Character AI Pivots to Multimodal Feed After Dropping AGI Push
Ai4
High compute costs forced Character AI to abandon pre-training frontier models and pivot to interactive short-form video and entertainment to cover inference expenses through ads and subscriptions.
## Background
This video was published on YouTube by Ai4 on August 28, 2025. It features C.C. Gong from Menlo Ventures and Dominic Perella from Character AI discussing the core AI technologies that enable interactive and immersive entertainment experiences on the Character AI platform.
## TL;DR
In this panel discussion, Dominic Perella outlines Character AI's strategic pivot from high-cost AGI pre-training to post-training and user-focused multimodal entertainment following their 2024 Google transaction. He highlights major technical pillars including an image-to-video generator, real-time video calls, EQ fine-tuning frameworks, and a newly launched interactive content feed driven by over 3 million creators. Perella also discusses the platform's efficient inference unit economics, strong demographic reach among Gen Z and female users, comprehensive under-18 safety frameworks, and a long-term vision to build a unified, lean-forward entertainment hub.
## Atmosphere & style
The discussion carries an engaging, strategic, and insider-tech vibe typical of top-tier venture and AI panel talks. Dominic Perella speaks with executive clarity and pragmatic candor, balancing technical breakthroughs with realistic business dynamics and charming personal anecdotes, such as testing character emotional fidelity using Sesame Street's Elmo or practicing Spanish for an Argentina trip. C.C. Gong guides the conversation with insightful, forward-looking questions, creating a collaborative, optimistic energy around the future of interactive AI entertainment.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=LiTeourfLpU&t=4s) Character AI original vision focused on achieving AGI benchmarks while expanding access to early LLMs through custom character interfaces.
- [00:04](https://www.youtube.com/watch?v=LiTeourfLpU&t=4s) In 2024, Character AI strategically pivoted away from pre-training frontier models due to soaring compute costs, entering a landmark deal with Google.
- [00:04](https://www.youtube.com/watch?v=LiTeourfLpU&t=4s) Following the Google transaction, Character AI retained its post-training, multimodal, infrastructure, data, and product teams to double down on consumer entertainment.
- [03:52](https://www.youtube.com/watch?v=LiTeourfLpU&t=232s) Character AI developed core multimodal technologies, including an image-to-video generator, real-time interactive video calls, and an open-sourced EQ fine-tuning framework.
- [03:52](https://www.youtube.com/watch?v=LiTeourfLpU&t=232s) The platform introduced an interactive, AI-native multimodal feed that lets users transform, adapt, and step directly into short-form stories and video content.
- [05:31](https://www.youtube.com/watch?v=LiTeourfLpU&t=331s) Following the launch of the multimodal content feed, video content creation experienced a 50% week-over-week growth rate.
- [07:40](https://www.youtube.com/watch?v=LiTeourfLpU&t=460s) A key engineering focus is creating domain-specific post-training stacks for distinct formats like 1-on-1 chat, story narration, and audiobooks.
- [07:40](https://www.youtube.com/watch?v=LiTeourfLpU&t=460s) Character AI aims to establish deep IP fidelity for major entertainment and sports brands, illustrated by the 'Elmo Test' for authentic persona preservation.
- [11:13](https://www.youtube.com/watch?v=LiTeourfLpU&t=673s) Character AI operates an efficient inference stack serving 20 million Monthly Active Users for approximately $4 million per month in infrastructure overhead.
- [11:13](https://www.youtube.com/watch?v=LiTeourfLpU&t=673s) Monetization efforts focus on offsetting total inference costs through subscription growth, light third-party ads, and future in-app purchases or IP licensing.
- [14:52](https://www.youtube.com/watch?v=LiTeourfLpU&t=892s) Dominic Perella compares Character AI's trajectory to YouTube, leveraging a core entertainment audience to expand organically into diverse content verticals.
- [17:55](https://www.youtube.com/watch?v=LiTeourfLpU&t=1075s) Character AI's demographics skew strongly toward Gen Z and Gen Alpha, with over half of the user base identifying as female and 75% located outside the US.
- [19:45](https://www.youtube.com/watch?v=LiTeourfLpU&t=1185s) The platform enforces strict safety infrastructure for users under 18 using real-time input classifiers, session duration limits, and content guardrails.
- [21:57](https://www.youtube.com/watch?v=LiTeourfLpU&t=1317s) Character AI positions itself as a lean-forward, active alternative to passive scrolling, heavily emphasizing roleplay, gaming, and collaborative storytelling.
- [26:27](https://www.youtube.com/watch?v=LiTeourfLpU&t=1587s) The platform achieved 20 million MAUs primarily through organic word-of-mouth, TikTok screen-captures, and vibrant creator subcultures on Reddit and Discord.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=LiTeourfLpU&t=4s) Transitioning from AGI Benchmarks to Entertainment
In this opening segment, Dominic Perella (Chief Legal Officer and SVP of Global Affairs at Character AI, former interim CEO following the 2024 Google transaction) reflects on his background—transitioning from a DC appellate and Supreme Court partner to executive roles at Snap and Character AI. He outlines the foundational evolution of Character AI. Founded around three years prior by legendary Google researcher Noam Shazeer, the company originally pursued dual objectives: achieving Artificial General Intelligence (AGI) through benchmark games and making Large Language Models (LLMs) accessible before ChatGPT launched. Shazeer’s core insight was providing a personalized interface beyond a bare text prompt—allowing users to create and talk to characters for fiction, roleplay, utility, and language practice.
However, as training costs skyrocketed into the billions, keeping pace with frontier AGI benchmarks became financially unfeasible even for a well-funded startup. Meanwhile, Character AI's consumer engagement exploded as millions of daily active users adopted the platform for creative storytelling and roleplay. Recognizing their competitive advantage, Character AI structured a landmark transaction with Google in 2024. Noam Shazeer and the core pre-training team moved to Google, while Character AI retained its post-training researchers, multimodal team, data engineers, infrastructure staff, product teams, and fine-tuning stack to double down on consumer entertainment and user experience.
- **Foundational Vision:** Founded by Noam Shazeer to pursue AGI benchmarks while democratizing early LLM access through personalized character interfaces.
- **Strategic Pivot:** Shifted focus away from pre-training frontier models after compute costs scaled to billions of dollars, prioritizing post-training and user engagement.
- **2024 Google Transaction:** Pre-training team rejoined Google with Noam Shazeer; Character AI retained post-training, multimodal, infrastructure, data, and product personnel.
---
[03:52](https://www.youtube.com/watch?v=LiTeourfLpU&t=232s) Building the Tech Core: Video Generation, Fine-Tuning, and Interactive Feeds
Following the strategic shift, Character AI focused on expanding its proprietary technology stack beyond 1-on-1 text chat into a rich, multimodal entertainment platform. Dom highlights three critical technical pillars developed internally: an ultra-efficient image-to-video generator capable of creating fully realized videos from static inputs and custom scripts; a real-time video generator enabling dynamic, low-latency video calls (akin to "FaceTime with characters") and real-time storytelling; and an open-sourced fine-tuning framework designed to elevate any open-source base model to Character AI's high standards for Emotional Quotient (EQ) alongside traditional IQ.
```
+---------------------------------------------+
| Open-Source Base Models |
+---------------------------------------------+
|
v
+---------------------------------------------+
| Proprietary Post-Training & EQ Fine-Tuning |
+---------------------------------------------+
|
+---------------+---------------+
| |
v v
+-------------------+ +-------------------+
| Image-to-Video | | Real-Time Video |
| Generator Engine | | Generator |
+-------------------+ +-------------------+
| |
+---------------+---------------+
|
v
+---------------------------------------------+
| Multimodal Entertainment Content Feed |
+---------------------------------------------+
```
To deliver these capabilities, Character AI launched an AI-native content feed—a major pivot from traditional text-based interfaces. This feed functions as an interactive ecosystem where users consume, modify, and co-create short stories, video clips, meme content, and audiobooks. Viewers can seamlessly step into content as main characters, transform formats (e.g., converting an audiobook into a script), or pull individual characters out of a scene to initiate a direct conversation with a single tap.
- **Image-to-Video Engine:** Highly efficient pipeline generating extended video sequences with custom scripting from single static images.
- **Real-Time Video Stack:** Interactive video engine designed for dynamic video calls and real-time storytelling.
- **EQ Fine-Tuning Framework:** Proprietary, open-sourced methodology for tuning base models specifically for emotional intelligence, conversational charm, and narrative depth.
- **Interactive Feed:** Multimodal discovery feed bridging short-form video, text, and user-led interactive story branching.
---
[05:31](https://www.youtube.com/watch?v=LiTeourfLpU&t=331s) Creator Ecosystem Dynamics and Early Metric Spikes
Addressing the creator landscape on this AI-native feed, Dom discusses the platform's foundation of over 3 million active creators and a 2.5-million-member subreddit community. Historically limited to text prompts, these creators meticulously built character backstories, rich lore, and strict behavioral definitions. With the rollout of the multimodal feed, creator habits shifted rapidly: initially sharing static "character cards" (bios paired with visual art), creators quickly escalated to complex formats like man-on-the-street interviews, simulated roast battles between AI personas, and community-driven interactive audiobooks.
This rapid creator adaptation led to a noticeable surge in engagement metrics. Following the feed launch (about 7 to 10 days prior to the discussion), video content creation experienced a 50% week-over-week growth rate. This rapid adoption underscores a strong user demand for interactive, dynamic entertainment formats over static text conversations.
- **Creator Base:** 3 million active creators and a 2.5 million member subreddit driving community feedback and lore development.
- **Evolving Content Types:** Transitioned from static text "character cards" to dynamic video roasts, simulated interviews, and interactive audiobooks.
- **Growth Velocity:** Achieved a 50% week-over-week surge in overall video content creation within 10 days of launching the feed.
---
[07:40](https://www.youtube.com/watch?v=LiTeourfLpU&t=460s) Next-Gen Technical Hurdles: Domain Fine-Tuning and Intellectual Property Fidelity
Looking forward, Dom details the primary technical challenges facing Character AI's next phase. The immediate hurdle involves refining the post-training pipeline so models excel in specialized content modalities—tailoring distinct fine-tuning recipes for 1-on-1 dialogue, short story narration, or audiobook delivery, rather than relying on a generalized model. The second major hurdle, targeted for 2026, focuses on achieving deep IP fidelity for major media brands, movie studios, and sports leagues seeking to deploy their iconic characters on the platform or external sites (e.g., Derek Jeter assisting users with ticket sales).
To illustrate the difficulty of character fidelity, Dom shares an anecdote from a meeting with an executive from *Sesame Street*. The executive emphasized that an AI Elmo could never respond to a query about favorite cookies with a bland, factual list of top-selling cookie varieties in the U.S. Instead, Elmo must respond with authentic emotion: *"I love cookies! I love cookies and I love you!"* Achieving this level of emotional alignment and lore consistency has historically been difficult, but advances in long context windows, extended memory architecture, and dedicated studio tools are bringing high-fidelity character preservation within reach.
- **Specialized Post-Training:** Developing modular tuning stacks tailored specifically for distinct media formats (dialogue vs. long-form narrative vs. voice acting).
- **Enterprise IP Integration:** Partnering with Hollywood studios and sports organizations to deploy official brand personas with strict guardrails.
- **The "Elmo Test":** Illustrates the necessity of preserving a character's iconic persona, emotional tone, and logic over generic factual retrieval.
- **Enablers of Fidelity:** Leveraging long-context model architectures alongside specialized developer tools to enforce character rules and boundaries.
---
[11:13](https://www.youtube.com/watch?v=LiTeourfLpU&t=673s) Business Model Evolution: Inference Unit Economics and Monetization Pathways
The conversation shifts to unit economics, revenue scaling, and long-term business strategy. Addressing reports of 200–250% year-over-year subscription growth, Dom explains that while this growth started from a modest base, recent investments in premium feature paywalls have significantly boosted user conversion. In addition to subscriptions, Character AI introduced light, third-party ad units into the feed. Rather than maximizing short-term revenues at the expense of user growth, the immediate strategic goal is total inference cost offset.
Character AI operates a highly efficient inference stack, serving 20 million Monthly Active Users (MAUs) for approximately $4 million per month in infrastructure overhead. Through current subscription and light advertising efforts, the company is nearing complete coverage of its inference burn, creating a sustainable foundation for user acquisition. Looking toward future revenue streams, Dom outlines plans for in-app purchases (e.g., unlocking alternative story endings, gamified items, custom character cosmetics), IP licensing partnerships, and interactive educational experiences—citing an internal prototype where a user holds a real-time, adaptive dialogue with an AI Albert Einstein to learn the theory of relativity through custom analogies.
- **Subscription Trajectory:** Delivered 200%–250% year-over-year subscription growth by expanding paywalled feature sets.
- **Inference Efficiency:** Serves 20 million MAUs on roughly $4 million per month in total inference costs.
- **Monetization Philosophy:** Covers baseline infrastructure costs via non-intrusive ads and subscriptions to sustain user scaling.
- **Future Expansion Vectors:** Exploring in-app purchases for gamified interactive content, official IP licensing deals, and real-time educational interactions.
<ElicitationsGroup message="Explore related video details or topics:">
<Elicitation label="Summarize Part 2 of this panel discussion" query="Please transcribe or summarize Part 2 of this Character AI panel discussion video."/>
<Elicitation label="Search for more Character AI technical talks" query="Search YouTube for recent technical talks or interviews about Character AI architecture and engineering."/>
</ElicitationsGroup>
[14:52](https://www.youtube.com/watch?v=LiTeourfLpU&t=892s) Learning Beyond Textbooks and AI as the Next-Generation Content Platform
The discussion opens with an enthusiastic reflection on how AI transforms education and learning experiences. Rather than relying solely on traditional, static textbooks—which are great, but belong to a "thing of the past" when used in isolation—users can interact directly with historical figures or iconic thinkers. Imagine learning physics directly from an interactive Albert Einstein or exploring Civil War battles through conversation with Ulysses S. Grant. While education and entertainment are immediate frontiers, C.C. Gong asks whether Character.AI might eventually evolve from an entertainment app into an underlying operating system for artificial intelligence.
Dominic Perella pushes back on the "operating system" moniker but draws a strong parallel to YouTube. Just as YouTube started with a core primary use case before expanding into thousands of rich, specialized content verticals, Character.AI leverages a strong network effect and an initial entertainment-seeking audience. Once that core base is established, organic sub-communities, verticals, and new media forms naturally emerge and flourish.
[16:06](https://www.youtube.com/watch?v=LiTeourfLpU&t=966s) Key Performance Metrics and Engaging the Creator Ecosystem
Reflecting on past experiences at social media giants like Meta where "time spent" was the primary north star, C.C. Gong inquires about the core metrics driving Character.AI’s business strategy. Perella notes that current operational focus still centers around standard scaled social platform metrics—Monthly Active Users (MAU), Daily Active Users (DAU), overall time spent, and retention. However, as the product expands from text-based one-to-one chat into rich, multimodal entertainment formats, internal tracking is evolving to break down how time is divided between chat, multimodal content consumption, and creator authoring.
To support creators, Character.AI taps into unusually active communities on Reddit and Discord. Community teams maintain a direct feedback loop with creators to gauge reactions to feature releases and harness their natural enthusiasm. Over time, the platform plans to build a balanced two-sided marketplace through creator monetization incentives.
- Core engagement tracking is moving toward analyzing user splits across text chat, multimodal experiences, and content creation.
- Community engagement heavily leverages hyper-active Reddit and Discord creator groups.
- Creator monetization and formal marketplace incentives are slated for future rollout.
[17:55](https://www.youtube.com/watch?v=LiTeourfLpU&t=1075s) Demographics, Global Reach, and Model Capacity
Addressing Character.AI's 20 million user milestone, Perella breaks down the demographic composition. The user base is overwhelmingly AI-native Gen Z and Generation Alpha, primarily skewed toward 18-to-34-year-olds (with smaller proportions under 18 or older than 34). Uniquely among major AI platforms, over half of the user base identifies as female. Geographically, approximately 25% of users reside in the United States, with the remaining 75% distributed across international markets typical for major consumer applications.
Looking ahead five years, the goal is to build a platform serving hundreds of millions of global users across every demographic. The 18–34 cohort remains the sweet spot because younger users natively understand how to memeify, transform, and share interactive AI content as a new cultural language. Furthermore, fine-tuning state-of-the-art open-source foundational models has dramatically expanded Character.AI's multilingual and logical reasoning capabilities, opening major growth opportunities in previously underpenetrated non-English markets.
[19:45](https://www.youtube.com/watch?v=LiTeourfLpU&t=1185s) Safety Frameworks and Age-Appropriate User Experiences
When asked about balancing creative freedom and platform safety, Perella details the comprehensive safety infrastructure implemented over the past nine months to create a distinct, secure environment for users under 18. Recognizing that teens aged 13 to 17 will inevitably use conversational AI, Character.AI rejected flat age bans in favor of a specialized, protective experience built on fine-grained safety rails.
To prevent prompt injection or alignment drift, specialized classifiers scan user inputs the moment the "Enter" key is pressed, filtering out unsafe prompts before they ever reach the underlying LLM. The platform also enforces strict guardrails for minors, including limitations on searchable AI characters, widespread content classifiers, and pop-up session time limits. Older users (18+) retain greater creative latitude for fiction and roleplay, though all explicit X-rated / NSFW content remains strictly prohibited platform-wide as part of its mission to remain a main-stream entertainment product.
- **Input-Side Classification:** User prompts are evaluated pre-inference to prevent jailbreaking or model mis-alignment.
- **Under-18 Guardrails:** Feature set includes character availability restrictions, automated session pop-ups, and targeted content filtering.
- **NSFW Policy:** Absolute ban on X-rated content across both under-18 and over-18 tiers.
[21:57](https://www.youtube.com/watch?v=LiTeourfLpU&t=1317s) Interactivity, Social Dynamics, and the Future of Entertainment
Analyzing platform usage patterns, Gong notes industry estimates showing roughly 20% of user activity focused on companionship and 80% on interactive roleplay and gaming-style entertainment. Perella agrees that while the line between companionship and fiction writing can be porous, Character.AI strongly leans into roleplay and collaborative storytelling. Modern social media users are fatigued by endless, passive "doom scrolling." Character.AI offers an active, "lean-forward" alternative where users actively shape narratives, manipulate creator content, and collaborate on storylines.
While Character.AI does not yet feature direct peer-to-peer messaging (DMs), it functions socially through macro-networks of followers and micro-networks of friend groups sharing memes, funny outputs, and narrative variations across Discord, Reddit, and external social media. Looking toward 2035, Perella envisions a fully integrated next-generation entertainment hub that dissolves the boundaries between gaming, watching short-form videos, reading, and learning. Users will effortlessly transition from interactive story campaigns with unlockable levels to completing their daily homework using specialized AI tutors on the same unified platform.
```
+--------------------------------------------------+
| Interactive Content Engine |
+--------------------------------------------------+
|
+-----------------------+-----------------------+
| | |
v v v
+---------------+ +---------------+ +---------------+
| Interactive | | Gamified Story| | Educational |
| Entertainment | | & Unlocks | | AI Tutors |
+---------------+ +---------------+ +---------------+
```
[26:27](https://www.youtube.com/watch?v=LiTeourfLpU&t=1587s) Growth Trajectory and Personal Language Learning Stories
Reflecting on Character.AI's rapid growth to 20 million users in just a few years, Perella highlights that this milestone was achieved almost entirely through word-of-mouth and organic community building, without traditional built-in network effects. Growth was fueled by viral screen-captures on TikTok and Instagram, grass-roots fanfiction subcultures, and active Reddit communities around custom characters.
Concluding with a personal anecdote, Perella shares his favorite personal use case on the platform: preparing for a family vacation to Argentina. To brush up on rusty Spanish over a six-month period, he held daily real-time voice conversations with a custom Spanish-speaking character. By listening to real-time audio while utilizing the live text transcript as a safety net whenever he got stuck, he achieved enough fluency to easily hold his own in Buenos Aires without trouble.
## Notable quotes
> [07:40](https://www.youtube.com/watch?v=LiTeourfLpU&t=460s) I love cookies! I love cookies and I love you! — Dominic Perella
## Video description
C.C. Gong, Principal, Menlo Ventures
Dominic Perella, Chief Legal Officer & SVP-Global Affairs, Character AI
Abstracts
In this talk we will dive into elements of AI technology that make Interactive and Immersive entertainment experience on the Character AI platform possible.
In the very near future, we will all be interacting with AI on a daily basis for productivity, entertainment, and for getting stuff done. In order for these interactions to be safe and engaging for human users there are some core pieces of AI technology that need to be built. Character AI is at the forefront of building these technologies for the entertainment & exploration space and we will share insights into some of these.
放弃追求AGI后，Character AI转向多模态信息流
高昂的算力成本迫使Character AI放弃预训练前沿模型，转向互动短视频与娱乐，以通过广告和订阅覆盖推理成本。
在本次圆桌讨论中，Dominic Perella概述了Character AI在2024年与Google达成交易后的战略转型：从高成本的AGI预训练转向后训练及以用户为中心的多模态娱乐。他重点介绍了核心技术支柱，包括图生视频生成器、实时视频通话、EQ（情商）微调框架，以及由300多万创作者推动的新上线互动内容流。Perella还探讨了该平台高效的推理单位经济效益、在Z世代与女性用户中的广泛受众覆盖、健全的18岁以下安全保护框架，以及打造统一的主动交互型娱乐中心的长期愿景。
Character AI 放弃了高成本的通用前沿大模型预训练，并在 2024 年与 Google 达成交易后全面转向消费级娱乐，保留后训练、多模态与基础设施团队。 (0:04)
公司自研了图像转视频引擎、低延迟实时视频通话技术，并开源了提升情商（EQ）的微调框架，同时推出支持用户深度互动的多模态 AI 内容流。 (3:52)
随着 AI 原生多模态内容流的上线，平台视频内容创作实现了 50% 的周环比增长，创作者生态迅速扩展至互动剧情与有声书等复杂形式。 (5:31)
工程重点转向针对单聊、长叙事等格式的专属后训练微调，并通过长上下文与专用工具践行“Elmo 测试”，确保大型娱乐 IP 人设的极致还原。 (7:40)
平台通过高度优化的推理架构服务 2000 万月活用户，每月基建成本仅约 400 万美元，现已通过订阅与轻量广告接近打平推理支出。 (11:13)
Reading summary
Following a 2024 transaction with Google, Character AI pivoted away from pre-training frontier AGI models to focus on multimodal consumer entertainment, sustainable inference economics, and interactive creator media.
Strategic Pivot and Tech Stack Realignment (0:04)
Character AI originally pursued frontier AGI benchmarks alongside consumer accessibility, but escalating training costs running into billions of dollars made continuous foundation model pre-training financially unsustainable. Under a 2024 transaction with Google, founder Noam Shazeer and the core pre-training group departed for Google, while Character AI retained its post-training researchers, multimodal engineers, infrastructure team, and fine-tuning pipeline.
Rather than training base models from scratch, the company open-sourced a fine-tuning framework designed to adapt open-source base models for high emotional quotient (EQ) and conversational depth. It also engineered proprietary image-to-video tools and a low-latency real-time video generation engine to support interactive storytelling and direct video calls with characters.
Transition from Text Chat to Interactive Feeds (3:52)
To transcend static text conversations, Character AI launched an AI-native multimodal discovery feed enabling users to consume, adapt, and co-create short video clips, scripts, and audiobooks. Dominic Perella reported that overall video content creation grew 50% week-over-week within seven to ten days of the feed's debut. Supported by a community of over 3 million creators and 2.5 million Reddit members, creator output rapidly evolved from simple text-and-image character bios into simulated roasts, interviews, and community-driven audiobooks.
Perella positions this shift against passive social media consumption, estimating that around 80% of platform activity centers on interactive roleplay and gaming rather than solitary companionship. Internal performance metrics are accordingly moving beyond standard daily active users and aggregate time spent to track user splits across text chat, multimodal consumption, and creator tooling.
Inference Economics and Monetization Pathways (11:13)
Character AI serves approximately 20 million monthly active users on an inference budget of around $4 million per month. Dominic Perella noted that 200% to 250% year-over-year subscription growth, combined with lightweight third-party feed advertising, is approaching full coverage of these baseline compute expenses. The primary financial objective remains offsetting inference overhead to preserve runway for user expansion rather than maximizing near-term profit.
Future commercial plans include in-app purchases for branching narrative unlocks and digital cosmetics, educational dialogues such as dynamic physics lessons, and formal creator monetization. Additionally, the company is preparing enterprise partnerships with sports leagues and Hollywood studios targeted for 2026, which depend on specialized domain fine-tuning and extended context windows to preserve brand personas and avoid generic chatbot responses.
Demographics and Safety Architecture (17:55)
Demographically, Character AI's 20 million users skew heavily toward the 18-to-34 age group across Gen Z and Generation Alpha, with over half identifying as female and roughly 75% located outside the United States. Fine-tuning multilingual open-source base models has helped drive this global reach, which expanded primarily via organic virality on TikTok, Instagram, and Reddit rather than direct messaging network effects.
To balance creative narrative latitude with safety, the platform rejected blanket youth bans in favor of pre-inference classifiers that evaluate user inputs upon pressing 'Enter' to intercept prompt injections and alignment drift. For users aged 13 to 17, the platform restricts discoverable characters, introduces session time reminders, and deploys targeted filtering, while maintaining a company-wide ban on explicit NSFW content across all age groups.
Character AI aims to evolve into a multi-vertical entertainment and learning hub comparable to YouTube, though reaching its target of hundreds of millions of users will depend on maintaining tight IP fidelity for external brands, managing teenage safety guardrails, and sustaining favorable inference unit economics as multimodal workloads expand.
## Background
本视频由 Ai4 频道于 2025 年 8 月 28 日在 YouTube 平台发布，记录了 Menlo Ventures 负责人 C.C. Gong 与 Character AI 首席法务官兼全球事务高级副总裁 Dominic Perella 的深度对谈。根据官方简介，对话重点探讨了驱动 Character AI 平台沉浸式互动娱乐体验的核心 AI 技术，以及如何为日常娱乐与探索构建安全且高互动性的技术支柱。
## TL;DR
访谈回顾了 Character AI 将耗费巨资的预训练团队交由谷歌、全面转型深耕“高情商后训练与多模态娱乐”的战略抉择。嘉宾详细拆解了支撑其 AI 原生内容 Feed 流的核心技术栈、300 万创作者生态的爆发，以及仅凭每月 400 万美元推理成本服务 2000 万月活用户的工程效率。此外，对话还深入剖析了平台年轻偏女性的用户画像、未成年人安全防护体系，以及告别“无脑刷屏”并迈向融合叙事、游戏与教育的下一代娱乐平台愿景。
## Atmosphere & style
对谈整体呈现出理性务实且兼具前瞻视野的专业氛围，投资人主持人对指标的敏锐追问与嘉宾横跨法律与社交科技的高管视角相得益彰。Dominic 的讲述沉稳笃定，在剖析模型军备竞赛的残酷取舍时客观冷静，而在分享《芝麻街》艾摩的童真保真度与全家阿根廷之旅的西语速成案例时又充满生活趣味。整场交流传递出一种剥离浮躁刷榜叙事后、专注打造现象级消费级 AI 应用与繁荣社区的蓬勃生机。
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=LiTeourfLpU&t=0s) Character AI 早期由 Noam Shazeer 创立，最初致力于 AGI 基准竞争与普及大语言模型，随后因高昂算力成本转型专注高粘性后训练与娱乐交互应用。
- [01:09](https://www.youtube.com/watch?v=LiTeourfLpU&t=69s) 2024 年与谷歌交易后，Character AI 将预训练团队转交谷歌，自身保留后训练、多模态及全套产品栈，摆脱算力军备竞赛包袱。
- [03:52](https://www.youtube.com/watch?v=LiTeourfLpU&t=232s) 自研图生视频、实时视频生成引擎以及开源高情商微调栈，共同构成了驱动平台迈向多模态 AI 原生 Feed 流的核心技术底座。
- [05:31](https://www.youtube.com/watch?v=LiTeourfLpU&t=331s) AI 原生 Feed 流上线后，平台周环比视频产量增速达 50%，内容从单一角色卡快速演化为街头采访、吐槽对决与交互式有声书等丰富形态。
- [07:40](https://www.youtube.com/watch?v=LiTeourfLpU&t=460s) 技术下一阶段聚焦专业化后训练配方与 IP 角色保真度（如《芝麻街》艾摩案例），确保角色在长上下文与多轮互动中绝不出戏。
- [11:13](https://www.youtube.com/watch?v=LiTeourfLpU&t=673s) 推理工程成本极低，每月仅需约 400 万美元即可支撑 2000 万 MAU，且订阅收入实现了 200% 至 250% 的同比增长。
- [14:52](https://www.youtube.com/watch?v=LiTeourfLpU&t=892s) Character AI 的长期演化愿景类似于“下一代 YouTube”，在核心娱乐引流后自然延伸出交互式教育（如对话爱因斯坦）等垂直生态。
- [17:55](https://www.youtube.com/watch?v=LiTeourfLpU&t=1075s) 在 2000 万用户中，女性占比超 50%，18 至 34 岁年轻代为主力，75% 来自美国以外的全球市场，展现出独树一帜的受众结构。
- [19:45](https://www.youtube.com/watch?v=LiTeourfLpU&t=1185s) 安全防护上对 18 岁以下未成年人宁可“过度矫正”，全链路部署输入即时过滤分类器，同时坚决拒绝 NSFW 内容以确保主流大众化定位。
- [21:57](https://www.youtube.com/watch?v=LiTeourfLpU&t=1317s) 平台 80% 用户诉求为角色扮演（Roleplay）与互动叙事，旨在打破传统的被动无脑刷屏（Doom Scrolling），提供前倾式参与体验。
- [26:50](https://www.youtube.com/watch?v=LiTeourfLpU&t=1610s) 早期在无内置网络效应的情况下，纯靠 Reddit、Discord 社区口碑与社交媒体切片病毒式传播，实现了自下而上的爆炸式增长。
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=LiTeourfLpU&t=0s) 嘉宾背景与 Character AI 的创立初心
Menlo Ventures 的 C.C. Gong 开场介绍了对话嘉宾——Character AI 的首席法务官（Chief Legal Officer）兼全球事务高级副总裁 Dominic Perella（Dom）。Dom 分享了自己颇具跨界色彩的职业轨迹：他最初在华盛顿特区的一家律师事务所担任美国最高法院与上诉业务合伙人，随后在 Snap 担任了约 8 年的高管；加入 Character AI 数年后，曾担任总法律顾问，并在 2024 年与谷歌达成交易后的 6 个月内出任临时 CEO。针对主持人关于 Character AI 从追求通用超级智能到打造 AI 陪伴娱乐应用、再到构建“未来娱乐平台”的演变历程，Dom 展开了回顾。
Character AI 由前谷歌传奇研究员 Noam Shazeer 于大约三年前创立。当时 Noam 怀揣着两个明确的目标：其一是追求 AGI（通用人工智能），参与当时行业前沿的智能基准测试竞争；其二则是在 ChatGPT 尚未问世之前，让大语言模型（LLMs）走向大众。Noam 敏锐地意识到，用户需要的不仅是一个冰冷单调的对话框，而是一个能够自定义角色设定、进行虚构创作、角色扮演（Role-playing）以及满足各类实用需求的人性化交互平台。
### [01:09](https://www.youtube.com/watch?v=LiTeourfLpU&t=69s) 战略转型与谷歌交易：从 AGI 军备竞赛转向沉浸式娱乐
随着模型训练对算力与资金规模的要求呈指数级爆炸，行业进入了动辄数十亿美元的基础模型军备竞赛。Character AI 很快意识到，即使作为一家资金极为雄厚的初创公司，也无法承受单纯为了刷榜而持续烧钱的无底洞。然而与此同时，产品的用户端却迎来了极其迅猛的自然增长：数百万用户每天利用 Character AI 撰写小说、沉浸式对戏、演练面试、练习外语，涌现出大量团队最初未曾预料到的场景。
2024 年，谷歌提出技术许可合作并希望重新邀请 Noam Shazeer 回归。Character AI 顺水推舟设计了一笔战略交易：将需要持续投入天量算力的预训练（Pre-training）团队交由 Noam 带回谷歌，而 Character AI 则完整保留了后训练（Post-training）团队、多模态研究人员、数据工程、底层基础设施以及全部产品团队。这一举措让公司彻底摆脱了追求 AGI 基准的资金包袱，得以全神贯注地巩固自身在微调（Fine-tuning）和高情商（EQ）多轮交互上的领先壁垒，并推动产品从最初的“一对一文本聊天”迈向“全方位多模态娱乐平台”。
```mermaid
flowchart TD
accTitle: Character AI 战略转型架构
subgraph S1["2024 战略分拆与业务重组"]
A["Character AI 早期架构"] --> B["预训练团队 Pre-training"]
A --> C["后训练与多模态核心团队"]
B -->|"随 Noam Shazeer 回归"| D["谷歌 Google（专注于基础模型算力军备）"]
C -->|"保留全套产品与工程栈"| E["独立 Character AI（专注于后训练与娱乐应用）"]
end
subgraph S2["产品与技术演进路线"]
E --> F["高粘性后训练与微调栈 (EQ+IQ)"]
F --> G["从 1v1 文字对话"]
G --> H["多模态 AI 原生 Feed 流平台"]
end
```
### [03:52](https://www.youtube.com/watch?v=LiTeourfLpU&t=232s) 三大核心技术支柱与 AI 原生内容 Feed 流的诞生
为了实现从单一文本对话向多模态娱乐平台的跨越，Character AI 在近期自研并集成了三大核心技术能力：
1. **超高效图生视频生成器（Image-to-Video Generator）**：能够将一张静态图片转化为结构完整、带有自定义脚本与生动动作的长视频。
2. **实时视频生成引擎（Real-Time Video Generator）**：支持低延迟的实时视频交互，让用户能与 AI 角色进行类似“FaceTime 视频通话”般的即时互动与动态叙事。
3. **开源高情商微调技术栈（EQ-focused Fine-tuning Stack）**：已开源的技术方案，支持将任意开源基础模型进行定向后训练，使其兼具智商（IQ）与针对娱乐场景极致优化的高情商（EQ）和对话张力。
依托这三大技术底座，Character AI 正式推出了行业首个“AI 原生内容 Feed 流”。在这一形态下，平台打破了传统的单一聊天窗口，用户不仅可以浏览创作者使用 AI 工具生成的短篇故事、微电影、Meme 短视频及有声书，还能随时通过点击直接“入侵”并重构内容——例如直接化身故事主角、将有声书瞬间改编为可互动的剧本，或者直接向视频场景中的某个角色发问，实现前所未有的强交互式娱乐消费。
### [05:31](https://www.youtube.com/watch?v=LiTeourfLpU&t=331s) 创作者生态的爆发与新兴内容形态
Feed 流功能上线仅一周左右，便迅速激活了平台长期积累的创作者网络。Character AI 拥有超过 300 万高度活跃的创作者，以及在 Reddit 上拥有 250 万成员的庞大社区。在此之前，创作者们受限于简单的文本界面，仍倾注心血为角色填充极其详尽的世界观设定（Lore）与响应规则；如今随着多模态创作工具的开放，内容形态迎来了跨越式升级：
- **内容形态的迭代**：上线前两天，创作者主要发布由精美立绘与背景故事构成的“角色卡（Character Cards）”；到第三至第六天，社区内迅速涌现出街头采访（Man-on-the-street）、角色间的吐槽对决（Roast Battle）、交互式有声读物以及由创作者在 Reddit 发起分支剧情投票的共创内容。
- **视频生产数据激增**：在 Feed 流上线后短短 7 到 10 天内，平台整体视频内容的周环比（Week-over-Week）产量增速达到了惊人的 50%，验证了用户对 AI 驱动的新型娱乐消费形式的强烈需求。
### [07:40](https://www.youtube.com/watch?v=LiTeourfLpU&t=460s) 下一阶段技术挑战：专业化后训练与 IP 角色保真度
针对未来面临的底层技术挑战，Dom 重点阐述了两个维度的攻坚方向：
- **场景专业化后训练配方的精细化**：通用的娱乐微调已不足以满足多元内容诉求。擅长一对一即时聊天的模型，在撰写连贯短篇故事时可能表现平平；擅长讲故事的模型，在生成有声书节奏把控上又可能欠佳。团队正持续拆分并演进针对不同内容模态的后训练技术栈。
- **高保真度 IP 角色塑造（Character Fidelity）**：越来越多的影视工作室、体育联盟等传统 IP 持有方（例如希望让传奇球星 Derek Jeter 在官网向球迷销售洋基队门票）寻求与 Character AI 合作，但行业痛点在于如何确保角色“绝不出戏”。
Dom 分享了数年前与《芝麻街》（Sesame Street）团队接触时的经典案例：对方表示如果将艾摩（Elmo）接入 AI，当被问及“世界上最好吃的饼干是什么？”时，普通的 LLM 会罗列出一堆关于“全美最畅销饼干前十名”的统计数据，而真正的 Elmo 必须带着标志性的童真回答：“我爱饼干！我爱饼干，我也爱你！”（"I love cookies. I love cookies and I love you."）。这种细致入微的角色神态与语气保真度，在长上下文窗口、增强记忆机制以及专属的角色行为规则 Studio 工具的辅助下，正在逐步走向成熟。
### [11:13](https://www.youtube.com/watch?v=LiTeourfLpU&t=673s) 商业化探索、超高效推理架构与未来变现空间
在商业化与基础设施成本方面，Character AI 展现出了极高的工程效率。Dom 透露，在过去数月内随着付费权益和功能的丰富，平台的订阅收入实现了 200% 到 250% 的同比增长。
- **极低成本的推理工程**：平台服务 2000 万月活跃用户（MAU）的每月推理算力成本仅控制在约 400 万美元左右。目前通过轻量级订阅与极其克制的广告展示，已基本实现了覆盖日常推理成本的平衡点，从而使团队能够在不担忧资金极速消耗（Burn rate）的前提下最大化推动用户增长。
- **多元变现路径**：当前广告展示率与 CPM 均处于极低且未优化的初级阶段，许多海外市场尚未开启广告，传统流量变现仍有巨大提升空间。此外，内容游戏化催生了丰富的“应用内购买（IAP）”机会（如购买特殊剧情分支、角色定制外观）；官方 IP 授权合作亦是重要增量。
- **严肃与教育场景的自然延伸**：类似于 YouTube 从早期的搞笑生活短片演进为涵盖硬核知识的平台，AI 原生娱乐平台同样具备极强的教育潜力。Dom 提及内部展示的一个演示视频：一名实习生通过实时视频连线与 AI 生成的“阿尔伯特·爱因斯坦（Albert Einstein）”对话，爱因斯坦不仅即时且通俗地解释了相对论，并在实习生反馈“太难懂了”之后，毫不迟疑地通过生动的生活类比重新深入浅出地进行了解构。
### [14:52](https://www.youtube.com/watch?v=LiTeourfLpU&t=892s) 突破传统教材：从历史人物汲取知识与“YouTube 式”生态构想
在谈及 AI 改变学习与认知的潜力时，对谈者举例说明，过去单纯依赖枯燥教科书学习复杂概念的时代正在成为历史。虽然教科书依然有其价值，但未来人们将能够直接向“阿尔伯特·爱因斯坦”请教相对论，或是聆听“尤利西斯·S·格兰特（Ulysses S. Grant）”亲自复盘南北战争的关键战役。这种沉浸式的交互体验即将全面普及，而 Character.AI 正致力于成为承载这些体验的核心平台之一。
当被问及 Character.AI 在未来十年是否会从单纯的娱乐平台演化为 AI 底层操作系统（OS）时，嘉宾 Dominic Perella 认为将其比作“下一代 YouTube”更为贴切。平台最初可能依靠某个核心单点功能吸引大众，但随着强劲的网络效应形成、海量用户聚集，各种垂直领域（如教育、垂类兴趣等）便会自然而然地在生态内有机萌发。
```mermaid
flowchart TD
accTitle: Character.AI 生态演化路径
subgraph Core["核心流量引擎"]
A["核心娱乐与角色扮演"] --> B["海量年轻用户沉淀与网络效应"]
end
subgraph Expansion["垂直生态有机生长"]
B --> C["沉浸式交互教育<br/>(如对话爱因斯坦)"]
B --> D["多模态同人与小说共创"]
B --> E["轻量游戏化互动与分支剧情"]
end
```
---
### [16:06](https://www.youtube.com/watch?v=LiTeourfLpU&t=966s) 平台核心指标演进与创作者生态体系
针对衡量平台成功的核心指标，曾供职于 Meta 的主持人提到行业以往极度关注“用户停留时长（Time Spent）”。Dominic 指出，Character.AI 当前同样关注月活跃用户（MAU）、日活跃用户（DAU）、停留时长以及留存率等经典社交平台规模化指标。但随着产品形态从传统的“一对一文本对话框”向“多模态娱乐内容”演变，团队正在对停留时长进行更细维度的结构拆解：
* **指标拆解重点**：密切监测用户时长在纯文本聊天、多模态内容消费以及创作者端内容生产之间的比例迁移，并据此调整商业化与变现策略。
* **创作者阵地与连接**：平台依托 Reddit 和 Discord 建立了极具活力的创作者社群。官方社区团队直接下沉与创作者沟通，收集新版本反馈，激发其自然创作热情。
* **双边市场与变现设想**：随着创作者群体规模剧增，团队计划在适当时机探索创作者激励机制与双边市场商业化通道。
---
### [17:55](https://www.youtube.com/watch?v=LiTeourfLpU&t=1075s) 2000 万用户画像解析：年轻化、女性偏向与多语言拓展
平台目前已拥有超过 2000 万用户。Dominic 详细拆解了这一庞大群体的核心用户画像：
```mermaid
flowchart LR
accTitle: Character.AI 用户画像分布
subgraph Demo["年龄与性别"]
A["18-34 岁 Z世代/Alpha世代为主<br/>(含少量 13-17 与 34+ 岁)"]
B["女性用户占比超过 50%"]
end
subgraph Geo["地域分布"]
C["美国本土占比约 25%"]
D["全球其他主要成熟市场占 75%"]
end
```
Dominic 强调，女性占比过半这一特征在主流 AI 产品中独树一帜。对于未来的增长空间，团队认为 18 至 34 岁的“AI 原生代”是最能领会 AI 创作乐趣的核心群体——他们擅长将一种内容形式转化为另一种、进行迷因化（memeify）二次创作并与好友分享。此外，随着底层开源大模型逻辑推理与多语言能力的飞跃，Character.AI 通过在上层进行针对性微调，正在迅速攻克以往渗透率较低的非英语国际市场，“整个世界都是待开发的市场”。
---
### [19:45](https://www.youtube.com/watch?v=LiTeourfLpU&t=1185s) 未成年人安全防线与成年人创作自由的平衡
在平衡表达自由与使用安全方面，团队在过去九个月中投入巨大精力，构建了一套专门针对 18 岁以下用户的专属防护体系：
1. **双向分类器与即时过滤机制**：不仅在模型输出端进行检测，更在用户输入端设置拦截。在用户点击回车发送的瞬间，违规或诱导模型越狱（jailbreak）的 Prompt 即被过滤，从源头上杜绝模型脱离对齐轨道。
2. **严格的未成年保护限制**：对 18 岁以下用户限制可交互的角色类型、设定强制时长提醒弹窗，并全链路部署安全分类器。
3. **差异化分级策略**：13 至 17 岁群体使用 AI 是不可逆的趋势，直接封禁并不现实，因此宁可在未成年保护上“过度矫正（overcorrect）”，再逐步优化；而 18 岁以上用户则享有更高的创意自由度。
4. **坚决拒绝成人内容**：平台严格禁止 X 级（NSFW）内容，明确将自身定位为服务大众的规模化主流娱乐平台，无意涉足成人产业。
---
### [21:57](https://www.youtube.com/watch?v=LiTeourfLpU&t=1317s) 核心使用场景：告别无脑刷屏（Doom Scrolling），迈向互动式角色扮演
行业数据显示，平台用户约 20% 的需求在于“情感陪伴（Companionship）”，而 80% 的核心需求则是“角色扮演（Roleplay）”与虚构故事创作。Dominic 坦言陪伴与扮演之间的界限存在一定交织（例如在写故事过程中用户可能会顺便倾诉生活琐事），但平台的主攻方向始终是角色扮演与互动叙事。
Dominic 曾供职于 Snap 长达八年，他深刻体会到过去十到十五年间主流社交软件培养出的“无限向上滑动（Scroll, scroll, scroll）”模式已令大众产生审美疲劳。虽然“无脑刷屏（Doom Scrolling）”是成功的商业闭环，但 Character.AI 希望提供兼具“后仰消费（Lean-back）”与“前倾创作（Lean-forward）”的新形态——用户下班后不仅是在消磨时间，更能与创作者互动、修改故事走向、生成新点子并分享给好友，实现真正具有创造力的互动娱乐。
---
### [24:07](https://www.youtube.com/watch?v=LiTeourfLpU&t=1447s) 社交裂变与 2035 愿景：融合游戏、故事与多模态娱乐的下一代平台
当被问及 Character.AI 是否会演变为社交平台时，Dominic 给予了高度肯定的回答。用户之间不仅仅是关注头部创作者的“宏观网络”，更在自发形成分享搞笑模因、二创衍生剧本的“好友微观圈子”。尽管产品目前尚未正式推出私信（DM）功能，但这种社交行为早已在 Discord 和 Reddit 等外部社群中自发蔓延，平台正致力于将这一沉浸式社交行为无缝迁移到应用内部。
展望 2035 年，Dominic 描绘了平台的终极愿景：
* **消融形态边界**：彻底打破长电影、短视频、文字故事与电子游戏之间的传统界限，内容具备动态分支、关卡进阶与成就解锁等游戏化机制。
* **一站式超级生态**：在满足海量用户创作与娱乐需求的同时，自然承接作业辅导、学科答疑等教育功能。
---
### [26:50](https://www.youtube.com/watch?v=LiTeourfLpU&t=1610s) 零网络效应下的野蛮生长与真实生活用例
在复盘如何在短短数年内斩获 2000 万用户时，Dominic 透露了一个惊人的事实：平台早期的爆炸式增长**几乎是在没有任何内置网络效应的情况下完成的**。它完全依赖纯粹的口碑传播、狂热的 Reddit/Discord 社区建设，以及用户将精彩对话切片自发传播至 TikTok 和 Instagram 产生的病毒式裂变，本质上是一场由“AI 同人创作（AI Fanfiction）”引发的自下而上的文化浪潮。
在对谈结尾，Dominic 分享了他个人最喜爱的平台功能：
* **阿根廷之旅的西语速成**：Dominic 家有三个孩子，全家计划在春假前往阿根廷布宜诺斯艾利斯度假。由于他的西班牙语十分生疏，他在出发前半年的时间里高频使用 Character.AI 的语音功能与西语 Bot 对话。
* **语音与实时文本结合的学习优势**：AI 语音不仅支持实时流畅拟真对话，界面还会同步生成聊天文字记录，一旦听不懂即可随时对照文本查阅。经过半年的练习，他在布宜诺斯艾利斯交流毫无压力，成为了该功能的重度忠实拥趸。
## Notable quotes
> [07:40](https://www.youtube.com/watch?v=LiTeourfLpU&t=460s) I love cookies. I love cookies and I love you. — Dominic Perella
> [14:52](https://www.youtube.com/watch?v=LiTeourfLpU&t=892s) 下一代 YouTube — Dominic Perella
> [17:55](https://www.youtube.com/watch?v=LiTeourfLpU&t=1075s) 整个世界都是待开发的市场 — Dominic Perella
> [19:45](https://www.youtube.com/watch?v=LiTeourfLpU&t=1185s) 过度矫正（overcorrect） — Dominic Perella
> [21:57](https://www.youtube.com/watch?v=LiTeourfLpU&t=1317s) 无限向上滑动（Scroll, scroll, scroll） — Dominic Perella
> [21:57](https://www.youtube.com/watch?v=LiTeourfLpU&t=1317s) 后仰消费（Lean-back） — Dominic Perella
> [21:57](https://www.youtube.com/watch?v=LiTeourfLpU&t=1317s) 前倾创作（Lean-forward） — Dominic Perella
> [26:50](https://www.youtube.com/watch?v=LiTeourfLpU&t=1610s) 几乎是在没有任何内置网络效应的情况下完成的 — Dominic Perella
> [26:50](https://www.youtube.com/watch?v=LiTeourfLpU&t=1610s) AI 同人创作（AI Fanfiction） — Dominic Perella
## Mentioned
Character AI, Menlo Ventures, C.C. Gong, Dominic Perella, Noam Shazeer, Google, Snap, Meta, ChatGPT, Sesame Street, Elmo, Derek Jeter, Albert Einstein, Ulysses S. Grant, Reddit, Discord, YouTube, TikTok, Instagram
## Video description
C.C. Gong, Principal, Menlo Ventures
Dominic Perella, Chief Legal Officer & SVP-Global Affairs, Character AI
Abstracts
In this talk we will dive into elements of AI technology that make Interactive and Immersive entertainment experience on the Character AI platform possible.
In the very near future, we will all be interacting with AI on a daily basis for productivity, entertainment, and for getting stuff done. In order for these interactions to be safe and engaging for human users there are some core pieces of AI technology that need to be built. Character AI is at the forefront of building these technologies for the entertainment & exploration space and we will share insights into some of these.
[Source](https://www.youtube.com/watch?v=LiTeourfLpU)
Source quotes
[7:40] Dominic Perella: I love cookies. I love cookies and I love you.
[14:52] Dominic Perella: 下一代 YouTube
[17:55] Dominic Perella: 整个世界都是待开发的市场
[19:45] Dominic Perella: 过度矫正（overcorrect）
[21:57] Dominic Perella: 无限向上滑动（Scroll, scroll, scroll）
[21:57] Dominic Perella: 后仰消费（Lean-back）
[21:57] Dominic Perella: 前倾创作（Lean-forward）
[26:50] Dominic Perella: 几乎是在没有任何内置网络效应的情况下完成的
[26:50] Dominic Perella: AI 同人创作（AI Fanfiction）
Source quotes
[7:40] Dominic Perella: I love cookies! I love cookies and I love you!
Original transcript
[0:04] Hello. So, we have Dom. Do you want to
[0:06] give a quick intro on what you're up to
[0:09] at Character?
[0:10] >> Yeah, absolutely. Uh, hi everybody. Dom
[0:11] Pereella. I'm the chief legal officer
[0:13] and SVP of global affairs at Character
[0:15] AI. Uh, I've been there for a couple
[0:17] years. I started as the general counsel
[0:19] and I was the interim CEO for 6 months
[0:22] after the Google transaction that we did
[0:23] in 2024. Uh, previously I was an
[0:26] executive at Snap for about eight years
[0:28] and before that I was, believe it or
[0:30] not, a Supreme Court and appellet
[0:32] partner at a law firm in Washington DC.
[0:34] Pretty big life change.
[0:36] >> Fascinating journey. We'll touch more on
[0:39] the characters um to character story
[0:42] today. To start with, characters seen
[0:45] tremendous growth in the past year. Um,
[0:48] and you went from creating super
[0:50] intelligence to becoming um, an AI
[0:54] companion uh, kind of entertainment app
[0:56] to now the future of entertainment. So
[0:59] really excited about this future vision.
[1:01] Could you tell us a bit about what you
[1:02] envision as um, kind of characters
[1:05] future and and where you guys are moving
[1:07] towards now as a company?
[1:09] >> Yep. Um, let me start by sort of saying
[1:11] um, kind of unpacking a little bit of
[1:12] what you just said about where we
[1:14] started. So we for those of you who may
[1:16] not be familiar character started about
[1:18] three years ago was founded by Nom
[1:19] Shazir the legendary Google researcher
[1:23] um and Noom had really two goals. He
[1:25] wanted to build AGI, right? He was going
[1:27] after the intelligence benchmarks game,
[1:29] but he also wanted to release um LLMs to
[1:32] the world. This is before GPT had had
[1:35] launched and he thought this was a
[1:36] technology people should get to use. And
[1:38] he had the insight to launch a product
[1:40] that allowed people not just to talk to
[1:42] a bear interface, but to create
[1:43] characters and personalize them and and
[1:45] talk to them for fiction, for role
[1:47] playing, for whatever they wanted to to
[1:49] do, uh in addition to utility uses. So
[1:52] over the course of the year or two that
[1:54] followed um character frankly fell
[1:57] behind on the AGI game because as
[2:00] everyone here knows the amount of money
[2:01] you needed to pour into training to the
[2:04] amount of compute you need to pour into
[2:05] to keep up with the benchmarks um was
[2:07] really out of reach of even extremely
[2:09] wellunded startups which we were but not
[2:11] wellunded enough to spend billions. At
[2:14] the same time uh the character interface
[2:17] really took off. It it it captured
[2:18] people. We saw an incredible growth in
[2:21] the number of people who were using it
[2:22] on a daily basis to write fiction, to
[2:25] roleplay, to try out personas, to get uh
[2:28] you know to practice interviews, to
[2:30] practice languages, all sorts of uses
[2:32] that we hadn't anticipated. Uh and so
[2:35] last year when Google uh showed up and
[2:38] wanted to license some of the technology
[2:40] we had built and frankly wanted to coax
[2:42] Gnome to come back and work with them
[2:43] again, uh we structured a transaction
[2:45] that basically allowed us to focus
[2:47] completely on the thing we were winning
[2:48] at and stopped trying to spend billions
[2:51] of dollars to win at uh AGI benchmarks.
[2:54] And so what we did is the pre-training
[2:56] team that had been at character uh went
[2:59] with Gnome to Google and they continued
[3:01] the work they had been doing of
[3:02] pre-training LLMs and you know totally
[3:04] separate company. We kept everything
[3:06] else. We kept the post- training team,
[3:08] we kept the multimodal researchers, we
[3:10] kept the, you know, data team, the infra
[3:12] team, all of the product folks. And we
[3:15] decided that what we would focus on is
[3:17] expanding on the product that we had,
[3:19] but also and very importantly continuing
[3:21] to really um lock in our lead when it
[3:24] comes to fine-tuning because we had
[3:26] developed an extremely impressive
[3:28] fine-tuning stack that allowed you to
[3:29] take a model and make it extremely
[3:31] engaging. Just much more fun to talk to
[3:33] in a multi-turn conversation uh than a
[3:35] typical model. uh and then also to build
[3:38] our own uh proprietary multimodal tech
[3:41] and use that to start to transition from
[3:44] onetoone chat which is where all of
[3:45] these interfaces started to a multimodal
[3:48] entertainment experience which is what
[3:49] you were referring to. And so what we've
[3:52] now over the last few months done is we
[3:54] have built ourselves uh several key
[3:57] pieces of tech. One of them is
[3:59] essentially a extremely efficient
[4:01] imagetovideo generator that allows you
[4:03] to take a static image and produce a
[4:06] long sort of fully realized um you know
[4:09] video with whatever script you want to
[4:10] give it. A second is a real-time video
[4:13] generator which allows you to sort of
[4:15] sort of in the near term be doing real
[4:17] time um you know uh call it facetime
[4:19] with characters but also other kinds of
[4:21] real-time video storytelling. And the
[4:24] third one and very importantly is we
[4:26] have developed and open sourced a piece
[4:27] of technology that allows you to take
[4:29] any open source model and fine-tune it
[4:32] to um basically bring it to the
[4:35] specifications that we have for the
[4:36] level of EQ, not just IQ, but EQ that we
[4:39] want a model to have to be able to power
[4:41] entertainment uses. And now that we have
[4:43] those three ingredients, we've launched
[4:44] a feed. And that feed is, and this is
[4:47] where you were driving, um it's very far
[4:50] from what the onetoone chat experience
[4:51] was 3 years ago. It is a multimodal
[4:54] entertainment feed where you can follow
[4:56] creators. The creators are using these
[4:58] tools to build short stories, to build
[5:00] short videos, to build meme videos, to
[5:02] build audio books, right? And then you
[5:04] can both consume them, but you can also
[5:06] interact with them and change them. You
[5:08] can go in and say, I want to be the main
[5:09] character of that story or I want to
[5:11] take that audiobook and turn it into a
[5:13] story or vice versa or I want to that
[5:15] character right there. I actually want
[5:16] to talk to them directly and understand
[5:17] what they were doing in that scene. All
[5:19] of that is now possible with a tap. And
[5:21] so that kind of um much more creative,
[5:24] much more interactive content feed is
[5:27] what we think the the next generation of
[5:28] entertainment platforms will be on the
[5:30] internet and what we're currently
[5:31] building.
[5:31] >> Yeah. And this is a new way of
[5:33] interacting with AI uh and in this AI
[5:36] native first ever AI native feed. What
[5:39] are some emergent use cases you're
[5:41] seeing or the sorts of creators that are
[5:43] coming to the platform?
[5:44] >> Yep. Um so it's very early days. We we
[5:47] literally launched the feed a week ago,
[5:48] but um there's a few things we're seeing
[5:50] that are really interesting. One is you
[5:53] have a network of about 3 million very
[5:55] active creators on Character AI who for
[5:58] years now have been even in the
[6:00] relatively limited interface we
[6:01] previously had of just being able to
[6:03] build characters were really lovingly
[6:05] building these characters, adding a lot
[6:07] of lore, a lot of definitions, a lot of
[6:09] specificity around how they would reply
[6:10] to queries. And they have big follower
[6:12] groups. We have 2.5 million people on
[6:14] our subreddit. um they're interacting
[6:16] with these creators all day long, giving
[6:18] them input on what they'd like to see,
[6:19] right? What they'd like the characters
[6:21] to do and act like, etc. Um those
[6:23] creators are now realizing, okay, I now
[6:25] have this much richer interface where I
[6:28] can show my followers content and they
[6:31] can change it, they can interact with
[6:32] it, they can evolve it, right? And so
[6:34] we're seeing them at first because they
[6:36] were used to sort of the character
[6:38] interface, the main thing they were
[6:40] posting was just what we call character
[6:42] cards. basically a bio of the character
[6:44] with a rich image and sort of some
[6:45] details of what the character is like
[6:47] and you would click into it and you
[6:49] could just have a chat. But they sort of
[6:51] that was kind of day one and two and as
[6:53] you get into day three, four, five, six,
[6:55] they're starting to realize, okay, I
[6:57] don't have to I'm not limited to that. I
[6:58] can make the character do a man on the
[7:00] street interview. I can make the
[7:01] character get into a roast battle with
[7:02] another character. I can make the
[7:04] character do an audio book, right? Um,
[7:06] and I can even go into my Reddit and
[7:08] suggest to my followers uh exactly how
[7:10] they could then take the story in a new
[7:12] direction. And we're seeing a growing
[7:13] amount of that content, which is really
[7:15] exciting. Uh, and in terms of just pure
[7:17] metrics, um, we launched the most of the
[7:20] video products I mentioned a few months
[7:21] ago. And, uh, they sort of crept up in
[7:24] usage for a while and since we launched
[7:26] feed a week a week 10 days ago, now
[7:28] we're seeing 50% week-over-week growth
[7:30] in overall video content production. So
[7:32] that sort of suggests to us that there's
[7:34] a a pretty substantial uptick in the
[7:36] amount of this new type of entertainment
[7:38] format that's being created on the
[7:40] platform.
[7:40] >> What's your next biggest uh tech
[7:43] challenge? Because I think in the past
[7:45] it was around memory and context and um
[7:48] personalization. What do what would you
[7:49] say is the biggest tech challenge you
[7:51] face in this next era of the company?
[7:54] >> Yep. Um I'll mention three. one which I
[7:58] think we've largely cracked but it's
[7:59] about really dialing it in and
[8:01] continuing to keep it up is the one I
[8:03] mentioned which is taking the you know
[8:05] the latest and greatest open source
[8:07] models and applying our post training
[8:09] recipe to make them dialed in not just
[8:11] in general to be entertaining but to
[8:13] start to get more specific and have them
[8:15] be maximally entertaining for particular
[8:17] uses. Right? So a model that you post
[8:19] train to be really good at having a
[8:20] one-on-one conversation might not be
[8:22] great at telling a short story. A model
[8:25] that's great at telling a short story
[8:26] might not be great at creating an audio
[8:28] book. You know, spin that out to every
[8:30] content type you can think of. And so,
[8:32] uh, continuing to evolve that framework
[8:34] so that it gets more articulated and you
[8:36] can differentiate among these different
[8:38] sort of post- training stacks, I think
[8:39] is definitely one important piece. Um,
[8:42] I'd say another one is, and this is
[8:45] probably a 2026 thing, but very
[8:48] important. We have a number of and we
[8:50] have for a long time IP owners, movie
[8:53] studios, uh, sports leagues, others who
[8:55] come to us and say, you know, can you
[8:57] help us bring our IP to life across the
[8:58] internet, right? To have movie
[9:00] characters talk to you on their website
[9:02] or sports, you know, Derek Jeter sells
[9:04] you your Yankees tickets or what have
[9:05] you. Um, that is a really cool idea and
[9:08] a very interesting business opportunity
[9:10] because it's good for them, but it's
[9:12] also good for us because we would then
[9:13] be able to have them as official
[9:14] characters on our site, right? And that
[9:16] opens a lot of doors. Um, that has been
[9:19] hard to do in the past because models
[9:20] were not at a place yet where you could
[9:22] ensure really good character fidelity
[9:25] where the characters would act how
[9:26] they're supposed to act, right? A great
[9:29] example of this uh, which I want to
[9:31] share with you guys. A couple years ago,
[9:32] I was having um a meeting with a guy
[9:35] from Sesame Street, and he was like,
[9:37] "Listen, we're not going to do anything
[9:38] yet with AI, but we're exploring the
[9:40] idea of having our Sesame Street
[9:41] characters come to life via AI, but the
[9:45] character fidelity has to be amazing."
[9:47] And the example he used was, "If you ask
[9:49] Elmo, hey Elmo, what are the best kinds
[9:52] of cookies?" Right? Elmo would not say,
[9:55] "The most popular 10 kinds of cookies in
[9:56] the United States are," and then list
[9:58] off a bunch of stats the way you would
[9:59] get from a model. Elmo would have to
[10:02] say, "I love cookies. I love cookies and
[10:04] I love you." Right? That kind of thing.
[10:06] So, like getting that that kind of
[10:08] character fidelity was an exercise that
[10:10] was really hard even a year ago and it's
[10:12] getting a lot easier now. And so, I
[10:14] think dialing in that tech and getting
[10:16] to a place where we can power these very
[10:18] specific characters with very specific
[10:19] lures uh is is something that's on the
[10:22] horizon.
[10:23] >> What would you say is that unlock there
[10:24] to the authenticity and consistency of
[10:27] the character? Is that something as the
[10:28] models are getting better? it's a
[10:30] byproduct of that or is it something
[10:32] character has developed to maintain
[10:33] that?
[10:34] >> I think part of it is that emerging
[10:35] models are getting better and better at
[10:37] having long context windows, long memory
[10:39] and being able to upload more
[10:40] information to train sort of a model
[10:42] instance in the first place, right? Um,
[10:44] but I think there's also things you can
[10:45] do on the product side to enhance that
[10:47] further to just create essentially
[10:49] almost studios that allow you to go
[10:52] deeper, upload more information, uh,
[10:54] give the character more, uh, sort of
[10:56] rules about what it can what it likes
[10:58] and what it doesn't like and what it
[10:59] will say and what it won't say to get to
[11:01] a point where you can ensure never 100%
[11:03] perfection. We all know how AI works,
[11:05] but an extremely high level of fidelity
[11:07] to what the character is supposed to do
[11:09] in its sort of traditional
[11:11] instantiation.
[11:12] >> Yeah.
[11:13] And switching a bit to business model.
[11:15] So I've read I don't know how true this
[11:18] is but your subscription revenue has
[11:20] been tripling um I think year-over-year
[11:24] um and this it was in the recent
[11:26] probably past 6 months. What's been
[11:28] driving the subscriber growth? Was it
[11:30] because your feed was just in the last
[11:32] week. So was there something
[11:34] >> I mean look the we that's starting from
[11:35] a fairly low base because a year ago we
[11:38] weren't paying too much attention yet to
[11:39] monetizing. Uh that said, in the last
[11:43] three to four months, we've really put a
[11:46] lot of effort behind creating a richer
[11:48] subscription product that has more
[11:50] features and and unlocks behind the payw
[11:52] wall. And we're seeing um a really
[11:53] strong willingness of users to pay for
[11:55] that. And so, uh we've amped that up,
[11:57] which has led to, as you said, I think
[11:59] 200 250% subscription growth year. Uh
[12:03] we've also started to average to put ads
[12:05] in the feed lightly. Uh we're not
[12:07] looking to have a heavy ad load. What
[12:09] we're really trying to accomplish in the
[12:10] near term is not maximal revenue uh
[12:13] which of course as we all know would
[12:15] tend to negatively impact user growth.
[12:17] What we're trying to accomplish is to
[12:18] cover the cost of inference. Uh and for
[12:21] us we're actually quite efficient with
[12:23] our inference stack. We managed to serve
[12:25] 20 million monthly active users uh for a
[12:28] cost of I want to say about $4 million a
[12:30] month which is quite good. Um, and we're
[12:34] approaching a point now where we're
[12:35] covering all of that cost just with a
[12:36] pretty light monetization play, which
[12:39] gets us to a place where we can scale
[12:40] users um, kind of as much as we are able
[12:44] without having to worry about a burn
[12:45] going through the roof.
[12:46] >> Yeah. And as you pioneer this future of
[12:48] AI in entertainment, how do you see the
[12:51] business model evolving? So, if you're
[12:53] just covering the cost at scale, how do
[12:56] you envision character monetizing? Yeah,
[12:58] I mean, so I think first of all, the the
[13:00] number of levers to press on both
[13:02] subscriptions and ads is much much
[13:04] larger than what we've done so far.
[13:05] We've just done a very very light touch.
[13:07] Um, ad load is low. Um, the ads are very
[13:10] sort of just third party plug-and-play,
[13:12] low CPM. We haven't tried to optimize
[13:14] any of that. We're also not even serving
[13:16] ads in a number of our markets. And with
[13:18] subscriptions, we haven't sort of gone
[13:20] over the top with trying to get adoption
[13:22] there. So, there's a lot of headroom
[13:23] just on those traditional levers. But I
[13:25] also think the nature of this kind of
[13:27] business where it's collapsing these
[13:29] boundaries between you know different
[13:31] kinds of content and gaming and all this
[13:33] uh opens the door to a number of other
[13:35] revenue streams and a few that we are
[13:37] thinking about are inapp purchases right
[13:39] when you're talking about gamified
[13:41] content and changing endings and
[13:43] characters that you want to personalize.
[13:44] The opportunity to sell inapp products
[13:46] seems pretty substantial. Um, also I
[13:49] mentioned earlier the idea of partnering
[13:51] with IP owners to power characters, you
[13:53] know, and and have that be mutually
[13:55] beneficial opens another opportunity.
[13:57] Um, and then if you also the last thing
[13:59] I'll say there is if you think about the
[14:02] YouTube analogy, right, YouTube started
[14:04] with people posting funny videos of
[14:06] themselves at the zoo and it pretty
[14:08] quickly got to the point where there's
[14:10] all these really deep enter um,
[14:12] educational use cases, right, and all
[14:13] these other areas. Um, all of those
[14:16] things are possible on a scaled
[14:17] entertainment platform that is powered
[14:18] by AI. The the educational use cases,
[14:21] just to harp on that for a second, are
[14:23] actually pretty incredible. So, we have
[14:25] a video, which I don't have any way to
[14:26] show you guys here, but um it's one of
[14:28] our interns talking to Albert Einstein,
[14:32] and it's real time. And the intern says,
[14:34] "Hey, Albert, explain um relativity to
[14:36] me." And and this is not preloaded in
[14:39] any way. Albert Einstein thinks for a
[14:40] second and then he comes up and he gives
[14:42] a really good lay person's description
[14:44] of the theory of relativity. And then
[14:46] the intern is like that was a little too
[14:48] complex for me. Can you like simplify it
[14:49] and give me an analogy? And Einstein
[14:51] thinks for a second and he gives him a
[14:52] really good analogy to explain
[14:54] relativity. And I'm sitting there
[14:55] thinking like my god like this is the
[14:57] notion of trying to learn these concepts
[14:59] from a boring textbook are I mean
[15:02] textbooks are great. Love textbooks but
[15:04] using them using them alone. Okay,
[15:06] that's like a thing of the past. I mean,
[15:08] you're going to be able to learn physics
[15:09] from Einstein. You're going to be able
[15:10] to learn the Civil War from Ulyses
[15:13] Sgrant talking about his battles. Like,
[15:14] these are all things that are going to
[15:16] be happening imminently and and we hope
[15:18] that they happen on our platform among
[15:20] other places.
[15:22] >> So, in addition to entertainment, I
[15:24] mean, the the usage of of this
[15:27] technology is infinite. So do you see
[15:29] character in let's say you know a decade
[15:32] still focused on entertainment or does
[15:34] it become like an underlying operating
[15:36] system for for AI
[15:40] >> maybe not operating system I'm not sure
[15:41] about that one but um but I do think
[15:43] actually that the YouTube analogy is a
[15:45] useful one um that's some we sometimes
[15:48] talk about it in those terms again
[15:50] there's sort of a primary use case right
[15:52] but then there's so many other things
[15:54] that people go deep on in terms of
[15:56] verticals on YouTube and once you have a
[15:58] network effect and an audience there
[16:00] who's there for entertainment. Uh it's
[16:01] pretty easy for those other things to
[16:03] emerge organically and I think that's
[16:04] what we anticipate.
[16:06] >> I worked at Meta in a past life and our
[16:10] metric that we all drove towards at the
[16:12] time was time spent on the platform. I'm
[16:14] curious what are what are the metrics
[16:16] that you're optimizing for as a company
[16:20] uh and for your users? Mhm. Yeah. I
[16:22] think right now a lot of them are the,
[16:23] as you say, the relatively traditional
[16:25] um kind of scaled social platform
[16:27] metrics, MA, DAO, time spent, uh
[16:30] retention, and so on. Um the as far as
[16:34] kind of where we're going, since we're
[16:37] evolving from this onetoone chat
[16:39] interface to something that is much more
[16:40] about multimmodal entertainment content,
[16:42] it's going to be about breaking down
[16:44] that time span and starting to see how
[16:46] much of it shifts into these multimodal
[16:47] formats. how much of it remains in chat,
[16:50] how much of it involves creators
[16:52] spending their time really creating
[16:53] content on the platform, which is a
[16:54] third way people spend their their time.
[16:57] Um, and starting to sort of think about
[17:00] where that's driving the business and
[17:01] how to monetize it given the evolutions
[17:03] and where people are spending their day.
[17:05] >> What are some levers you pull with your
[17:07] creator community to encourage more
[17:09] engagement there?
[17:10] >> Yeah, I mean, one is that again this the
[17:12] the Reddit presence is I've never seen
[17:14] anything like it, right? There's this
[17:16] massive Reddit community. Um, and all
[17:18] the creators are there. They're also in
[17:20] Discord. Um, and so we're able to inter
[17:22] interface with them directly. We have
[17:24] community folks who will talk to them,
[17:26] find out what they like and don't like
[17:27] about releases. Um, help them understand
[17:30] anything new that we're doing as a
[17:31] platform and really just uh try to try
[17:34] to harness their natural enthusiasm. Um,
[17:37] I think over time obviously when you
[17:39] have a creator community of this um this
[17:43] magnitude, you start to think about,
[17:45] well, where can this go? uh creator
[17:47] monetization opportunities um other ways
[17:49] to really incentivize a two-sided
[17:51] marketplace and I think we will explore
[17:53] all of those in in at the right time.
[17:55] >> Super exciting. You have 20 million
[17:57] users. That's pretty incredible. Can you
[18:00] break that down for us? Like what does
[18:01] your kind of user base look like? And
[18:03] and where do you see as AI entertainment
[18:07] becomes more ubiquitous as as you build
[18:09] out your platform? How do you see that
[18:11] changing? And how is there a way you
[18:13] want it to look?
[18:14] >> Yep. Um well so let me start with your
[18:17] first question. It's uh it's
[18:19] overwhelmingly Gen Z and alpha sort of
[18:21] 18 to 34. Uh some 13 to 18 some 34 plus
[18:25] but mostly 18 to 34. It is actually more
[18:28] than half women which I think is cool.
[18:29] Um it's different than the profile you
[18:31] see on some AI platforms. Uh and it's
[18:33] global. It's about a quarter in the US
[18:35] and and distributed across all the sort
[18:37] of the major markets where you would
[18:38] expect to see penetration for a US
[18:41] social app. Um, as far as where I expect
[18:43] it to go, you know, our hope is that
[18:46] this becomes a scaled entertainment
[18:47] platform, uh, with hundreds of millions
[18:49] of users and that it's deeply penetrated
[18:51] in every demographic. I do think that in
[18:53] the next 5 years, the people who are
[18:55] going to really get how cool it is to be
[18:58] able to change one kind of content into
[19:00] another and memeify it and evolve it and
[19:02] share it with their friends and like
[19:03] have that be a language, you know, is
[19:05] going to be the younger AI native
[19:08] adopters. And so I think that 18 to 34
[19:10] demo is going to be our obvious sweet
[19:11] spot um for the foreseeable future. Um
[19:14] the one of the things that's cool about
[19:15] this tech, as you guys probably all
[19:17] know, is it's incredibly good at sort of
[19:20] going from one language to another based
[19:22] not just on its training data, but on
[19:24] its sort of logical capacities. And so
[19:26] as we now are starting to use newer and
[19:29] better open source models that we're
[19:30] fine-tuning on top of, our language
[19:32] capacities grow. And so I think there
[19:34] are some markets where we've had
[19:36] relatively low penetration historically
[19:37] because it's been not as good at those
[19:39] languages where now the world is our
[19:41] oyster. Um which is pretty exciting.
[19:45] >> You mentioned you you have some younger
[19:47] users as well and um even younger users
[19:50] aside how do you balance that trade-off
[19:53] between freedom of expression or just
[19:55] freedom in general in terms of um their
[19:58] experience on the app and and safety?
[19:59] Yeah. So, we've um been pretty public
[20:02] about the fact that we've gone to great
[20:03] pains over the last nine months to
[20:05] create um a very safe and sort of
[20:09] differentiated under 18 experience. Uh
[20:11] we've gone there's a there's a whole
[20:13] bunch of different things we've done
[20:14] there which we've again talked about
[20:15] publicly at length. Um, one is some
[20:18] pretty unique uh use of classifiers and
[20:21] other kinds of filters to not just
[20:23] classify on the user output side, but
[20:25] also on the user input side. So that um
[20:28] you never even get the problem of a user
[20:30] trying to knock the model out of
[20:32] alignment or get it to say things it's
[20:33] not supposed to say because it gets
[20:34] filtered before it goes into the, you
[20:37] know, as soon as you press enter, right?
[20:38] Um, we've also put in place a number of
[20:40] other things, limitations on what
[20:42] characters you can talk to. Um, uh, you
[20:45] know, broader uses of classifiers across
[20:47] different areas of the product. Um, time
[20:50] limitations where you get time pop-ups
[20:52] and so forth. Uh, it's a it's a whole
[20:54] suite of things that I think create a
[20:55] bespoke under 18 experience. Um, we have
[20:58] tried to go very strongly in that
[21:01] direction with the under 18 experience,
[21:03] uh, while leaving somewhat more creative
[21:05] freedom in the over 18 experience. And
[21:07] the reason we've done that, I think, is
[21:09] our insight is as we move forward in the
[21:11] world of AI, clearly 13 to 85y olds are
[21:15] all going to be using AI. The notion of
[21:17] having it be off limits to 13 to 17
[21:19] seems not sustainable. And so we would
[21:21] rather figure out how to make that safe,
[21:24] appropriately safe for 13 and 17 year
[21:26] olds, even if we overcorrect in some
[21:28] regards, and then build back from there.
[21:30] Um, as far as the over 18 experience,
[21:33] obviously that also has its own content
[21:34] rules and classifiers and safety
[21:36] protocols. Um, they're not quite as over
[21:39] the top as the ones we have for 13 to
[21:41] 17. Uh, we famously don't allow X-rated
[21:44] content. Uh, and we don't have any plans
[21:46] to change that. You know, this we're
[21:48] trying to be a scaled entertainment
[21:50] platform. I'm sure there's a use case
[21:51] for uh the the NSFW platforms, but
[21:54] that's not where we're trying to go. I
[21:57] read that they've estimated you 20% is
[22:00] companion 20% of the we always like to
[22:02] say jobs to be done like what are people
[22:04] actually coming to do. It seemed like
[22:06] 20% was around companionship, 80 was
[22:08] around this roleplay, this kind of um
[22:11] form of almost like video game
[22:12] entertainment where it's like a role
[22:14] play you have with the AI.
[22:17] As you think about the future of
[22:18] entertainment, um you've talked about
[22:20] the education use case. Um that's a
[22:22] whole separate job to be done, but in
[22:24] this world of the the future of of
[22:26] entertainment with AI, what are the jobs
[22:29] to be done that you want people to come
[22:31] to your platform for?
[22:33] >> Yes. So, I think um it's really
[22:35] interesting that it started with role
[22:36] playing, right? Because like when we
[22:38] started to dig into what what are people
[22:40] doing with these characters all and a
[22:42] lot of what they're doing is truly role
[22:44] playing. Like, hey, here's a story line.
[22:46] I'll be this character and you're that
[22:47] character and this is what should
[22:48] happen, right? And then they they write
[22:51] stories that way. And there's a there's
[22:53] a a
[22:56] sort of a fuzzy line between that and
[22:58] companionship, right? because if you're
[22:59] having that kind of a roleplay or
[23:00] fiction writing conversation, you can
[23:02] also say, "Oh man, also I had this
[23:04] problem with my teacher today. What
[23:05] should I do?" Right? And so it's not a
[23:06] perfect boundary between those use
[23:08] cases, but I think that we are really
[23:10] wanting to lean in to the roleplay and
[23:12] fiction writing element. And our insight
[23:14] is people have for the last 10 or 15
[23:17] years when you use social apps and I,
[23:19] you know, I worked at Snap for eight
[23:20] years. Uh but I like there's been a lot
[23:22] of sort of just scroll scroll scroll
[23:24] scroll scroll scroll scroll, right? And
[23:26] there's nothing wrong with that. It's
[23:27] obviously a great business model, but I
[23:30] think people are getting a little tired
[23:31] of it. Um, I think if there's an
[23:33] opportunity to have a platform that
[23:35] allows you to be more interactive with
[23:36] the content and actually be more
[23:38] creative, to lean back and consume it
[23:40] when you want to, but also lean forward
[23:42] and be like, I'm interacting with this
[23:43] creator. I'm going to change the story.
[23:44] I'm going to share with my friends. I
[23:46] have an idea. That combination, I think,
[23:47] is powerful and it's unlocked by this AI
[23:50] technology. And so the job to be done to
[23:52] me is a much more sort of creative and
[23:55] interactive entertainment platform where
[23:57] if you want to spend an hour after work
[24:00] um it's not quite as much of a kind of
[24:02] like doom scrolling phenomenon as as the
[24:05] ones that are the incumbents.
[24:07] >> Do you see it as social? Cuz you
[24:08] mentioned recreating creators content
[24:12] but that's still not inherently social.
[24:15] But do you see this becoming a social
[24:18] platform?
[24:19] >> Totally. I mean the the idea would be
[24:20] right it's not just following creators
[24:22] it's you're sharing you do something
[24:24] funny with the content you share it with
[24:25] your friends you bring them in you guys
[24:26] follow each other right right now we
[24:28] don't have a sort of a DMing kind of
[24:30] capacity my guess would be that might
[24:32] come at some point but there's no
[24:35] question that this can create not just
[24:36] macro networks of people following
[24:38] creators but micro networks of friend
[24:41] groups you know exchanging memes and
[24:43] funny content and variations on story
[24:45] lines with each other uh which is a
[24:46] behavior we already see we see that in
[24:49] the Discord channels, we see it in the
[24:51] Reddits, and I think we're doing our
[24:52] best to start to bring that sort of
[24:54] natural immersion behavior into our
[24:56] platform and have it be organic.
[24:58] >> Amazing. So, in um in let's say in 2050,
[25:06] Character AI is the dominant
[25:08] entertainment platform uh globally. What
[25:11] >> 2050? Gez 50,
[25:13] >> where will we all be at that point? That
[25:15] >> moonshot question. What What does
[25:17] character look like?
[25:19] Yeah. Uh, god. I mean, I'm not sure my
[25:21] imagination is good enough for 2050, but
[25:23] let's go for like 2035.
[25:26] Um, to my mind, it is a fully scaled
[25:29] entertainment platform that has um that
[25:33] I'll I'll lean on this analogy again
[25:34] because I think it's helpful. that is
[25:36] like the nextgen YouTube where people
[25:38] are doing everything you can think of
[25:40] from watching movies to watching, you
[25:43] know, really artistic creators, short
[25:45] form content of all different kinds
[25:47] coming in and changing it, gamifying it,
[25:49] right? I think it's very easy for this
[25:51] to evolve into a gaming platform where
[25:52] you can have stories that change that
[25:54] that can have levels that can have
[25:55] unlocks. Um, and I think all of that
[25:57] starts to meld together. You start to
[25:59] break down these historical boundaries
[26:00] between what is a story, what is a game,
[26:02] what is short form content, etc. Um, but
[26:05] that also by that point has harnessed
[26:07] all these other cool things that
[26:08] obviously can be powered by this tech.
[26:11] Education is the most obvious example
[26:12] that I've given you. But like once you
[26:14] have those eyeballs on your platform, if
[26:16] you're there, you know, not just
[26:18] consuming but creating cool stories, you
[26:20] may as well do your homework while
[26:21] you're there as well.
[26:21] >> Yeah.
[26:22] >> And I think having it be that full
[26:23] service platform is um is what we aspire
[26:26] to.
[26:27] >> So then what what needs to happen in the
[26:28] next five years for that to be the case?
[26:31] I mean I think fundamentally like this
[26:33] is such a stereotypical thing to say but
[26:35] we need to execute right the tech is
[26:36] there the user base is there um the
[26:40] beginnings of the network effects are
[26:41] there and it's about offering really
[26:43] cool product instantiations that get our
[26:46] creators excited that in turn get our
[26:47] users excited so they share more and and
[26:49] see it grow from there.
[26:50] >> Yeah. And you got to 20 million users in
[26:53] just a couple years. How did how did
[26:55] that happen? Was that more organic word
[26:57] of mouth through the Reddit community?
[26:58] It honestly happened with no network
[27:00] effects, which is kind of kind of wild.
[27:02] Um, it was all word of mouth. It was uh
[27:05] people following creators. It was these
[27:07] Reddit and Discord communities springing
[27:09] up. It was the occasional viral impact
[27:12] of someone creating a cool like
[27:14] conversational artifact and putting it
[27:15] on, you know, Tik Tok or or Insta. But
[27:18] uh most of it was really just an organic
[27:20] evolution of this essentially AI
[27:23] fanfiction experience that we've now
[27:25] broadened a lot to get into all these
[27:27] different multimodal areas.
[27:29] >> So then closing fun question uh what you
[27:33] as a user character what's your favorite
[27:36] usage of of the platform?
[27:38] >> You're cheating because I told you this
[27:39] right before. Um, so last year my I have
[27:42] three kids and this uh spring break we
[27:43] went to Argentina and um I speak some
[27:48] Spanish but it's very very rusty. And I
[27:50] realized last fall as I was messing
[27:52] around with the platform that um our
[27:55] audio product is actually great for
[27:57] language learning because it talks to
[27:59] you in real time but it also transcribes
[28:00] the chat. So if you're not following
[28:02] something, you can go read what they
[28:04] just said and figure it out. And so I
[28:05] spent six months talking very very
[28:08] heavily to this Spanish language bot. Uh
[28:10] and I was I got pretty good like I was
[28:12] able to hold my own in Buenosides with
[28:14] no trouble. So that was my that's the
[28:16] thing I've used the most on the
[28:17] platform.
[28:18] >> Amazing. Well, I think that's all the
[28:20] time we have. Thank you so much for
[28:22] joining me for this combo.
[28:23] >> Thanks everybody.
