Higgsfield Shifts Video Generation to Agents That Provision Own GPUs

Higgsfield - Building Agents That Provision Their Own Compute | DevDay 2026 — OpenAI

OpenAI

Producing coherent long-form video requires deterministic agent loops that self-evaluate outputs and dynamically spin up dedicated GPU infrastructure rather than relying on static hosted models.

AI-assisted summary of Higgsfield - Building Agents That Provision Their Own Compute | DevDay 2026 — original source. Check the original for full context.

Source published:

Key ideas

  • Higgsfield engineered an agentic system that dynamically provisions dedicated GPUs on demand to handle fine-tuned models and intensive rendering pipelines, rather than relying solely on fixed API endpoints. (0:00)
  • Achieving consistent 10-minute automated video generation required over $30,000 in experimental R&D to build deterministic execution pipelines with autonomous self-evaluation loops. (0:00)
  • During a 10-hour live broadcast in New York City with streamer N3on, foundation models performed continuous, photorealistic real-time world stylization and reconstruction. (5:00)
  • Higgsfield open-sourced its internal automated video editing skill, which was built using Astra to create product tutorials and can now be adopted by external developers. (10:00)

Detailed notes

## Background Published by OpenAI on YouTube on October 7, 2026, this DevDay session explores how Higgsfield develops AI agents that select models and provision their own compute infrastructure. According to the description, the presentation details when to deploy dedicated GPUs versus hosted models, methods for achieving predictable throughput, and key failure modes in the pipeline.

## TL;DR Higgsfield co-founder and CEO Alex Mashrabov details the architectural shift from generating isolated assets to orchestrating coherent, multi-minute interactive media through autonomous agent pipelines. To bypass the heavy tool fragmentation and high production costs of traditional studios, Higgsfield employs planning and self-evaluating agents that route tasks to hosted API models or dynamically spin up dedicated GPUs for fine-tuned rendering. The talk highlights practical applications, including a 10-hour continuous real-time AI live stream in New York City, secure enterprise perimeter infrastructure, and open-source recursive video editing skills powered by foundation models like Astra.

## Atmosphere & style Delivered from the outdoor stage at DevDay, the talk carries a fast-paced, pragmatic, and inspiring startup energy. Mashrabov pairs technical system design with candid development anecdotes—such as spending over $30,000 to test long-form video agent loops—grounding futuristic concepts in concrete engineering practice. The conversation feels buoyant and forward-looking, evoking an empowering sense that the technological barriers to studio-scale media production have permanently dissolved for general developers.

## Bulletin - [00:00](https://www.youtube.com/watch?v=8804JYm81cc&t=0s) Higgsfield engineered an agentic system that dynamically provisions dedicated GPUs on demand to handle fine-tuned models and intensive rendering pipelines, rather than relying solely on fixed API endpoints. - [00:00](https://www.youtube.com/watch?v=8804JYm81cc&t=0s) Achieving consistent 10-minute automated video generation required over $30,000 in experimental R&D to build deterministic execution pipelines with autonomous self-evaluation loops. - [05:00](https://www.youtube.com/watch?v=8804JYm81cc&t=300s) During a 10-hour live broadcast in New York City with streamer N3on, foundation models performed continuous, photorealistic real-time world stylization and reconstruction. - [10:00](https://www.youtube.com/watch?v=8804JYm81cc&t=600s) Higgsfield open-sourced its internal automated video editing skill, which was built using Astra to create product tutorials and can now be adopted by external developers.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=8804JYm81cc&t=0s) Higgsfield operates as an end-to-end multimedia platform combining third-party foundation models like Astra with proprietary image and video models. - [00:00](https://www.youtube.com/watch?v=8804JYm81cc&t=0s) While single-image generation is easily commoditized, generating coherent, multi-minute video experiences requires complex deterministic planning and self-evaluating agent architectures. - [00:00](https://www.youtube.com/watch?v=8804JYm81cc&t=0s) Developing an automated video editing skill capable of producing consistent 10-minute outputs required an experimental R&D investment exceeding $30,000. - [00:00](https://www.youtube.com/watch?v=8804JYm81cc&t=0s) The agentic architecture dynamically switches between hosted API endpoints for standard tasks and provisioning dedicated GPUs for heavy, fine-tuned video models. - [05:00](https://www.youtube.com/watch?v=8804JYm81cc&t=300s) Agentic computer use compresses complex creative toolchains, replacing workflows that traditionally required 15 specialists operating across 10 to 12 software suites. - [05:00](https://www.youtube.com/watch?v=8804JYm81cc&t=300s) A live 10-hour interactive street broadcast with streamer N3on proved the feasibility of real-time, photorealistic world reconstruction and stylization. - [05:00](https://www.youtube.com/watch?v=8804JYm81cc&t=300s) Small engineering teams of three to five people are now able to ship full-scale interactive media experiences and mobile applications without studio-level resources. - [05:00](https://www.youtube.com/watch?v=8804JYm81cc&t=300s) Enterprise rollouts require strict guardrails and isolated perimeter GPU infrastructure to guarantee compliance, low latency, and deterministic output quality. - [10:00](https://www.youtube.com/watch?v=8804JYm81cc&t=600s) Advances in model reasoning and planning are steadily eliminating post-production and rendering bottlenecks, fundamentally democratizing media creation. - [10:00](https://www.youtube.com/watch?v=8804JYm81cc&t=600s) Higgsfield developed its automated video tutorial pipeline recursively: Astra was used to build the production skill, which then produced tutorials teaching developers how to use Higgsfield. - [10:00](https://www.youtube.com/watch?v=8804JYm81cc&t=600s) The entire Higgsfield platform, spanning autonomous GPU provisioning, continuous video generation, and public APIs, was built within an 18-month timeframe.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=8804JYm81cc&t=0s) Introducing Higgsfield and Autonomous Compute Provisioning

The session opens on the DevDay outdoor stage with an introduction of Alex Mashrabov, co-founder and CEO of Higgsfield. Mashrabov previously co-founded the startup that became Snapchat Cameos and later directed generative AI efforts at Snap before founding Higgsfield. Today, Higgsfield operates as an end-to-end multimedia AI platform where creative marketers construct full advertising campaigns using foundation models alongside Higgsfield’s proprietary image and video models. Beyond enterprise marketing, the team observed an unmet demand for interactive and long-form creative media, prompting the release of the Higgsfield API. Mashrabov highlights that coupling modern foundation models like Astra with the Higgsfield API allows a solitary developer to produce an entire interactive show—a feat that previously required large studio operations.

Moving beyond static image generation, Mashrabov explains that while getting a compelling single asset or concept image is relatively easy, orchestrating a coherent, continuous multi-minute experience is exceptionally difficult. To solve this, Higgsfield engineered an agentic "supercomputer" system equipped with self-evaluation mechanisms that decide which model fits a given sub-task and whether an output meets production thresholds. Mashrabov recalls spending over $30,000 on an experimental video editing agent designed to generate consistent 10-minute videos. The core technical hurdle was shifting from pure creative generation to deterministic execution pipelines: the agent starts from an abstract concept, devises an execution plan, self-evaluates candidate outputs against strict quality criteria, and dynamically spins up dedicated GPU infrastructure on demand to run fine-tuned rendering and video models.

```mermaid flowchart TD accTitle: Higgsfield Agentic Compute Pipeline accDescr: Flowchart illustrating how the Higgsfield agent moves from an initial creative concept to execution planning, dynamic model selection, GPU provisioning, and self-evaluation.

subgraph Planning ["1. Planning & Decisioning"] A["Creative Concept"] --> B["Execution Plan"] B --> C{"Model Selector Agent"} end

subgraph Compute ["2. Dynamic Compute Engine"] C -->|"Standard Task"| D["Hosted API Models"] C -->|"Heavy / Fine-Tuned Video"| E["Spin Up Dedicated GPUs"] end

subgraph Quality ["3. Evaluation & Output"] D --> F["Autonomous Self-Evaluation"] E --> F F -->|"Meets Quality Criteria"| G["Continuous Final Output"] F -->|"Fails Quality Gate"| B end ```

- **Creative Execution Overhead:** Single-image generation is easily commoditized, but generating consistent long-form video requires deterministic planning and rigorous self-evaluating agent loops. - **R&D Scale:** Proving the capability of a fully automated video editing skill capable of generating 10-minute coherent videos required an experimental investment exceeding $30,000. - **Infrastructure Autonomy:** Higgsfield's agent does not merely query fixed endpoints; it directly provisions dedicated GPU instances when running fine-tuned models and intensive rendering pipelines.

---

[05:00](https://www.youtube.com/watch?v=8804JYm81cc&t=300s) Computer Use, Real-Time Streams, and Enterprise Guardrails

Mashrabov delves deeper into how models like Astra bridge the divide between creative ambition and practical execution by handling computer use workflows. In traditional media and gaming pipelines, realizing a high-fidelity visual concept exacts a punishing "production tax": studios rely on teams of 15 or more specialists, each operating across 10 to 12 separate software suites. By acting across applications on the creator's behalf, agentic computer use condenses those disparate tool chains into unified workflows accessible to small teams. This capability eliminates the barrier where visual ambition is capped solely by production budgets.

To illustrate real-time performance and scalability, Mashrabov details an experimental production with the popular live streamer N3on and Higgsfield spokesperson Adil. Walking through the streets of New York City and engaging with bystanders, the pair conducted an interactive AI stream that ran continuously for over 10 hours. Astra reconstructed and stylized the surrounding world live with photorealistic consistency throughout the broadcast. Beyond direct-to-consumer and mobile app experiences created by small three- to five-person teams, Mashrabov points out that larger enterprise adoptions demand strict control. For enterprise deployments, the agent is enclosed within explicit guardrails and runs across a secure perimeter GPU infrastructure, ensuring predictable latency, compliance, and deterministic output quality.

- **Eliminating Tool Fatigue:** Replacing manual workflows across 10 to 12 distinct creative software packages by delegating tool invocation directly to agentic models. - **The 10-Hour Live AI Stream:** Creator N3on and Adil broadcasted for over 10 uninterrupted hours in New York City, demonstrating continuous, real-time world reconstruction. - **Small-Team Leverage:** Teams of just 3 to 5 engineers are shipping end-to-end mobile apps and interactive web experiences without conventional studio infrastructure. - **Enterprise Perimeter Security:** Implementing strict operational guardrails and dedicated perimeter GPU instances to provide the reliability and predictability required for corporate go-to-market rollouts.

---

[10:00](https://www.youtube.com/watch?v=8804JYm81cc&t=600s) Skating to Where the Puck Is Going and the Democratization of Media

Reflecting on the pace of progress, the interviewer draws a hockey metaphor, emphasizing that Higgsfield succeeds by "skating to where the puck is going"—building architectures for capabilities that foundation models are just beginning to support. Mashrabov agrees, noting that rapid jumps in reasoning, planning, and decision-making capabilities are steadily dissolving the technical bottlenecks that historically limited media production. The barrier is no longer access to rendering farms or specialized post-production teams; interactive media creation has become fundamentally democratized for general software developers.

To accelerate this ecosystem, Higgsfield has open-sourced internal agent skills, such as an automated AI video editing skill engineered to produce explanatory YouTube content. In a self-referential twist, Mashrabov explains that Higgsfield utilized Astra to construct an automated video production skill, deployed that exact skill internally to generate educational tutorials teaching users how to build with Higgsfield, and now sees external developers adopting the same open-source workflow to generate automated video explainers for their own products. Wrapping up the session, Mashrabov encourages the audience to take advantage of this moment, reminding them: "In just 18 months, we built Higgsfield, and now is the best time to build yours."

- **Recursive Skill Development:** Higgsfield leveraged Astra to build an automated explainer video skill, used it to make videos explaining Higgsfield products, and open-sourced it for the broader developer community. - **Rapid Velocity:** The entire Higgsfield platform—spanning the dynamic GPU provisioning engine, long-form continuous video capabilities, and public APIs—was developed in only 18 months.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=8804JYm81cc&t=0s) coupling modern foundation models like Astra with the Higgsfield API allows a solitary developer to produce an entire interactive show—a feat that previously required large studio operations — Alex Mashrabov

> [00:00](https://www.youtube.com/watch?v=8804JYm81cc&t=0s) while getting a compelling single asset or concept image is relatively easy, orchestrating a coherent, continuous multi-minute experience is exceptionally difficult — Alex Mashrabov

> [00:00](https://www.youtube.com/watch?v=8804JYm81cc&t=0s) spending over $30,000 on an experimental video editing agent designed to generate consistent 10-minute videos — Alex Mashrabov

> [05:00](https://www.youtube.com/watch?v=8804JYm81cc&t=300s) studios rely on teams of 15 or more specialists, each operating across 10 to 12 separate software suites — Alex Mashrabov

> [05:00](https://www.youtube.com/watch?v=8804JYm81cc&t=300s) By acting across applications on the creator's behalf, agentic computer use condenses those disparate tool chains into unified workflows accessible to small teams — Alex Mashrabov

> [10:00](https://www.youtube.com/watch?v=8804JYm81cc&t=600s) skating to where the puck is going

> [10:00](https://www.youtube.com/watch?v=8804JYm81cc&t=600s) rapid jumps in reasoning, planning, and decision-making capabilities are steadily dissolving the technical bottlenecks that historically limited media production — Alex Mashrabov

> [10:00](https://www.youtube.com/watch?v=8804JYm81cc&t=600s) In just 18 months, we built Higgsfield, and now is the best time to build yours. — Alex Mashrabov

## Mentioned Alex Mashrabov, Higgsfield, Snapchat Cameos, Snap, Astra, Higgsfield API, N3on, Adil, New York City, YouTube, DevDay

## Video description Explore how Higgsfield agents choose models and provision their own compute. Learn when to use dedicated GPUs or hosted models, how to deliver predictable throughput, and where the approach can fail.

[Source](https://www.youtube.com/watch?v=8804JYm81cc)

中文简报

Higgsfield将视频生成转向可自主调配GPU的Agent架构 制作连贯的长视频需要能够自我评估输出并动态启动专用GPU基础设施的确定性Agent循环，而非仅依赖静态托管模型。 Higgsfield联合创始人兼首席执行官Alex Mashrabov详细阐述了一项架构转变：从生成孤立的素材，转向通过自主Agent管线编排连贯、长达数分钟的交互式媒体。为了规避传统工作室严重的工具碎片化问题和高昂的制作成本，Higgsfield采用具备规划和自我评估能力的Agent，将任务路由至托管API模型，或动态启动专用GPU以进行微调渲染。本次演讲重点介绍了实际应用场景，包括在纽约市进行的10小时连续实时AI直播、安全的企业边界基础设施，以及由Astra等基础模型驱动的开源递归视频编辑技能。 Higgsfield设计了一套Agent系统，可根据需求动态调配专用GPU以处理微调模型和密集渲染管线，而非仅仅依赖固定的API端点。 (0:00) 为了构建具备自主自我评估循环的确定性执行管线，实现连贯一致的10分钟自动化视频生成，研发实验投入超过了30,000美元。 (0:00) 在纽约市与主播N3on进行的10小时实况直播中，基础模型实现了连续、逼真的实时世界风格化与重建。 (5:00) Higgsfield开源了其内部的自动化视频编辑技能，该技能基于Astra构建，最初用于制作产品教程，现可供外部开发者采用。 (10:00)

Reading summary

Alex Mashrabov, CEO of Higgsfield, argues that producing coherent long-form AI media requires shifting from static asset generation to deterministic agentic systems that autonomously evaluate output quality and spin up dedicated GPU infrastructure on demand. Autonomous Infrastructure Provisioning and Quality Evaluation (0:00) Mashrabov notes that while generating single concept images is relatively straightforward, orchestrating continuous, multi-minute video experiences presents severe execution challenges. To overcome this, Higgsfield engineered an agent pipeline that devises an execution plan from an abstract concept, selects task-appropriate models, and dynamically provisions dedicated GPU instances for fine-tuned video models and rendering tasks instead of relying solely on fixed API endpoints. The pipeline incorporates an autonomous self-evaluation mechanism that assesses candidate outputs against quality criteria, re-routing failed attempts back into execution planning. Mashrabov reports that validating this deterministic workflow required an experimental research investment exceeding $30,000 to construct an agent capable of generating consistent 10-minute videos. Collapsing Studio Tool Chains and Real-Time Streaming (5:00) Conventional production pipelines impose a heavy operational burden, frequently requiring teams of 15 or more specialists coordinating across 10 to 12 separate creative software applications. Mashrabov explains that delegating tool execution to foundation models like Astra condenses these disjointed workflows, enabling teams of three to five engineers to build and ship interactive web and mobile media experiences without traditional studio infrastructure. To test continuous real-time execution and world reconstruction, content creator N3on and Higgsfield spokesperson Adil broadcasted a live, interactive AI stream through New York City for over 10 consecutive hours. Throughout the broadcast, Astra processed the live video feed to continuously stylize and reconstruct the surrounding environment with photorealistic consistency. Enterprise Perimeter Controls and Recursive Development (10:00) For enterprise deployments, Higgsfield implements explicit operational guardrails and runs workloads across secure perimeter GPU infrastructure, which Mashrabov states is necessary to guarantee compliance, predictable latency, and deterministic output quality for corporate rollouts. To expand developer adoption, Higgsfield open-sourced internal production workflows using a recursive development cycle: the team utilized Astra to engineer an automated video production skill, deployed that exact skill internally to generate educational tutorials explaining Higgsfield products, and released the workflow for external developers to produce their own product explainers. Mashrabov notes that the entire platform—encompassing the dynamic compute engine, public APIs, and video generation—was developed over an 18-month timeframe. Mashrabov concludes that advancing reasoning and planning capabilities in foundation models have removed access to specialized rendering farms and post-production studios as the primary bottleneck in media creation, making end-to-end interactive media development viable for individual software engineers.

Detailed notes — zh

## Background 该视频由 OpenAI 于 2026 年 10 月 9 日在 YouTube 官方频道发布，记录了 DevDay 2026 现场的精彩对谈。根据官方简介，本期内容聚焦于 Higgsfield 的智能体如何自主选择模型并调配自有算力，探讨何时使用专属 GPU 或托管模型、如何保障可预测的吞吐量，以及该方案在实际应用中可能面临的挑战与边界。 ## TL;DR Higgsfield 联合创始人兼 CEO Alex Mashrabov 详细介绍了他们如何构建能够自主申请并调配专属 GPU 基础设施的“超级计算机”智能体系统。通过结合 OpenAI Astra 的计算机使用（computer use）能力与确定性渲染工作流，团队成功消除了传统创意产业中沉重的“制作税”，使几人规模的微型团队也能完成过去需十余人协作的复杂媒体项目。分享中还着重展示了与主播 N3on 合作完成的长达 10 小时纽约街头实时世界重构 AI 直播，并剖析了企业级可控性与安全护栏设计。最后，Alex 分享了团队开源的 AI 视频编辑技能，激励现场开发者抓住模型能力高速跃迁的历史机遇打造自己的产品。 ## Atmosphere & style 对话在 OpenAI DevDay 的户外舞台上展开，兼具前沿技术的硬核务实与硅谷开发者大会特有的活跃开放氛围。主持人的提问敏锐且极具共鸣，而 Alex 的阐述则条理清晰、充满创业者的热忱与笃定，在坦诚透露数万美元研发成本等工程现实的同时，又毫无保留地描绘了互动媒体的未来图景。整场交流节奏紧凑、干货密集，尤其是披露 10 小时全实时 AI 直播的瞬间，为现场带来了强烈的技术震撼与极具感染力的行动号召感。 ## Bulletin - [00:00](https://www.youtube.com/watch?v=8804JYm81cc&t=0s) Higgsfield 推出多媒体智能体平台与 Higgsfield API，结合 OpenAI Astra 等模型，使单人或微型团队即可制作出以往需要 15 人以上团队交付的广告营销与互动媒体内容。 - [02:15](https://www.youtube.com/watch?v=8804JYm81cc&t=135s) 针对长视频一致性难题，Higgsfield 构建了能够自主选择模型并启动专属 GPU 基础设施的“超级计算机”智能体系统，支持执行确定性渲染工作流与输出质量自我评估。 - [05:00](https://www.youtube.com/watch?v=8804JYm81cc&t=300s) Higgsfield 联手知名主播 N3on 在纽约街头完成了持续超过 10 小时的实时 AI 世界重构直播，验证了 Astra 驱动下超低延迟创意媒体工作流的可行性。 - [08:30](https://www.youtube.com/watch?v=8804JYm81cc&t=510s) 为满足企业级客户对可控性与安全性的要求，Higgsfield 建立了包含 GPU 边界基础设施的护栏机制，并开源了基于 Astra 构建的自动化 AI 视频编辑技能。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=8804JYm81cc&t=0s) Higgsfield CEO Alex Mashrabov 曾创立 Snapchat Cameos 并在 Snap 负责生成式 AI，现致力于通过智能体自主调配模型与 GPU 算力重塑多媒体制作体验。 - [00:00](https://www.youtube.com/watch?v=8804JYm81cc&t=0s) Higgsfield API 结合 OpenAI Astra 等基础模型，彻底降低了创意门槛，使个人创作者能够端到端交付互动节目与商业化营销内容。 - [02:15](https://www.youtube.com/watch?v=8804JYm81cc&t=135s) 生成单一概念资产相对容易，但要交付连贯、长周期的媒体体验极难；团队曾花费超过 30,000 美元才摸索出一套可稳定输出 10 分钟长视频的编辑技能。 - [02:15](https://www.youtube.com/watch?v=8804JYm81cc&t=135s) 为将生成式视频投入生产环境，系统必须进化为能够进行严密自我评估并依据质量标准自主决策调用何种模型的智能体系统。 - [02:15](https://www.youtube.com/watch?v=8804JYm81cc&t=135s) Higgsfield 的“超级计算机”架构赋予了智能体按需调配和启动专属 GPU 的能力，以保障高级渲染等确定性工作流的执行效果与一致性。 - [05:00](https://www.youtube.com/watch?v=8804JYm81cc&t=300s) 传统制作流程存在高昂的“制作税”，往往需要 15 人以上掌握 10 至 12 种不同工具的专业团队；Astra 的“计算机使用”与子代理能力大幅压缩了这一协作成本。 - [05:00](https://www.youtube.com/watch?v=8804JYm81cc&t=300s) Higgsfield 与主播 N3on 合作完成了持续逾 10 小时的纽约街头实时世界重构直播，标志着超高速 AI 视频模型已迈入实时互动阶段。 - [08:30](https://www.youtube.com/watch?v=8804JYm81cc&t=510s) 在面向大型企业的落地场景中，“可控性”是核心关切，Higgsfield 通过安全护栏机制与带有 GPU 的边界基础设施保障企业级部署合规与稳定。 - [08:30](https://www.youtube.com/watch?v=8804JYm81cc&t=510s) Alex 建议开发者“滑向冰球即将到达的位置”，紧随模型决策与执行能力的快速迭代，提前布局下一代互动媒体应用。 - [08:30](https://www.youtube.com/watch?v=8804JYm81cc&t=510s) Higgsfield 开源了内部基于 Astra 开发的 AI 视频编辑技能，形成了团队用其制作产品教程、外部开发者同样用其为各自产品制作视频的生态自举循环。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=8804JYm81cc&t=0s) Higgsfield 平台与 Higgsfield API 介绍 在 OpenAI DevDay 的户外舞台上，主持人热情欢迎了 Higgsfield 的联合创始人兼 CEO Alex Mashrabov。Alex 此前曾联合创立了后来演变为 Snapchat Cameos 的产品，并随后在 Snap 领导生成式 AI 团队。如今，他正致力于利用智能体（Agents）自主选择模型并启动新的 GPU，从而解锁创建和体验媒体的全新方式。 Alex 介绍道，Higgsfield 是一个领先的多媒体 AI 平台，创意营销人员可以在上面结合 OpenAI Astra 等现代模型以及 Higgsfield 提供的高级图像和视频模型，来制作广告营销活动。更重要的是，这些由 AI 生成的广告能够实实在在地帮助企业实现营收和销售额的增长。与此同时，团队注意到许多人在构建前所未知的新型创意媒体和互动媒体方面有着强烈的需求与向往。为此，他们推出了 Higgsfield API，帮助开发者交付这些创新体验并建立新的业务。通过将 OpenAI Astra 等模型与 Higgsfield API 相结合，过去难以想象的工作流得以实现——如今仅仅一个人的团队就能制作并发布一档互动节目。 ### [02:15](https://www.youtube.com/watch?v=8804JYm81cc&t=135s) 构建能够自主调配 GPU 的“超级计算机”智能体 当被问及是什么促使团队构建这套智能体系统时，Alex 指出，在行业中很容易获得像概念图这样的单一优质资产，但要围绕它交付连贯、持续的体验却非常困难。因此，Higgsfield 推出了“超级计算机”（Supercomputer）。他提到，最近在另一个播客中他曾分享过，团队利用这套系统开发了一项全新的视频编辑技能，能够保持一致性地制作长达 10 分钟的视频，这在没有 Astra 模型的情况下是根本无法想象的。然而，为了实现这一个体验，他们花费了超过 30,000 美元。如果要将此类技能投入生产环境，系统必须针对不同用例就“使用哪个模型”做出大量决策，这也要求智能体系统（Agentic systems）具备极强的自我评估能力，以便定义什么样的输出质量标准是可以接受的。 针对主持人关于“从创意火花到最终成品中间经历了什么”的提问，Alex 强调了基础模型（从早期的模型演进到 5.5 版本乃至 Astra 模型）在创造力方面的独特之处。一切虽然始于想法和概念，但将其转化为非常具体的执行计划并明确定义最终产出标准至关重要。在此过程中，Higgsfield 发现拥有确定性的工作流（deterministic workflows）非常有价值，这些工作流可能包含使用视频模型或微调视频模型进行高级渲染。这正是他们将其称为“超级计算机”的原因：他们必须构建一个能够自行启动专属 GPU 基础设施的智能体，这对于在各种用例中交付高度一致的执行结果起到了决定性作用。 ```mermaid flowchart LR accTitle: Higgsfield 智能体从创意到成品的执行工作流 accDescr: 展示从最初的创意火花到通过 Astra 制定计划、启动专属 GPU 并执行确定性渲染工作流以输出最终产品的过程。 subgraph Plan["1. 创意与规划"] A["创意火花与概念"] --> B["通过 Astra 制定执行计划与质量标准"] end subgraph Infra["2. 基础设施与执行"] B --> C["智能体按需启动专属 GPU 基础设施"] C --> D["执行确定性工作流 (微调视频模型渲染)"] end subgraph Eval["3. 评估与交付"] D --> E["智能体自我评估输出质量"] E --> F["一致性的长视频或互动媒体成品"] end ``` ### [05:00](https://www.youtube.com/watch?v=8804JYm81cc&t=300s) 利用 Astra 消除“制作税”与 10 小时实时 AI 直播案例 在探讨 OpenAI 模型在技术栈中的其他作用时，Alex 表示 Astra 代表了真正解锁“计算机使用”（computer use）能力的下一代模型。传统的创意工作流通常需要一个超过 15 人的团队，并且团队成员必须熟练掌握 10 到 12 种不同的工具。这在创意想法与最终执行之间形成了巨大的“制作税”（production tax）——这种现象在游戏行业尤为明显，视觉质量门槛的提高直接导致预算呈指数级膨胀。而 Astra 能够通过子代理（sub-agents）等机制代表创作者调用这 10 到 12 个程序，从而让小团队也能取得巨大的成功。 谈及构建过程中的重大技术决策与权衡，Alex 回顾了连续视频生成和自动视频编辑的用例。为了实现实时体验，团队必须全力优化运行速度。虽然在活动当天刚刚发布了超高速模型（ultra-fast model）让实时体验变得更加触手可及，但在那之前，Higgsfield 就已经与极具人气的直播主 N3on 合作，解锁了史上首场 AI 直播。在这场直播中，两位真实的创作者——Higgsfield 发言人 Adil 和主播 N3on——走在纽约市的街头与路人互动，而 Astra 模型则实时重构了整个世界，使其看起来逼真自然。当主持人惊讶地确认这是否真的是实时直播时，Alex 证实这场直播持续了超过 10 个小时。这种全新的媒体形式在过去根本无法想象，而现在全球任何地方仅由两三个人组成的开发者团队，都能交付这种体验并在其之上建立业务。 - **传统创意团队规模**：超过 15 人，需精通 10–12 种软件工具。 - **使用 Higgsfield + Astra 后的团队规模**：仅需 1–5 人的微型团队即可端到端交付产品。 - **标杆案例**：与主播 N3on 及发言人 Adil 在纽约街头完成超过 10 小时的实时世界重构 AI 直播。 ### [08:30](https://www.youtube.com/watch?v=8804JYm81cc&t=510s) 企业级可控性、围栏机制与开源视频编辑技能 面对台下的开发者听众，Alex 分享了更多令人振奋的应用场景。目前已有许多三到五人的团队正在端到端地交付直接面向消费者（D2C）的移动应用和网站。但当视角转向更大规模的企业级用例时，“可控性”（controllability）变得至关重要。为此，Higgsfield 确保智能体不仅围绕自身建立明确的护栏（guardrails），还能通过带有 GPU 的边界基础设施（perimeter infrastructure）提供安全部署。这极大地加速了 Higgsfield 自身的企业级市场推广（go-to-market），并且现在这些能力都已通过 Higgsfield API 向其他开发者开放。 主持人用冰球运动作比喻，称赞 Higgsfield 总是“滑向冰球即将到达的位置”（skate to where the puck is going），在模型尚未完全成熟时就提前布局，从而实现从 2 小时到 10 小时直播的跨越。对此，Alex 建议开发者要为下一次 AI 跃迁做好准备，因为模型迭代极快，甚至当天早些时候就已经看到了关于决策能力（decisioning）等方面的新特性。如今，互动媒体创作正迎来前所未有的民主化，过去受限于技术瓶颈的想法在 Astra 的加持下终于成为可能。 最后，Alex 兴奋地分享了一个有趣的“奇点”（singularity）现象：Higgsfield 开源了他们自己开发的新技能，例如用于制作 YouTube AI 解说视频的“AI 视频编辑技能”。他们在内部利用 Astra 帮助构建了这项技能，并用它制作教学视频来教大家如何使用 Higgsfield 产品；而现在，外部开发者们也开始使用同一个开源技能来为他们自己的产品制作教程视频。在演讲结束之际，Alex 用一句话鼓励全场开发者：“我们仅用 18 个月就建成了 Higgsfield，而现在正是你们构建自己产品的最佳时机。” ## Notable quotes > [00:00](https://www.youtube.com/watch?v=8804JYm81cc&t=0s) 通过将 OpenAI Astra 等模型与 Higgsfield API 相结合，过去难以想象的工作流得以实现——如今仅仅一个人的团队就能制作并发布一档互动节目。 — Alex Mashrabov > [02:15](https://www.youtube.com/watch?v=8804JYm81cc&t=135s) 在行业中很容易获得像概念图这样的单一优质资产，但要围绕它交付连贯、持续的体验却非常困难。 — Alex Mashrabov > [02:15](https://www.youtube.com/watch?v=8804JYm81cc&t=135s) 团队利用这套系统开发了一项全新的视频编辑技能，能够保持一致性地制作长达 10 分钟的视频，这在没有 Astra 模型的情况下是根本无法想象的。然而，为了实现这一个体验，他们花费了超过 30,000 美元。 — Alex Mashrabov > [02:15](https://www.youtube.com/watch?v=8804JYm81cc&t=135s) 这正是他们将其称为“超级计算机”的原因：他们必须构建一个能够自行启动专属 GPU 基础设施的智能体，这对于在各种用例中交付高度一致的执行结果起到了决定性作用。 — Alex Mashrabov > [05:00](https://www.youtube.com/watch?v=8804JYm81cc&t=300s) 传统的创意工作流通常需要一个超过 15 人的团队，并且团队成员必须熟练掌握 10 到 12 种不同的工具。这在创意想法与最终执行之间形成了巨大的“制作税”（production tax） — Alex Mashrabov > [05:00](https://www.youtube.com/watch?v=8804JYm81cc&t=300s) 两位真实的创作者——Higgsfield 发言人 Adil 和主播 N3on——走在纽约市的街头与路人互动，而 Astra 模型则实时重构了整个世界，使其看起来逼真自然。 — Alex Mashrabov > [08:30](https://www.youtube.com/watch?v=8804JYm81cc&t=510s) 当视角转向更大规模的企业级用例时，“可控性”（controllability）变得至关重要。为此，Higgsfield 确保智能体不仅围绕自身建立明确的护栏（guardrails），还能通过带有 GPU 的边界基础设施（perimeter infrastructure）提供安全部署。 — Alex Mashrabov > [08:30](https://www.youtube.com/watch?v=8804JYm81cc&t=510s) 滑向冰球即将到达的位置 — 主持人 > [08:30](https://www.youtube.com/watch?v=8804JYm81cc&t=510s) 他们在内部利用 Astra 帮助构建了这项技能，并用它制作教学视频来教大家如何使用 Higgsfield 产品；而现在，外部开发者们也开始使用同一个开源技能来为他们自己的产品制作教程视频。 — Alex Mashrabov > [08:30](https://www.youtube.com/watch?v=8804JYm81cc&t=510s) 我们仅用 18 个月就建成了 Higgsfield，而现在正是你们构建自己产品的最佳时机。 — Alex Mashrabov ## Mentioned Higgsfield, Higgsfield API, OpenAI, OpenAI DevDay, Alex Mashrabov, Snapchat Cameos, Snap, OpenAI Astra, Supercomputer（超级计算机）, N3on, Adil, 纽约市, YouTube ## Video description Explore how Higgsfield agents choose models and provision their own compute. Learn when to use dedicated GPUs or hosted models, how to deliver predictable throughput, and where the approach can fail. [Source](https://www.youtube.com/watch?v=8804JYm81cc)

Source quotes

[0:00] Alex Mashrabov: 通过将 OpenAI Astra 等模型与 Higgsfield API 相结合，过去难以想象的工作流得以实现——如今仅仅一个人的团队就能制作并发布一档互动节目。 [2:15] Alex Mashrabov: 在行业中很容易获得像概念图这样的单一优质资产，但要围绕它交付连贯、持续的体验却非常困难。 [2:15] Alex Mashrabov: 团队利用这套系统开发了一项全新的视频编辑技能，能够保持一致性地制作长达 10 分钟的视频，这在没有 Astra 模型的情况下是根本无法想象的。然而，为了实现这一个体验，他们花费了超过 30,000 美元。 [2:15] Alex Mashrabov: 这正是他们将其称为“超级计算机”的原因：他们必须构建一个能够自行启动专属 GPU 基础设施的智能体，这对于在各种用例中交付高度一致的执行结果起到了决定性作用。 [5:00] Alex Mashrabov: 传统的创意工作流通常需要一个超过 15 人的团队，并且团队成员必须熟练掌握 10 到 12 种不同的工具。这在创意想法与最终执行之间形成了巨大的“制作税”（production tax） [5:00] Alex Mashrabov: 两位真实的创作者——Higgsfield 发言人 Adil 和主播 N3on——走在纽约市的街头与路人互动，而 Astra 模型则实时重构了整个世界，使其看起来逼真自然。 [8:30] Alex Mashrabov: 当视角转向更大规模的企业级用例时，“可控性”（controllability）变得至关重要。为此，Higgsfield 确保智能体不仅围绕自身建立明确的护栏（guardrails），还能通过带有 GPU 的边界基础设施（perimeter infrastructure）提供安全部署。 [8:30] 主持人: 滑向冰球即将到达的位置 [8:30] Alex Mashrabov: 他们在内部利用 Astra 帮助构建了这项技能，并用它制作教学视频来教大家如何使用 Higgsfield 产品；而现在，外部开发者们也开始使用同一个开源技能来为他们自己的产品制作教程视频。 [8:30] Alex Mashrabov: 我们仅用 18 个月就建成了 Higgsfield，而现在正是你们构建自己产品的最佳时机。

Source quotes

[0:00] Alex Mashrabov: coupling modern foundation models like Astra with the Higgsfield API allows a solitary developer to produce an entire interactive show—a feat that previously required large studio operations [0:00] Alex Mashrabov: while getting a compelling single asset or concept image is relatively easy, orchestrating a coherent, continuous multi-minute experience is exceptionally difficult [0:00] Alex Mashrabov: spending over $30,000 on an experimental video editing agent designed to generate consistent 10-minute videos [5:00] Alex Mashrabov: studios rely on teams of 15 or more specialists, each operating across 10 to 12 separate software suites [5:00] Alex Mashrabov: By acting across applications on the creator's behalf, agentic computer use condenses those disparate tool chains into unified workflows accessible to small teams [10:00] skating to where the puck is going [10:00] Alex Mashrabov: rapid jumps in reasoning, planning, and decision-making capabilities are steadily dissolving the technical bottlenecks that historically limited media production [10:00] Alex Mashrabov: In just 18 months, we built Higgsfield, and now is the best time to build yours.

Original transcript

[0:00] *music* Thank you. All right. Thank you. All right, everybody, welcome. Welcome to the outdoor stage. All right, thank you. Today, our next guest coming up here, I'm so excited to have you hear from him. He co-founded a product... that became Snapchat cameos and then went on later to lead generative AI at Snap. But now what he's doing is he's using agents to choose models and spin up new gpus that unlocks new ways to create and experience media so i would love a warm dev day welcome for the co-founder and ceo of higgs field alex mashubov Thank you. Thanks for being here, Alex. Thank you very much. Alex, for the people, I'm a huge fan, but I'm not sure everybody knows. How do you describe Higgs Field these days and what are people building with it? So Higgsfield is a leading multimedia AI platform where creative marketers can build their advertising campaigns and so on using modern models like Astra, but also advanced image and video models provided by Higgsfield. Most importantly, these AI-generated ads become a way how actually businesses can grow their revenue and their sales. But also we felt that there are some unmet needs, as a lot of people have admiration, around building new forms of creative media, interactive media, which were not possible before. That's why we launched Higgs field API. which can actually help developers to ship those experiences and build new businesses. And that's where we have seen that models like Higgs failed. models like OpenAI Astra in conjunction with Higgs field API can deliver those workflows where like essentially a team of just one person can ship an interactive show, which would be impossible to even imagine a few years ago. Yeah, yeah. And so you went through and you built this agent, right? You built an agent. What caused you to have to think through that? Thank you. - Absolutely. What we have seen very often is that it's very easy to get a good asset like concept image but very difficult to deliver a continuous experience around that. Thank you. And that's why we shipped supercomputer. Recently on another podcast, I actually mentioned that we were able to deliver absolutely new skill, like, for example, video editing skill which can go and make 10 minute long video consistently. That would be impossible to imagine without Astra model. But we spent... over $30,000 to deliver this experience. And you can imagine that it takes... that's for-- to put this for production purposes, such skills, there are a lot of decisions which need to be made on which model to use and which use case. And that's also where a genetic systems have to be really good at self-evaluation. to define a specific criteria which quality of the output is acceptable. So it's choosing those models as it goes. What goes through that? What happens when you start? You have this creative spark. Let's go high level. Creative Spark, then it comes out of the end product. What is that? Yeah, so that's where I think also that's a really good question, because I think One of the areas where the Sol model and the Astra model are truly unique, and it starts probably from Sol 5.5, is creativity. Like, obviously it all starts with idea and a concept, but then it's very important to bring this concept to build a very specific plan of execution, and define a very specific criteria, what's the final outcome is? And that's where we actually have seen that having deterministic workflows is also very valuable. Like those workflows may actually include advanced rendering using like video models, fine-tuned video models and so on. That's why it's a supercomputer. We actually had to build-- agent which can launch its own GPU infrastructure. And that has been critically important for us to deliver a really consistent execution across various use cases. Awesome. Um, [5:00] So we talked a little bit about where the OpenAI models fit in. Anywhere else in that stack where that's happening? So what I have seen with Astra is that it's probably the next generation of models which actually unlock computer use cases. And that's so important. When we think about the creative workflow, traditionally it required a team of over like 15 people who have to be proficient at like maybe 10, 12 tools out there. So essentially it put a lot of like huge production tax, which goes between the creativity and the idea and execution. And that's why we have seen this a lot happening in the gaming industry, where really the bar in terms of the visual quality gets high and the budgets actually grow enormously. And that's where I feel like Astra can actually allow smaller teams to be way more successful as really Astra can call all these 10 12 programs on creator behalf Gotcha Through sub stuff Awesome So when you're building this, what are some of the bigger decisions or trade-offs that you had to make as you were thinking through this? one of the going back to going back to the previous examples is it's we were so happy that we were able to unlock this net absolutely new use cases related to continuous video generation and automatic video editing and this requires multiple really multiple decisions which needs to be made. We really had to make sure that we can optimize the speed to really enable real-time experiences. Today, obviously, there was a very exciting announcement of ultra-fast model. So it's now becoming more accessible to everyone. But back then, we were working on, like with Neon a very popular streamer to unlock the first AI stream RICK VISCOMI BELAROVSKY Tell me about that Yeah that truly unique So it essentially two real creators One is Adil Higgsville spokesperson and Neon, very popular streamer. They were walking through New York City and really interacting with other people and so on. And Astro was able to really reconstruct the whole world So it feels realistic. Live? Yeah, live. And I think the stream went over 10 hours. 10 hours? Yeah, and imagine this is a new form of media, which would be impossible to imagine before that. And now everyone, every developer in the world, just a team of two or three people can ship these new experiences and build businesses on top of that. Yeah, yeah, I like that. I think so that's a great customer use case that we're talking about there. Is there something that as you were building Higgs field, we've got an audience full of developers here, is there a specific moment that you can think of that you could help to inspire these devs of building it? Because that's amazing going from 10 hours. What has been truly unique There are many of these truly unique moments where a team of really teams of three or five people can ship product end to end. And we have seen a lot of that actually happening with direct to consumer apps like mobile apps and websites. where essentially it's absolutely new consumer type of the experiences, which could be enabled by Higgs field API and OpenAI models. But also I wanted to step back really towards more larger scale use cases where actually controllability becomes important. And that's where it was so important for us that the agent does not just create very certain guardrails around the agent and can provide a secure deployment through a perimeter infrastructure with GPUs. And this really helped us to accelerate our own enterprise go-to-markets. And now as this becomes available in Higgs field APIs, [10:00] Others can build those experiences as well. Great. Yeah, that's-- I think that there's like-- You're doing this thing that we always recommend this, is that if you are building with the latest models, you should always, to use a hockey metaphor, you should always kind of skate to where the puck is going. We always say that, and you all do that so well. You're building these experiences. The model's not quite there yet. You're going from two to ten hours, and now all of a sudden it's possible, and everybody's able to access this. be building today to get ready for the next jump, the next AI jump? Look, the models improve so quickly. Even earlier today, we have seen that there are going to be new capabilities regarding to decisioning and so on. And what truly excites me is that Now the interactive media creation gets absolutely democratized like really never before. Historically, it all was really bottlenecked. by just by just ability to And with Astray, it becomes finally possible to unlock that. So what we have seen also, what truly excites me, is that we have to do a lot of things. We open sourced our own new skills, like the AI video editing skill, like to help to make AI YouTube explainer videos. We use it at Higgsfield as well. So this skill was actually very valuable for us to build more education about Higgsfield products. So it's very funny that basically Astra helped us to build a skill, which we use internally to build more videos, actually use Higgs field, right? So it's like singularity. But what's important is that we are actually seeing developers to build-- now they use the same skill to build tutorials about their own products as well. That's awesome. That's a brand new use case that you're unlocking there. I think what you're building is incredible. And I hope that everybody's feeling inspired to build a little bit. Anything-- any last words you want to give here to the audience? Thank you. In just 18 months, we built Higgsfield, and now is the best time to build yours. Awesome. Thanks everybody. Thank you. Thank you. Bye. .