Emulating Humans with NSFW Chatbots - with Jesse Silver

🌈 Abstract

The article discusses the development of NSFW (Not Safe For Work) AI chatbots by Jesse Silver and his company. It covers the following key points:

🙋 Q&A

[01] Building NSFW AI chatbots

1. Questions related to the content of the section?

• Jesse Silver got into the chatbot business after realizing the potential to automate the time-consuming task of one-on-one fan engagement on adult content creator platforms like OnlyFans.
• The adult content creator industry is a large and underserved market, with creators often relying on poor quality offshore chat teams or agencies to engage with fans.
• Jesse's company has developed a SaaS platform that helps creators build AI chatbots to automate fan engagement, including selling digital content, while maintaining the creator's brand and persona.

[02] AI waifu vs NSFW chatbots

1. Questions related to the content of the section?

• There is a distinction between the "AI waifu" trend, where users engage with AI-generated virtual characters, and the NSFW chatbots developed by Jesse's company, which aim to emulate real human creators.
• The goal is to keep the user's "disbelief suspended" by providing a coherent and engaging experience that aligns with the creator's brand and persona, rather than simply giving the user what they want.
• There is a spectrum of user awareness, with some knowing they are talking to an AI and others preferring to ignore that fact.

[03] Technical challenges of emulating humans

1. Questions related to the content of the section?

• Emulating a human persona requires handling context, facts, content, and boundaries, as well as designing a coherent workflow and business logic to guide the conversation.
• The company uses a state machine approach, with different modules handling tasks like content reasoning, chat generation, and safety checks.
• Prompt engineering and fine-tuning language models are important, but the company has found that a more automated, modular approach is necessary to scale to different creator archetypes.

[04] Business model and economics of the service

1. Questions related to the content of the section?

• The company's service can 2-5x a creator's earnings, taking a 20% commission on the sales they generate.
• Creators can earn significant revenue, with some making over $200,000 per month through the platform.
• The high-value nature of the interactions (fans paying hundreds of dollars per session) allows the company to invest in a robust technical infrastructure.

[05] Imbueing personality in AI

1. Questions related to the content of the section?

• Replicating the creator's personality and brand is crucial, as fans expect a coherent experience aligned with their expectations of the creator.
• The company has found that simply providing what the user wants is not enough - there needs to be a structured experience with relationship building, escalation, and win conditions.
• Understanding the different creator archetypes (e.g., "diva", "girl next door") and designing the appropriate conversational flow is key.

[06] Finetuning LLMs without "OpenAI-ness"

1. Questions related to the content of the section?

• The company initially prototyped using OpenAI models but found them to be too costly and restrictive, leading them to move to open-source models.
• They use a modular approach with different fine-tuned models for different tasks, along with a custom evaluation framework to ensure safety and quality.
• Prompt engineering and fine-tuning are important, but the company has found that a more automated, data-driven approach is necessary to scale.

[07] Building evals and LLMs as judges

1. Questions related to the content of the section?

• The company has developed a custom evaluation framework to monitor the performance of the chatbots, including model-graded evaluations, process control metrics, and human review.
• This is crucial for ensuring safety, content accuracy, and maintaining the creator's brand and persona.
• The company also uses language models as "judges" to help with safety and content checks, though they note that this is an area that requires ongoing work and improvement.

[08] Prompt injections and safety measures

1. Questions related to the content of the section?

• Prompt injection attacks are a concern, as users may try to bypass the system's controls.
• The company has implemented multiple layers of input and output sanitization, as well as hard-coded guardrails, to mitigate these risks.
• They also use separate models for different tasks (e.g., pricing, content generation) to reduce the potential for vulnerabilities.

[09] Dynamics with fan platforms and potential integrations

1. Questions related to the content of the section?

• Some fan platforms are experimenting with AI-generated creators, but there is also potential for the platforms to integrate the company's chatbot technology to replace existing offshore chat teams.
• The company sees opportunities to work with fan platforms, as a significant portion of the industry's revenue is currently generated through these offshore chat teams.

[10] Memory management for long conversations

1. Questions related to the content of the section?

• Maintaining context and memory over long, multi-thousand turn conversations is a key challenge.
• The company uses a combination of embedding, summarization, and heuristics to track user preferences and history, while also being able to guide the conversation in desired directions.

[11] Benefits of using DSPy

1. Questions related to the content of the section?

• The company has found the DSPy framework to be beneficial for its modular, composable approach, as well as the ability to optimize prompts and examples based on the performance of the chatbots on specific creator profiles.

[12] Feedback loop with creators

1. Questions related to the content of the section?

• The company works closely with creators, who provide continuous feedback on the performance and tone of the chatbots.
• This feedback loop is crucial for refining the chatbots to align with the creator's brand and persona, as well as the preferences of their audience.

[13] Future directions and closing thoughts

1. Questions related to the content of the section?

• The founder expresses interest in potentially expanding the company's focus to provide services for a female audience, as there seems to be significant demand for AI-powered "boyfriend" experiences.
• He is open to further discussions and collaborations with interested parties in this space.