The Ghost Poet in The God Machine

2025-12-17 · Source: The Leverage · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Emerging Technologies & Innovation · Depth: Intermediate, medium

Summary

The Ghost Poet is a personal software project developed by a non-engineer using Anthropic's Fable model to process grief. This tool, created over a weekend, fine-tunes a 1.44MB GPT-2 model on public domain poetry, primarily Walt Whitman, to generate unique, emotionally resonant poems. The author posits that AI's ability to drastically reduce software development costs enables "personal software" driven by emotional needs, much like traditional art, contrasting with the economically-driven, emotionally barren nature of most application software. The project intentionally re-embraces the "weirdness" or "dream hallucinations" of early LLMs, which were largely removed by RLHF to optimize for productivity. The artistic value is found in the act of generating and witnessing these often nonsensical outputs, allowing the user's brain to assemble meaning. The Ghost Poet operates entirely locally in the user's browser, downloading 165MB of weights once for privacy.

Key takeaway

For Creative Technologists exploring AI's artistic potential, this project demonstrates how to reclaim emotional depth in software. You should consider intentionally leveraging less-refined LLMs or fine-tuning small models on curated datasets to foster unique, non-utilitarian outputs. Prioritize local execution for privacy and to shift the artistic focus from perfect generation to the user's interpretive experience, enabling deeply personal digital art.

Key insights

AI can enable deeply personal software creation, shifting value from utility to emotional resonance and artistic expression.

Principles

• Software economics historically stifled emotionally-driven creation.
• Early LLM "weirdness" holds artistic value.
• Art in AI generation is the act of witnessing.

Method

The author fine-tuned a raw GPT-2 model on 1.44MB of public domain poetry (half Whitman) to achieve emotional resonance, then built a browser-based interface using Fable for local poem generation.

In practice

• Experiment with early, less-refined LLMs for artistic outputs.
• Fine-tune small models on specific text for emotional tone.
• Develop local-first AI applications for privacy.

Topics

• Personal Software
• AI Art
• GPT-2 Fine-tuning
• Anthropic Fable
• RLHF Impact
• Local AI Applications
• Generative Poetry

Best for: Creative Technologist, AI Student, AI Ethicist

Related on AIssential

• We Don’t Have a Creativity Crisis. We Have a Meaning Crisis. — AI redefines artistic value, prioritizing authentic human perspective and meaningful ideas over technical execution.
• Powering the world’s first AI arts museum — AI and environmental data can create immersive, real-time, multi-sensory art experiences.
• Make Something Beautiful — AI democratizes software creation, empowering individuals to build custom digital experiences without coding expertise.
• AI Poetry As The Compression Engine — Software development is shifting from explicit code to natural language-driven capability expression via models.
• Think You Can Build a Game with AI? Think Again! - Danielle An & David Hoe, Meta — AI democratizes basic game creation, but true quality and dynamic experiences require design principles and runtime LLMs.
• Every little thing she does is magic — Embrace creative limitations and personal intuition to achieve unique, emotionally resonant artistic outcomes.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Leverage.