satya Nadella > [16:23](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=983s) 仅需占用你 15 分钟的有效时间 — Manos Koukoumidis > [16:23](https://www.youtube.com/watch?v=j0kMTC6Lkfo&t=983s) 下一代 AI 时代的赢家，绝不会是那些仅仅租用智能的企业；真正的赢家，属于那些学会将构建、专用化并持续复合自研智能转化为企业永续进化资产与 IP 的组织。 — Manos Koukoumidis ## Mentioned Manos Koukoumidis, Oumi, Ai4, OpenAI, GPT, Anthropic, Claude, Claude Code, Claude Sonnet, Claude Opus, Opus 4.6, Google, Gemini, Google F4, Cursor, Intercom, Fin, Harvey, Legora, Satya Nadella, Microsoft, Alex Karp, Palantir, Apache 2.0, Qwen 3.5 4B, Meta, Apple Intelligence, DMG ## Video description Manos Koukoumidis - CEO and Co-Founder, Oumi Most organizations began their AI journey by renting generic intelligence through frontier models. It was the fastest way to get started, but they quickly discovered three fundamental limitations: generic models rarely deliver the highest quality on domain-specific tasks, inference costs become difficult to justify at scale, and the intelligence powering their business remains under someone else's control. Leading enterprises are now shifting toward specialized AI they own. Until recently, building domain-specific models required months of work and scarce AI expertise. Today, AI-native platforms automate model evaluation, data generation, training, deployment, and continuous improvement, reducing development cycles from months to hours and enabling domain experts to build production-grade AI. If building specialized AI becomes as accessible as modern software development, how should enterprise AI strategy evolve? What competitive advantages become possible when every organization can build, own, and continuously improve intelligence tailored to its business? What will separate the winners from everyone else? [Source](https://www.youtube.com/watch?v=j0kMTC6Lkfo)

Source quotes

[0:04] Manos Koukoumidis: 企业如果无法掌控自身的“AI 闭环（AI loop）”，缺乏构建、拥有并持续复利自身智能的能力，终将被历史淘汰。 [0:04] Manos Koukoumidis: 如果某项 AI 能力是企业的核心业务与差异化优势所在，企业必须亲自构建、自主拥有，并使其智能在内部持续复利。 [2:15] 律所管理层: 通用生成式 AI 确实抬高了行业门槛与基线，但客户为我们付费不是为了得到平均水平的‘地板级交付’，而是为了顶级的卓越表现。 [2:15] Satya Nadella: 反向信息悖论：企业在支付极其昂贵的 API 溢价的同时，反而在免费为模型厂商提供业务经验，帮助对方学习自己的商业机密，相当于企业为此付出了“双重代价”。 [2:15] Alex Karp: 前沿模型实验室正在系统性地窃取客户的知识产权与商业核心竞争优势（Alpha）。 [6:37] Manos Koukoumidis: 在面对直接决定业务成败的关键智能时，企业真的能够承受得起靠‘租用’过日子吗？ [9:46] Satya Nadella: 企业必须掌控自己的技术命运，避免被单一供应商锁定。 [16:23] Manos Koukoumidis: 仅需占用你 15 分钟的有效时间 [16:23] Manos Koukoumidis: 下一代 AI 时代的赢家，绝不会是那些仅仅租用智能的企业；真正的赢家，属于那些学会将构建、专用化并持续复合自研智能转化为企业永续进化资产与 IP 的组织。

Source quotes

[0:04] Manos Koukoumidis: on the wrong side of history [0:04] Manos Koukoumidis: AI is eating software [2:15] Manos Koukoumidis: raises the floor for everybody [5:10] Manos Koukoumidis: paying twice [9:46] Manos Koukoumidis: Build a model that does whatever you put in that prompt. [11:46] Stefan: Claude Code, but for AI engineering. [16:23] Manos Koukoumidis: Own the intelligence that your business runs on. You can literally build it in hours if not in minutes and then let it compound for you into your asset forever. [0:04] Manos Koukoumidis: Build and own your core differentiators [0:04] Manos Koukoumidis: Optimize unit economics for high-volume tasks

Original transcript

[0:04] We are at the inflection point of the biggest  shift that is happening in enterprise AI. [0:10] Enterprises are moving at an increasing  pace from generic intelligence to renting [0:16] to specialized intelligence that are owning.  They're moving from consuming the APIs of GPTs, [0:22] Claude's, or Gemini that I was building at Google  towards open models that they are fine-tuning, [0:29] specializing to the domain and tasks that they  can fully own. Enterprises are realizing that [0:33] if they don't own their AI loop, if they  don't own the capability to build, own, [0:38] and compound their own intelligence, they're  going to end up on the wrong side of history. [0:44] I've been talking the last couple of years  to, I don't know, dozens and dozens of [0:50] senior leaders and C-level executives of some of  the largest organizations, and the thing that I [0:54] especially increasingly lately keep asking is  what is the value accruing for my business? [1:01] Is it software, the application experiences, is  it AI? I'm hearing that AI is eating software, [1:07] and if it's AI, what exactly AI, you  know, where where should I be putting my [1:11] my focus for my organization, my investment? Well,  for AI, it's going to be AI, but it's going to [1:18] I'm not going to be a matter of building the most  AI. It's going to be a matter of building the AI [1:23] that is critical for your success. And there's  two two key rules for this. The first one is [1:29] if there's AI that is critical and the  differentiating thing for your business, [1:33] you should be building, owning, and compounding  that intelligence yourself. If it's something [1:37] that's generic, let's say for example, you are  a bank and you're doing software development, [1:41] that's fine. Just use what for example,  Anthropic with Claude code is giving you, [1:45] because coding is not what differentiating your  business from every all the other business are [1:49] writing software as well. But for example, if  it's like risk analysis, mortgage processing, [1:52] something that's the thing that differentiates  your business, that's the AI you should be owning. [1:58] The The second rule is um if it's AI that is  really costing you, it's affecting your economics, [2:04] then you should also consider using AI that is  specialized for the task. If one already exists, [2:09] just use it. If not, build this  so you can optimize your costs. [2:15] So, that's why we're seeing this shift uh the  last couple of years. You know, in the first [2:19] 2023 when ChatGPT came out and also late 2025,  most organizations still pretty pretty much [2:26] renting, you know, using the GPTs, Gemini's,  and Claude's, using those black box APIs. [2:31] Especially 2026, even earlier, but we have  seen that an accelerating pace, especially [2:35] the last 2 months, organizations quickly moving  to building and owning their own intelligence. [2:41] As a couple quick examples from Mars, uh  Cursor said, "Hey, we can't be depending [2:45] on Claude or any of these companies. Uh we  build our own coding agent that is as good as [2:52] the Claude's and the GPTs. It cost us  10 times less to use." And then a week I [2:58] think about a week or 2 later, uh Intercom  came out to say, "Hey, we also built Fin, [3:02] better uh CX customer support agent than Claude  and GPTs, five times cheaper for us to operate." [3:09] Enterprises are realizing that this is the  shift increasingly that they all need to do [3:14] because if they don't do that, they end up they're  going to end up getting replaced. I'm going to [3:18] break it down three reasons why this is they're  realizing this. The first one is differentiation. [3:23] If you're using the same AI as your competitors,  prompting the same models, you are not really [3:29] differentiating. And as that AI becomes the bigger  and bigger part of the experience of this what you [3:34] provide to your customers, and that big part  is the same as your competitors, the less your [3:39] differentiation becomes over time. That's why  you should see your differentiation going down. [3:43] If you're building if you go and build your own  intelligence based on your own institutional [3:46] knowledge, and you put in production and you  keep improving it, making better and better [3:51] with every interaction your customers have with  it. That's how you increase your differentiation. [3:57] So, that's the way for any enterprise to build  the moat because the next moat for enterprise [4:02] going to be the specialist intelligence that they  own. There's many examples of this. So, circle in [4:07] analysis a couple months back. That's a legal  firm. The highest in gross revenue legal firm, [4:13] not a tech company, a legal firm. They said,  "You know what? Okay, all these G&A models, [4:17] they are raising the floor for everybody, but  we're an exceptional legal firm. We're not hired [4:21] for the floor. So, we need alongside our best  lawyers to also have the best legal AI. We're [4:26] not going to just use the same models, you know,  the Harveys and the Legoras, which pretty much [4:30] is GPT and Claude's as everybody else. So, that's  why they went on building their own intelligence. [4:36] Also, sorry, a couple weeks later, maybe  actually that's a month and a half ago. [4:39] He came up with a blog post. If you haven't  read it, highly recommend you to read it. [4:43] Actually, it's a next post that says the  frontier without an ecosystem is not stable. [4:47] Pretty much what he's saying, if you're  an enterprise, you should build, own, [4:50] compound your own intelligence. You cannot  outsource the learning of your organization. [4:55] What differentiates organizations is the learning  that their humans and AI is going to be doing. [5:00] That's what you should be owning. You cannot  outsource it. You should not be renting it. You [5:03] should be building it and owning yourself. Um and  the other reason is that besides of you building [5:10] that competitive advantage, if you're not building  this and you're just consuming those APIs, [5:14] slowly but steadily your business intelligence,  your edge, is leaking to the frontier labs. [5:21] Many ways in this is happening even though  they say we're not training on your data, [5:24] this is not a strong guarantee. Because they see,  they monitor what things are failing and then go [5:28] and fix it. They may not use your own data, but  there's ways to work around this and create data [5:33] that is similar and still make their models that  they're going to release next better for your [5:40] business. And then that's something  that goes to all of your competitors. [5:43] One way to tell is that if you see contracts  that don't explicitly say that, "Hey, [5:46] you cannot use the aggregated and all of this,  you know, Um explicitly they say no, we cannot [5:51] use in this the aggregated way your information  about your business to make our products better. [5:56] Likely one way or another this is happening. By  the way, this is also what Satya called with a [6:01] post that he pasted he posted a week later calling  it reverse information paradox pretty much saying [6:08] you're paying high premium prices to  use those APIs and you're paying them [6:15] for them to be learning the intelligence that your  business is running on. So, you're paying twice. [6:21] Alex Karp said the same  thing. They said, you know, [6:23] our customers they are very angry because  those frontier allows they're stealing their [6:29] their pretty much their IP, their  intelligence, their alpha as he called it. [6:33] So, that's how you're losing your  differentiation if you're renting these APIs. [6:37] It's the easiest thing to do, but it's the lazy  and the thing that will end up costing you the [6:42] most in the medium to long term not very not very  long term. The second reason why enterprises are [6:47] using their own intelligence, are building their  own economics. If you're using these big genetic [6:52] models, these are models that are trained to do  a thousand things when you only care about one, [6:57] but you're paying for all these thousand things.  You're paying for a model that is five to 10 [7:01] times or even hundred times more expensive  than it needs to be. That's why, you know, [7:06] if you're using these big models your economics  are very bad. And if you, you know, you have two [7:11] options when you costing you so much either you  increase your prices, but then you get out priced [7:15] by your competitors who have a more cost effective  solution like what Cursor did or what Intercom [7:19] did. Or you are increasing your price and then  you lose the customers, right? I mean, bad bad [7:23] economics for AI just very detrimental for your  business. Whereas if you build your own AI day [7:28] one, you can have a model that can cost you five  to 10 times less or even more. And then as you [7:33] compound that in products you keep improving it  and your differentiation becomes higher, you can [7:37] price it higher and then that's how your  margins get even even better across time. [7:42] The second reason why enterprises are moving  to building their own intelligence besides [7:46] differentiation, besides economics, is control  of the technologies that's so critical for their [7:51] success. They're saying, "If If this this is so  critical for me, I I cannot lose access anytime [7:58] just because of geopolitical reasons what happened  with Fable or because Anthropic one day decide to [8:02] start competing with me in drug discovery or in  financial services and they decide to take away [8:07] access. Uh or they may deprecate those models and  then they're forcing me to migrate to models that [8:13] are twice as expensive, not any better, and I have  to do the work to migrate. That's the problem of a [8:17] bank I'm working with. And last but not least,  especially if you're regulated industries, [8:21] but I would say for any industry, is privacy,  security, knowing where your data is going, [8:26] right? You can control. If you own that  model, you can control where you deploy it [8:30] and where your data is going. If you don't own it,  you have no control. So, the thing I would ask is, [8:35] can you really afford to rent the intelligence  that's so critical for your success? You can't. [8:42] And better to have a strategy now as opposed to  incur the cost when the changes come upon you." [8:50] Okay. Now, you may say, "Okay, if this is very  compelling, but why haven't enterprises done [8:53] this change yet?" It's because of operational  challenges. We have talked to some of the largest [8:58] organizations you can think of that told us, you  know, "We want to build our own intelligence. [9:02] It's hard. It will take us months per use  case. Then we need to maintain those models. [9:07] We don't even have the right expertise or maybe  we have limited amount of people that can build [9:10] this. So, what do we do?" The solution that all  enterprise need is to operationalize the building [9:16] of their intelligence, is to have their own AI  factory, a way that they can take the definition [9:20] of the problem they want to solve, the data they  may have, if any, and their domain expertise, [9:26] and start building efficiently those custom  AI specialized AI models they can own. [9:30] And the thing that will really differentiate  them is if they build that into a flywheel. [9:34] That's also what Satya called out I wrote and  a half ago, as compounding your intelligence, [9:38] so that every interaction in production  makes that model better and better for you. [9:42] And that's how you build your IP and you're off  to the races and nobody can catch up with you. [9:46] So, that's the way to build a robust mode for  any enterprise. So, at the same time you need [9:52] something that can be easily governed. Uh  you know, you need to move fast with the [9:56] speed of AI innovation, but you need specially  regulated industry, but again, also beyond a way [10:01] you can also govern this. So, you need a solution  that can be auditable, flexible, transparent. [10:07] Such a solution did not exist. That's why we  built it. Uh we built the first of its kind [10:12] fully automated AI factory. To give you an idea,  you can start with a prompt. If you're using GPT, [10:18] Claude, or Gemini, you can get that prompt  and bring it and say, "Build a model that does [10:24] you know, whatever you put in that prompt." It  will automate everything for you. It will tell [10:27] you what are the right ways to evaluate different  models, close ones and open ones, just to see them [10:32] side by side. It will analyze those evaluations,  even tell you what exactly those models fall [10:36] short, not just metrics, but deep deep analysis.  So, and this so you can understand as a human [10:41] where they fall short, and then you will automate  the creation of training data to address these [10:46] failures that it automatically identified, these  failure modes, tell you what's the right way to [10:51] train, algorithms, hyperparameters, train for you,  and repeat the process again and again until you [10:55] have the best model possible, specialized to the  task. And then you can do one-click deployment [10:59] with GPUs auto scaling, or you can export the  model, deploy on prem, or wherever you need to be. [11:04] The key thing I want to emphasize here was that  it's not just a black box automation. A key thing [11:09] that I think it's critical for any enterprise to  be in control, Satya also called it a month and [11:13] a half ago, is you need something that is fully  flexible, transparent, and does not lock you in. [11:19] So, that's why for every step of the process,  not just the data, not just the weight, but how [11:25] every step was executed, how your data was how the  data was synthesized, how the model was evaluated, [11:30] the all these recipes, they're fully  transparent for you to inspect and change. [11:34] Uh and then you own everything. You own the model  weights, the data, and also how the recipes, [11:38] how everything was created. You can run  them against an Apache 2 library. So, [11:42] again, full control, no lock-in. That's how  your AI infrastructure should look like. [11:46] So, to see a quick demo, I'll show you a quick  demo from Stefan. vastly different token costs. [11:51] On the left, a coding query is sent directly to  Claude. On the right, the same query is routed [11:57] to a smaller, cheaper model more appropriate for  the query's complexity. The result is stark. This [12:03] model just routed that coding request to a model  100 times smaller. Same answer, 10 times cheaper, [12:10] running on hardware you already own. Routing  everything through one big rented model is the [12:15] expensive default. Every AI request you send to  a frontier API costs metered tokens and you break [12:22] the feedback loop. But, this doesn't have to be.  Fine-tuning platforms give you the option to ditch [12:28] the frontier for everything and use inference more  intelligently. You still do the engineer's job, [12:34] evals, test sets, data curation, failure analysis,  and iteration. Umi automates that whole workflow. [12:40] Think Claude code, but for AI engineering. And  in this video, I'll show you how. We'll build [12:46] a coding model router from a single prompt, no  data required. Let's log in to the Umi platform. [12:52] You can chat with Umi just like you would  chat with Claude code. Only Umi is specialized [12:56] for AI development and we call this agentic AI  engineering. Let's enter our prompt and the prompt [13:03] is as follows. I want to build a model router  specifically for deciding between coding models. [13:08] Given a coding question, the router must analyze  its complexity, category, and requirements, [13:13] then recommend the most cost-effective open model  or Claude Sonnet and Opus for the most complex [13:19] tasks to answer it. You must balance quality  against cost. Use cheaper, smaller models for [13:24] simple tasks and reserve expensive, larger models  for complex ones. So, after we enter the prompt, [13:29] the Umi agent is now going to ask clarifying  questions in in this case about the output format. [13:36] We'll choose JSON and specify two fields, model  chosen and the rationale. And it also asks about [13:43] which models the router needs to decide between.  Now, at this stage, the Umi agent decides that [13:48] it has enough information about the task and it  devises a plan. And the plan is as follows. So, [13:55] define the metric, synthesize test data, evaluate  the baseline models, synthesize training data, [14:01] then fine-tune the model, and evaluate  the model after fine-tuning. Finally, [14:05] doing a review and deciding on next steps.  So, you can see that the different steps [14:11] ask additional clarifying questions and launch  jobs on Umi managed infrastructure. It walks us [14:16] through the entire process of model development.  We've chosen Qwen 3.5 4B as the base model, [14:23] the one being fine-tuned, and Opus 4.6 as the  model that synthesizes the training data. However, [14:29] the platform supports a comprehensive range of  open and closed-weight models for fine-tuning, [14:35] data synthesis, and evaluation by LLM judges.  After training, the evaluation metrics show that [14:41] we were able to from our fine-tuning drastically  improve accuracy and other metrics over the small [14:47] base model and also match Opus 4.6 performance  with a model that is 100 times smaller and 10 [14:54] times cheaper to run. In the example in this  video, we distilled Opus 4.6 into a small model, [15:00] and so we expect the student, our base model,  to match the teacher, Opus 4.6. However, [15:05] when fine-tuning to external data, the fine-tuned  model typically outperforms the frontier model by [15:10] 10% or more. There's an example of that right  here. Check out the rent versus own calculator [15:16] at umi.ai/calculator to get an estimate of how  much you can save by building your own custom AI [15:22] for your specific use case. So, we've turned  requiring meter tokens from a frontier model [15:28] forever to a fixed cost of a few dollars  to train a small language model. But, [15:33] we can do even better both before the model  is put into production and afterwards, [15:38] valuable feedback signals exist. If only they  could be captured and utilized. Fortunately, [15:43] Umi has several features for doing just that and  we call this the self-improvement loop. Let's [15:49] examine the failure modes of the fine-tuned model.  And what this means is the high-level groupings [15:54] of the types of prediction errors observed.  Umi takes those failure modes and curates a [16:00] training set targeted at exactly those gaps. Umi  can continuously improve your models, performing [16:06] the same steps as we've just shown after the  model's been put into production to learn from [16:10] failed predictions, detect distribution drift,  and more. Stop throwing away valuable learning [16:16] signals when you build with Anthropic or OpenAI  and start earning the self-improvement loop. [16:23] I'll cut I'll cut it a little bit short in the  interest of time. Uh so, as I told you, it's very [16:28] easy. Literally, it can take you 15 minutes of  your time to build your own specialized AI model. [16:32] It may take you hours on the clock, let's say a  day, but it may only take 15 minutes of your time, [16:37] literally. So, uh Umi's being built by a very  strong team who used to many of us build Gemini [16:42] at Google. Actually, I started at F4 at Google.  Um and many other people from Meta, Microsoft, [16:47] and Apple Intelligence. We have an exceptional uh  support from some of the most reputable academics [16:53] in AI uh because they really love the mission  we're pushing forward. Uh we have worked with [16:58] many companies. I mean, I'll go very very quickly  show you a couple of results, you know, with one [17:01] of the top 10 international banks. They wanted a  model that runs ideally on device and we managed [17:07] to build a model small enough to run on device,  reducing errors by 66% and the cost, of course, [17:12] went to zero because if it's running on device  like mobile devices, there's no cost for them [17:16] for GPUs. And the best part, it took us 2 hours to  build it. Uh DMG, same case, we improved accuracy [17:23] with a model that was in the end compared to the  GPTs they were using a hundred times cheaper. [17:29] A hundred times, a hundred X. Uh also another top  10 CX company uh improved quality 3.5% increase [17:37] in accuracy, 50 less 50% less false positives, 33  times less cost. Not 33%, 33 times less cost. Omis [17:46] has also been used by many other organizations  that are using our platform. Uh and to close [17:52] this and help the uh it seems almost right on  time. The key thing I want you to take is that [17:57] the winners of the next AI era will not be the  enterprises that are renting their intelligence. [18:03] It will be the enterprises that learn how they can  operationalize building their own intelligence, [18:09] specializing their intelligence, and compounding  it in production it and into a never improving IP [18:16] and asset. Because that's going to be the  most robust mode for any organization. So, [18:24] own the intelligence that your business  runs on. You can literally build it [18:28] in hours if not in minutes and then let it  compound for you into your asset forever. [18:34] Uh we have I have a link here for uh a webinar  that we're doing here exactly next week. [18:39] Uh and also oh you can see it over here,  but uh we are also at booth 1341. So, [18:46] if you want, please drop by. We can show you a  demo and you can pick up some swag. Thank you all.