Latent Space: The AI Engineer Podcast

Emulating Humans with NSFW Chatbots - with Jesse Silver

Overview

The company has built an AI-powered chat system for adult content creators that handles approximately 50,000 daily conversations for over 150 creators, generating $2 million in monthly revenue with a business model taking 20% commission only on new sales.

Their AI system is designed to maximize creator earnings by building authentic fan relationships through strategic interaction patterns that maintain the creator's unique brand, gradually escalate engagement, and implement intermittent reward mechanisms rather than immediate transactional exchanges.

The technical architecture uses open-source models within a complex workflow of specialized modules (state machine, reasoning, content understanding, chatting) with sophisticated evaluation frameworks and safety systems to prevent harmful content and prompt injections.

The solution addresses a significant market opportunity ($10 billion annually across platforms) where approximately 80% of creator revenue comes from fan interactions, offering a superior alternative to problematic offshore teams and agencies that often charge $2-4,000 monthly while delivering poor service.

Content

Company Background and Market Opportunity

Jesse Silver discusses founding an AI chatbot company focused on the adult content creator market
Inspiration came from a friend making $5k/month on OnlyFans through time-consuming fan interactions
Identified a significant market opportunity in fan platforms:

- OnlyFans does about $6 billion in annual gross merchandise value - Smaller platforms generate around $4 billion annually - Approximately 80% of creator revenue comes from fan chat interactions

Business opportunity emerged because:

- Creators lack sophisticated software tools due to industry stigma - Existing alternatives (offshore teams, agencies) are ineffective - Goal: Automate fan interactions and relationship building for creators

Company progress:

- Developed prototype and pitched at AVN conference - Secured $50k in initial gross merchandise value - Launched refined product in December - Currently working with over 150 creators - Handling approximately 50,000 daily conversations - Generating over $2 million in monthly creator account size

Fan Platform Dynamics and User Behavior

Users on fan platforms typically engage for three primary reasons:

- Alleviating loneliness - Seeking convenient content - Playing power/fantasy games with perceived stakes

Customer base characteristics:

- Approximately 80% of fan platform customers are men - Monthly customer turnover ranges from 50-80% - Surprisingly, about 10% of customers remain engaged for over a year

Interaction frequency and user preferences:

- Some AI interactions involve multi-thousand turn conversations - Different platforms (like Character AI) may have varying user engagement models - Women might prefer more conversational interactions compared to men - Average user session is 7-8 minutes - Some fans spend significant amounts (e.g., $3-5K daily from 100 fans)

Business Model and Value Proposition

Provides an AI-powered workflow/agent to interact with creators' audiences
Clients typically earn 20-150K monthly
Product aims to 2-5X creators' earnings
Takes 20% commission only on new sales, not existing revenue streams
Goal is to provide better fan interactions than manual chat teams

Key product functions:

- Builds relationships with fans - Handles text-based media interactions - Fulfills customer requests - Negotiates custom content - Replicates creator's personality/tone

Core philosophy:

- Respect creators' income streams - Provide fans a believable, enhanced interaction experience - Maximize creator earnings through intelligent AI interactions

AI Interaction Design and Strategy

Importance of maintaining a creator's specific brand and fan interaction expectations
Need to avoid "puncturing the fantasy" of a creator's established persona
Goal is to create nuanced, escalating interactions rather than immediate transactional exchanges

Interaction strategy principles:

- Build relationship before explicit content - Create a gradual escalation of interaction - Ensure fans feel emotionally engaged, not just transactionally served - Maintain ongoing interest by not immediately satisfying all fan desires - Implement strategic interaction logic to prevent giving away content for free - Create intermittent reward mechanisms - Set boundaries for non-paying or rude users

Different interaction types discussed:

- Diva, girl next door, dominant/submissive dynamics - Language models can play roles and have basic psychological profiling capabilities - Creators may adapt their approach based on perceived desires of fans/clients

Technical Implementation and Architecture

Using open-source models instead of closed commercial models like OpenAI
Complex workflow with multiple modules:

- State machine for different interaction modes - Reasoning modules - Content understanding modules - Chatting modules - Each module potentially uses a different fine-tuned model

Technical development progression:

- Started with low-code prototyping tools - Transitioned to DSPy for: * On-the-fly optimization * More elegant workflow representation * Easier model fine-tuning - Moved towards automated, scalable interaction generation

Backend and scaling:

- Backend designed to potentially handle millions of conversations monthly - API connects with fan platforms - Workflow generates and sends interactions to fan platforms - Latency target of responding within approximately 2 minutes - Handle multiple concurrent conversations (up to 50 simultaneously)

Data Collection and Fine-Tuning

Data gathering methods include:

- Bootstrapping with intelligent models - Scraping successful interaction histories - Model-graded evaluation of interactions

Challenges in data collection:

- Messy data sources - Inconsistent human chat team performance - Need to filter out undesirable interaction patterns

Approach to personality profiling:

- Not using standard psychological frameworks - Focus on generating interactions that align with creator's specific brand - Prioritize coherent role-based interactions - Analyzing previous successful sales/interaction patterns - Examining profile tone and communication history - Comparing similar creator profiles

Evaluation and Safety Systems

Developed a sophisticated evaluation framework with multiple components:

- State machine elements tracking conversation states - Dedicated behavioral evaluations - "Golden sets" for significant model changes - Model-graded evaluations (focusing on safety and response quality) - Manual sample review by a small, fractional ops team

Safety mechanisms:

- Implement reasoning modules with multiple layers of safety controls - Prevent harmful content (gore, inappropriate interactions) - Block prompt injections and malicious user attempts - Handle platform-specific and creator-specific safety limits - Hard-coded protections around pricing and content generation - Additional models to test and validate AI-generated content

Prompt injection mitigation:

- Acknowledge prompt injection is not fully "solved" - Use multi-step input and output sanitization - Implement reasoning modules to detect and block potential injection attempts - Proactively test system by having creators/agencies try to break it

Industry Context and Comparison to Alternatives

Offshore teams and agencies:

- Offshore teams can cost $2-4,000 per month - Human chat teams typically paid $3/hour plus 5% of sales - Often low-quality, low-compensation teams handling interactions - Agencies typically handle three key areas: chatting, content, and traffic - Many agencies have poor reputations, with creators reporting issues like: * Doxing * Losing fans * Financial misconduct

Platform landscape:

- Some fan platforms showing interest in AI creators - Potential industry hesitation due to authenticity concerns - Platforms could potentially verticalize and bring chat services in-house - Platforms are building comprehensive fan profiles across multiple creators

Creator Feedback and Future Directions

Creators frequently communicate with and provide feedback on their AI bots
Creators are deeply invested in how their personal brand is replicated in AI interactions
Success is measured by creators being unable to distinguish between their own text and the AI's output

Future interests:

- Potentially developing products targeting women users - Noted that Character AI has many repeat users, particularly women - Open to collaborating with others interested in the AI character/interaction space - Exploring contextual, long-term engagement AI experiences

The speaker is willing to be contacted via Twitter for further discussions about AI character development

More from Latent Space: The AI Engineer Podcast

Explore all episode briefs from this podcast

View All Episodes →

Listen smarter with PodBrief

Get AI-powered briefs for all your favorite podcasts, plus a daily feed that keeps you informed.

Download on the App Store