Loti AI Pressures Hosts, Payment Rails to Force Deepfake Takedowns

Loti AI: Building Delete for the Internet With Loti AI — Ai4

Ai4

Automated likeness protection scales efficiently only by replacing heavy monolithic models with cheap mathematical pre-filters and escalating directly against infrastructure and payment processors when rogue sites resist.

AI-assisted summary of Loti AI: Building Delete for the Internet With Loti AI — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published by the Ai4 channel on YouTube on August 27, 2025, this presentation features Luke Arrigoni, CEO and Founder of Loti AI. The description outlines the startup challenges of running Loti's Watchtower platform across multiple state-of-the-art systems to discover, track, and remove unauthorized online content at massive scale.

## TL;DR Luke Arrigoni breaks down how Loti AI functions as an automated 'delete' button for the web, defending individuals and brands against deepfakes, impersonation scams, false endorsements, and non-consensual intimate imagery. The platform ingests hundreds of millions of images and videos daily across mainstream platforms and thousands of illicit sites using a constellation of roughly 40 specialized machine learning models and vector databases. When infringements are detected, the system sends automated legal demands, swiftly escalating to DNS registrars, CDNs, and credit card processors if hosting providers fail to comply within 48 hours. To remain financially viable at this volume, Loti leverages mathematical pre-filtering, model quantization, and predictable, cost-engineered architectures over heavyweight monolithic models.

## Atmosphere & style The presentation carries a fast-paced, pragmatic, and direct tone, blending technical architecture insights with aggressive enforcement tactics. Arrigoni speaks with candid confidence and zero corporate fluff, unapologetically describing how the company systematically pressures bad-faith operators and payment processors to defend victims. It feels like an authentic operational briefing from the startup trenches, illustrating how rigorous compute optimization and decisive legal automation intersect to police the modern internet.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=4s) Loti AI functions as an automated likeness protection platform and web-scale "delete button," indexing face and voice data across the internet to execute takedowns. - [00:04](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=4s) The service protects A-list talent, enterprise IP holders, and everyday consumers against impersonation scams, manipulated endorsements, repurposed social videos, and non-consensual intimate imagery (NCII). - [00:04](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=4s) Production-grade likeness detection relies on a constellation of around 40 specialized ML models handling normalization (such as pitch and yaw adjustment) rather than a single monolithic architecture. - [01:55](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=115s) Extremely high precision is required to maintain credibility with major platforms like Meta, ensuring automated legal takedown notices are honored without dispute. - [01:55](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=115s) The infrastructure processes roughly 600 million images per day plus video feeds, monitoring over 13,000 illicit websites to achieve takedown turnarounds within 24 hours. - [01:55](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=115s) Vector database embeddings and large video models enable the system to identify structural scam patterns and mass-takedown hundreds of mutated variants from one flagged video. - [04:39](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=279s) When bad-faith hosting sites ignore initial 48-hour notices, Loti triggers an automated escalation ladder targeting registrars, DNS providers, CDNs, and financial payment processors. - [04:39](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=279s) Loti maintains platform partnerships by cleaning up scams and deepfakes across major social networks for free, monetizing directly through represented clients and agencies. - [06:19](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=379s) To prevent unsustainable inference costs, Loti uses lightweight mathematical pre-filters to detect candidate faces before executing heavy PyTorch video models. - [06:19](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=379s) Emphasizing early model quantization and small, specialized "boring" models reduced unit compute costs and established a major operational advantage.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=4s) Likeness Protection: Building "Delete" for the Internet

Luke Arrigoni introduces Loti AI and its core mission: ingesting the entire internet to perform face and voice recognition and execute automated takedowns. Operating in the likeness protection space, Loti AI functions essentially as a "delete" button for the web. While the presentation slides do not explicitly state it, Loti AI runs one of the largest production vector databases on Earth, indexing virtually every item across the web that contains at least one face or voice.

```mermaid flowchart TD accTitle: Loti AI Ingestion and Takedown Pipeline subgraph Ingestion["1. Scale Ingestion & Detection"] A["Internet Scraping & Indexing<br/>(~600M images/day + video)"] --> B["Face / Voice Preprocessing<br/>(Pitch, yaw, frame math)"] B --> C["Constellation of ~40 ML Models<br/>(High precision verification)"] end

subgraph Understanding["2. Content Analysis"] C --> D["Vector Database & Large Video Models<br/>(Pattern & scam matching)"] end

subgraph Enforcement["3. Automated Remediation"] D --> E["Automated Takedown Notice<br/>(Direct platform API / notice)"] E --> F{"Compliant within 48h?"} F -- Yes --> G["Content Removed<br/>(~1-day turnaround)"] F -- No --> H["Escalation Hierarchy<br/>(DNS, Registrars, CDNs, Hosts, Payment Processors)"] end ```

The system primarily serves high-profile celebrities, everyday consumers, and enterprise clients holding valuable intellectual property (such as Pixar's Buzz Lightyear). The pipeline detects various unauthorized uses: * **Impersonation & Scams:** Fake accounts or deepfakes pretending to be figures like Keanu Reeves to run cryptocurrency scams. * **False Endorsements:** Manipulated media placing unendorsed products in celebrities' hands (e.g., holding up a branded t-shirt). * **Consumer Likeness Theft:** Unauthorized use of personal videos (e.g., "Get Ready With Me" clips repurposed to sell fly-by-night shampoos or workout routines paired with vitamin promotions). * **Non-Consensual Intimate Imagery (NCII):** Rapid tracking and removal across adult content networks and revenge porn repositories.

Under the hood, Loti deploys a pipeline consisting of roughly 40 distinct machine learning models. Arrigoni emphasizes that real-world production systems cannot rely on a single monolithic model: preprocessing models must first normalize faces, adjusting parameters such as pitch and yaw before matching occurs. Once an unauthorized asset is flagged, an automated ("agentic") system contacts the host site, warning them of severe legal liability and requiring deletion within two days.

---

[01:55](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=115s) High Precision, Customer Dashboards, and Scam Detection

Because Loti works closely with top entertainment talent agencies—including CAA, WME, and UTA—as well as A-list talent, the system requires near-perfect precision. Model accuracy must be airtight so that takedown requests are not rejected by major platforms like Meta. If a claim asserts that Tom Hanks appears in an image, the system must guarantee it to prevent losing credibility with hosting platforms.

In addition to mainstream platforms, Loti continuously monitors and indexes over 13,000 illicit websites commonly banned on public library networks, maintaining one of the largest specialized collections of such material to ensure fast takedowns.

* **Scale and Speed:** Ingests approximately 600 million images per day (alongside heavy video streams), achieving successful removals within roughly 24 hours. * **Client Dashboard:** Acts like a personal, private monitoring hub dividing findings into adult content, impersonations, and deepfakes with real-time status tracking. * **Vector Content Understanding:** Utilizes large video models and vector embeddings to identify structural scam patterns. When one scam video is analyzed, vector similarity enables instant identification and automated removal of hundreds of identical or mutated variants.

---

[04:39](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=279s) Escalation Strategies and Bullying the Bad Actors

Loti maintains strong relationships with major social media companies by identifying scams and cleaning up their networks free of charge (monetizing instead through its own clients). However, when fringe and uncooperative sites fail to comply with initial 48-hour takedown notices, Loti initiates an aggressive, automated escalation ladder:

1. **Hosting and DNS Infrastructure:** Direct outreach to domain name registrars, DNS providers, and hosting services. 2. **Content Delivery Networks (CDNs):** Escalations to caching and distribution networks. 3. **Financial Choke Points:** Contacting credit card payment processors (e.g., Visa processing banks) to alert them to active merchant liability.

Arrigoni notes that when handling non-consensual imagery or illicit scams on bad-faith websites, they have no problem playing the role of aggressive enforcers to ensure the content disappears.

---

[06:19](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=379s) Cost Engineering: Boring Models and Resource-Constrained Elegance

Processing hundreds of millions of images and full-motion video feeds creates massive compute bottlenecks. Running naive, frame-by-frame deep learning inference across entire video catalogues would lead to unsustainable AWS compute bills.

To overcome this, Loti adopted several counter-intuitive architectural strategies: * **Mathematical Pre-filtering:** Rather than spinning up heavy PyTorch models on every frame, the team developed lightweight mathematical methods to detect potential faces early, bypassing full video processing. * **Choosing "Boring" Models:** Arrigoni advises engineering teams against reflexively chasing the newest complex models. Using smaller, specialized, and more predictable architectures dramatically drives down inference costs. * **Early Quantization:** Long before open-source releases like DeepSeek made aggressive model optimization a boardroom talking point, Loti was forced by early budget constraints to embrace model quantization and lightweight networks.

By engineering inference down to the cheapest unit cost on the market, the company turned its compute efficiency into a competitive advantage during fundraising and operational scaling.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=4s) Loti AI functions essentially as a "delete" button for the web.

> [00:04](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=4s) Loti AI runs one of the largest production vector databases on Earth, indexing virtually every item across the web that contains at least one face or voice.

> [00:04](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=4s) real-world production systems cannot rely on a single monolithic model — Luke Arrigoni

> [00:04](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=4s) Once an unauthorized asset is flagged, an automated ("agentic") system contacts the host site, warning them of severe legal liability and requiring deletion within two days.

> [01:55](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=115s) Because Loti works closely with top entertainment talent agencies—including CAA, WME, and UTA—as well as A-list talent, the system requires near-perfect precision.

> [01:55](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=115s) If a claim asserts that Tom Hanks appears in an image, the system must guarantee it to prevent losing credibility with hosting platforms.

> [01:55](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=115s) maintains one of the largest specialized collections of such material to ensure fast takedowns.

> [01:55](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=115s) When one scam video is analyzed, vector similarity enables instant identification and automated removal of hundreds of identical or mutated variants.

> [04:39](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=279s) Loti maintains strong relationships with major social media companies by identifying scams and cleaning up their networks free of charge

> [04:39](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=279s) when handling non-consensual imagery or illicit scams on bad-faith websites, they have no problem playing the role of aggressive enforcers to ensure the content disappears. — Luke Arrigoni

> [06:19](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=379s) Running naive, frame-by-frame deep learning inference across entire video catalogues would lead to unsustainable AWS compute bills.

> [06:19](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=379s) Arrigoni advises engineering teams against reflexively chasing the newest complex models. Using smaller, specialized, and more predictable architectures dramatically drives down inference costs. — Luke Arrigoni

## Mentioned Luke Arrigoni, Loti AI, Watchtower, Buzz Lightyear, Pixar, Keanu Reeves, Tom Hanks, CAA, WME, UTA, Meta, Visa, AWS, PyTorch, DeepSeek

## Video description Luke Arrigoni, CEO and Founder, Loti AI

Abstracts

Loti’s Watchtower presents challenges to us as a startup because we have to run multiple state of the art systems to search, find and remove content at scale.

[Source](https://www.youtube.com/watch?v=ROR0_ZkC62Y)

中文简报

Loti AI向托管商与支付渠道施压以强制下架深度伪造内容 自动化肖像保护唯有以低成本数学预过滤器取代笨重的单体模型，并在违规网站抵制时直接向基础设施及支付处理机构升级施压，方能实现高效扩展。 Luke Arrigoni 详细阐述了 Loti AI 如何作为网络的自动化“删除”按钮运作，保护个人和品牌免受深度伪造、冒名诈骗、虚假代言以及非自愿亲密影像的侵害。该平台依托由约 40 个专用机器学习模型与向量数据库构成的体系，每日从主流平台及数千个非法网站摄取数亿张图片和视频。一旦检测到侵权，系统便会自动发送法律诉求；若托管服务商在 48 小时内未予配合，则会迅速向 DNS 注册商、CDN 和信用卡处理机构升级交涉。为了在如此体量下保持财务可行性，Loti 摒弃了笨重的单体模型，转而采用数学预过滤、模型量化以及成本受控且可预测的架构。 Loti AI 打造针对互联网人脸与声音资产的自动化“删除”平台，依托超大规模生产级向量数据库和由约 40 个专用 ML 模型构成的流水线，精准检测冒名诈骗与侵权内容。 (0:04) 系统每日吞吐约 6 亿张图像并持续监控逾 1.3 万个非法网站，在确保极高精度的同时实现 24 小时内快速下架，并通过向量嵌入批量清除变体诈骗视频。 (1:55) 针对 48 小时内拒不下架的恶意站点，Loti 会启动自动化升级打击阶梯，直接施压域名注册商、DNS、CDN 乃至银行信用卡支付处理机构以切断其支撑。 (4:39) 为应对海量流媒体推理带来的高额云成本，Loti 采用轻量数学预过滤、选择可预测的小型模型及早期模型量化策略，大幅压缩了单次检测的单位算力成本。 (6:19)

Reading summary

Loti AI operates an automated likeness protection and takedown platform that ingests roughly 600 million images daily to identify and remove unauthorized uses of faces, voices, and intellectual property. According to founder Luke Arrigoni, functioning as an effective "delete button" for the web requires high-precision verification, infrastructure-level enforcement, and strictly optimized inference costs. Multi-Model Detection Across Likeness Infringements (0:04) Loti AI targets a broad spectrum of unauthorized content, including deepfake cryptocurrency scams impersonating figures like Keanu Reeves, manipulated commercial endorsements, repurposed consumer clips, and non-consensual intimate imagery (NCII). To serve enterprise IP holders, everyday users, and talent agencies such as CAA, WME, and UTA, the platform indexes both mainstream networks and over 13,000 illicit websites, achieving removals within an estimated 24-hour window. Pipeline Precision and Vector-Based Scam Identification (1:55) Because false claims can damage standing with major platforms like Meta, Arrigoni stresses that the pipeline demands near-perfect precision rather than relying on a single monolithic model. Loti deploys roughly 40 specialized machine learning models that execute normalization—adjusting facial pitch, yaw, and frame geometry—before running verification. In addition, large video models and vector databases allow the platform to analyze the underlying structure of scam campaigns, enabling automated takedowns across hundreds of mutated variants once an initial template is identified. Infrastructure Escalation and Financial Choke Points (4:39) Loti provides free network cleanup for major social platforms to maintain cooperative partnerships, monetizing directly through its own clients. When fringe or bad-faith websites ignore initial 48-hour automated takedown notices, the company initiates a multi-stage escalation protocol. The system targets supporting infrastructure by alerting domain registrars, DNS providers, and content delivery networks (CDNs), and will escalate directly to credit card processors to confront uncooperative sites with merchant liability. Inference Optimization Through Lightweight Architecture (6:19) Ingesting hundreds of millions of daily media assets creates severe computational bottlenecks that would make continuous, frame-by-frame deep learning inference economically unsustainable. To lower unit compute costs, Loti relies on mathematical pre-filtering to identify potential faces before engaging heavy frameworks. Arrigoni advocates using predictable, specialized "boring" models alongside early model quantization, turning aggressive compute efficiency into a core operational and fundraising advantage. Loti AI presents an automated model for cross-web content removal by pairing large-scale vector indexing with infrastructure-level legal pressure, though its operational speed remains tied to the responsiveness of external registrars, CDNs, and payment networks when target sites resist voluntary compliance.

精炼阅读摘要

Loti AI 创始人 Luke Arrigoni 提出，公司正在构建针对互联网全网的肖像与声音保护体系，通过大规模抓取、向量数据库索引以及多模型流水线，充当互联网的“删除按钮”来执行自动化维权下架。 肖像侵权识别场景与系统架构 (0:04) Loti AI 的核心业务是全网抓取并建立向量索引，覆盖包含面部或声音的绝大多数公开网络内容。其保护对象涵盖知名演艺明星、普通消费者以及持有高价值知识产权的企业客户（例如皮克斯旗下的巴斯光年等 IP）。系统主要检测与处理四类侵权场景：假冒名人的加密货币诈骗、将未经代言的商品合成到名人手中的虚假背书、盗用普通人日常视频（如化妆或健身片段）用于推销劣质商品的肖像盗用，以及非自愿亲密影像（NCII）。 Arrigoni 强调，生产环境无法依赖单一巨型模型完成全部工作。Loti 的技术流水线整合了约 40 个专用的机器学习模型。在进行身份匹配前，系统会先利用预处理模型对人脸的俯仰角和偏航角等几何姿态参数进行数学归一化计算。当检测到侵权资产后，自动化代理系统会立即向托管网站发送包含法律责任警告的通知，要求其在 48 小时内移除相关内容。 高精度要求与全网抓取规模 (1:55) 由于 Loti 为 CAA、WME、UTA 等好莱坞头部经纪公司及顶流艺人提供服务，系统对识别准确率提出了极高要求。向 Meta 等主流大型平台提交下架请求时，必须保证证据确凿，若错误指认则会损害平台的下架信誉。除了主流社交媒体平台，Loti 持续监控并索引超过 1.3 万个在公共网络中常被屏蔽的非法站点，维护着专项影像数据库以实现快速识别。 在数据处理与执行效率方面，系统每日抓取和处理约 6 亿张图像并伴随大量视频流，侵权内容通常在发现后 24 小时左右完成下架。客户可通过私有控制面板实时追踪成人内容、冒充和深度伪造的下架进展。同时，系统结合大型视频模型与向量嵌入技术分析诈骗内容的结构模式；一旦捕获单个诈骗样本，即可利用向量相似度迅速检索并自动下架数百个衍生变体。 多层级升级施压与基础设施阻断 (4:39) 在合作模式上，Loti 免费协助主流社交媒体平台发现诈骗行为并清理违规内容，其商业变现主要来自需要维权保护的客户，以此维系与主要平台的协作关系。然而，对于那些在收到 48 小时初始通知后拒绝配合的边缘站点或恶意网站，系统会自动启动阶梯式的施压流程。 Arrigoni 介绍，该升级流程包括直接联络域名注册商（Registrars）、DNS 解析商以及底层服务器托管商；若仍无响应，则向内容分发网络（CDN）进行通报要求切断缓存；在必要时，Loti 会进一步介入金融支付链条，向 Visa 处理银行等信用卡支付处理机构提示商户涉及非法业务的法律风险。在打击非自愿成人影像和恶意诈骗时，公司采取强硬手段迫使目标站点下线侵权内容。 低成本工程化与模型轻量化设计 (6:19) 每天分析数以亿计的图像与连续视频流，给算力成本带来了巨大挑战。如果直接对视频目录进行逐帧的深度学习推理，会导致高昂且难以维系的云计算开销。为此，Loti 采用了偏向工程实用性的低成本架构方案。 团队放弃在所有视频帧上频繁调用重量级 PyTorch 模型的做法，转而利用轻量级数学方法在前置阶段初步筛查人脸，大幅减少进入全量视频处理的负载。Arrigoni 建议工程团队避免盲目追求最新复杂的庞大模型，而是选用更小、更专用且行为可控的“基础”模型以压低推理开销。此外，由于早期受到严格预算限制，Loti 在 DeepSeek 等开源量化热潮出现之前便推进了模型量化与精简网络，将单位推理成本压制在行业较低区间，进而转化为运营扩展与融资时的竞争优势。 该方案展示了基于多模型与底层基建施压的肖像侵权治理路径，但其长期运行高度依赖对超大规模网络数据的持续抓取，以及域名解析、CDN 与支付通道等第三方基础设施服务商对侵权通报的响应意愿与配合度。

Detailed notes — zh

## Background 该内容由 YouTube 频道 Ai4 于 2025 年 8 月 27 日发布，主讲人为 Loti AI 创始人兼 CEO Luke Arrigoni。视频简介指出，Loti AI 的核心产品 Watchtower 在作为初创企业推进时面临巨大工程挑战，必须同时运行多个顶尖系统以实现超大规模的侵权内容搜索、发现与自动化下架。 ## TL;DR Loti AI 致力于为互联网打造一键“删除键”，保护好莱坞名流、个人创作者及知名 IP 免受虚假代言、诈骗和非自愿露骨图像的侵扰。团队运营着全球顶尖规模的向量数据库，每日摄取并分析约 6 亿张图像当量，并通过约 40 个精细协同的机器学习模型实现超高 F1 分数的精准识别。当遭遇违规甚至成人网站拒绝删除内容时，系统会启动层层加码的施压手段，直至切断其 DNS、CDN 乃至信用卡结算通道。面对海量视频带来的算力成本危机，创始人强调回归“枯燥朴素的小模型”与极致量化，走出了一条在严苛资源约束下倒逼出极致工程性价比的技术道路。 ## Atmosphere & style Luke Arrigoni 的演讲风格兼具极客的硬核务实与硅谷创业者的幽默犀利，台风松弛且毫无花哨的公关辞令。他一面直白调侃行业对“单一超级大模型”的营销吹嘘，一面坦然自嘲为被贫穷逼出来的“量化界嬉皮士”，展现出对底层工程细节的极致自豪。整场分享从对网络黑产与侵权乱象的强力铁腕回击，平滑过渡到对算力成本控制的深度工程反思，听起来既像是一场大快人心的维权反击纪实，又是一堂给 AI 工程师的清醒务实避坑课。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=4s) Loti AI 的核心使命是为整个互联网构建“删除键”，通过大规模索引全网人脸与声音实现侵权内容的识别与下架。 - [00:04](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=4s) 后台运行着全球规模最大的生产级向量数据库之一，持续覆盖互联网上包含人脸或声音的多媒体内容。 - [00:04](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=4s) 系统底层由约 40 个专业机器学习模型串联协作，包含人脸俯仰角与偏航角校正等辅助模型以服务核心识别任务。 - [01:55](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=115s) 模型判别采用极其严苛的 F1 分数标准，确保高置信度以承受 Meta 等主流社交平台的反向抗辩。 - [01:55](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=115s) 专门索引并监控了超过 13,000 个成人与边缘站点，高效打击未经授权的非自愿露骨内容（NCI）及 IP 侵权。 - [01:55](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=115s) 系统每日摄取与分析约 6 亿张图像当量数据，侵权内容的平均下架周期约为一天。 - [04:02](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=242s) 利用大视频模型（LVM）和向量内容理解解析音视频深层意图，能通过向量数学批量锁定结构相似的诈骗网络。 - [04:02](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=242s) 构建了层层递进的升级施压机制，若侵权源站拒绝配合，将直接联合域名商、DNS、CDN 直至信用卡结算机构实施网络断流与封锁。 - [06:16](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=376s) 为破解海量视频逐帧推理带来的算力成本危机，团队选择回归古典数学算法与轻量模型，而非盲目堆砌庞大架构。 - [06:16](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=376s) 早期因资源匮乏而全面采用激进的模型量化与小型化技术，印证了在严格工程约束下往往能倒逼出极致的高性价比创新方案。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=4s) 肖像保护：为整个互联网打造“删除键” Luke Arrigoni 开门见山地向现场观众抛出 Loti AI 的核心使命：为整个互联网构建一个“删除键”（Delete for the internet）。为了实现全网人脸与声音识别并完成侵权内容下架，团队在后台默默运行着目前全球规模最大的生产级向量数据库之一，持续索引互联网上每一份至少包含一张人脸或一段声音的内容。这项名为“肖像保护”（Likeness Protection）的业务，主要服务于好莱坞明星、普通创作者以及拥有知名 IP 的大型企业（例如像“巴斯光年”这样的经典角色）。 ```mermaid flowchart TD accTitle: Loti AI 内容检测与自动化下架全流程 subgraph Ingestion ["数据摄取与预处理"] A["全网抓取：图像与视频"] --> B["视频快速抽帧与预处理模型"] B --> C["姿态校正模型：俯仰角 / 偏航角"] end subgraph Recognition ["多模型协作与向量识别"] C --> D["人脸 / 声音识别模型集群"] D --> E["大规模向量数据库比对"] E --> F["高置信度判定：深度伪造 / 冒充 / 侵权"] end subgraph Takedown ["自动化履约下架"] F --> G["自动化下架系统发出通知"] G --> H{"目标站点是否合规？"} H -- 2日内合规 --> I["内容成功下架"] H -- 拒绝下架 --> J["升级施压：DNS / CDN / 支付通道"] end ``` 在日常网络环境中，各类侵权与欺诈层出不穷：有未经授权让基努·里维斯（Keanu Reeves）手持印有陌生品牌的 T 恤虚假代言，有冒充名人私聊诈骗加密货币，也有恶意的深度伪造（Deepfake）。Loti AI 建立了一个由大约 40 个机器学习模型串联组成的协作管线（Constellation of models）。Luke 调侃道，在机器学习行业，市场营销人员往往只喜欢吹嘘某一个耀眼的超大模型，但在工程前线，工程师必须开发几十个辅助模型来服务核心模型——例如在识别人脸前，必须先调用多个模型对人脸的俯仰角（Pitch）和偏航角（Yaw）进行校正对齐。而在最终的执行环节，系统通过全自动化的工作流（如果为了拿融资蹭热点，也可以管它叫“Agentic AI”）主动向侵权网站发出严正通知，告知对方所托管的内容将带来严重的法律责任，并限定其在两天内彻底删除。 - **核心数据资产**：运行全球最大的生产级向量数据库之一，覆盖互联网上海量的面部与声音特征。 - **协同模型管线**：底层部署约 40 个专业机器学习模型，涵盖人脸角度校正、音视频预处理与特征匹配。 - **自动化执行机制**：全自动发起侵权通告，向目标网站出具有法律效力的通知并限期 48 小时下架。 --- ### [01:55](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=115s) 极致精度标准、监控看板与每日 6 亿张图像摄取 在娱乐产业深耕的 Loti AI，客户名单涵盖了 CAA、WME、UTA 等顶级好莱坞经纪公司及一众一线明星，但 Luke 强调团队在本质上就是一群对模型精度极度执着的“技术宅”（Nerds）。由于面对大众客户时通常只展示通俗的“准确率”，但在技术同行面前，他直言团队的核心指标是极高的 F1 分数。这种对精度的极致追求源于现实压力：要让 Meta 这样的社交巨头无条件配合下架，技术判定的门槛高得惊人——“如果我告诉你们这张图里是汤姆·汉克斯（Tom Hanks），那我最好百分之百正确，否则平台以后绝不会再受理你的任何下架请求。” 面向普通消费者和企业，Loti AI 同样处理各种棘手场景。普通用户经常遭遇自己精心拍摄的“跟我一起化妆”（Get Ready With Me）生活视频被盗用，沦为劣质洗发水或维生素保健品的虚假带货素材；更有甚者遭遇非自愿露骨图像（NCI，即报复性色情内容）的侵害。为此，Loti AI 专门索引了超过 13,000 个在大多数公共图书馆网络中被屏蔽的成人与边缘网站，建立了极具规模的相关资料库，能以极高效率执行清除；甚至连偶尔流落到这些站点的“巴斯光年”IP 侵权内容也能一并拔除。 系统每天需要摄取分析大约 6 亿张图像（包含视频解压出的帧画面），通常在一天之内就能促成内容下架。用户登录后会看到一个类似“私人微型 NSA”的监控仪表盘，直观分类展示成人内容、身份冒充和深度伪造等各类侵权动态，后台发现的侵权线索会化为源源不断的信息流实时推送到用户眼前。 - **合作机构与严苛门槛**：对接 CAA、WME、UTA 等好莱坞巨头；模型判别标准必须承受住 Meta 等顶级社交平台的反向抗辩。 - **特殊站点索引**：索引并监控超过 13,000 个成人及边缘站点，实现违规露骨内容与侵权 IP 的闪电下架。 - **处理吞吐量与时效**：日均处理约 6 亿张图像当量数据，侵权内容平均下架周期约为一天。 --- ### [04:02](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=242s) 向量内容理解与层层升级的“技术霸凌” 除了基础的特征比对，Loti AI 还利用向量内容理解（Vector Content Understanding）实现深层语义识别。通过大视频模型（Large Video Model）对音视频语义进行端到端解析，系统能够迅速推断出画面背后的实际意图（如某种典型的诈骗套路），并借助底层的向量数学计算，将网络上所有结构相似的连带诈骗内容一网打尽。Luke 幽默地表示，虽然向客户推销时只需包装成“我们懂得识别网络诈骗”，但在同行面前，这本质上就是一个优雅且完全可以用白板推导的向量数学问题，其真正的难点在于如何将其推向超大规模的工业级落地。 在具体推动下架的过程中，Loti AI 与主流社交平台保持着极佳的关系，因为平台免于法律问责，而且 Loti 向被侵权的客户收费而非向平台收费。而针对那 13,000 个游离于主流规则之外的网站，Loti AI 则准备了一套极为强硬的“升级施压”（Escalation）阶梯机制： | 施压阶段 | 目标对象 | 施压手段与核心筹码 | |:--- |:--- |:--- | | **第一级：源站通知** | 目标侵权网站站长 / 运营方 | 发送自动化法律通告，指出法律责任，限期 48 小时删除 | | **第二级：网络基建** | DNS 服务商、域名注册商（Registrar） | 投诉侵权与违法托管，要求暂停域名解析服务 | | **第三级：托管与分发** | CDN 服务商、主机托管商（Hosts） | 指控违规内容违背服务条款，切断网络加速与服务器托管 | | **第四级：金融命脉** | 信用卡收单机构（Credit Card Merchants） | 警告其正在为违规站点结算 Visa/Mastercard，转嫁连带法律风险 | 如果侵权站点置之不理，Loti AI 会毫不犹豫地层层加码，直接向域名注册商、DNS 服务商、CDN、主机托管商施压，甚至直接找到网站绑定的信用卡结算商，警告对方正在为违法侵权内容处理资金结算，瞬间将对方拖入合规噩梦。谈及这种铁腕手段，Luke 坦言：“面对那些未经许可散布他人裸照或大搞诈骗的网站，我们觉得在这种情况下当个‘霸凌者’（Bully）完全合情合理。” --- ### [06:16](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=376s) 工程困境的反思：拥抱“枯燥模型”与资源约束下的优雅架构 在谈及技术架构的重大挑战时，Luke 分享了创业过程中最深刻的一堂工程课：海量视频如果逐帧解码运行超大模型推理，算力成本将是天文数字，“当你在算力推理成本上一算账，心里立刻会凉半截——‘这成本绝对过不了任何人的审批’”。面对算力黑洞，Loti AI 并没有随大流去堆砌更前沿、参数量更庞大的热门架构，而是果断回归古典的纯数学算法和基础图像处理手段，专注攻克一个细分核心点：在不全面逐帧解压视频的前提下，极度高效地锁定画面中的人脸。 这套反直觉的技术路径让团队在融资时展现出了强大的底气。Luke 可以理直气壮地对投资人说：“我确实需要你们帮我支付巨额的 AWS 账单，但我能证明，全天下没有任何人能用比我更便宜、更高效的方案完成这项任务。”因此，他给在场 AI 工程师的核心建议是：不要盲目追逐所谓更酷炫的大模型，往往一个“更无聊、更朴素的模型”（A more boring model）才是破解高昂工程难题的真正解药。 在演讲尾声，Luke 提到了当时引发全行业震动的 DeepSeek 现象。当开源小模型架构和低成本训练技巧爆火、所有人都在被老板追问“我们是否也在用这些技术”时，Loti AI 团队内部却感到既欣慰又亲切。他笑着感慨，因为早期极度贫困、算力捉襟见肘，他们早在量化技术流行之前就为了活下去而全盘采用了激进的模型量化与小型化技术，堪称“量化界的时髦嬉皮士”（The hipsters of quantization）。严苛的资源约束非但没有扼杀创新，反而倒逼出极其精炼、高性价比的优雅工程解法。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=4s) 为整个互联网构建一个“删除键”（Delete for the internet）。 — Luke Arrigoni > [00:04](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=4s) 如果为了拿融资蹭热点，也可以管它叫“Agentic AI”。 — Luke Arrigoni > [01:55](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=115s) 如果我告诉你们这张图里是汤姆·汉克斯（Tom Hanks），那我最好百分之百正确，否则平台以后绝不会再受理你的任何下架请求。 — Luke Arrigoni > [04:02](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=242s) 面对那些未经许可散布他人裸照或大搞诈骗的网站，我们觉得在这种情况下当个‘霸凌者’（Bully）完全合情合理。 — Luke Arrigoni > [06:16](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=376s) 当你在算力推理成本上一算账，心里立刻会凉半截——‘这成本绝对过不了任何人的审批’。 — Luke Arrigoni > [06:16](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=376s) 我确实需要你们帮我支付巨额的 AWS 账单，但我能证明，全天下没有任何人能用比我更便宜、更高效的方案完成这项任务。 — Luke Arrigoni > [06:16](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=376s) 不要盲目追逐所谓更酷炫的大模型，往往一个“更无聊、更朴素的模型”（A more boring model）才是破解高昂工程难题的真正解药。 — Luke Arrigoni > [06:16](https://www.youtube.com/watch?v=ROR0_ZkC62Y&t=376s) 因为早期极度贫困、算力捉襟见肘，我们早在量化技术流行之前就为了活下去而全盘采用了激进的模型量化与小型化技术，堪称“量化界的时髦嬉皮士”（The hipsters of quantization）。 — Luke Arrigoni ## Mentioned Luke Arrigoni, Loti AI, 基努·里维斯 (Keanu Reeves), 汤姆·汉克斯 (Tom Hanks), 巴斯光年 (Buzz Lightyear), CAA, WME, UTA, Meta, AWS, DeepSeek, Visa, Mastercard ## Video description Luke Arrigoni, CEO and Founder, Loti AI Abstracts Loti’s Watchtower presents challenges to us as a startup because we have to run multiple state of the art systems to search, find and remove content at scale. [Source](https://www.youtube.com/watch?v=ROR0_ZkC62Y)

Source quotes

[0:04] Luke Arrigoni: 为整个互联网构建一个“删除键”（Delete for the internet）。 [0:04] Luke Arrigoni: 如果为了拿融资蹭热点，也可以管它叫“Agentic AI”。 [1:55] Luke Arrigoni: 如果我告诉你们这张图里是汤姆·汉克斯（Tom Hanks），那我最好百分之百正确，否则平台以后绝不会再受理你的任何下架请求。 [4:02] Luke Arrigoni: 面对那些未经许可散布他人裸照或大搞诈骗的网站，我们觉得在这种情况下当个‘霸凌者’（Bully）完全合情合理。 [6:16] Luke Arrigoni: 当你在算力推理成本上一算账，心里立刻会凉半截——‘这成本绝对过不了任何人的审批’。 [6:16] Luke Arrigoni: 我确实需要你们帮我支付巨额的 AWS 账单，但我能证明，全天下没有任何人能用比我更便宜、更高效的方案完成这项任务。 [6:16] Luke Arrigoni: 不要盲目追逐所谓更酷炫的大模型，往往一个“更无聊、更朴素的模型”（A more boring model）才是破解高昂工程难题的真正解药。 [6:16] Luke Arrigoni: 因为早期极度贫困、算力捉襟见肘，我们早在量化技术流行之前就为了活下去而全盘采用了激进的模型量化与小型化技术，堪称“量化界的时髦嬉皮士”（The hipsters of quantization）。

Source quotes

[0:04] Loti AI functions essentially as a "delete" button for the web. [0:04] Loti AI runs one of the largest production vector databases on Earth, indexing virtually every item across the web that contains at least one face or voice. [0:04] Luke Arrigoni: real-world production systems cannot rely on a single monolithic model [0:04] Once an unauthorized asset is flagged, an automated ("agentic") system contacts the host site, warning them of severe legal liability and requiring deletion within two days. [1:55] Because Loti works closely with top entertainment talent agencies—including CAA, WME, and UTA—as well as A-list talent, the system requires near-perfect precision. [1:55] If a claim asserts that Tom Hanks appears in an image, the system must guarantee it to prevent losing credibility with hosting platforms. [1:55] maintains one of the largest specialized collections of such material to ensure fast takedowns. [1:55] When one scam video is analyzed, vector similarity enables instant identification and automated removal of hundreds of identical or mutated variants. [4:39] Loti maintains strong relationships with major social media companies by identifying scams and cleaning up their networks free of charge [4:39] Luke Arrigoni: when handling non-consensual imagery or illicit scams on bad-faith websites, they have no problem playing the role of aggressive enforcers to ensure the content disappears. [6:19] Running naive, frame-by-frame deep learning inference across entire video catalogues would lead to unsustainable AWS compute bills. [6:19] Luke Arrigoni: Arrigoni advises engineering teams against reflexively chasing the newest complex models. Using smaller, specialized, and more predictable architectures dramatically drives down inference costs.

Original transcript

[0:04] Good afternoon everyone. How's everyone [0:06] doing today? [0:07] All right, cool. So, we're going to talk [0:10] about basically how we ingest the entire [0:12] internet for face and voice recognition. [0:14] Do some takedowns. Uh what's not on the [0:17] slide, but what kind of cool is we found [0:19] out we are running one of the largest [0:20] production uh vector databases on Earth [0:23] right now, which is just like every item [0:25] on the internet with at least one face [0:28] or voice on it. All right. So what do we [0:30] do? Likeness protection. How we build [0:31] build delete for the internet. [0:34] We solve a problem for mainly [0:36] celebrities but also some consumers and [0:38] some larger enterprises that have [0:40] intellectual property. Think Buzz [0:41] Lightyear. And what we do is we find [0:44] impersonations, explicit deep fakes, [0:47] false endorsements. Ke Reeves holding up [0:48] a t-shirt with someone's brand on it [0:50] that he's never seen before or someone [0:52] pretending to be him and trying to get [0:54] crypto. We find all of these accounts, [0:56] all of these images, these videos. And [0:58] then we issue takedowns all of them. [1:01] We built about 40 machine learning [1:02] models to ingest the internet and we [1:04] make the takedown portion entirely [1:06] automated. So there's uh you can think [1:08] about like a constellation of models [1:10] that are in like a pipeline that have to [1:12] work together because if you're going to [1:13] do something like face recognition, you [1:15] can't just do a single model. You have [1:17] to have models that prep the faces ahead [1:19] of time. There's things like pitch and [1:20] yaw you have to adjust for on a face. [1:22] And I'm sure everyone here who works on [1:24] models understands and appreciates [1:25] there's that one big model that your [1:27] marketing person likes to talk about, [1:29] but down in the ground, we all have to [1:30] build dozens of models to make that one [1:32] model work. And so we're no different. [1:34] And then the takeown part is you could [1:37] think about it as if we're really [1:38] wanting to be buzzwords, but I don't [1:40] think we're trying to raise money here, [1:41] uh, aenic. We go out to a website and we [1:44] pull down something. It's basically an [1:47] automated flow to tell a site you have [1:49] something that's going to cause you a [1:50] lot of liability and you have two days [1:52] to remove it. [1:55] Our superpowers is that we're nerds. Uh [1:57] we work in entertainment, we work for CA [1:59] or with CA, WME, UTA, these types of [2:02] folks. Uh we work with top tier Aclass [2:05] celebrities. But the reality is [2:07] underneath the covers, we're just trying [2:09] to build really precise models. I have [2:11] accuracy up here because I normally show [2:13] this slide deck to people that don't [2:14] know what an F1 score is. But, uh, [2:17] mainly we have to be so good at knowing [2:19] that it is you in the picture or your [2:21] voice in that audio, even if it's a deep [2:24] fake, that we don't get challenged by [2:26] meta. Like, and that's a very high bar, [2:29] right? You basically have to say, if I'm [2:31] telling you that this is Tom Hanks in [2:33] this image, um, I better be right. [2:37] Otherwise, they won't ever take anything [2:38] down. So, we're we're really good at [2:39] that portion. You could think about our [2:42] consumer or our applications as we [2:44] mainly sell to celebrities and [2:46] consumers. So, for consumers, you could [2:48] think about someone that's like stolen [2:49] one of your get ready with me videos and [2:51] put their fly by night shampoo brand on [2:53] top of it or someone trying to sell [2:54] vitamins with whatever workout routine [2:57] you've posted or um NCI, right? You like [3:00] revenge porn or something like that [3:02] because we also index 13,000 sites that [3:04] are banned in most public libraries. We [3:07] have the largest collection of that [3:08] material actually strangely and we know [3:10] how to get all of that to come down very [3:12] quickly. And then enterprises and while [3:14] Buzz Lightyear might not be on one of [3:16] those sites occasionally unfortunately [3:18] he is we can take down those as well. [3:21] You can think about us doing roughly six [3:23] or so 600 or so million images a day. So [3:26] you have to also think about the videos [3:27] that are incorporated in that. And it [3:30] takes us about a day to get things to [3:31] come down. What you do is you get this [3:34] dashboard that looks something like [3:35] this, right? It's your private little [3:38] NSA, right? It's it's your version of [3:40] where you show up online. And this is me [3:42] being lazy and literally pulling up the [3:43] account and like screen capping it. Um, [3:46] but this is roughly the the kind of [3:48] levels that you can think about, which [3:49] is, you know, there's a section for [3:51] adult content, impersonations, and deep [3:53] fakes. And all of them have some degree [3:55] of takedowns that are required for all [3:57] of it. And there's just this ongoing [3:59] feed of everything we're finding all the [4:00] time. [4:02] And then we also do something very cool [4:04] but very not new which is vector content [4:07] understanding. Imagine running an image [4:09] or a video through a model like a large [4:11] video model that says I know what's [4:12] going on in here. And then you could do [4:15] some really cool takedowns. You can find [4:16] out, oh this thing is a scam and there's [4:19] a bunch that look like it because vector [4:21] math is cool. Of course when we sell [4:23] this to customers we just say we kind of [4:25] understand the scams that run and we can [4:27] automate those takedowns. But since I'm [4:29] with fellow nerds, I get to talk about [4:31] how this is actually just a math [4:32] problem. And you could probably all [4:34] solve it right now if I asked you to [4:35] whiteboard it. But still, it's cool to [4:37] do at scale. [4:39] Our rapid takedowns are fun, which is, [4:42] you know, we issue takedowns and with [4:43] our stuff come down. We enjoy really [4:45] great relationships with the social [4:47] media companies because we save them [4:48] liability like refining all the scams [4:50] and removing them. And we don't charge [4:52] them. We charge our customers. [4:54] and uh the 13,000 banned sites take it [4:58] down because they don't want to get sued [4:59] by our sometimes latigious customers. [5:02] And this is what it looks like. You [5:02] click on one of those and you say, "Hey, [5:05] um is it did it get taken down? How did [5:07] it get taken down?" And you know, [5:09] ultimately, um is it escalating? So, you [5:12] can think about an escalation like let's [5:14] say a site um let's say a site goes [5:17] through and and we we say like, "Hey, [5:20] you need to take it down." They don't. [5:21] Well, we can go to the DNS, we go to [5:22] name registar, we could go to CDN's, [5:25] uh, so hosts, we can even go to credit [5:27] card merchants and ask them like, "Hey, [5:29] by the way, there's someone that's [5:30] processing Visa on your site and this [5:32] liability that's presented to you [5:34] because you're processing their credit [5:35] cards." And it ends up being like the [5:38] worst day they've had. But we give them [5:40] time, like we say like, "Hey, you're [5:41] hosting naked pictures of someone that [5:43] needs to come down." We feel like we can [5:45] be a little bit of bullies in that [5:46] situation. Um, especially if we have the [5:49] relationship with a person that says, [5:50] "I'd rather not have that image up or if [5:52] someone's scamming someone, we feel like [5:54] it's okay to be a little bit of a [5:56] bully." Um, but that's really it. You [5:59] know, I'll leave that there. Um, I know [6:01] we have plenty of time, so I'll put that [6:04] there and I could talk about some [6:06] challenges we've had now that we all are [6:07] on the same page of you all had to hear [6:10] my crappy sales pitch and if you ever [6:12] need stuff deleted, you know where to [6:13] go. Um, but I guess I could talk about [6:16] infrastructure as a large problem too. [6:19] One of the the biggest issues we had was [6:21] how do we process videos really really [6:23] quickly uh because it was a fortune to [6:26] go frame by frame and so we got really [6:28] really good at solving one particular [6:31] part of the problem which is when you [6:32] give me a video how do I find faces in [6:34] it without processing the whole video [6:37] and trying to do that in the most [6:38] efficient way pro possible. And if [6:40] anyone here is trying to solve like big [6:41] models and you're running the numbers on [6:43] inference and you're like, "Oh crap, [6:46] that is definitely not going to get [6:47] approved by anyone." Um, the funny and [6:50] ironic part about our industry is if you [6:52] return to the older principles around a [6:54] pure math solution early on or any kind [6:57] of solution that's not necessarily [6:58] requiring you to boot up PyTorch, um, [7:01] then you might actually have some good [7:02] gains. So one of the biggest lessons we [7:04] had was a returning to old thoughts on [7:07] how do we make computation cheaper [7:10] enabled us to do something cool like [7:12] being able to fund raise and say I need [7:14] you to pay my enormous AWS budget but [7:16] proving that there's no one else on [7:17] earth that could do it cheaper than the [7:18] way I do it. So if anyone's in that same [7:20] spot strongly encourage you not to try [7:23] to find a cooler model but find a more [7:25] boring model to solve your problem. Uh [7:28] that tends to work really well. Also, [7:31] I'm sure everyone was got emails from [7:33] your colleagues and your bosses about uh [7:36] DeepSeek when it came out. You were all [7:38] asked the question, are we doing all [7:39] these things? Uh we got all those same [7:41] questions and we were we were excited to [7:43] say yes, we were broke and poor and we [7:46] had to do all of these things. And so [7:48] there's also another great lesson with [7:49] which is sometimes resource constraints [7:51] make really elegant solutions. Uh we [7:54] were doing you know quantization before [7:57] it was cool. We're like the hipsters of [7:59] quantization and small models. Uh so [8:02] that's another huge takeaway.